We are introducing a new feature called The Changelog. For the time being, we will update it quarterly because this industry changes enough over three months to justify a formal review. It will cover our core theses and the companies we track. It will also cover the micro and macro trends shaping our work. We will explain what changed and why, then identify new theses or areas where our conviction increased.

This regular review will help keep us honest and make industry changes easier to follow. Our research covers the full technology stack, so The Changelog will also give readers a consistent set of mental models they can use across our work.

What this first review found

This first edition covers six months of published research across our broad technology industry coverage. Our conviction in the broader AI thesis remains high and increased in several areas as we studied enterprise customers, their early deployments, and what they learned. We also maintain that demand for compute, and the supply chain around it, will exceed supply for the foreseeable future. At the same time, new evidence changed several views under the broader trend and produced eight new theses.

The broad AI demand thesis held across the six months. The evidence changed several views underneath it. Inference raised our conviction in agentic CPU demand, bringing along revived winners as a part of it, and the tension building across memory and networking. Our enterprise work increased our confidence that current adopters are seeing enough ROI to keep spending. Security also emerged as a new thesis as companies focused more on protecting their data and controlling what agents can access.

The thesis map below gives a high-level view of how each thesis moved during the period. Raised conviction means the evidence strengthened an existing thesis, while hold means our view remains intact at roughly the same level of conviction. Changed means the evidence altered the thesis or how we measure it. New means the thesis first emerged during this period, so we start with lower confidence and test it again in future editions.

Inside the full report

• The full raised conviction, hold, changed, and new thesis ledger.

• Why inference turned the merchant-versus-custom accelerator market into a TCO competition and raised our CPU forecast.

• Eight new theses, with the first report, current confidence, and evidence still needed.

• How chiplets deepened the advanced-packaging constraint and made Intel’s opportunity clearer.

• What enterprise ROI, token budgets, and security requirements now tell us about adoption.

• Why bitcoin miners may become better capacity partners for hyperscalers than pure AI neoclouds.

• A next-quarter monitoring table showing what would strengthen or weaken each view.

Read more

From our conversations we believe deployments are more likely to start next year and build into 2028, but the ecosystem is maturing enough for CXL to move from a standard into something that can be deployed with enough customers to meaningfully start to deploy in their AI compute infrastructure. The first and most logical customer base is the hyperscalers, which can put memory and inference workloads into custom compute clusters and use the surrounding ecosystem to qualify the architecture.

It is our conviction that solving the memory wall will take many different shapes by many different players but the same problem statement remains. We need more memory, and designers are up against how much memory can go on CPU/XPU/GPU package or near the package. Having ways to expand the available memory pool while keeping latency low enough for selected near-memory workloads is the promise of CXL.

Memory is becoming a fleet problem

The old server model ties memory capacity to one processor and its local channels. AI workloads make that boundary more expensive because context, KV cache, and orchestration state can grow faster than the memory attached to one compute device. A fleet can have plenty of memory in total while individual CPUs or accelerators still run short.

The practical distinction is important: “hot” describes the role memory plays in the workload, while local, off-die, off-board, and rack-level describe where that memory sits. HBM can remain the hot tier in an attached appliance when the fabric preserves the bandwidth and latency the workload requires. DDR can be split by role in the same way. Local DDR can serve the CPU’s most active data, while DDR4 or DDR5 behind a CXL controller can provide a larger attached tier and eventually a shared pool. That remote memory will not have the same latency as on-package HBM, but it can still be the highest-performance tier available outside the package.

CXL gives the system those placement choices through a coherent interface. The first commercial use is likely to be a card, module, or box that adds memory to one host. Later designs can attach several hosts to a pool, with software deciding which data belongs close to compute and which data can move into the attached memory. The value comes from expanding the working set and using capacity more fully without buying a full server for every increment of memory.

The graphic below frames the opportunity by workload rather than by device. HBM remains closest to compute for the hottest accesses, while local DDR/SOCAMM and CXL-attached memory support larger working sets, overflow KV cache, retrieval buffers, and shared state. CXL is not a fixed warm tier; its role depends on the memory type, topology, and workload.

Exhibit 1. Inference moves the memory problem from model weights to active state. CXL expands the set of places where that state can live.

Deployment flexibility

It is noteworthy how early we are in the rack scale compute era. By our estimates, using our accelerator installed base model, we believe rack scale accelerators in the range of 9-11% of the total AI accelerator installed base. The shift to rack scale solutions are the necessary catalyst that will enable CXL, and other solutions even if custom, to be designed by the customer. Knowing the data center customer is scaling their rack scale infrastructure helps as a catalyst for CXL whether that deployment is north south or east to west in its location. Enough of the ecosystem was present, through announcements and demonstrations at FMS, to make the deployment path more visible. Memory suppliers, interconnect companies, switch vendors, custom silicon providers, and networking ASIC companies were all discussing a path toward viability with much different language than earlier in the year. As of now, we expect the larger deployments are more likely to arrive in 2027 and build into 2028, yet the ecosystem is now moving from evaluation toward qualification, driven by a key set of customers.

We believe hyperscalers are the first catalysts for several reasons. They control rack specifications, work with ODMs, and can make system-level TCO decisions that are difficult in standardized OEM deployments. They also have access to a large legacy memory resource that is not always useful in its original server configuration, but could become valuable again in a CXL-based system. We explain the size of that resource and the qualification path in the full report. That flexibility gives hyperscalers room to qualify a CXL memory box, attach memory beside a CPU or custom ASIC, or build a dedicated memory system. The economic benefit is that they can add controller and system content around capacity they already own.

Our key take here is: We have increased confidence that the CXL standard will emerge as a preferred additional approach to infrastructure build out. Its economic and TCO benefits, detailed in the full report, along with its flexibility in implementation give it many advantages over other solutions. While we understand the tradeoffs, we believe the biggest customers in the world are positioned to drive the adoption of CXL and have distinct advantages over others in the market. The competitive advantage available to those customers through CXL could become evident.

Inside the Full Report

• Why hyperscalers are likely to become the first large CXL customers, and how their ODM model helps them qualify custom systems faster.

• How much existing memory could become reusable, what that changes for deployment timing, and the potential TCO benefit.

• Where CXL fits alongside HBM, local DDR and MRDIMM, HBF, proprietary memory attach, and networked-memory approaches.

• Our CXL market model through 2030, including the path from single-host expansion to rack-level pooling.

• The beneficiary map across controllers, memory suppliers, interconnect, systems, and software, plus the production signals needed to validate each opportunity.

Read more

In our prior report, The Hyperscaler Capacity Partner Hierarchy, we talked about the importance of partners for hyperscalers, particularly those with a specific type of business model that allowed them to keep a preferred margin profile when they could not secure infrastructure fast enough to convert their backlog. We think, as the market adopts behind-the-meter power and standards emerge, this can accelerate time to power and perhaps shift the speed at which hyperscalers can get land, shells, and other physical infrastructure in place (all things easier to secure and build), then work with a partner like Bloom, and others, to bring capacity online while they still work to get broader grid connectivity.

Grid connectivity is the hardest part of this equation. The hyperscalers that needed to monetize their compute were doing deals with those who had secured power via the grid, often on less favorable terms that impacted their margins, because they were desperate for power. As BTM scales and becomes more viable, it gives them another choice since it can operate without going through all the hoops needed to secure a grid connection. We are not saying they will stop pursuing grid connections. By adopting BTM, they can scale the easier part of the project, if anything in this process is actually easy: getting a shell built and all the surrounding infrastructure in place. That gives them first-party ownership instead of forcing them to go through partners simply because those partners have grid contracts done and ready to go. The maturation of BTM means the option now exists to add the grid later. That is why BTM can cost more and still be the better choice.

Inference changes the equation for BTM

From conversations a year ago, when BTM seemed more like a theoretical with potential but needed to be proved out, a few things have changed. First, the AI infrastructure mix moving from training toward more inference is a big driver. In training, the workloads are bursty and thus prone to massive power spikes versus more nominal and consistent power draw. Inference is less bursty and more manageable as a workload, meaning that having a range of redundancies in place to handle power spikes, which was a challenge for pure BTM, is less of an issue with inference. But even with training workloads now, the latest GPUs and rack-scale systems are putting more power control into their systems to help regulate power spikes and make them smoother. One data center operator we spoke with told us anything around 25% or lower for this kind of power spike was low enough that BTM can suffice, and that the latest rack-scale systems can keep those spikes at 25% or lower.

Public data supports the view from our conversations with those in the power industry. Microsoft production data measured a maximum two-second change of 37.5% of provisioned power for training, compared with 9% for interactive inference. NVIDIA’s GB200 documentation describes programmable power smoothing, while its newer Vera Rubin rack architecture adds more local energy buffering. We would not treat 25% as a universal engineering cutoff. It is one operator’s practical marker, and the broader point is that better rack-level power management gives the onsite generation system a smoother load to follow. Smoother workloads are one enabler of the BTM shift. Grid delays and the economics of bringing compute online earlier remain the larger forces.

Customers have moved from “can we use BTM?” to “how will we use it?”

We are not saying BTM becomes the standard or only source of power, only that BTM was not in a place where it was as viable as it appears now. That waas clear from Bloom’s call where they disclosed every major customer is now engaged to use BTM as a way to get to revenue faster. Customers have moved from BTM looks good on paper to all systems go to ramp BTM. We are in a cycle where the supply chain needs to scale and start to ramp to meet demand so customers can use BTM either to bridge the time to grid service or offset some grid needs and work toward more favorable power economics by using a hybrid architecture. BTM seems to be inflecting, for all the reasons we include, and now the main players need to start ramping their supply chains to meet demand.

From talking to operators in the field, we had consistently heard about the challenge that is electrical integration for the full site. While the equipment that goes into a BTM solution is one piece of the puzzle, the customer still needs to secure labor and testing, then bring together switchgear, transformers, UPS capacity, microgrid controls and a host of other things. The best-positioned companies are those that have already secured the fuel, have utility relationships and have sites ready to go.

Power is only part of the equation when analyzing who captures most of the value. This is why the ability to control the complete delivery schedule and stand behind the contract matters, because any delay has a direct impact on the customer’s time to revenue. The guarantee must cover an operating outcome. That shifts the risk of turning a power plan into operating compute away from the hyperscaler and onto the power or capacity partner.

Bloom is the clearest public proof point for now

Bloom went from essentially entering this market nine months ago, per the call, with one customer, to management saying its technology is now validated and approved by all major U.S. hyperscalers and more than a dozen other AI infrastructure customers. Those relationships span live deployments, booked and shipped systems under construction, and definitive agreements. Customer engagement has clearly inflected. The next proof is whether those approvals become repeat deployments and accepted, revenue-producing critical IT MW.

The reported results now sit behind that claim. Revenue reached $1.065 billion in Q2, up 165.5% from a year earlier, while GAAP gross margin increased to 33.4%. The next proof is repeat orders and accepted, revenue-producing critical IT MW.

A handful of large projects can make the market appear broader than it is. Developers may reserve several generation options for the same campus, while different suppliers count the same prospective demand. Repeat deployments provide stronger evidence. They show that the first installation worked well enough for the customer to use the architecture again. Accepted, revenue-producing critical IT MW then confirms that the project cleared every gate: fuel, permits, generation, electrical integration, commissioning, and usable compute. That would establish BTM as a repeatable procurement model rather than a collection of emergency power projects.

Inside the Full Report

• How BTM gives hyperscalers a faster path to owned compute

• The six gates between a power plan and usable capacity

• Where Bloom leads and how engines and turbines compete

• What Bloom’s Q2 confirms and what still needs proof

• Who captures value across the BTM delivery stack

• How BTM can scale and what would weaken the thesis

Read more

We have spent a lot of time tracking power and utilities because we continue to believe power will set the pace of the AI infrastructure buildout. Our recent conversations with industry sources keep coming back to the same problem. We have known about the power bottleneck for some time. What we are hearing now is that it has not gotten any better and, if anything, is getting worse. Data center demand continues to move faster than local utilities can support it. Grid studies can take six to twelve months, and those timelines are getting longer. A new substation can add another two to three years. A company can sometimes move faster if it pays for the substation work itself, but that makes the project more expensive. Even then, the equipment and skilled labor needed to finish the work are getting harder to secure on schedule. Permitting and local resistance can push the timeline out again. Based on what we are hearing, we think the power constraint is likely to stay with the industry through at least 2030 and probably longer.

As we have heard directly from hyperscalers, this timing problem has been shaping how they think about capacity for some time. Demand is coming in faster than utilities and the normal data center build cycle can support. That leaves them using outside partners that already have firm power or can bring a site online sooner.

