Expanded coverage in CS Atlas since Thursday

• Applied Optoelectronics Q3 2026: Orders Run Ahead of Parts, Power and Production Lines

• After the Midterm Selloff, AI Supplier Multiples Sit Below Our Growth Forecasts

• Cloudflare Q3 2026: Growth Speeds Up Again, and Agents Are Choosing Workers

We continue to believe that studying the enterprise adoption cycle of AI is one of the most fruitful exercises we can do, if not the most fruitful, to understand both current and future compute demand and the economics needed to fund the AI buildout. We emphasize that enterprise adoption, and budget expansion around AI, is the single thing that decides whether this early cycle gets funded. We have also been clear that what we are seeing right now is the traditional diffusion of innovation curve. The early adopters, in this case the more tech-savvy enterprises, adopt first, go through the hard lessons, and create a learning loop that others will follow. Our main concern for the longevity of this cycle has been twofold. First, the remaining 80% of the enterprise market will move more slowly to adopt, integrate and pay for AI at scale. Second, security will prove a harder hurdle to clear, forcing a much broader review and transformation of business processes before AI can scale safely in enterprises large and small.

Our CIO/CTO survey series is designed to capture the broader enterprise adoption cycle. With this report we introduce our first CISO survey, focused on the security gates that have to be cleared for agentic AI, the new and underappreciated role cybersecurity and resilience play in enterprise adoption of agentic AI, and how the people deciding how and when AI can scale look at security. We surveyed 200 US security leaders in September, alongside the 200 CIOs and CTOs in the second version of our E/AI Index.

The number we highlight in the survey comes from the approval step before production. Of the 143 security leaders who reviewed an AI project for production in the last 90 days, 49 (34%) delayed or rejected the most recent one mainly for security reasons, and only 16 of those 49 have funding approved to fix the problem. CIOs see the same thing from their side, with 31% naming security approval as the main requirement for their next AI expansion, ahead of the 26% who name returns approved by finance. Security now sets the pace of enterprise AI deployment, and with it the pace of the token and compute demand these customers create.

We have heard in our qualitative conversations that agents were running wild, ungoverned or unsecured, in the enterprise and we wanted to test this quantitatively. In our survey, 71 of the 200 organizations let business agents take actions in production, and 59 of those 71 (83%) have at least one of three basic gaps in credentials, rollback or the ability to cut off an agent's access. We think AI-related security spending is growing, that most of it goes through the security vendors enterprises already use, and that controls for agents are the next large budget line.

Inside the Full Report

• Investment implications: four conclusions ranked by how confident we are, covering budget size, how the money reaches vendors, and timing

• The approval funnel: how many AI projects security approved, held or rejected, how few held projects have funding, and where holds are most common by company size

• The missing controls behind security holds, and how closely they match where CISOs are adding budget

• How one production incident would change deployment plans and control budgets, from a randomized split of the sample

• Agents in production: control coverage across 127 organizations, and the credential and rollback gaps among the 71 with agents that take actions

• AI inside security operations: where it is allowed to act and what it has measurably changed

• AI labs against established security vendors, pricing, and the effect on managed security providers

• Companies covered: Palo Alto Networks, CrowdStrike, Microsoft, Zscaler, Netskope, Okta, SailPoint, Cloudflare, SentinelOne, Rubrik, Alphabet (Google Cloud), Amazon (AWS), Anthropic, OpenAI, and managed security providers

• 11 exhibits and two tables, including a map of likely winners ranked by our confidence

• A companion data packet in CS Atlas, with answer counts for every survey question and key measures by company size and industry

The Diligence Stack, built on the back of Creative Strategies decades of industry research, delivers analyst-grade research built on our own surveys, models, and industry work. Paid subscribers get the full report, access to CS Atlas, and the companion data packet. Click below to subscribe.

Read more

• SiTime Q3 2026: First Quarter With Renesas Timing, and a $1 Billion Run Rate

• Coherent Fiscal Q1 2027: A $20 Billion PhotonLink Deal, and a Capacity Build to Match

• Arm FQ2 FY27: Licensing Carried the Quarter While Phones Held Royalties Back

• Qualcomm FQ4 FY26: Data Center Revenue Starts as Handsets Shrink

The setup: the clock under every AI system

We like companies whose technology represents a shift from analog to digital. When a function moves from an analog part to a semiconductor, it picks up what semiconductors do well. It gets smaller, it can be programmed, it improves with each process generation, and it can be combined with other functions. That kind of shift does more than replace the old part. It brings new capabilities, and it usually opens up a bigger market than the one it replaced. We have been studying SiTime, talking with engineers building next generation infrastructure and spent time with SiTime management and now believe the “timing” is right to share our view of the company and introduce folks to a company and technology they may not be as familiar with.

Deep in nearly every part of computing sits an analog part most investors never think about, the quartz crystal oscillator. A thin piece of quartz vibrates at a precise frequency when voltage is applied, and the oscillator turns that vibration into the clock signal a chip uses to know when to send and receive each bit. Accelerators, CPUs, network cards, switch ports and optical modules all run off one, and the devices that make and distribute that clock are called timing devices. Quartz has done this job well for decades, but it is a cut piece of crystal. Each frequency needs its own cut, higher frequencies need thinner and more fragile crystals, and the part is sensitive to heat, shock and vibration. For most of the history of the data center that did not matter much. A server needed a handful of clocks, the parts were inexpensive, and few people outside the design team paid attention to them.