Calling a company an power/shell/ landlord, neocloud or data center developer only tells us so much. What we care about is whether the site can actually get power and be delivered on time. Where we may differ from many consensus is our belief that the best fit is often a partner that can bring the power and the building online, then stop there. The hyperscaler still controls the compute and the customer, which keeps more of the economics inside its own business.

Amazon and Microsoft will put the capacity gap back in focus

This report will publish during the same week Microsoft reports fiscal Q4 results on July 29 and Amazon reports Q2 results on July 30. We expect both companies to raise the capital-spending bar and show another increase in contracted cloud demand. The labels differ between backlog and remaining performance obligations, but the economic signal is the same: Google (reported), AWS, and Azure are booking demand faster than internal infrastructure can be delivered.

The chart below puts the imbalance on one scale. Across the companies shown (full universe including meta and neoclouds), backlog and RPO have grown faster than cost-adjusted capex. That pulled the ratio from 44% in 2024 to about 37% in 2026. We would not treat this as a measure of physical capacity coverage because capex is an annual flow and backlog is a point-in-time balance. Still, the direction is clear. Contracted demand has grown faster than the capital response, which is why we do not expect capex to slow anytime soon. Key point to remember - revenue always lags capex.

That leaves the economics of the partner as the main question. Hyperscalers will keep using outside capacity while they build more of their own. On Alphabet’s latest earnings call, management called that capacity a bridge and said it would create modest near-term margin pressure. Google is willing to pay the premium because the revenue is available now. Waiting for its own data centers would mean leaving some of that demand unserved. As Google brings more controlled capacity online, part of the premium should go away. The type of partner, and how much of the stack that partner owns, will determine how expensive the bridge becomes. Alphabet Q2 2026 call transcript

Our prior reports looked at the AI cloud from the supplier side. We mapped where hyperscalers and neoclouds compete, then separated neoclouds by the layers they own.

This report looks at the same issue from the buyer’s side. We think AWS and Azure face the same choice Google described. They can, and will, buy outside capacity now and accept some pressure on unit economics, then move more of the workload into controlled infrastructure as their own supply arrives. The question is who is best suited to the hyperscalers are a foundational capacity partner.

That is why we are spending more time on the partner type. A company that secures firm power and delivers the physical layer can remain an attractive partner after the scarcity premium fades. The hyperscaler still controls the server system and the customer. From here, we want to know which providers can deliver the same kind of project more than once.

External capacity is not one product

A complete compute service and a leased data center shell can both deliver capacity, but the buyer is paying for very different things. A managed compute provider supplies the building and the server system, then operates it. The price has to cover the cost of the hardware and the risk that the provider cannot keep it fully used. It also has to cover the chance that the equipment loses value faster than expected. All of that adds another supplier margin to the cost.

A third-party-owned shell stops at the physical layer. The landlord develops the site and delivers a building that can support the required density. The hyperscaler can still own the accelerators and decide how the network is designed. It also keeps the workload inside its own control plane, which is what the hyperscalers generally prefer. More of the cloud margin stays with the buyer, while someone else owns the slow, long-lived physical asset.

This is the capacity hierarchy we are trying to describe. When there is enough time, hyperscalers prefer to own and operate strategic capacity themselves. A leased shell can get them there sooner without giving up control of the compute. Managed third-party compute still makes sense when the value of serving demand now is greater than the premium being paid.

Exhibit 1. The Hyperscaler Capacity Hierarchy. Relative profiles are Creative Strategies analytical judgments, not reported provider margins. Source: Creative Strategies analysis of Alphabet public disclosures and public infrastructure agreements.

The best landlord owns more than land

Power delivery is the known hook making these partners attractive. The latest disclosures help answer the next question: what turns that power into a valuable contract? Applied Digital delivered another 75 MW on schedule, while Core Scientific had 437 MW billing by mid-July against 1.1 GW leased. More important for our thesis, Applied Digital said direct hyperscaler leases should lower its financing cost over the full term. Core Scientific’s mix of direct AMD leases and neocloud leases with AMD protections shows how the risk can change by project. The value sits in converting power into accepted capacity under a contract the market can finance with a specific margin profile. Applied Digital Q4 2026 earnings call Core Scientific Q2 2026 results Core Scientific and AMD partnership

Hut 8 may add another version. The company has leased 704 MW at Beacon Point to an unnamed investment-grade tenant. The Financial Times reported that NVIDIA is the tenant and may sublease the capacity to neocloud partners, although neither company has confirmed it. If accurate, NVIDIA would be securing physical capacity itself rather than waiting for partners to find it. Hut 8 Beacon Point announcement Reuters summary of the FT reporting

Chip vendors may be starting to secure physical capacity themselves. AMD is doing it directly with Core Scientific, while NVIDIA is reported to be leasing Hut 8 capacity that it may place with neocloud partners. The obvious reason is speed, but we think there may be a competitive angle as well. Hyperscalers are putting more custom ASICs into the infrastructure they control. Securing outside sites could help NVIDIA keep scarce power tied to NVIDIA systems before those sites are absorbed into hyperscaler builds. That is our interpretation, not something NVIDIA or Hut 8 has said, but it would expand the buyer pool for the physical layer.

For now, our perhaps out of consensus view, is we still think the best long-term position may sit with the company that controls the physical asset and stops before the compute layer. Hyperscalers will use neoclouds when speed is worth the premium. Chip vendors may now compete for the same sites. All of them still need the landlord to deliver.

Inside the Full Report For Clients and Subscribers

• An original model for comparing the cost of delay with the premium paid for managed compute.

• A buyer-side margin and capital sensitivity across owned, leased-shell, and managed-compute structures.

• Google and Meta case studies showing how the same buyer uses different structures for different time horizons.

• A contract-quality ladder separating direct hyperscaler leases from backstopped neocloud tenancy and uncontracted pipeline.

• A selected capacity map showing where supply sits, followed by public-company market, beneficiary, and partner-fit maps that separate economic role from announced MW.

• A deployable-MW proof ladder and monitoring framework for testing whether announced capacity can become revenue-producing infrastructure.

Read more

This retrospective draws on a wide range of research from third party sources as well our own. Much of that work saw AI coming. The largest misses came from underestimating how quickly the pieces around the accelerator would have to scale together. Something we now refer to as the GPU Tsunami.

How much the market moved

• Four years early: A 2022 forecast put the semiconductor market at $1 trillion in 2030 (historical framework). Our current model reaches $1.79 trillion in 2026 and $3.49 trillion in 2030. Adding adjacent AI infrastructure brings the combined 2030 silicon-systems pool to $3.95 trillion.

• 5.8 times: Our 2027 accelerator model is now $720 billion, compared with an old $125 billion AI-compute endpoint.

• 60%: One quarter of NVIDIA Data Center revenue reached about 60% of the old full-year 2027 AI-compute endpoint.

• 2.7 to 3.2 times: The 2025 CoWoS capacity estimate was revised from 25,000 to 30,000 wafers per month to more than 80,000.

• 79%: The 2027 AI-switching endpoint was revised higher in less than a year.

• 3.0 to 3.6 times: Current frontier rack power surpassed an old 2030 expectation four years early.

• 2.4 times: Our seven-company 2027 capex model is already 2.4 times an old global data-center endpoint.

We think its important to look back, as a post mortem, to understand where the analysis went wrong and what the collective industry (us analysts) underestimated, so we can recognize those patterns when they appear again. Some earlier calls were right, while others moved too quickly on adoption timing. We also missed how much the unit of analysis was changing as the market developed. We were particularly too conservative on the ASP expansion across the semiconductor supply chain. We fully modeled semiconductor output based on our foundry capacity model, but we did not fully appreciate the pricing leverage created by the AI infrastructure cycle. Our midpoint scenario implies the industry’s blended semiconductor ASP rises from approximately $0.75 today to more than $2.10 by 2028, nearly tripling in just three years. Between 2025 and 2028, we estimate semiconductor unit shipments increase only 25–30%, while the average semiconductor ASP increases nearly 185%, driving approximately 260% industry revenue growth. This represents a structural shift in semiconductor industry economics as value per device compounds substantially faster than industry unit shipments. We map every constraint in the below report leading to this ASP increase, in the era we now call the era of margin expansion.

What the archive saw, and where its boundary failed

We reviewed a broad set of historical forecasts to understand how expectations changed over time. Licensed historical work is described as third-party forecasts or archive estimates, while current comparisons use public outcomes and Creative Strategies models.

The scorecard separates outcomes, run rates, and model revisions. A newer forecast shows expectations moving; only an outcome proves the old forecast wrong.

Each row uses a different type of evidence, but the pattern is consistent. The largest forecast revisions appeared once the market moved beyond the accelerator and began modeling the complete AI system. Compute only becomes usable capacity when the full rack can be commissioned. The early forecasts treated the infrastructure around the accelerator as a supporting input. It turned out to be part of the market itself.

Compute demand escaped the original boundary

In March 2023, a third-party forecast assumed that 40 to 60 large-model builds over the following 12 to 24 months would create approximately $10 billion to $15 billion of incremental GPU TAM. A June 2023 custom-silicon model put AI computing semiconductors at roughly $43 billion in 2023, growing to $125 billion in 2027, with custom ASICs reaching as much as 30% of cloud AI semiconductor spending.

Those estimates were aggressive at the time, but their boundary was still too conservative. The model counted likely builds and the accelerators required, then extended the curve as custom silicon gained share. It did not capture the wider inference market that would form around installed compute. Ultimately, what most got wrong was underestimating the compute intensity of the workload that is agentic AI. We think many still under appreciate this today as well.

NVIDIA reported $75.2 billion of Data Center revenue in Q1 FY2027. Data Center includes systems and networking, so it is not like-for-like with semiconductor TAM. Even so, that quarter reached 60% of the old 2027 AI-compute endpoint. In one quarter, NVIDIA generated revenue equal to 60% of what the earlier model expected from the entire AI-compute market for all of 2027. Looking forward, our current model puts merchant GPU and XPU logic plus custom AI accelerators at $770 billion in 2027 and $1.36 trillion in 2030.

The models underestimated how much demand each new accelerator would create. Training built the installed base, while better performance expanded the market for inference. That pulled custom silicon and rack-scale systems into what had started as a GPU market. As supply grew, the TAM grew with it.

Packaging was the clearest under call

In August 2023, a third-party advanced-packaging forecast put industry CoWoS capacity at roughly 15,000 wafers per month, rising to 20,000 to 25,000 in the second half of 2024 and 25,000 to 30,000 in 2025. The forecast saw little likelihood that CoWoS would remain a meaningful bottleneck beyond 2024.

By January 2025, a later archive model expected TSMC CoWoS capacity to exceed 80,000 wafers per month by the fourth quarter of 2025. That was 2.7 to 3.2 times the 2023 estimate for the same year. In July 2026, TSMC still described packaging as tight enough to limit customer growth. TSMC Q2 2026 materials.

In retrospect, what was missed was the timing, pull-in, of chiplet designs as AI accelerators were the first evidence of the broad shift from monolithic to systems based chip design.

Packaging became part of the capacity boundary, which is the central idea in our report The GPU Tsunami: TSMC, Intel, and Samsung Foundry. A leading-edge wafer becomes useful AI capacity only after advanced packaging brings it together with HBM and the rest of the rack.

Memory became a much larger market than the models allowed (and structural)

Memory was not missing from the early forecasts. It was treated as a cyclical recovery. In late 2023, the public market forecast called for roughly $130 billion of memory revenue in 2024. Actual sales reached $165 billion, 27% above that estimate.

The longer-range forecasts were even more conservative. One public model published in 2024 put combined DRAM and NAND revenue at approximately $206 billion in 2025 and just $214 billion in both 2026 and 2027. The assumption was that the recovery would level off once pricing normalized.

Our current model looks very different. We estimate combined DRAM and NAND revenue of $230–240 billion in 2025, $550–570 billion in 2026, and $800–850 billion in 2027. Against the old forecast, the 2026 market is now modeled at 2.6 times the prior estimate. By 2027, the difference reaches approximately four times.

This is a forecast revision, not yet a realized outcome, yet trending in that direction. But it shows what the earlier models missed. AI did not create demand only for HBM. It increased the amount of server DRAM and enterprise NAND required around each accelerator. At the same time, HBM consumed capacity that would otherwise have served conventional memory. The result was a much broader shortage and a much larger market than the original forecasts allowed.

The bottleneck kept moving

The early work saw the constraint moving beyond the accelerator, but underestimated how quickly the network and site would control the deployment schedule.

A June 2023 forecast put 2027 AI-switching revenue at $8.5 billion. Eleven months later, that endpoint had risen 79% to $15.2 billion. Our broader networking-silicon model now reaches $105 billion in 2027. The categories differ, but reported revenue confirms the direction: Broadcom’s $10.8 billion of quarterly AI semiconductor revenue reached 77% of its old full-year FY2025 forecast. Broadcom Q2 FY2026 results.

Power moved faster still. A July 2024 note expected average rack density to reach 40 kilowatts by 2030. GB200 and GB300 systems reached 120 to 142 kilowatts four years early, or 3.0 to 3.6 times the old endpoint. These are frontier systems rather than fleet averages, but they set the next building standard. GB200 specification and GB300 reference architecture.

Capex followed. A May 2024 model put global data-center capex at $500 billion in 2027. Our seven-company cost-adjusted base now reaches $1.18 trillion, or 2.4 times that endpoint.

The bottleneck moved from the chip into the network and then the site. Each fix exposed the next constraint. That is why Counting Real AI Capacity and Gigawattonomics focus on commissioned systems and economic output per watt. Until the full system is commissioned, announced capacity remains intent.

The top line hid the infrastructure reallocation

A January 2024 semiconductor model forecast sales of $645 billion in 2024 and $718 billion in 2025. WSTS reported $630.5 billion and $795.6 billion, respectively. The model was 2.3% high for 2024 and 10.8% low for 2025. WSTS 2025 result. It caught the recovery but missed the next acceleration in memory pricing and AI-infrastructure mix.

Public cloud was also broadly on track. A June 2024 model forecast spending rising from roughly $675 billion in 2024 to $1.38 trillion in 2028. The surprise occurred inside the total as AI infrastructure grew faster and carried more capital intensity than conventional workloads.

A top-line forecast can land within normal error while the internal demand map changes enough to redirect capital, capacity, and profit. The semiconductor and cloud totals did not reveal how much of the next dollar would be pulled toward accelerators and the infrastructure required to deploy them.

Enterprise AI had two opposite forecast errors

Infrastructure estimates were generally too small. Enterprise adoption produced two errors in opposite directions. Spending and usage intensity ran above the early markers, while scaled-production timing was sometimes too aggressive.

Two 2023 third-party forecasts framed an $820 billion enterprise-software TAM and approximately $150 billion of GenAI software spending within three years.

Our current working range puts enterprise GenAI spending across software, services, and inference compute at $175 billion to $200 billion in 2026, rising to more than $600 billion by 2030. The boundary is broader than software alone, so the comparison is directional. Even with that caveat, the 2026 midpoint is about 25% above the old three-year marker, and the decade-end range is more than four times as large.

The survey archive explains the other error. A 2024 IT survey found 52% of respondents live with at least one AI use case, but only 10% in production at scale. Later CIO surveys measured production at 30% in 2024 and 39% in 2025, while an earlier CTO survey expected nearly universal use by year-end 2024.

The error was treating a company with one live use case as equivalent to a company that had reorganized a production workflow. An enterprise can pay for AI and increase token consumption while only a small number of workflows reach scaled production. In our report, From AI Usage to AI Earnings Power picks up there. The test is whether repeated use changes a measurable operating baseline enough to earn a durable budget.

Where the current Diligence Stack work goes next

The current research starts from this revised unit of analysis. Counting Real AI Capacity establishes what has actually been commissioned. Gigawattonomics then asks whether that capacity can earn an acceptable return from the power it consumes.

Below the site, The GPU Tsunami applies the same capacity discipline to foundries. Where AI Constraints Become Pricing Leverage asks which shortages can support durable economics rather than temporary scarcity pricing. The memory and enterprise reports carry the method into capacity allocation and workflow monetization.

Bottom line

As a firm that has been doing market models, forecasting, sizing, and more for over 40 years, we understand being conservative, and how forecasts are almost always wrong. A forecast is only as good as the underlying assumptions we stay in a state of constant learning, observing, and looking for the right past and present patterns in order to continually strengthen our assumptions on every part of the industry we research.

We shared a lot of our internal forecasts and models in the look back, knowing full well they may be wrong and also pose the scenario, within the historical view, that what if even today we are not bullish enough.

For subscribers

Below the paywall, we share our notes from attending AMD Advancing AI and having time with management to discuss products and strategy and Intel earnings with feedback from our calls with IR. The Intel note explains why the earnings recovery may arrive before an external foundry win, with custom ASICs emerging as another business to watch. The AMD note looks at Helios and whether its open rack-scale design can turn strong component technology into a coherent system.

Read more

We have been thinking about cybersecurity as a much deeper enterprise function than the word security tends to imply. The same framework can be extended to sovereign nations. Closed and open models, including models developed outside a company’s or country’s borders, are becoming capable enough that persistent threats will be a fact of operating life. Those threats will reach a company through its products and systems. They will also reach the company’s data and, increasingly, information tied to its employees. The question is how enterprises and governments defend themselves when the tools available to attackers keep improving.

Over the last few months, we have spent considerable time speaking with enterprise decision-makers about agentic deployments, including through our own CIO/CTO survey and conversations with large corporate customers. One concern kept surfacing: security, and cybersecurity in particular. As capable open models continue to advance, the cost and technical barrier for bad actors will come down with them, increasing the threat surface corporations have to defend. For that reason, we think security will be one of the earliest and most durable sources of pull-through from enterprise AI adoption. Security attaches to an AI project as it moves into production, then returns through the existing cyber budget at renewal.

While the early evidence indicates spending is unlikely to arrive as one neat new AI-security category. Instead, AI increases the burden on security controls companies already have in place. In conversations on this with stakeholders one of the first challenges we hear brought up is data access. Before an enterprise can put a model into production, it needs to know what information the model can retrieve and whether that information should be available to the person making the request. Identity becomes more important when an agent begins to act, and runtime controls enter the picture once those actions reach production systems. Some of this need will create new products, but a meaningful share may appear as deeper use of platforms customers already own. That makes the demand easier to see than the eventual revenue pool. The need for greater control can become obvious well before investors can measure where the value is accruing. We also have outlined the need for a new class of compute, to go with a new class of models, or model variants specific to security, cyber security.

Open models raise the market floor

The external threat environment pushes that same budget in a secondary direction. Our companion research on China’s AI stack argues that China does not need semiconductor parity at every layer to keep model development moving. Adequate sovereign compute, paired with competitive open-weight models, is enough to widen access to capable AI. For cybersecurity, capability diffusion does not wait for chip parity. Any open source model, regardless of where it comes from, expands the pool of models available outside controlled services. That linkage has a clear limit: model access alone does not create a successful attacker.

Strategic competition gives nation-states a reason to keep investing in that capability. Enterprises absorb much of the operating cost because their systems are common targets. Models can package parts of reconnaissance or exploit adaptation into tools that are easier to use, which may raise attempted attack volume faster than a human-led defense process can scale. Defenders gain from the same models, but they still need a control layer that can operate at machine (agent) speed.

Exhibit 1. AI Deployment Is Outrunning the Control Layer

The budget converts through existing control points

A direct survey in the source set points in the same direction: deployment is running WELL ahead of dedicated AI-security tooling. The gap is wider if the standard is the full control path around data access and agent behavior. Survey definitions differ, so we do not treat the size as a universal market estimate. The exact number is less insightful than the behavior it reveals. Enterprises are deploying AI before they can fully explain how it behaves and some of the unintended consequences of an unstructured deployment.

Vendors that already own enterprise context or an enforcement point start with an advantage. A data-security platform can attach to a copilot rollout because it controls what the model can retrieve. Identity becomes relevant once agents receive permissions to act. Broader security platforms can spread AI-assisted workflows across an installed base, although the economics remain unclear until customers pay more or use more of the platform.

For stakeholders, a product announcement only shows that a vendor is participating. Continued use through renewal shows whether customers are willing to keep paying. The evidence so far suggests AI will expand cybersecurity spending because every production workload creates more systems and activity to secure. Much of the early spending is likely to flow through vendors already embedded in how companies protect their systems.

The rise in external threats may create a separate revenue opportunity at the model layer. As attackers gain access to more capable models, large enterprises and sovereign governments will need defensive models that can operate at the same speed. We think frontier labs such as OpenAI and Anthropic could build (may already be building) or license restricted models specifically for cyber defense. Some of these capabilities may be too sensitive for broad public access, giving the labs a direct path into enterprise and government security budgets.

Inside the Full Report

• A two-pool analysis separating AI for Security from Security for AI, including the different maturity curves and budget sources.

• The seven-layer Enterprise AI Control Layer map, showing where identity, data, runtime, and access controls sit around production AI.

• A category conversion matrix distinguishing high-probability budget pull-through from capabilities likely to be bundled.

• A vendor-positioning exhibit that maps established control-path platforms, specialists, emerging options, and exposed point tools.

• A capability-diffusion framework connecting sovereign AI and Chinese open-weight releases to the enterprise security-spend ratchet, without assuming chip parity or proven attack causality.

• A falsification and monitoring framework centered on paid attach, dedicated budgets, machine identity, platform consolidation, and realized SOC productivity.

Read more

Demand after K3 launched indicates the interest, to at least test and evaluate. Moonshot neared its server limit and paused new subscriptions while it added capacity. A useful model brings more users, and every user adds load to servers and chips. Model progress therefore raises the need for more compute.

We think China sees AI as an arms race between countries. China will not make these decisions only on the basis of near-term financial returns, but they need their AI economy to thrive. The goal is to keep Chinese AI running even if access to outside suppliers is cut off. China will accept higher power use and lower performance per chip if that gives it more control. It may also build capacity before the business case looks attractive.

This is why Kimi K3 and GLM-5.2 belong in a China semiconductor primer. China’s models are improving faster than its chip factories (true of US labs as well to which we estimate AI software is at least two years ahead of hardware and US frontier model labs have to keep meeting hardware capabilities where they are). Each useful model creates more work for Chinese clouds and more reason to make that work run on Chinese chips. Open models can also run outside China, so overseas use may still benefit foreign cloud and chip companies—maybe. Inside China, though, better models create more demand for the local infrastructure below them which in turn helps local silicon roadmaps.

China does not need a chip that matches NVIDIA on every measure. In fact, they may never match leading US chip companies in capabilities. It needs enough local compute to keep its AI work moving if outside supply becomes harder to get. Its largest gaps are advanced chip production and HBM. China also still needs foreign tools to make leading chips. Those gaps are barriers, but they do not stop China from building a useful domestic AI system and economic upside for their local cloud providers.

The full report is our most current read on China compute, infrastructure, and models. We separate the parts China can already supply from the choke points it still cannot replace, then ask what would prove the local stack works at scale. Which is a key metric we are tracking.

Inside the full report

• A layer-by-layer map of where China can rely on local semiconductor supply and where foreign tools still set the limit.

• A China-versus-U.S. heat map for server CPUs and AI accelerators, including where Hygon and Huawei Kunpeng change the local picture.

• How Huawei’s full-system approach compares with independent accelerators and custom chips built by China’s largest internet companies.

• How China can use more chips to make up for weaker performance, and whether local switches and optical links are strong enough to support the larger system.

• Why Kimi K3, GLM-5.2, Qwen, and DeepSeek can create demand for Chinese infrastructure even when the models are open.

• Our planning ranges through 2030, the proof points we are watching, and the evidence that would make us change our view.

Read more

Power Sets the Deployment Schedule

If you have followed our research, we have tried to maintain as close a view as possible on returns on invested watts. Power availability determines when capacity can be deployed, while revenue per watt determines whether that capacity earns an adequate return. We have consistently framed infrastructure decisions through tokens per dollar and tokens per watt because power is the binding budget. When the power envelope is fixed, the platform that turns each watt into the most useful monetizable output creates the better economic return. Revenue per watt is therefore the grounding metric for compute capex.

We built the Gigawattonomics model to maintain that lens on a common facility boundary. It begins with one facility GW, bridges that to IT and accelerator-rack power after cooling and other site load, and then asks how much useful compute reaches a customer or internal product. The financial question follows: what can each active watt earn on a given platform, and how much gross profit remains after the infrastructure cost is carried? Revenue per watt ties pricing to utilization. Intelligence per watt is the strategic output, but it becomes valuable only when published performance turns into useful work. Maintaining the model as a time series lets us track which infrastructure and product roadmaps improve that economic conversion.

The Competitive Read

Our Gigawattonomics model informs how we read NVIDIA’s competitive position in particular. NVIDIA carries higher modeled capex per GW than the leading custom-silicon cases. Its newest systems also carry a premium to the announced AMD alternative. Architecture closes part of that gap; software maturity and deployment certainty have to close the rest. For customers serving a changing workload mix, those advantages can justify the premium because they improve time to revenue and reduce the risk that expensive capacity sits idle. The premium becomes harder to defend when a customer has enough internal volume to tune the full stack around a stable workload. As we have argued in prior reports, vendor versus custom is ultimately a workload-specific decision.