SiTime sits at the point where timing shifts from quartz to MEMS, a small vibrating silicon structure built with semiconductor processes. That shift opens up new capabilities in compute timing, and the upside gets larger as AI infrastructure, and computing in general, gets more complex. We also like companies that are relatively small today but growing into a much larger market, especially one that the shift itself opens up and expands. SiTime is running at about $1.35 billion a year against a timing market we see reaching $20 billion to $28 billion by 2030, and it checks all of these boxes. Most of the investors we speak with know the AI networking story well, but very few have spent any time on timing, and we think that is a gap worth closing.

AI is what makes timing more critical, because AI workloads no longer run inside one compute box. The race to build larger scale-up domains means a training or inference job now runs across thousands of accelerators that exchange data constantly. Operators want all of that compute to work as one fabric, so a single job can run on GPUs, custom accelerators and CPUs at once, and as power pushes capacity to new sites, it can span buildings and even data centers in different locations. None of that works unless every device is synchronized to the same time, and over time we expect nearly every piece of hardware in the data center to be synchronized.

When devices drift out of sync, the cluster slows down. A frequency jump can make a link drop and retrain, which can stall a job across the whole cluster. More often the cost shows up as lower utilization, and on a cluster of accelerators, lost utilization costs far more than the timing parts do. The requirement also gets harder with each generation, as lanes move to 200G and 400G and PCIe moves to generations 7 and 8. Synchronization, which used to be a telecom topic, is now a design requirement inside AI clusters.

The amount of timing grows faster than the compute itself, because the highest-value clocks sit in the network. In our Base case from 2025 to 2030, accelerators grow about four times, while optical link ends grow about seven times, switch capacity about 16 times and active electrical cables about 15 times. On top of that, content per system is rising: sync-grade parts sell for several times the price of a standard oscillator, the higher frequencies needed for more bandwidth also carry higher prices, and as optics move onto the package, the timing moves from the module to the switch, where designs need several times the content of today’s.

SiTime is the company we think is best placed to sell into that. Because its resonator is silicon and SiTime pairs it with its own mixed-signal circuits, compensation algorithms and packaging, the parts are small, can be programmed to different frequencies, and hold up to the heat, vibration, shock and power noise inside a dense AI rack. They also reach the higher frequencies that 1.6T optics need without the thin, fragile crystals quartz requires. In July SiTime bought the Renesas Timing Products Division, which makes the clock generators, buffers, jitter cleaners and network synchronizers that take the reference clock and distribute it across a system. That gives SiTime nearly the whole timing chain, from the reference clock to every chip, and we do not know of another supplier with a portfolio that covers the full chain in the data center. SiTime reported its first quarter with the Renesas business on November 4 and left the quarter running above $1 billion a year, which we believe makes it the largest pure-play timing company in the industry. We cover the quarter in our SiTime Q3 2026 earnings read.

This report is an introduction to the company and to the timing market for investors who have not looked at either. We size AI data center timing from the bottom up with our own accelerator, server CPU and networking forecasts, and we get to roughly $1.1 billion to $2.1 billion in 2026, growing to $8 billion to $14 billion in 2030, with a midpoint of about $11 billion. That would take AI data center from about 9% of a timing market of about $11.5 billion today to somewhere between 40% and half of a $20 billion to $28 billion market by 2030. That is where timing suppliers will win or lose share over the next five years, and the design wins for the 2027 and 2028 platforms are being decided now.

In the full report for subscribers:

• Who SiTime is, how its MEMS timing differs from quartz, and what the Renesas timing business adds to the product line

• The four levels where AI needs more timing, from the board to the rack, between racks and between sites

• A plain guide to the timing chain, from the reference clock to every chip, and where MEMS and quartz compete for each socket

• An attach matrix for an AI rack, showing where each type of timing device sits today and in next-generation designs

• Our bottom-up sizing of AI data center timing by layer of the rack through 2030, built on our accelerator, server CPU and networking forecasts

• The whole timing market across every end market, and how AI data center grows to 40% to half of it by 2030

• Why we think SiTime is well positioned, where the risks sit, and the signals we are watching

• Seven exhibits

Companies covered in this report: SiTime (SITM); Renesas Electronics (6723.T); Microchip Technology (MCHP); NVIDIA (NVDA); TXC (3042.TW); Seiko Epson (6724.T); Nihon Dempa Kogyo (6779.T); Kyocera (6971.T); Rakon (RAK.NZ); Murata Manufacturing (6981.T).

Go deeper in Atlas. Atlas subscribers can read our SiTime earnings note and query the accelerator, server CPU and networking forecasts behind this report. Start with one of these prompts:

• How optics on the package changes timing content

• What agent workloads do to timing per rack

• Our latest view on SiTime

Read more

• Samsung Q3 2026: Memory Capex Buys Fewer Bits Than It Looks

• Oracle Investor Day: A $20 Billion Answer to the Funding Question

• Microsoft Q1 FY27: Azure Accelerates While Capex Waits on Supply

• Alphabet Q3 2026: Cloud Doubles, and Capacity Is Still the Limit

• Seagate Q3 2026: Exabytes Are Allocated Through 2028, and Price Is Doing the Work

• SK hynix Q3 2026: Customers Are Buying Memory Years Out, and Our Model Lets the Shortage Ease Too Early

• Cadence Q3 2026: Agents Are Pulling Work Through the Core Tools, and Our Revenue Path Was Too Low

We have been steady in our stance that understanding the buyer for AI software, think AIaaS (intelligence as a service), is central to understanding the AI cycle, given they are the customers who will end up funding and sustaining the AI buildout. We continue to see sentiment, as well as mass confusion, around the ROI for AI as it is adopted in the enterprise. Our checks with enterprises date back to mid-2024, when enterprise software research quickly turned into AI adoption research, and most of 2024 and 2025 revealed insights around pilots being deployed, and the costs and learnings associated with those pilots. Pilots then moved to production, and most of our qualitative and quantitative work with IT decision makers has been around observations being found as AI has begun to scale in their enterprises.

The E/AI Index is our quantitative side of that work, and this is its second version. The first, published in May, surveyed 70 CIOs and CTOs and focused on budget architecture. This version surveyed 200 in September across North America and Europe. It keeps the core budget and deployment questions so we can track the move from pilot to production, and it adds new questions on verified production, full-cost ROI, token spend, and model choice. Every matched measure has moved forward since May, with AI budgets, agents executing workflows, and spending already shifted from software and services each up 7 to 9 points.

That confusion shows up most clearly in the question we hear from investors, which is whether enterprise AI pays for itself. What we found is that most enterprises do not yet have the tracking in place to answer it. Only 33 of the 200 CIOs measure return against the full cost of their AI programs, and 113 have not measured it at all. That is consistent with the top-level reports of low AI ROI, since an organization that is not tracking full cost has no return to report. When we look at the CIOs who are tracking, the story changes. Of those 33, 67% (22) are at break-even or better, and among CIOs who have scaled AI with operating controls and measure full cost, 92% (11 of 12) are

Our view, which is supported by hard data, is that the returns exist, that they are concentrated, and that they rise with production scale, the same progression from pilot to production our checks have followed since 2024. The returns show up first in software development at scale and in IT support once it spans several functions, and they are stronger where agents execute work across systems and where IT owns the AI budget.

Token spend is the other half of the debate we wanted more clarity on, and it is where we see the directional signal in the survey that connects most directly to the buildout these customers are funding. About half of CIOs are over their 2026 token plan at every stage of deployment, but the ones cutting projects because of it are mostly early. At scale, CIOs keep funding the overrun, and more of that funding is coming out of existing software budgets. That is where CIO returns start to show up in vendor revenue, and it is the part of the data we would watch most closely.

Inside the Full Report

• What has changed since the first E/AI Index in May, on matched questions

• The ROI evidence ladder, and how the picture changes among CIOs who track full cost

• Returns by production stage, including full-cost results at scale

• Which workloads pay as they scale, and returns by agent deployment

• Governed production: the approval models and budget owners tied to measured value

• Token spend against plan, project cuts, routing controls, and open-weight models

• Where the next AI dollar comes from, the expansion gate at scale, and the realized workforce effect

• Companies covered: Microsoft, GitLab, Atlassian, ServiceNow, Salesforce, Workday, Snowflake, Databricks, Amazon (AWS), Alphabet (Google Cloud), Oracle, OpenAI, Anthropic, Meta, IT services and systems integrators

• 10 exhibits, including a market beneficiary map, cross-tabbed by production stage, workload, governance, and token controls

• Full companion deep research packet in CS Atlas for subscribers to dive deeper with this report

The Diligence Stack delivers analyst-grade research built on our own surveys, models, and industry work. Paid subscribers get the full report and the companion packet in CS Atlas. Click below to subscribe.

Read more

Late last year, when nearly every conversation in the industry was about GPUs, we started making the case that the CPU was about to become strategic again. We formalized that view in March with Secret Agent CPU and built out the full framework in September with The Agentic CPU Guidebook. The simplest way we describe it is that the GPU does the thinking and the CPU goes and does the actions. An agent calls tools, runs code, queries databases, reads what comes back, and decides on its next step, and all of that execution lands on CPUs. Our view came from studying the workload, which is how we approach every compute cycle. And yet, a year in, we still get asked what people are missing about this demand, so that is where we started with Mohamed.

His answer lines up closely with how we model the market. He breaks the AI data center into three classes of CPU: the head node that feeds the accelerator, the cloud-native CPUs hyperscalers have run for the better part of two decades, and a new class of agentic CPUs he thinks of as virtual users. Those are the same three segments in our Server CPU forecast, which we call general purpose, accelerator hosts, and agent execution. The head node is the one everyone already understands, and the more interesting part is how he describes the third class. An agent works on a person’s behalf the way that person would at a terminal, making database calls, researching a purchase, booking travel, or writing a script, except the agent never sleeps. A workload that runs around the clock changes what an operator needs from the compute platform, and it is the main reason we expect this segment to grow much faster than the rest of the CPU market.

Our Server CPU model puts the market at $62 billion in 2026 and $251 billion in 2030 in our Base case, with a range of $195 billion to $298 billion across our Low and Stretch cases. Price does as much of the work as units in that forecast, because we expect manufacturing capacity to stay tight, and as long as it does, the CPU suppliers keep their pricing leverage. The part we think the market has not caught up to yet is power. Our current thinking is that individual agentic compute alone needs somewhere between 25 and 35 gigawatts of additional capacity by 2030, on top of what the accelerators draw. That range comes from our demand model, and it assumes half of the global internet user base uses a personal AI agent in some fashion. Put another way, that is new data center capacity built and powered for agent execution by itself, and we do not think most have it in their numbers.

Several parts of the conversation map directly to how we built that model. Mohamed says agentic CPUs mostly sit outside the scale-up domain, and his reasoning is that agent execution is a scale-out problem. You have lots of independent agents each running their own path, and in the case of Meta Muse and OpenAI’s dots, every user gets their own virtual machine. That fits how we size the segment, as dedicated agentic CPU racks attached to accelerator racks, instead of a fixed CPU-to-GPU ratio inside the accelerator system. It also means those CPU pods need their own scale-out network, to each other and to the GPU cluster, and that is a topology decision operators have not had to make before.