Google’s TPU platform is the most demanding custom-silicon comparison in our current work. Its modeled cost and compute density create a real hurdle for NVIDIA, especially inside workloads Google can keep highly utilized. The evidence is less complete once the comparison moves beyond Google’s internal environment. Production rack power and useful goodput are still not disclosed on a fully comparable basis. We would take the direction seriously without treating the current ranking as settled.

AWS has a different strategic objective with Trainium. It does not need Trainium to replace NVIDIA everywhere for the program to create value. A credible internal accelerator gives AWS another cost curve for cloud inference and more control over service pricing. Against the current Trainium-class systems, NVIDIA’s density and software breadth appear capable of earning back much of the capex premium. Future generations could change that math, which is why this needs to be maintained as a time series rather than published once and left alone.

AMD remains the merchant alternative with the broadest opportunity to benefit from customers that want a second source without taking on a fully custom stack. The announced Helios system screens well in some scenarios, but the public dense-versus-sparse performance basis remains ambiguous. We are not ready to make a precise cost-per-compute ranking until production power and comparable workload data are visible.

Custom silicon is most compelling when the workload is stable enough to justify control of the architecture. The buyer can trade software flexibility for lower technical capital per unit of useful output and capture the benefit inside its own product. That trade is harder than a capex table makes it look. Compiler work, developer migration, and lower utilization during the ramp can absorb a meaningful part of the hardware advantage. A merchant platform retains value because it spreads those costs across a broader customer and workload base.

Memory and the Revenue Identity

The other underappreciated part of the analysis is memory. Capex per deployed peak dense FP8 exaFLOP falls quickly across the new platform roadmaps, while capex per deployed HBM bandwidth improves much more slowly. This is why the report carries a separate bandwidth curve. Inference can become limited by data movement before peak arithmetic is exhausted, giving HBM suppliers exposure across NVIDIA and custom designs. The earnings benefit still depends on supply discipline. More capacity can preserve the technical value while compressing the margin available to suppliers.

The final distinction is how the watt gets paid and who owns the shell. One-time rack revenue should not be compared with annual compute rental or the gross profit created inside an integrated product. Each path has a different cost base and duration. Our Gigawattonomics model keeps those revenue identities separate, then applies the selected ownership structure before testing the return. In the leased-shell case, the operator expenses occupancy and measures payback against technical capital. The owned-campus case removes that lease and asks the same operating revenue to repay the full site. That keeps a strong revenue-per-watt result from being mistaken for a complete return on the campus.

Our Current View

Our current view: NVIDIA remains highly competitive even with higher capex per GW because customers buy usable output and deployment certainty. Custom silicon becomes more compelling as workload stability and internal scale rise. For a given workload, the economic winner should be judged by gross profit per active MW, not the lowest hardware bill.

Inside the full report

• A normalized facility-GW comparison of shipping and announced NVIDIA, Google TPU, AWS Trainium and AMD platforms.

• The capex-per-compute curve, including the adjustment from facility power to accelerator-rack power.

• A separate HBM bandwidth curve showing why the byte is not deflating at the same rate as the FLOP.

• The operating premium NVIDIA must earn back against Trainium and TPU alternatives, with announced systems kept outside the base case.

• Selected outputs from the v1.2 Gigawattonomics model across merchant rental, inference services and integrated custom silicon. The analysis shows how useful-goodput realization and ownership structure affect the return.

• A quarterly monitoring framework for fleet repricing, useful-life policy and gross profit per active MW as new platforms enter production.

Institutional Gigawattonomics Model Access

The full report presents our analysis and selected model outputs. Access to the maintained interactive model is licensed separately for institutional clients.

Model clients can change the operating assumptions, compare ownership structures and run their own platform cases against our maintained cost basis. We update the model as production pricing, power requirements and realized-goodput evidence become available.

Contact us to discuss institutional model access.

Read more

Business model dictates strategy. That has always been one of the more useful observations in technology analysis, and it provides the way to interpret Meta’s AI buildout.

Meta is an advertising company. It distributes its products to billions of people at no direct cost, monetizes their attention, and uses technology to increase the value of each user and each unit of engagement. That economic model should remain the starting point for evaluating how Meta deploys its AI infrastructure.

External Compute as an Extension of the Model

Press reports in early July suggested that Meta was building a dedicated compute business that could sell model access and raw AI infrastructure to outside customers. The market read that as a possible cloud pivot and a new revenue stream capable of helping fund the company’s infrastructure buildout.

The two parts of that story now sit in different places. Meta has confirmed the Meta Model API, giving developers paid access to Muse Spark 1.1 at $1.25 per million input tokens and $4.25 per million output tokens. Meta is therefore selling inference by the token for the first time through its own developer platform. Zuckerberg has also said that entering cloud computing is on the table and that companies regularly approach Meta about access to its models and available compute. (Meta AI⁠)

What Meta has not confirmed is the reported Meta Compute unit or a plan to sell raw infrastructure capacity at scale. Those plans remain based on news reports that Meta declined to comment on. We therefore model the API as a product, the broader cloud ambition as directionally confirmed, and the specific compute business as a thesis that still requires operating evidence.

The pricing provides some indication of how Meta may approach the market. Muse Spark costs roughly one-quarter of the input price and less than one-fifth of the output price of the flagship APIs offered by OpenAI and Anthropic. Meta appears willing to use its infrastructure scale to drive token volume, developer adoption, and utilization rather than protect a high initial API margin. That is consistent with a company whose existing business was built by distributing products broadly and monetizing the activity that followed. (developer.meta.com⁠)

Selling available capacity is not, by itself, a pivot away from advertising. Meta’s internal engagement and advertising workloads should remain the first priority on its infrastructure because that is where compute currently generates the clearest economic return. External sales can improve utilization around that internal demand. The API is the more strategically important step because it establishes a product, a price, and a direct developer relationship.

The possible connection between the two should be easy to see. The API creates external token demand, and a broader compute platform would provide the infrastructure through which Meta serves that demand. What remains unresolved is whether Meta intends to build a full cloud platform, sell selected blocks of excess capacity, or use low token pricing mainly to distribute its models and keep more of the inference workload on its own infrastructure.

Where the AI Return Should Appear First

The business-model lens also tells us where the more immediate evidence should appear. In the first quarter, ad impressions increased 19% and average price per ad rose 12%. Both improved at the same time. That combination suggests Meta is expanding the amount of monetizable activity across its platforms while also increasing the value advertisers place on each impression.

AI is likely contributing to both sides of that equation. Better recommendations can create more engagement and more available inventory. Better targeting, creative tools, measurement, and conversion performance can increase advertiser returns and support higher auction pricing. This is the clearest current expression of Meta’s AI economics: increasing the revenue generated from a user base that already exists at enormous scale.

Our full report builds an ARPU framework around that relationship. The model separates underlying growth from the incremental uplift that must come from continued improvements in engagement, ad relevance, conversion, and pricing. The thesis remains tied to the reported results. If AI is generating an attractive return, that contribution should continue to appear in impressions, price per ad, ARPU, revenue growth, and eventually margins. If the uplift weakens while infrastructure costs continue rising, the model has to adjust.

The Cost of the Compute Buildout

The scale of the spending makes that test more important. Reports citing internal planning point to approximately 7 GW of compute capacity this year and roughly 14 GW in 2027. A buildout of that size would make it difficult to maintain the earlier assumption that capital spending reaches a clear peak in 2027 and then declines quickly.

The better question is how the economics develop while the infrastructure base continues expanding. Depreciation is the mechanism through which the cost of the buildout reaches the income statement. On the heavier investment path, depreciation rises quickly and remains elevated for several years. Meta therefore needs the profit generated by AI to compound faster than the cost of carrying the infrastructure required to produce it.

The 2027 and 2028 Earnings Debate

That balance is likely to become one of the central earnings questions for 2027 and 2028. The answer depends partly on revenue growth and partly on asset utilization, useful lives, and the pace at which new generations of AI systems replace older ones. The industry still has a wide range of views on how long an AI server remains economically productive. Meta has lengthened the useful lives of some server and network assets. Those accounting decisions reflect assumptions about utilization and obsolescence, and they can materially affect reported earnings during the heaviest years of the buildout.

What the Current Valuation Requires

Our expectations analysis suggests that Meta’s current market value already assumes that a meaningful share of the modeled AI return arrives on schedule. Some additional value may also be forming around a compute business that the company has not yet confirmed.

The next phase of the thesis therefore requires more evidence. The first step is whether Meta confirms an external compute offering. From there, the relevant questions concern the duration and economics of the contracts, the identity and credit quality of the customers, the amount of capacity being committed, and whether demand comes from enterprises using the infrastructure directly or intermediaries reselling it.

The reported production timeline for Meta’s Iris silicon and the company’s next capital-spending update provide additional points of validation. Together, these disclosures should help establish how much infrastructure Meta intends to build, how much of it is supported by internal workloads, and whether external monetization is becoming a real part of the economic model.

Inside the full report:

• The capacity waterfall: how we turn reported capacity plans into sellable capacity, and why applying one revenue-per-GW figure to the whole pipeline is the most common modeling mistake on this stock

• Our rebased CapEx and free cash flow path through 2030, including the scenario where spending never flattens

• A year-by-year depreciation model showing what a shorter server life does to operating margin, and the single disclosure that would settle the useful-life debate

• The AI ARPU model: base versus uplift, calibrated to Meta’s reported pricing history, with the quarterly test that would prove it wrong

• The agentic commerce build, sized against Meta’s newly launched subscription pricing ladder

• An expectations map, in operating terms, showing what today’s $1.67 trillion market value already assumes across bear, base and bull paths

• The demand-side bear case almost nobody is modeling: what agentic discovery does to the feed auction that sets ad pricing

• The short quarterly KPI list that will decide the thesis

Read more

There is more leading-edge wafer demand than TSMC can serve alone. Our ballpark estimate is that demand is running at 110–120% of available capacity, with AI accelerators, custom ASICs, mobile, and CPUs all competing for the same advanced-node base. Samsung and Intel do not need to displace TSMC to matter. They need to become usable where TSMC is constrained, expensive, or strategically concentrated.

The other dynamic to appreciate is a change in what the customer buys. Wafer supply, process leadership, and cost per transistor are still important, but they describe a shrinking share of the purchase decision process. Foundry is moving from chip-level manufacturing to system-level capacity. Customers are buying a more complete path to deployable AI silicon, where wafer supply, advanced packaging, memory integration, qualification, delivery timing, and acceptable geography have to line up together.

Chiplets (disaggregated design) are what make this chips to systems platform shift evident. They are still rare by count, likely under 1% of total semiconductor units and perhaps 10–20% of HPC/AI processor shipments, but they already carry 20–35% of semiconductor revenue and an even higher share of leading-edge compute value. In the angstrom era we expect more high-performance silicon to be built as systems of dies, memory, and packaging rather than as single monolithic chips. A dynamic that benefits Intel and Samsung.

TSMC remains the default. It leads in market share, profit share, yield, ecosystem depth, and packaging scale. AI has made that position more valuable. When compute revenue is gated by silicon supply, customers have less room to resist price increases. That is why advanced-node pricing power has become more visible.

Packaging is the second moat. The company that controls the packaging slot helps decide which customer ships. TSMC sells the wafer and the slot as one platform, and our model suggests packaging is the faster-growing layer of that platform. The market still tends to score foundry on wafer share. We think the scoreboard now includes packaging capacity, HBM integration, chiplet assembly, supply assurance, and geography.

That is where Samsung and Intel become more interesting. They are usually framed as second sources, but that misses the jobs customers may hire them to do. The question is not only who has the best transistor. It is what problem the customer needs solved.

Samsung is the integration option. Logic, HBM, packaging, and Taylor give it a turnkey AI pitch. For custom ASIC teams, automotive programs, and customers constrained by HBM access, one accountable supplier has value. The gaps are still real: advanced-node yield needs to improve, and external packaging capacity remains far below TSMC’s. Tesla and Groq help validate the direction, but they are not yet volume proof.

Intel is the packaging and geography option. External foundry revenue is still small, so the near-term wedge is not broad wafer share. It is EMIB-T, advanced packaging, and US supply assurance. A hyperscaler can send packaging work to Intel without committing a flagship die to an unproven external node. That is why judging Intel only on 18A wafer share misses the first part of the ramp.