On design, Arm built AGI CPU around a balance of single-threaded performance, scalability, and efficiency, and Mohamed was clear that leaning too hard on any one of the three leaves you with a compromised product. That is close to the wide, fast, and fed framework in our guidebook, where every design trades cores, per-core speed, and memory against a fixed budget of power and area. He also pushed on two areas we think investors are underestimating. The first is security. Hyperscalers that run only their own workloads never needed the isolation the cloud providers built with DPUs, because they trusted every workload. Agents change that, since an operator can no longer know whether a given agent is nefarious, so those techniques are now spreading to anyone running a data center. DPUs sit outside our CPU number, and our Networking model builds DPU demand from accelerator servers, so DPUs attached to agentic CPU racks would be additive. The second is inference. Mohamed’s point is that whether work stays on the CPU or moves to an accelerator comes down to the quantum of compute, and when the job is small, shipping it off costs more than running it locally. As better CPUs reach the installed base, models can make different choices about where work runs. That is a real possibility, but we want to see evidence of the shift before we count it.

The question we hear most from investors on Arm is about supply. Arm has never had to secure memory, substrates, and boards before, and it is doing it for the first time in a market where nearly everything is short. We closed the interview on that question, and Mohamed did not pretend it is easy. He said Arm is comfortable it will hit its numbers, and he pointed to decades-long relationships with the fabs and memory makers, plus executives Arm has brought in from NVIDIA, Broadcom, and Marvell. That is fair, and we would add one thing. Being in the supply chain means Arm is now having the same conversations its customers are having, and we see that as a positive. Our models do not see capacity easing soon, and that is the environment where Arm’s new supply chain muscle gets tested.

In the full interview:

• Why Mohamed separates the AI data center into head nodes, cloud-native CPUs, and agentic CPUs, and why he thinks of agents as virtual users.

• What Arm optimized AGI CPU for, and why it balanced single-threaded performance, scalability, and efficiency.

• Where agentic CPU racks sit relative to the GPU scale-up domain, and what that means for networking.

• When inference stays on the CPU and when it moves to the accelerator.

• How DPUs went from a cloud provider tool to the control plane for the AI data center.

• How Arm thinks about selling silicon, compute subsystems, IP, and architecture licenses to the same customers.

• How Arm is handling the supply chain as a first-time merchant silicon vendor.

Go deeper with Atlas: Subscribers with CS Atlas access can work through the assumptions behind our Server CPU forecast, compare CPU vendor approaches from the guidebook’s technical companion, and test the agentic rack ratio against their own view. A few questions to start with:

• What assumptions get us to a $251 billion server CPU market in 2030, and what would change that forecast?

• How many agentic CPU racks does each accelerator rack need in our model, and how sensitive is the forecast to that ratio?

• How do Arm, NVIDIA, AMD, and Intel differ in their approach to agentic CPUs?

• When do agentic workloads need more cores versus faster cores or more memory?

• How do DPUs fit into agent isolation and security, and who supplies them?

The executive interview video archive also lives in CS Atlas under the Videos link on the left-hand bar.

Companies covered in this report: Arm Holdings (ARM); NVIDIA (NVDA); Advanced Micro Devices (AMD); Intel (INTC); Amazon (AMZN); Alphabet (GOOGL); Meta Platforms (META); OpenAI (private).

Paid subscribers get the full video interview with Mohamed Awad and the complete edited transcript below, along with every Diligence Stack report and CS Atlas access to our Server CPU forecast and the Agentic CPU Guidebook companion. If you are not yet a paid subscriber, upgrade to keep reading.

Read more

• Texas Instruments Q3 2026: Analog Turns Into a Capacity Story, and Our Models Are Behind It.

• Lam Research Q3 2026: A Constrained $200 Billion, and NAND Gets Its Own Trade Ratio.

• SpaceX Q3 2026: The Newest Hyperscaler Prices Its Watts at Spot.

• Micron’s AI Memory Roadmap: Bandwidth Moves Onto the Compute Die, and Wafers Stay Tight.

• Why GPUs Can Arrive Before the Data Center Is Ready.
  

On October 19, Meta published a detailed engineering account of Prometheus, its gigawatt-class training cluster in New Albany, Ohio, and the same day we sat down with Santosh Janardhan, who leads Meta’s infrastructure, and Ian Buck, who runs NVIDIA’s hyperscale and HPC business. We wanted to understand what it took to build, and more importantly what the two companies think it was for.

Meta did not have one building large enough for what it wanted, so it joined together several structures (scale across), and wrote the software to make them look like one cluster to its researchers. It trained models while it was still adding capacity, and it had to drive down a failure rate that, by our arithmetic, meant something broke every few seconds across a large job at its mid-2025 peak. Teams from both companies rotated through Ohio for months to trace those failures to their source, and Meta says it has since cut the interruption rate by 99%.

The engineering is impressive, and what we find most interesting is the economic argument underneath it. Our models have power holding the energized accelerator fleet to roughly 72% of what demand calls for by 2030. When watts are that scarce, the amount of useful training a company gets out of each one becomes a competitive variable in its own right, and Meta has now published enough to measure it. The payoff shows up in products: a company that gets more research out of the same power can train a better frontier model, adapt it into more products, and distill it into smaller models that are cheaper to serve. Meta’s third-quarter results, reported this week, show both sides of that model. Mark Zuckerberg pointed to Muse Wonder, Meta’s first frontier model and the latest trained on Prometheus, as evidence of “velocity across the board,” and more than 10 million people in the United States and Canada set up a Muse agent in its first six weeks. On the other side, capital spending of $34.9 billion in the quarter pushed free cash flow below zero.

Key insights from the conversation and Meta’s disclosures:

• Meta reports goodput, the share of training time that produces useful progress, above 85% for pretraining its Muse Wonder model and above 90% for reinforcement learning. Mean time between interruptions is above 54,000 GPU-hours in the latest runs, against a target of about 100,000.

• On a gigawatt-class site, ten points of goodput can deliver substantial energy savings, but the gain does not translate one-for-one into power savings. Depending on operating conditions, the equivalent power benefit could be roughly 50–75 MW, and no operator can bring new power at that scale online quickly in the current market.

• Reliability carries a hardware cost. Meta runs large training jobs on 64 of the 72 GPUs in each NVL72 rack and keeps the rest as spares, so roughly 11% of the GPUs in each rack are there for resilience. This enables both optimal power-of-two GPU sizes and resilience for fast failure recovery.

• The same cluster that trains the frontier model also produces the product versions of it. Meta post-trains one model differently for agents, coding and the assistant, then distills it down, and as Janardhan put it, “you have to create a bigger model before you distill it down.” Training capacity sets how many of those products Meta can build and how cheaply it can serve them.

• What Meta learned does not reset with each cluster. GB300 picked up on the reliability curve where GB200 left off, and Janardhan described starting over with every new cluster as “a fundamental disadvantage.”

• We think NVIDIA benefits more than most investors assume, because the fixes that came out of Prometheus land in its platform for every customer at the same time our models have custom silicon gaining unit share.

The full analysis sizes what goodput is worth in dollars on our own unit costs, explains why we think the Street does not model it yet, and works through what Prometheus means for Meta, NVIDIA, Arista and Broadcom, including where Meta’s lead stays in-house and where it spreads to rivals through NVIDIA.

Companies covered in this report: Meta Platforms (META), NVIDIA (NVDA), Arista Networks (ANET), Broadcom (AVGO), CoreWeave (CRWV)

What subscribers get in the full report:

• What ten points of training goodput is worth in dollars on a gigawatt-class site, built from our own shell and accelerator unit costs, and why it is the cheapest capacity a power-capped operator can add.

• Our 2030 power cap, with how much of the accelerator fleet we think gets energized, how many accelerators end up waiting for power, and why that makes useful compute per dollar of capex the number to watch.

• Why the Street does not model goodput yet, and what we would track in its place as Meta spends $135 to $145 billion this year.

• The product case in three parts (better models, more of them, and cheaper ones to serve), tied to Muse adoption and Meta’s serving-cost targets from its third-quarter results.

• What owning training and serving capacity does to unit costs, with a heavy personal-agent user priced at public API list rates against our modeled cost to serve.

• Meta’s energized capacity path through 2030 from our DC Capacity Tracker, including Prometheus and Hyperion, and why the reliability lessons carry to each new site.

• Company by company views on Meta, NVIDIA, Arista and Broadcom, including why we think NVIDIA gains more from Prometheus than most investors assume, and where Meta’s lead spreads to rivals.

• Three exhibits: Prometheus reliability from the mid-2025 peak to Meta’s target, the power cap and the value of goodput, and the owned versus rented cost of serving an agent.

Go deeper in Atlas. Subscribers get the Atlas companion packet for this report, with the capacity, power, accelerator share and networking tables behind it, Meta’s reported reliability metrics, and the full set of attributed quotes from our conversation. Ask the Atlas Agent to work with any of it, or start from one of these prompts:

• Size what training goodput is worth in a power-capped market

• Track Meta’s capacity and reliability against our signals

Read more

We have been early on the demand side of the AI funding debate this year, tracking who pays for AI alongside how much compute gets built. That work now runs through our models. Bull and bear cases for AI monetization run in parallel against the accelerator demand we forecast and the pricing and gigawatt capacity we track. As enterprise and consumer software spending on AI begins to ramp, we want to know as much as anyone whether AI can fund itself.

The bear case rests on a range of consensus estimates (which we model as well): the industry will spend roughly $9.6 trillion on AI data centers between 2024 and 2030, while the AI revenue we can track over 2026 to 2030 adds up to about $4.1 trillion.

The bear case is fair on one point: the spending is real, and most of it lands years before the revenue does. We think the comparison still asks the wrong question. Hard assets are paid for once and expensed over their useful life. AI revenue is recurring and, so far, compounding. The fairer test is whether each year’s AI revenue covers each year’s cost of owning the fleet: depreciation on the hardware, the power to run it, and a reasonable return on the capital tied up in it. A lender would underwrite the fleet the same way, and we run that test here with the connected models in the CS Forecast Stack.

On that basis, our view is that AI does pay for itself, but later than the optimists hope and for different reasons than most of the debate assumes. We forecast in ranges, and we show them here, because the spread is as informative as the central number.

In our central scenario, revenue covers about two-thirds of the bill today and about 85% by 2030. We estimate the annual cost of owning the global AI fleet rises from about $0.41 trillion in 2026 to about $2.27 trillion in 2030. Tracked AI revenue, measured as an end-customer run-rate, rises from about $275 billion to about $1.93 trillion over the same period. Coverage improves every year, from 67% to 85%, and if revenue keeps growing faster than the fleet after 2030, the two lines cross in the early 2030s. Across our Low and High scenarios, 2030 coverage spans roughly 50% to 135%, so adoption carries most of the uncertainty, and in the Low scenario we would not expect a crossover this decade.

The unit economics already work. A gigawatt of AI capacity costs roughly $34.5 billion to build and equip. When that gigawatt is rented for inference, it recovers its cost in about three years at our central rental price, and in roughly two to six years across our price range. The problem is the average gigawatt, which takes about seven years to pay back, longer than the six-year life we give the hardware.