The swing factor is TSMC. If CoWoS and SoIC scale fast enough to absorb the bottleneck, scarcity fades and challenger urgency falls. That same data point is TSMC upside and Samsung/Intel downside. The remaining opportunity for Samsung and Intel sits in the lanes where they can compete on technical merit and become first choice for a specific job, rather than simply second source to the default platform.

If interested in the full foundry model, please inquire.

What’s in the full report:

• The TSMC platform revenue bridge, 2025–2030: leading-edge wafer revenue and advanced packaging modeled separately, with the attach ratio that most coverage misses.

• The 2028 scenario model: bear, base, and bull platform revenue for TSMC, Samsung, and Intel, with the yield and packaging variables that swing the challenger numbers threefold.

• The advanced packaging capacity roadmap: CoWoS/SoIC, EMIB/EMIB-T, I-Cube/SAINT, and OSAT spillover, year by year through 2030.

• How to assess the three foundry exposures: the evidence each name reprices on, and on which quarterly clock.

• A signal and read-through table for the next four to six quarters: which data updates move which thesis, and in which direction.

• A confidence-weighted map of twelve customer engagements: what is confirmed, what is only reported, and what should not be modeled at all.

• What would change our view: the specific falsifiers for each name, including the one TSMC data point that would damage the thesis most.

Read more

For most of the last twenty-five years, unit shipments ran well ahead of industry revenue. The industry scaled unit shipments faster than dollars. Then AI hit.

The chart below shows this shift in action. Semiconductor revenue is moving toward a much steeper path than unit shipments, which means more of the growth is coming through ASP, mix, complexity, and supply-chain scarcity rather than device volume alone. That spread is what we refer to as the pricing-leverage zone.

The underlying semiconductor supply chain is not (can not) scaling at the same pace as AI demand. From 2025 to 2030, semiconductor industry revenue grows roughly 22%–24% per year, while unit shipments grow roughly 5% per year. Revenue is growing about 4–5x faster than units. Every part of the supply chain ecosystem from the obvious things like HBM, advanced packaging, substrates, board materials, power analog, passives, test, metrology, and specialty materials all have to scale together. When the unit base cannot expand fast enough, the constrained layers below the shipment line capture more economics.

This report maps where that happens. Our first report showed where AI capacity slips. This report shows where those same constraints become margin expansion across the semiconductor supply chain.

Our central argument is that margin quality depends on the constraint behind the price. Many suppliers can raise price in a cycle this tight. Fewer can convert those increases into durable margin, and the difference shows up in the evidence we are tracking. Durable pricing leverage forms where customers have limited alternatives, qualification cycles run long, and delay carries a direct deployment cost and risk to revenue. Those conditions push buyers into long-term agreements, prepayments, and reservation fees, and the result appears as financial flow-through in blended ASP and gross margin. Temporary tightness produces price action too, but through weaker channels: spot spikes, input-cost pass-through, utilization recovery, and restocking.

Applied across the supply chain, the framework concentrates the highest-quality pricing power in memory and HBM, advanced packaging and substrates, and the qualified high end of high-speed board materials. AI-rack power analog and server-grade passives grade as real but SKU-specific and architecture-dependent. Power discretes, mature-node foundry, commodity passives, and undifferentiated materials carry weaker margin signals — more cyclical, closer to pass-through, or dependent on utilization rather than pricing.

The map is built to be tracked regularly. Confirming evidence sits near the top of the evidence ladder: LTAs, prepayments, price floors, allocation, and gross-margin flow-through. Disconfirming evidence shows up in the areas we are tracking as well: capacity additions, China supply, spot rollover, shorter contract durations, and easing lead times.

What’s in the full report

• A five-level pricing evidence ladder that grades every price move by how much confidence it deserves — from lead-time chatter to LTAs, prepayments, and gross-margin flow-through

• Observed pricing trends by segment: a July 2026 baseline of triangulated price ranges across the chain — memory contract moves, substrate and CCL increases, analog catalog actions, MLCC spot, and foundry ASP — with direction graded for each

• A constraint-intensity vs. margin-quality matrix covering eleven segments, separating durable pricing power from pass-through, restocking, and utilization recovery

• The pricing leverage heat map: where price action is most likely to become durable margin across memory, packaging, board materials, power, passives, foundry, test, and materials

• Segment-by-segment analysis with names to track, the pricing mechanism behind each, and the key signal that confirms or breaks it

• A beneficiary map organized by how value is captured — structural price owners, mix beneficiaries, utilization recovery, pass-through, and cost absorbers

• The negative-evidence file: where customer pushback, China capacity, spot rollover, and EV digestion cap pricing that looks stronger than it is

• A confirmation/disconfirmation dashboard, anchored to the July baseline, to track every thesis as the cycle moves from capacity announcements to margin reality

Read more

We have been tracking a constraint stack that now runs much deeper than accelerator allocation. The map in this report pulls those constraints into one framework and asks a simple question: which layers determine when announced capacity becomes billable compute?

Capacity has to be counted by synchronized deployability. A purchase order becomes revenue only when every layer arrives at the same site in the same window. That means the accelerator and memory. It also means the transformer, cooling loop, networking fabric, power equipment, EPC capacity, and crew that commissions the rack. Chip allocation set the pace earlier in the cycle. The limiting layer has since moved deeper into the physical and operational stack and a timeline slip in any one of these deeper chokepoints could slow down many other parts of the cycle.

That is why announced megawatts, GPU orders, and hyperscaler capex need to be read through a deployment lens. Those announcements are helpful from the view that they measure intent. The work we maintain a focus on is determining how much of that intent becomes real capacity, which layer slows the conversion, and which suppliers control the scarce inputs the rest of the stack depends on.

Semiconductors remain the heart of the system. Advanced packaging, HBM, and the substrate and validation layers around them remain the critical silicon constraints. The map adds what decides whether that silicon ever becomes billable: power and cooling, construction and permitting, networking and commissioning. A campus can hold its GPUs, racks, transformers, and substation and still produce no revenue, because the cooling loop has not passed first-fill or the only crew that can certify the rack is on another job.

Central to much analysis is the tracking of GW or ones ability to monetize each unit of power. This is why we point out that announced megawatts can be misleading. A project can be announced, land-secured, and sitting in an interconnection queue while still years from billable compute. Some of these slip. Others get resized, repriced, or pushed into a later vintage. Treating the full announced pipeline as near-term supply overstates capacity, and it understates the operators who locked up power and equipment early.

We define and track capacity differently. A real megawatt is power-secured, equipment-procured, commissioned, AI-ready, and billable. Scored that way, the 2027 and 2028 deployable pool looks smaller and more concentrated than the public pipeline implies. Much of that advantage was set by procurement decisions made back in 2024 and 2025, before most investors were watching the full stack.

AI infrastructure is a synchronized systems problem, and a GPU order is one gate in a longer chain. We track fourteen gates between a GPU order and billable compute. Most of them sit in the physical and operational layers, with GPU and ASIC allocation near the end of the line rather than the front. Ranked by deployment severity, the hardest gates now sit outside the chip.

The full report separates constraints that delay deployment from constraints that mainly raise cost. A transformer shortage holds revenue offline; a copper-foil price increase only raises the cost of a board. A substrate shortage can stop accelerator shipments, while an MLCC spike can be real pricing power without blocking a single megawatt. Read every tight component as a capacity constraint and you end up with the wrong operator ranking and the wrong supplier map.

We map the stack by role: constraint owners, constraint-exposed companies, scarcity arbitrageurs, and deployment-risk reducers. The main question this cycle is who can convert announced capacity into real capacity, and who controls the scarce inputs the rest of the stack depends on.

What paid subscribers get in the full report

• The Real MW framework: how we discount announced capacity into power-secured, equipment-procured, AI-ready, and billable capacity.

• The fourteen-gate deployment map: the full chain between a GPU order and revenue-generating compute.

• A ranked Top 10 deployment-gating constraint table: including interconnection, transformers, turbines, switchgear, UPS, substations, advanced packaging, HBM, liquid cooling, commissioning labor, and permitting.

• The paper MW to real MW funnel: why the announced pipeline shrinks materially as projects move toward revenue.

• The deployment dependency chain: how semiconductor and physical infrastructure constraints clear on different timelines.

• The power stack broken into four separate bottlenecks: generation, transmission/interconnection, equipment, and regulation/politics.

• The below-accelerator constraint layer: PCBs, CCL, glass fabric, copper foil, PMIC/BCD capacity, probe cards, sockets, bonding/debonding tools, metrology, and cooling interfaces.

• A deployment-gate vs. cost-inflation classification: which constraints actually delay capacity and which mainly raise cost.

• The constraint ownership map: constraint owners, constraint-exposed companies, scarcity arbitrageurs, and deployment-risk reducers.

• A framework for where constraints become pricing power: the bridge into our next report on bottleneck leverage across the AI supply chain.

• The monitoring dashboard: what to track next across transformer lead times, turbine orders, switchgear and UPS availability, substrate LTAs, HBM tool backlogs, cooling interfaces, commissioning labor, prepayments, slot reservations, and price floors.

Coming next in Part 2: where constraints become pricing leverage

This constraint map sets up the next layer of work. Once we know where the bottlenecks sit, the follow-up question is which constrained layers can convert scarcity into price, margin, prepayments, reservation fees, long-term agreements, mix uplift, or lower discounting.

The next report will separate true scarcity rents from input-cost pass-through, mix-driven ASP uplift, and temporary restocking. A transformer shortage, an HBM shortage, an ABF substrate shortage, a cooling-validation bottleneck, and an MLCC price spike should not be treated the same. Each has a different mechanism, duration, and margin implication.

The first report maps where capacity slips. The next report bridges who has leverage and price accordingly due to their constraints.

Read more

We attended Qualcomm's Investor Day, spent time with management, and participated in the Q&A with Akash Palkhiwala, Cristiano Amon, and Tony Pialis. We came away believing the data center opportunity is now considerably more tangible than many previously appreciated. Positioning wise, we still view Qualcomm first as a semiconductor engineering company, but data center has become the next platform leg of that engineering story.

Qualcomm used Investor Day to put a different revenue curve in front of investors. The company had already been moving beyond handsets through automotive, IoT, PC, XR, and edge AI, but that version of the story, on the surface, still looked like offset work. Automotive and IoT could help absorb Apple modem loss, Samsung mix pressure, memory-driven Android weakness, and periodic QTL renewal concern. That improved the quality of the business, but it did not force investors to place Qualcomm in a different category.

Chart for visual effect. We detail our entire model and assumptions for each revenue case in the full report and our estimate scenarios carry out to 2030.

As the chart above visualizes, the data center framework changes Qualcomm’s growth trajectory. Management raised its FY29 non-handset QCT target to $40B, roughly twice the prior FY29 target, and put a more than $15B FY29 data center target inside that number. CFO Akash Palkhiwala also made the mix change direct: by FY29, handsets are expected to fall toward roughly one-third of QCT revenue. That is the cleanest stat from the day on how the diversity story has evolved with data center now in the picture. Qualcomm can still be one of the most important mobile silicon companies in the world, while the investment argument increasingly depends on whether it becomes an edge-to-data-center AI compute platform.

We came away from the event and follow-up Q&A believing the data center narrative is more concrete than many investors assumed going in. The FY27 anchor is custom silicon-led, with two global hyperscale customers each contributing at least $1B according to management’s Q&A comments and both with multi-generational programs planned. The product cadence then layers in HBC-based AI acceleration and C1000 server CPUs. This sequence reduces the burden on any one product line. Custom silicon creates the first revenue line, CPU creates the more lasting hold on the socket, HBC gives Qualcomm an inference architecture of its own, and Alphawave adds the I/O, die-to-die, SerDes, optical, and chiplet assets that make the platform story more credible.

The custom Arm CPU point deserves more attention than it has received. Agentic AI increases host-side work inside the data center. Tool calls, retrieval, API routing, state management, security, scheduling, and accelerator coordination all run through the CPU complex. If hyperscalers want a standard Arm server CPU path, Arm CSS is available. If they want a more specialized Arm CPU with custom cores, chiplets, high-speed I/O, memory attach, and implementation help, the external partner list narrows quickly. Qualcomm’s Oryon work, architecture-license position, mobile-to-auto CPU experience, and Alphawave connectivity assets give it a credible claim to that role.

HBC is the second technical piece to understand. High Bandwidth Compute is Qualcomm’s answer to the inference memory bottleneck. Traditional accelerator systems spend power and packaging budget moving data between compute and external memory stacks. Qualcomm’s approach places the XPU under DRAM stacks so the compute sits closer to memory. The claim is SRAM-like performance with DRAM-class density, with better bandwidth per watt and capacity per watt across different inference workloads. The business read-through is cost per token, rather than a generic accelerator benchmark. We view this as interesting, needing further proof, but we know much of the industry has been circling around how to do near memory compute, mostly in RND, so this could be validation for that approach which will also make LPDDR a strategic part of compute packaging.

Our scenario model frames the change. The bear case takes Qualcomm to roughly $61.5B of FY29 revenue with $10B from data center. The base case, which largely follows management’s Investor Day framework, reaches roughly $73.6B of FY29 revenue with $15B from data center. The bull case reaches roughly $89B of FY29 revenue with $22B from data center, driven mainly by custom Arm CPU absorption and HBC/connectivity attach above the initial guide. In the base case, Qualcomm grows well beyond the pre-event low/mid-$40B revenue framing. In the bull case, the business is roughly twice its FY25 revenue base by FY29.

That is the reason we frame this as Qualcomm’s second platform moment. The first platform was mobile. The second is the attempt to extend Qualcomm’s compute, connectivity, and low-power design DNA into data center AI infrastructure while keeping the edge portfolio compounding. Data center is the growth driver. Automotive, industrial, robotics, personal AI, PC, XR, and QTL make the bridge less fragile. The diligence question is now whether $15B is a ceiling, or the first visible layer of a larger custom silicon platform business.

Inside the full subscriber report

• The pre- and post-Investor Day revenue bridge: why the old model looked like offset work and the new model changes the category.

• A full scenario model through FY31, including bear/base/bull revenue paths and the EV/sales read-through at each path.

• A data center stack that separates custom Arm CPU, custom ASIC services, HBC acceleration, and Alphawave connectivity/IP.

• Why custom Arm CPU may be the more lasting upside layer if hyperscalers move beyond standard Arm CSS building blocks.

• An explanation of HBC and why its economic value is tied to memory movement, bandwidth per watt, and cost per token.

• A full FY29 business breakdown showing how auto, industrial, robotics, personal AI, PC, XR, and QTL contribute around the data center ramp.

• What would change our view. The operating variables that would make us more constructive or force us to reduce the data center multiple credit.

Read more

That is where the cycle starts to get more interesting for all industry stakeholders. The obvious AI beneficiaries have already been heavily debated and largely mapped by us and others. The more interesting question to us is which parts of the semiconductor supply chain are being pulled into a new growth curve even though they do not screen like AI businesses at first glance. Power semiconductors and analog control sit near the top of that list.

A GPU creates the load, but it cannot use electricity in the form it arrives from the grid. Power has to be converted, stepped down, regulated close to the die, sensed, protected, and monitored continuously. As AI racks move from roughly 100kW-class systems toward several hundred kilowatts and eventually toward megawatt-class designs, that power tree becomes more complex and more semiconductor-rich. Content per rack rises because the system needs more point-of-load regulation, more intermediate bus conversion, more protection, more telemetry, and more analog control to make the accelerator usable at density.

That point becomes easier to see when you look directly at the power delivery hardware around a next-generation GPU tray. The photo below shows dense capacitor banks sitting immediately adjacent to a Vera Rubin GPU tray. Those cans are not the regulators themselves, and the voltage markings should not be added up as a direct power calculation. A 16V or 63V marking is a component rating, not the rail’s wattage. But the density is still insightful for our thesis. Many dozens of 100µF-class polymer capacitors sit beside each GPU power zone, likely supporting a mix of intermediate and local rails, absorbing fast current swings, reducing ripple, and giving the voltage regulators enough local energy storage to keep the GPU stable during abrupt workload transitions.

This is a Vera Rubin + Vera CPU compute trey in an HP Cray design. OEM/ODM builds will vary.

That is the physical version of the thesis. AI power content is not just moving into the PSU, the sidecar, or the facility layer. It is moving onto the tray, potentially the substrate, and closer to the accelerator. Each generation requires more regulated, sensed, buffered, protected, and telemetry-managed power within inches of the die. For analog and power semiconductor suppliers, rising rack power translates into a larger and more complex control problem. Every additional watt moving through the rack has to pass through layers of conversion, regulation, sensing, protection, and telemetry before it can be turned into usable compute.

Analog and power are different from the parts of the semiconductor industry investors tend to associate with fast AI scaling. Logic can often ride leading-edge process roadmaps, large platform concentration, and aggressive foundry capacity plans. Analog and power scale through a more physical supply chain: mature-node capacity, high-voltage process know-how, 200mm and 300mm power-device fabs, SiC and GaN material availability, thermal packaging, passives, magnetics, test, and long customer qualification cycles. The dirty secret is that analog often scales less cleanly than logic—it is also much harder from a design and engineering standpoint. It stays closer to the physics, where noise, heat, voltage behavior, layout, and process variation can determine whether a part works at spec. That makes the supply response slower when demand suddenly accelerates. Huge opportunity for agentic EDA here:

That makes power and analog a fascinating layer in the giga cycle. ASPs can rise when supply tightens, but the better point is that these components become part of the deployment math—and—a highly specific engineering challenge. A rack can have the right accelerator, memory, and networking, and still be limited by whether the power delivery architecture can support the density. That is the part of the GPU tsunami we focus on in this report: how AI rack density pulls power semiconductors and analog control into a demand cycle that is still under-discussed relative to how much it affects the rest of the build.

While this report is focused on power and analog attach to compute racks, we have a full report on 800 VDC beneficiary ecosystem.

What’s in the full report:

For paying subscribers, the full report goes deeper into the model, the rack architecture, and the supplier map. Inside we cover:

• Our direct AI datacenter power semiconductor and analog control TAM model from 2025 to 2030.

• Why stage 2 point-of-load power and intermediate bus conversion capture the largest share of direct semiconductor value.

• How rack density changes the BOM as systems move from 100kW-class racks toward 600kW and 1MW-class architectures.

• Why 48V and 54V distribution start to run out of room as current, copper, heat, and rack volume become limiting factors.

• How 800V DC should be understood as a capacity architecture rather than only an efficiency upgrade.

• The “power tree” from grid to core, including AC/DC, sidecars, IBC, point-of-load, protection, telemetry, and buffering.

• The analog control layer: hot swap, eFuse, current sensing, isolation, digital power control, PMICs, and telemetry.

• The role of BBU, CBU, and power smoothing as AI workloads become more dynamic.

• A socket-level beneficiary map separating silicon / analog control, power modules / sidecar components, and datacenter power systems.

• The risks and watch items that could change the slope of the thesis, including architecture timing, multi-sourcing, qualification cycles, and whether power vendors begin disclosing AI-specific revenue separately.

Read more

The first report started where the capacity models stop. Sizing AI as a supply problem, GPUs, power, packaging, capex, captures how much gets built, not whether the highest-value workloads are allowed to run. The data with the best returns is usually the data legal, compliance, and security teams will not expose to a multi-tenant cloud, so capacity can be financed and energized and still sit unused. We called that second variable permission. The argument was that confidential computing sits on that boundary as a conversion and pricing layer rather than a security line item: hardware-rooted trusted execution plus remote attestation turns trust into an artifact a compliance team can file and an auditor can check, and once that artifact exists, approval behavior changes and blocked workloads become consumed infrastructure. The lift runs two ways in our framework, a trust premium on workloads already heading to cloud and the larger conversion of regulated demand that could not run at any price before, which is why, for this equation particularly, the useful unit moves from tokens per watt to protected tokens per watt and a dollar cleared by compliance should behave differently in a price war than a dollar of experimentation.

The second report took the same permission lens to the supply side and found that AI server demand is no longer one market. Once you model by who owns the hardware rather than where it sits, the buildout separates into three: hyperscaler-owned capacity, the neocloud and third-party AI factory layer, and the enterprise or private AI factory that companies, governments, and regulated industries run inside their own walls. A single GPU cluster financed by a neocloud, contracted by a hyperscaler, consumed by a model lab, and booked by an ODM lands in four reporting streams, and adding them together produces a market that does not physically exist. Our base case carries total demand from roughly $228 billion in 2025 toward $845 billion in 2030, with hyperscalers still the anchor near two-thirds of the market and the marginal dollar of growth migrating outward. The enterprise and private segment is the least observable of the three and the one we hold with the widest range, precisely because its growth is governed by the permission economics the first report priced. That is where the two notes become the same question asked from opposite ends.

The third report went inside that enterprise market and asked who captures the value as token generation moves closer to enterprise data. The workloads that qualify are the persistent ones, agents that run continuously and hold state, retrieval against proprietary corpora, fraud and compliance pipelines, the cases where steady utilization, data sensitivity, latency, and governance line up at once. For those deployments the private AI factory should not be read as a server sale. The server is the entry ticket, and the durable economics sit in what attaches behind it: storage, networking, power and cooling, security, confidential compute, management, software, financing, and services. That reframing turns the investment question into a revenue-quality question, because the same AI server dollar can sit on thin GPU pass-through or pull a higher-quality attach stack behind it, and presence somewhere in that stack is not the same as a strong position in it. Dell and HPE are the cleanest public test cases, and they attack the opportunity from opposite ends, Dell through AI-factory scale and storage pull-through and HPE through a networking-led private-cloud and operations layer. That distinction is the one the scorecard is built to make legible.

Put those three lines together and the question for an enterprise moving into production changes shape. It is not which single venue wins. Production AI runs hybrid and multicloud, across owned capacity, neocloud, and the hyperscalers at once, so the operative question is which vendors hold the operating-stack pieces, the governance, security, data, orchestration, and confidential and sovereign capabilities the series identified, that let an enterprise run agentic AI in production wherever it sits. The scorecard grades eighteen of those capability domains for breadth and ownership on our framework, not for revenue or valuation. Read that way it makes one structural point first: accelerated compute is the only domain where every vendor scores a full three, so the metal is a cleared baseline and the entire spread opens up above it, in exactly the operating and data layers the three reports said decide whether a workload reaches production. Note, this is purely a capability graph not one scoring the quality of the capability.

Breadth favors the hyperscalers, with Azure separating at 53 of a possible 54 and its lone soft spot in autonomous network operations, the one domain where HPE scores a three. That same plane explains HPE’s 47 against Dell’s 40, a gap that sits almost entirely in orchestration, observability, networking, security, and FinOps while Dell’s owned strength runs through storage and data-protection resilience, which is the deployable-versus-operable split from the third report rendered as two different shapes rather than one ranking. The AI-native clouds play in a more specific lane for now, with CoreWeave and Nebius strong on compute, orchestration, and scheduling and thin on the data-platform and governance domains that regulated production demands, coverage profiles of 35 and 36 that read as infrastructure depth without enterprise breadth. IREN sits earliest at 22, its strength concentrated in the physical layer with the operating-stack build still ahead of it.

The point we want to land is what stack breadth represents. A vendor with owned or tightly integrated capability across compute, storage, data, networking, security, observability, and cost governance has more ways to capture regulated, sovereign, and enterprise workloads — and more ways to keep the attach revenue around those workloads. That is the line this series has followed from the start: revenue quality improves when the vendor captures more of the operating stack, and deteriorates when AI infrastructure remains a hardware pass-through cycle.

The scorecard is the compressed version of that argument. As production AI moves into a hybrid and multicloud world, the winning stacks will be the ones that make token generation governable, secure, metered, operable, and recoverable. The hardware still matters. The question is who owns enough of the environment around it to make the revenue durable.

As we have been laying out, most enterprise AI analysis still starts from the assumption that the enterprise is a pure cloud customer. That remains true for many workloads today, including AI, but it no longer captures the full direction of how we see this playing out. We are increasingly confident that certain AI workloads will work their way back toward infrastructure the enterprise owns, controls, or has dedicated access to. Our conversations from Dell Technologies World and HPE Discover reinforced that view.

We are not talking about every AI workload, and we are not arguing for a broad reversal back to the old on-prem server cycle. The workloads that start to qualify are the ones that become persistent internal processes: agents running in support, code assistants embedded into developer workflows, fraud systems scoring transactions continuously, or knowledge systems sitting close to proprietary enterprise data. Once those workflows run every day, token generation starts to behave less like experimental software consumption and more like an operating input.

That changes the buying conversation. Finance begins to care about recurring token cost. Security begins to care about where data moves and who can touch it. Infrastructure teams begin to care about utilization, latency, governance, and whether the workload is predictable enough to justify dedicated capacity. Public cloud still remains the better answer for burst, experimentation, frontier model access, and workloads where elasticity matters more than control. But the more repeatable the workload becomes, the more the enterprise begins to ask a familiar infrastructure question: if we use this capacity constantly, should we keep renting it by the unit or control more of the stack ourselves?