The gap sits in absorption. In our central 2030 forecast, about 42% of AI capacity is paid for by revenue we can track. Roughly 37% trains models, and the remaining 21% runs inference that nobody bills as AI revenue: ad ranking, search, recommendations, free consumer tiers, and assistants bundled into other products. Very little of it sits idle. Much of it is monetized in ways that never appear on an AI revenue line, and our first estimate of its value, roughly $0.35 trillion to $0.48 trillion a year, would lift 2030 coverage to about 100% to 106%.

Absorption and pricing close the gap, and accounting only narrows it. An eight-year hardware life lifts 2030 coverage to 96% and 10% less capex to 92%, while half the fleet earning rent, or agents lifting revenue per gigawatt by about 20%, each clears 100% on its own.

For investors and stakeholders, the question becomes which of those levers moves first. Our thesis is that agentic workloads move both demand levers: they raise the share of capacity doing billable work, and they are increasingly priced on work completed instead of tokens consumed (we have modeled price by tokens consumed as well). The weak link is the frontier labs, where compute absorbs roughly 80% of revenue in our forecast and outside capital funds most of the buildout they anchor.

What subscribers get in the full note

• How we build the annual bill, and why a return on capital belongs in it

• Who pays: the 2030 revenue mix, and why consumer subscriptions are only about 4% of it

• Unit economics per gigawatt, and why $50 billion cost estimates do not break the case

• Our first sizing of the unbilled 21%, and how far it pulls the crossover forward

• What the Low and High scenarios assume, and what would have to be true for each

• Who carries the risk, from hyperscalers and neoclouds to labs and suppliers

Three connected models. This analysis ties together three models in the CS Forecast Stack. The AI Accelerator Market Outlook sets the spending path and the installed fleet, including how much of it trains models. AI Factory Power & Economics turns that fleet into gigawatts, energy, and the economics of a rented gigawatt. The AI Revenue Tracker supplies revenue, lab finances, and the Low and High scenarios. Atlas subscribers can open each model, see every forecast in all four scenarios, and query them with the Atlas research agent.

Companies discussed, by where they sit in the risk: hyperscalers (Microsoft, Alphabet, Amazon, Meta, Oracle); neoclouds (CoreWeave, Nebius); frontier labs (OpenAI, Anthropic, xAI); suppliers (NVIDIA, AMD, Broadcom, TSMC, SK hynix, Samsung, Micron).

Read more

• TSMC Q3 2026: The Foundry Checked the Power Before It Poured the Concrete

• ASML Q3 2026: Pricing Power Arrives and the Order Book Reaches 2030

• Applied Materials Investor Breakfast: Yield Becomes Capacity

• Silicon to Systems: Every Fix Moves the Bottleneck, from Applied’s EPIC Center panel

• NVIDIA’s 800 VDC Wall: GaN Gets the Attention, SiC Gets the Dollars

In January we argued that this equipment cycle would run longer and higher than the industry’s history suggests. Demand for tools had decoupled from wafer volume, for three reasons. Gate-all-around transistors and backside power were cutting tool reuse at 2nm. HBM was consuming three to four times the wafer capacity of standard DRAM for the same bits. And advanced packaging had become a second bottleneck next to lithography. On that basis we put 2026 wafer fab equipment at $125 to $129 billion in our base case, with upside to $145 billion.

We were too cautious in our approach but between earnings from ASML and TSMC as well as many of our conversations with the semicap ecosystem at Semicon West this week we are now leaning to our more bull case as the base case. Our WFE Core model now has 2026 at about $153 billion in the Base case, up roughly 28% from 2025, and at about $157 billion in our High case. Both numbers sit above the bull case we published nine months ago. The direction of the January call held, but we underestimated how quickly memory makers and leading-edge foundries would commit capital once AI demand proved it would last, and how much more equipment each new wafer would need.

Three events this week gave us a clear read on where the cycle stands. At its investor breakfast, Applied Materials described yield as the largest change its CEO has seen this year. Leading-edge logic and DRAM fabs are running at roughly full utilization, and cleanroom space is setting the pace of tool installs. ASML reported a record quarter and raised its 2026 outlook to roughly €44.5 billion. It also told investors it had negotiated price increases on its most critical tools for the first time in its history. A day later TSMC raised its 2026 growth outlook to about 44%, said AI demand is still far above what it can supply, and told investors its capital spending from 2026 through 2028 will be significantly higher than in the prior three years.

The common thread is a change we have been tracking all year: the hardest problems in AI hardware have moved from the transistor to the system. We think this, and its implications for chip vendors, is being entirely misunderstood by the market. Accelerators already sit at the reticle limit, packages are approaching the largest size anyone can build and cool, and the remaining gains come from moving each bit a shorter distance at lower energy and from getting heat out of the chip. Each of those fixes becomes a new process step, a new material or a new tool, and each one opens a new kind of chip design that needs still more steps. We described the engineering side of this in our note from the EPIC Center panel, where engineers from NVIDIA, Qualcomm and Broadcom walked through the constraints they face. This report puts numbers on what those constraints mean for equipment spending in what we believe is a structural growth cycle in WFE that will materially change the future shape of the semiconductor industry.