That is the on-prem qualification. The private AI factory is not a universal destination for enterprise AI. It is the infrastructure response to workloads where utilization, data sensitivity, governance, and workflow value line up. Where those variables line up, local token generation becomes easier to justify. Where they do not, public cloud remains the default.

The Workload Has To Earn Its Way On-Prem

We again need to emphasize, this shift is early and we are outlining problems we hear and challenges faced by enterprise customers as they see agentic AI get deployed in their enterprise. This is not a sweeping “AI moves back on-prem” call, and it is not just a new label for the old enterprise server cycle. Public cloud has real advantages: elasticity, model availability, global reach, lower operating burden, and the ability to experiment without building dedicated infrastructure. For many enterprises, most AI will stay there.

However, as enterprises have begun to deploy agentic AI, it is clear classes of workloads have brought them on a path to ask new questions and think about longer term strategies. When a workload runs continuously, touches sensitive data, and creates enough internal value, the economics start to change. A support agent running all day, a developer assistant embedded into the engineering workflow, or a fraud system scoring transactions continuously has a different utilization profile than a pilot project. At steady utilization, the customer eventually asks the same question it has asked in every compute cycle: is this cheaper to rent by the unit, or control directly because we use it constantly?

That is data-center utilization logic applied to tokens. The language is new because the unit is new, but the underlying infrastructure math is the same. The asset used occasionally is usually easier to rent. The asset used constantly eventually invites an ownership or dedicated-capacity discussion.

Why The Local Case Is Getting Better

Local inference is getting cheaper as accelerators improve and smaller models become good enough for defined production tasks. The software stack for running inference on owned hardware is also becoming more deployable. All important factors because many enterprises do not want to engage in non-ROI enabling exercises. They need something their infrastructure teams can operate, govern, secure, and support.

Data gravity strengthens the local case. Many enterprise workflows already sit close to internal databases, file systems, identity layers, permission structures, and proprietary data. Moving that data to an external model can be expensive, slow, and operationally awkward. In some cases, the challenge is less cost than approval. For regulated, sovereign, or IP-sensitive workloads, the ability to prove where the data runs, who can touch it, and how the system is governed can determine whether the project moves into production at all.

That is why cost-per-token is only part of the discussion. Control, auditability, latency, data movement, and workflow integration all matter. None of those variables moves every workload into controlled infrastructure, but where they line up, the local-token case becomes much easier to assess.

Agents Are The Utilization Case

The workload category we are most focused on is persistent internal agents. These systems do not behave like occasional chatbot sessions. They run continuously, call tools, query databases, maintain state, and interact with enterprise systems across many steps. That makes them one of the cleaner utilization cases for owned or controlled AI capacity.

It also makes the governance problem harder. A chatbot can be controlled at the application boundary. An agent that retrieves, reasons, acts, and writes back across internal systems creates a different kind of control challenge. It touches more systems, creates more audit requirements, and raises the importance of identity, permissions, observability, and policy enforcement.

That does not mean every agent needs to run locally. We would be careful about making that leap. The more grounded point is that the highest-control agentic workflows are likely to be among the first places enterprises ask for dedicated AI infrastructure. When agentic AI becomes a real production architecture, rather than a proof of concept layer, it could pull token generation toward controlled capacity faster than any single regulated vertical does.

The Server Is Only The Anchor

The private AI factory should not be modeled as a server sale as it once was. The server anchors the deployment, but the investment question is the attach stack around it: storage, networking, data protection, power and cooling, security, operating software, financing, lifecycle services, and integration work. For every dollar of AI server hardware, the diligence question is how many additional dollars follow, what margin they carry, and how repeatable the deployment becomes.

This is where backlog size can mislead. A large AI server backlog may say more about supply-chain access than durable earnings power. Backlog quality depends on what attaches to the compute sale and whether that attach turns into a broader operating layer around local token generation. A vendor that sells the GPU box and stops there may get revenue, but the quality of that revenue is different from a vendor that captures storage, networking, services, software, financing, and lifecycle management behind the same deployment.

That distinction is important because private AI infrastructure could either become another lower-margin hardware cycle or a more durable enterprise platform cycle. The difference will come down to attach, utilization, repeat deployments, and margin behavior as the market scales.

Dell And HPE Are The Clearest Test Cases

Dell and HPE are the two clearest public test cases, but they are approaching the market from different parts of the stack. Dell is the scale, storage, financing, and full-rack local-token platform case. Its advantage starts with distribution, supply-chain scale, AI server volume, and a large installed storage footprint. The strategic read is that Dell is trying to turn AI server demand into a broader enterprise AI Factory motion, with compute pulling storage, data management, services, financing, and lifecycle attach behind it.

HPE is taking a different route. We read HPE as more of a networking-led private cloud integration case. Juniper, Aruba, GreenLake, and its sovereign enterprise focus give it a different wedge into the same customer problem. In HPE’s version of the thesis, the network and operating layer become part of the control plane for private AI. That matters more if agentic systems force enterprises to rethink identity, policy, traffic flow, observability, and governance around AI workloads.

Both companies are aligning around the same broader shift, but from different starting points. Dell starts with the AI factory and tries to attach the stack behind it. HPE starts with networking, private cloud operations, and governance, then tries to pull compute and storage alongside that control plane. The full report goes deeper on where each company is advantaged, where the execution risk sits, and how the broader vendor map forms around this local-token stack.

Inside the full report

▪ A workload-by-workload suitability matrix, with our directional estimate of how much of each workload’s token generation lands on owned capacity versus public cloud.

▪ The full private AI factory attach stack: the layers a single server sale pulls, who benefits in each, and the evidence that would confirm the attach is real.

▪ The Dell versus HPE scorecard across nine dimensions, and where Lenovo’s hybrid and edge model fits without diluting the comparison.

▪ Why Cisco’s networking opportunity is a security and observability control-plane story rather than switching alone, and how HPE through Juniper attacks the same problem from the other direction.

▪ The agentic storage shift: why persistent AI turns storage from a passive repository into part of the inference loop, and reprices the data layer.

▪ How to separate confidential-compute silicon from the software that monetizes it, and which layer actually captures the recurring revenue.

▪ A revenue-quality ladder for grading any beneficiary’s AI revenue, plus a diligence checklist of what to watch and the bear case that mirrors it.

Read more

Most AI server numbers in circulation still try to measure one market. We have increasing conviction that is no longer the right way to analyze what is happening. The buildout is separating into three markets, and the line between them is better explained by ownership than location.

Ownership is the cleaner starting point.
Cloud versus on-premises, the framing inherited from the last server cycle, has become the least useful question you can ask about an AI rack. What decides the economics now is who owns or finances the hardware, who consumes the compute, and where the capacity sits in the value chain. Those questions increasingly return three different answers for the same rack.

This is where the double counting starts.
Walk a single cluster through the value chain and you can watch it happen in real time. A neocloud raises debt against a long-term contract, buys GPU systems from a contract manufacturer, and installs them in a leased, powered building. A hyperscaler contracts that capacity for several years. A model lab consumes the compute through the hyperscaler. The manufacturer books a cloud AI server sale, the neocloud reports the asset and its backlog, the hyperscaler reports the lease commitment, and the model lab announces a multi-gigawatt pipeline.

One cluster, turns into four reporting streams. Add them together, which is what a lot of market sizing quietly does, and you get a total addressable market that does not physically exist.

The three ownership buckets.
Modeling by owner fixes this, because every dollar of server hardware has exactly one owner even when four parties touch it. That gives us three mutually exclusive buckets:

• Hyperscaler-owned hardware: capacity the largest cloud platforms build and run for themselves.

• Neocloud and third-party AI factories: GPU clouds and hosted-compute operators that own infrastructure and rent it out.

• Enterprise and private AI factories: systems companies, governments, and regulated industries buy and run inside their own walls.

Consumption still earns its place, just in a separate view that tells us how hard the capacity is working and how durable the demand is. It should never be folded back into the size of the market.

Power capacity needs the same discipline.
Gigawatts are also not generic demand signals. A contracted megawatt tied to a hyperscaler campus, a neocloud balance sheet, or a sovereign/private AI factory carries different deployment risk and should not be interpreted the same way. AI server sizing needs a more precise view of power capacity because megawatts determine what can physically deploy, while ownership determines where the hardware is counted.

The marginal dollar is moving outward.
Hyperscalers remain the center of gravity and we believe they will stay there, but the marginal dollar of growth is migrating outward. Neoclouds have gone from a rounding error to a financed capacity layer large enough that ignoring it, or attributing its hardware to the hyperscalers who rent it, throws the whole model off. Enterprise demand is the hardest segment to see cleanly, but it is also the segment where the economics are changing in a way we think the market is under-modeling.

Why we size enterprise/private AI factories as a new layer of server infrastructure.
A class of enterprises will not want every token to be a metered cloud token. As AI moves from experimentation to persistent internal workflows, token costs become an operating variable that companies will manage. Some workloads will stay in public cloud because the flexibility is worth paying for. Others will move closer to the enterprise because sustained utilization, improving local compute, smaller and more efficient models, and maturing on-prem software stacks make owned capacity more attractive.

The goal is not to bring every AI workload inside the company. The goal is to generate more local tokens where the workload is steady, sensitive, and valuable enough to justify the infrastructure.

Where this connects to permission.
Enterprises own AI infrastructure where governance, sovereignty, latency, sustained utilization, and token-cost control make ownership the approvable option. The workloads with the best returns are often the same workloads a compliance team is least willing to run anywhere else. Capacity can be financed and powered, and still sit unused if no one inside the customer is allowed to put sensitive data on it.

Sizing the capacity is only key one question. Whether the regulated and sovereign share of it gets permission to run is the other. They are the same question asked from opposite ends of the spectrum.

The scale is large and growing fast, a market in the hundreds of billions of dollars a year and on a path toward the trillion-dollar range by the end of the decade. The level matters less than the structure, because the structure is what tells us which businesses—and which infrastructure players—capture the growth. This in turn informs where the revenue is durable, and where the same hardware is being counted twice. That structure, and the model behind it, is what the full report is for.

Inside the full report

The subscriber edition turns this framing into a working market model. It includes:

• The full 2025–2030 forecast for all three ownership segments, with explicit low, base, and high scenario ranges rather than a single point number.

• The capex-to-server-hardware bridge that reconciles disclosed hyperscaler capex to deployed, owner-based server TAM, with every adjustment shown so you can argue the assumptions.

• Three segment deep dives covering hyperscaler, neocloud, and enterprise/private, each built from the company backlogs, capex guidance, contracted power, and channel data behind it.

• The ownership-versus-consumption matrix, the discipline that keeps model-lab pipelines and leased capacity from being counted twice.

• Architecture and cost-stack economics, including cost per gigawatt across NVIDIA and custom-ASIC systems and why the architecture mix is a first-order driver of the dollar TAM.

• Forecast confidence grades by segment, naming the single variable most likely to move each one.

Read more

This report argues that confidential computing sits at exactly that boundary, and that it functions as a conversion and pricing layer for AI infrastructure rather than another security budget line. To be clear about what this is: an AI infrastructure revenue-quality report, not a cybersecurity report.

The mechanism

Enterprises spent two decades building controls, certifications, and audit language around two states of data: at rest and in transit. AI puts pressure on that model because the value is created in a third state, while the data is in use. During inference and agent execution, prompts, model weights, retrieved context, and agent credentials all sit decrypted in memory, where the cloud operator, hypervisor, or another privileged software layer could in principle reach them. Hardware-enforced trusted execution environments, now present in server CPUs and recent GPU generations, are designed to close that gap. Remote attestation then adds the piece that matters commercially: a signed evidence artifact a compliance team can file, an auditor can check, and a regulator can review. Once that artifact exists, approval behavior can change. Approval behavior is what turns blocked workloads into consumed infrastructure.

For our broader thesis on how enterprise adoption plays out, with the overarching platform/control plane layer, confidential AI is a key component. See our report here:

Two findings worth sharing

The first is about pricing power. Our research indicates confidential compute instances have commanded premiums on the order of 20 to 30 percent over comparable standard instances in regulated environments, a working estimate we are still tracing to SKU-level pricing. The interesting part is what kind of premium it is. A scarcity premium on GPUs erodes as capacity gets added; that is just supply. A trust premium rests on regulation, audit requirements, and customer risk tolerance, none of which follow semiconductor supply curves down. Generic compute will always face commoditization risk, but verified isolation gives providers a feature customers can pay for.

That is where protected tokens per watt becomes a useful lens. Standard AI infrastructure is judged by how much inference it can deliver per watt, per rack, and per dollar. Regulated AI adds another requirement: how much of that inference can run inside a trust boundary the customer can approve. The customer is not only buying throughput. They are buying usable throughput. That difference is what gives confidential compute a pricing and retention argument even if baseline AI compute becomes more competitive.

The second is what we call the double dollar, and it is the value equation in the full report. The premium is the smaller half of the money, because it applies to workloads that were coming to cloud anyway. The larger opportunity is conversion: regulated workloads that did not run at all, at any price, because nobody could approve the data exposure. In our illustrative base case, the conversion effect contributes roughly twice what the premium does, which means anyone modeling the premium alone is missing most of the lift. The full scenario math, conservative through high, sized per $100 billion of CSP AI compute revenue, is in the subscriber report. And the thesis survives even if the premium compresses, because the conversion and retention effects come from workloads clearing compliance rather than from a SKU markup. An investor does not need to believe in a permanent premium at the top of the observed range. They only need to believe confidential architecture unlocks production workloads that standard infrastructure could not capture.

The insight we would leave readers with is this: not all AI cloud revenue is equal. A dollar cleared by legal, compliance, and security should be more durable in a price war than a dollar of developer experimentation. Confidential mix is one of the few observable markers that lets stakeholders separate generic AI consumption from workloads with permission, auditability, and switching friction attached. The market is still modeling AI infrastructure through capacity. We think it should also be modeling permission. Protected tokens per watt is one way to describe that next layer.

What the full report covers

The subscriber report runs the full diligence framework, including:

• The blocked-workload evidence base — insights from commentary from banking, healthcare, defense, and real estate on which AI workloads are stalled, what asset is at risk, and what unblocks them.

• Exhibit 2: the dollar-lift model — our per-$100B CSP revenue sensitivity framework across conservative, base, and high scenarios, with the full equation and the input the model is most sensitive to.

• Exhibit 3: the beneficiary map — five tiers from silicon to adjacent AI security, graded by directness, materiality today, and exactly what to track for each.

• Sovereign AI as a control problem — why local data centers alone fail the sovereignty test regulators are actually applying, and the regulatory calendar behind the demand.

• Private AI factories as governance infrastructure — the order-book evidence, agent tokenomics, and the attach economics that decide whether OEM AI revenue deserves better than a hardware multiple.

• Where the model changes — company by company: the CSPs, NVIDIA, Dell and HPE, AMD and Intel, and the confidential middleware layer.

• The honest valuation answer — whether any of this moves a stock today, where the lens is most material, and the forward checklist that tells you when the valuation question goes live.

• What to watch — the five verification items that graduate this thesis from working view to conviction.

Read more

For most of its history, EDA has been valued as an access business. The model was built on seats: how many engineers need the tools, how firmly the workflows hold them, and how much advanced-node complexity lifts renewal value. Those drivers still hold. What has emerged underneath them is a different question, whether an agent stops being a convenience layered on top of the design flow and becomes a productive worker operating inside it. That is the shift that could pull the category off its historical pricing logic, because it changes what the customer pays for from the right to use the tools to the work the tools complete.

The lens that shapes EDA from here is verified design throughput. A customer in advanced silicon works against a shrinking window in which a fault can still be absorbed, and a fault found late costs far more than the same fault found early. A bug caught near tapeout converts directly into a respin and the schedule slip behind it, and in a competitive node race that slip becomes a roadmap problem before it becomes anything else. An agent that runs inside the design environment and calls the same verification and signoff engines that already gate tapeout does the thing that compounds for an incumbent. It pulls more billable work back into infrastructure the vendor already owns, and it does so at the stage of the flow where the customer is least willing to economize.

That changes how to read the concern that has hung over the group. Software investors have spent two years asking whether AI compresses seat-based pricing, and EDA keeps getting swept into that same trade. The pressure is real for most application software. It is the wrong read here, because EDA does not monetize the way SaaS does, and the constraint that actually binds sits elsewhere. What gates a design organization is its ability to bring a correct chip to tapeout as complexity and schedule pressure keep climbing, and that has little to do with how many engineers sit in front of a license. Once the work moves from manual iteration to autonomous tool usage, a smaller and more productive team can pull more compute and more verification cycles through the same tools, the reverse of the headcount-linked decline the seat-compression thesis assumes. The revenue question shifts from seats sold to throughput verified, and throughput is a consumption variable rather than a headcount one.

Verification is where the agentic case should prove out first, because it is the part of the flow where rising complexity turns into measurable schedule risk. Custom and analog design is the more differentiated secondary wedge, where the scarcity is institutional knowledge rather than digital complexity, and where native agents embedded inside established flows could monetize design history that has never been easy to encode. The two carry different proof burdens, and that difference is most of what separates the near-term call from the long-term one.

That same split separates the two companies. Cadence holds the cleaner near-term agentic case and should be able to prove it first, with its strength sitting closest to the verification flow where throughput turns visible soonest. Synopsys may own the larger long-term platform if the Ansys integration lands, though that path carries more execution risk and a longer runway to proof. Both can compound from here, and the evidence to underwrite each differs: for Cadence, verification throughput showing up in usage and renewal economics; for Synopsys, integration milestones that convert into design wins rather than roadmap claims.

The full report is where we size this and gauge it against what the market already pays. Our base case puts the opportunity at $2.5B to $3.0B of incremental annual core EDA revenue by 2030, inside a wider $1.5B to $5.0B range where customer acceptance of consumption pricing, rather than technical capability, is the swing variable. The report builds that revenue bridge step by step, lays out the Cadence versus Synopsys assessment map, works through how verification and custom design actually monetize, and runs the agentic dollar lift against each company’s current enterprise value to ask whether the opportunity is already priced in and where it defends or extends the multiple. It also sets the risks, from pricing resistance to China exposure, against the thesis, and names the contract-level signals worth watching before any of this shows up in reported numbers.

Inside the full report

• A full framework for why agentic EDA should be evaluated through verified design throughput rather than seat access.

• Creative Strategies’ revenue bridge estimating the potential incremental annual core EDA opportunity by 2030, including base case and sensitivity range.

• Why verification is the first monetization proof point, and what to watch in regressions, emulation demand, cloud EDA usage, and module attach.

• Why custom and analog design may be the more differentiated wedge as scarce expertise, proprietary IP history, and node migration become larger bottlenecks.

• A Cadence vs. Synopsys underwriting map, including where Cadence may prove agentic attach first and where Synopsys may have a broader long-term silicon-to-systems opportunity after Ansys.

• A valuation test that runs the agentic dollar lift against each company’s current enterprise value, using EV/Sales revenue bars to show what Cadence and Synopsys must earn to support today’s multiple, and where the lift defends or expands it.

• The risks investors need to underwrite, including pricing pushback, hyperscaler internal tools, open-source pressure, China/export controls, and Synopsys integration execution.

• The contract-level evidence that would confirm or disprove the thesis: renewal uplift, agentic SKU attach, usage budgets, production deployment in custom/analog, and margin durability.

• The key conclusion: the market does not need to believe in fully autonomous chip design for EDA to deserve a different revenue lens. The real question is whether agents create more monetizable work inside workflows Cadence and Synopsys already control.

Read more

This report continues our work on neoclouds, although we are coming at the question from a different angle. The last report focused on backlog quality, monetizable MW, customer concentration, contract duration, financing structure, and the way hyperscaler demand was turning into external infrastructure commitments. That still feels like the right starting point. The next step is a full competitive SWOT, which is a useful exercise for separating near-term capacity leverage from the harder question of platform durability.

Our thesis remains that neoclouds are a direct proxy for hyperscaler urgency. The hyperscalers are still the demand center, but the neoclouds give us a useful read-through because their backlog and financing structures help quantify how that urgency is being translated into external AI infrastructure demand. We continue to think this is the cleanest way to study the space. A large hyperscaler contract, a GPU-backed debt facility, a power reservation, or a multi-year take-or-pay structure tells us something about the pressure inside the cloud market even before that pressure fully appears in reported cloud revenue.

The report also puts some limits around the neocloud narrative. CoreWeave, Nebius, and IREN are each trying to become more cloud-like, and we believe those efforts are rational. CoreWeave still has the strongest neocloud software and orchestration layer today. Nebius is pushing in both directions, down into owned infrastructure and up into inference and agent software. IREN starts with the clearest ownership of the physical bottleneck, while the Mirantis acquisition gives it a more credible path to build managed GPU services above that power base. All three are taking steps beyond GPU rental, which they need to do if they want the market to value them as more than capacity intermediaries.

We would still be careful about treating them as future hyperscalers. The big three cloud platforms have breadth that took years to build: identity, security, governance, data platforms, developer services, application integration, global operations, support, compliance, and procurement relationships. That breadth is what makes them difficult to displace across the long tail of enterprise workloads. The neoclouds can be very good GPU clouds and have a meaningful role in AI factory capacity, while still facing a much harder path in competing for the broader enterprise cloud control plane.

That distinction is the central point of the report. Neoclouds are well positioned where capacity speed, GPU availability, financing creativity, and power access are the binding constraints. Those are real advantages in this phase of the cycle. Training demand, burst capacity, and frontier model infrastructure can move toward the provider that can deliver the most usable compute at the right time and price. That is where the neoclouds have earned their relevance. Whether those are long term differentiators is an area we explore in this report.

Production inference creates a different test. Once AI workloads move into enterprise deployment, customers start to care more about reliability, identity, governance, data proximity, application workflows, security posture, support maturity, and cost per token. That favors providers with deeper platforms. It also raises the importance of custom silicon in a cost per token or all you can eat world. We know from our CIO and CTO work that token cost is becoming one of the central considerations for agentic AI spend. AWS, Google, and Microsoft all have more infrastructure and software depth to apply against that problem, and AWS and Google in particular have more mature custom silicon paths through Trainium, and TPUs.

That is why the scorecard in the full report separates stack presence from business quality. Azure leads our stack-presence score because Microsoft owns enterprise distribution, identity, M365, Dynamics, OpenAI access, and Copilot pull-through. AWS remains the infrastructure trust and custom silicon benchmark. GCP remains the data gravity and TPU economics specialist. Oracle sits in the middle because OCI has credible bare-metal GPU and RDMA capacity, plus a real database and enterprise applications estate, although its AI software middle is thinner than the big three.

The neoclouds split by scarce asset. CoreWeave is the backlog and orchestration case. The bull case is revenue visibility and software depth, while the diligence work is customer concentration, lease duration, financing cost, and whether inference becomes a larger part of the mix. Nebius is the most hyperscaler-like of the group, with large Microsoft and Meta commitments, a push toward self-owned infrastructure, and software assets like Token Factory and Tavily. The test is whether those pieces become repeatable consumption economics. IREN is the power-first case. Its advantage is physical and harder to replicate, but the multiple depends on whether managed AI cloud services can turn that power base into recurring revenue.

The broader conclusion to us is that AI cloud demand is segmenting by workload. Training can follow price and availability. Production inference should stay closer to platforms with identity, data, reliability, and cost-control advantages. Regulated enterprise workloads will put more weight on governance and support. AI-native startups may continue to value speed, GPU access, and modern tooling before procurement standardization starts to matter. Power-constrained capacity has its own logic because the scarce input starts before the GPU cluster is deployed.

The full-stack map does not give us one winner. It gives us a better way to analyze and value the next phase of the cycle. For hyperscalers, the issue is whether distribution, custom silicon, data gravity, and enterprise trust turn AI demand into durable consumption. For neoclouds, the issue is whether power, orchestration, and capacity access remain scarce enough to support repeatable economics as hyperscaler self-supply catches up. We still think the neoclouds have an important role in AI factories. We also think the path from GPU cloud to full cloud platform is a much harder climb than the current backlog numbers alone suggest.

What subscribers get in the full report

• A nine-layer full-stack framework for comparing cloud providers across power, data centers, custom silicon, GPU compute, networking, orchestration, model access, applications, and enterprise distribution.

• A directional scorecard for AWS, Azure, GCP, Oracle OCI, CoreWeave, Nebius, and IREN.

• Company-level SWOTs for each provider, with a specific diligence focus and read-through for every name.

• A breakdown of why Azure leads through enterprise distribution, AWS through infrastructure trust and custom silicon, and GCP through data gravity and TPU economics.

• A bridge-tier analysis of Oracle OCI and whether database and apps gravity can convert into AI workload attachment.

• A neocloud comparison across CoreWeave, Nebius, and IREN, including backlog quality, power position, software depth, customer concentration, financing, and durability.

• A workload map showing where training, production inference, regulated enterprise, data and analytics, AI-native startups, office agents, and power-constrained capacity are likely to route.

• The central diligence question for the next phase: which AI cloud providers own a scarce layer that stays valuable after capacity becomes easier to procure.

Read more