In the full report:

• Why we keep our Base case as the anchor while the evidence leans toward High, and the three signals that would move our 2030 anchor from about $278 billion to about $308 billion

• How this week’s numbers from ASML, Applied Materials and TSMC line up against our Base and High cases

• Memory’s rise to more than half of device WFE, and the DRAM tool spending path behind it

• The Angstrom-era transitions, from gate-all-around and backside power to 4F² and 3D DRAM, and what each adds to tool spending per wafer

• The silicon-to-systems constraints (DRAM on logic, system-on-wafer and co-packaged optics) and the tool categories they land in

• Where the dollars land by tool category, region and vendor, with 2030 revenue paths for ASML, Applied Materials, Lam Research, KLA and ASM International

• Five exhibits and the six signals we are tracking over the next two quarters

Companies covered in this report: ASML (ASML), Applied Materials (AMAT), Lam Research (LRCX), KLA (KLAC), Tokyo Electron (8035.T), ASM International (ASM.AS), TSMC (TSM), Samsung Electronics (005930.KS), SK Hynix (000660.KS), Micron (MU), Intel (INTC), NVIDIA (NVDA), Broadcom (AVGO), Qualcomm (QCOM)

Subscribe to read the full analysis, including our scenario forecasts through 2030 and the segment, tool and vendor detail. All reports and data are powered by the CS Forecast Stack.

Go deeper in Atlas. Atlas subscribers can query the forecast tables and evidence behind this report, across all four scenarios, and subscribers with model access can open the WFE Core Model V2.3 directly. Ask the Atlas Agent to break down any case by segment, tool category, region or vendor, or start from one of these prompts:

• Compare Base and High in the WFE model

• Check the three signals that would move our anchor

Read more

Semiconductors in the post-device era

For four decades, the semiconductor industry grew on device adoption cycles. PCs, then smartphones, then tablets and wearables each required consumers and businesses to buy new hardware before they could get the benefit of the new thing. Also, importantly, each cycle left behind a larger installed base of connected users. Every prior transformational technology worked this way. Electricity required wiring a home, the internet required a PC and a modem, and mobile required new hardware via a handset, then smartphone. AI is different, and we think it is one of the reasons this cycle may play out differently than the ones before it. The installed base of billions of PCs and smartphones that those cycles built is the distribution channel for AI software, so people can use it today on hardware they already own, with no adoption cycle of new and costly devices to wait for. This is an entirely new dynamic we have not had before with any prior transformational technology where demand is already here we just need the compute that can sustain diffusion at scale. Hence, the build happens in the data center instead, where the compute to serve billions of daily users has to be manufactured, energized and deployed. We think of this as the post-device era: the device cycles set the table, and AI software diffuses across what they built with an economic TAM magnitudes larger than any prior cycle.

That changes what a weak device market means for the industry. In every prior cycle, PCs and smartphones were the wafer volume that filled leading-edge fabs, so a down device cycle hit the whole semiconductor industry. Foundries lost utilization, equipment orders were pushed out and memory makers cut production. A PC downturn as severe as the one under way would once have been extremely bearish.

Today the fabs have a different anchor. Data Applied Materials presented at the opening of its EPIC Center this month shows data center chips rising from roughly 12% of leading-edge foundry-logic wafer starts in 2020 to about a third in 2026, crossing mobile for the first time, and heading toward roughly half by 2030. PCs fall from about a third of leading-edge logic starts to roughly a fifth over the same period. In leading-edge DRAM the crossover came earlier, in 2023, and data center now accounts for more than half of wafer starts. When the fabs are anchored by data center demand that is short of supply, wafers released by a PC downturn do not sit idle. They move to data center compute.

Exhibit 1. Applied Materials, EPIC Center presentation, October 2026; redrawn by Creative Strategies with approximate values.

This report looks at how large that shift could be. The processors, SoCs and discrete GPUs inside a PC sit on the same leading-edge TSMC nodes that AI accelerators and server CPUs need, and a foundry that is sold out on AI work will place released wafer starts with whichever customer is short of supply. So the question we set out to answer is how much leading-edge capacity a weaker PC market actually releases, how much of it can become accelerators and server CPUs that ship to unforeseen upside, and who has the supply chain prowess to use it. Our short answer is that accelerators gain less than the headline suggests, server CPUs gain more than current models assume, and the companies best positioned to capitalize include one most investors would not expect.

In the full report for subscribers:

• Our PC forecast for this year and next, plus a downside case where 2027 falls a further 10% to 15%, and why we think the risk sits below consensus.

• -The wafer math. How much leading-edge capacity a weaker PC market actually frees, broken out by node and by year, and how much of it can realistically move.

• Why accelerators gain less than the headline suggests. Packaging caps the upside in every case we model, so the freed wafers act as supply insurance, not new deployments.

• Why server CPUs are the cleaner beneficiary, and how much of the projected server CPU supply gap the freed wafers can close.

• Where the rest of the supply chain still has to deliver, from CoWoS-class packaging and memory to substrates and power.

• Our ranking of who is best positioned to capitalize, including the one company where we think the market has the PC downturn most wrong.

• What would change our view, and the signals we are watching.

Read more

In “Bet on Bits, Not History,” we argued that the memory revenue floor is bits at the bottom of the cycle times the price at the bottom, and that the floor has moved several times higher than anything in the industry’s history. Our Micron fiscal Q4 note showed the company-level evidence: customer agreements covering more than 35% of Micron’s revenue through 2030, about $150 billion of remaining performance obligations valued at minimum contract prices, HBM repriced higher for 2027, and no supply relief until 2028. Both notes left one question open. How are valuations being modeled?

Our answer is that the market has moved off history on the level of earnings but still prices more than history on the depth of the next downturn. The three suppliers trade at roughly 5.4 to 6.8 times run-rate earnings, a classic peak-cycle range that says investors expect earnings to fall. Set against our Core Memory model, the multiples also say how far. As a hypothetical, assuming a 40% mid-cycle net margin, Micron is priced for a downturn one and a half to two times the size of the last two cycles, Samsung for one two and a half to three times that size, and SK hynix for one three to three and a half times that size. The gap is widest at the supplier with the most HBM in its revenue, the opposite of what our thesis says the market should be doing.

This note is not a call on when or if a downturn comes or how deep it runs. Memory may be at or near a peak, and among the investors we talk to, the working assumption is that today’s margins cannot hold and that a sharp correction arrives near the end of the decade. Our argument from Part 1 is that the next low will not look like the lows of the past, because the floor has moved. So for the sake of argument, we take each market value, assume a through-cycle margin well below today’s, since we expect memory to give back margin to the rest of the accelerator stack, and ask how deep a downturn the price already implies. The SK hynix gap is also not mainly a disagreement with our forecast. Published consensus earnings for SK hynix in 2028 land close to what our forecast produces at a mid-cycle margin, and the stock trades at roughly four times that consensus figure.

How we read a memory valuation

The analysis runs the valuation backward. We carry each market value forward two years at a 10% cost of equity and divide it by the 12 to 15 times the market would pay for normalized, through-cycle earnings. That gives the normal-year net income the market is capitalizing. We compare it with what our model says each company earns in 2028 at a mid-cycle net margin, and call the revenue shortfall the implied downturn. It is what the price assumes, not what we forecast.

The right comparison is a through-cycle average, not the trough. Memory revenue fell by roughly one-third from 2018 to 2019 and about 40% from the 2021 peak to the 2023 bottom, but average revenue from each peak year through the year before the next peak ran about 19% below the peak in both cycles. We use 19% as the historical benchmark and roughly 38% as a downturn twice that size.

The multiple, the cost of equity, and the margin are our choices, and the margin matters most. For the example, we assume a 40% through-cycle net margin for all three suppliers, well below today’s 70%-plus and well above the losses of 2023, as an assumption for the argument, not a margin forecast. Micron’s own guidance argues for more. Management expects margins at the floor prices in its customer agreements to sit meaningfully above any prior cycle peak, and its net margin at the fiscal 2018 peak was in the mid-40s. That points to 50%, so we show both, and a higher margin deepens every implied downturn below.

Starting well below today’s margin reflects what our accelerator and memory models show when they are run together. In one sensitivity in our forecast, HBM passes half of accelerator package value by 2030, and the memory in each package comes to cost more than its logic. That package is sold by accelerator vendors earning gross margins in the 70s and built with memory from suppliers earning net margins above 70%. We do not think buyers will pay for both for long, so something has to give, and we think it is memory. Accelerator vendors own the platform and the customer and can qualify a second and third HBM supplier each generation, while a memory supplier sells a part qualified into someone else’s platform. The 40% assumption builds that scenario in.

Similar multiples, different businesses

The market is valuing the three suppliers within a narrow band on current earnings.

Samsung’s preliminary September-quarter figures, released on October 8, put group operating profit at about KRW 107 trillion on sales of about KRW 195 trillion, up about 20% and 14% from the June quarter and slightly ahead of published consensus. That lowers Samsung’s multiple from about 6.4 times to about 5.4 times, in line with what we expected when this note first ran. SK hynix still annualizes the June quarter and reports late this month, and on current expectations its multiple falls into the same 5 to 5.5 times range. Micron annualizes guidance for the quarter ending November. Samsung’s value also includes foundry, mobile, display, and other businesses, but memory produced almost all of its June-quarter operating profit, and segment detail arrives with full results on October 29. The market is applying one peak-cycle range across the group, which means it is pricing the cycle and not the differences between the companies.

For subscribers, the full note covers:

• The normalized earnings each market value supports, and the downturn each price implies at a 40% and a 50% margin

• The accelerator and memory model numbers behind our margin assumption

• What each stock is worth if the next downturn is half the size of history

• Why the market’s ranking runs opposite to HBM mix

• Why the next low should sit well above past lows, including our thesis on memory earning on older nodes the way logic does

• What would prove us wrong

Subscribe to The Diligence Stack with Atlas to read the full note, with the exhibits and the Atlas companion packet behind it.

Read more

Salesforce reported Agentforce ARR above $1.5 billion in its fiscal second quarter, up more than 240% year over year, and it prices an Agentforce action at $0.10 in Flex Credits. In the same quarter, 82 trillion of the 104 trillion records flowing into its Data 360 platform arrived through zero copy, up 731% from a year earlier. ServiceNow said its AI annual contract value passed $1 billion, with a $1.5 billion target for year-end. Vendors are doing two things at once, metering what agents do inside their systems while opening read access to data that lives outside them.

We rebuilt our sizing around this added dynamic. On our base case, recurring enterprise spend on agent data services grows from roughly $3 billion to $4 billion in 2026 to roughly $25 billion to $37 billion in 2030, with our base case near $30 billion. Metered access paid to systems-of-record vendors is about half of that spend today and about 30% by 2030, growing more than fivefold in dollars while losing share to retrieval, memory and write-back services. The larger bill comes first, because one-time spending to make enterprise data agent-ready runs at roughly $20 billion to $40 billion this year and $65 billion to $145 billion by 2030, still about three times the recurring spend in that year on our base case.

The correction (update) to our May piece concerns bytes. When we built the enterprise side from the bottom up, the data agents leave behind in a company turned out to be small: the indexes they search, their working memory and the records and work product they create add up to a little over 2 exabytes across all enterprises by 2030 in our base case scenario. That figure rests on assumptions we have not yet been able to measure, chief among them how much output agents generate and how long enterprises keep it, and both will rise as more autonomous agents get write access. Stressing them together, with ten times the work product, seven-year retention of fuller agent traces, and three times the index footprint and agent memory, takes the total to roughly 13 exabytes. That is a real increase, but it is still less than a tenth of the roughly 180 exabytes of flash our memory model dedicates to inference context, where long conversations and agent sessions spill out of memory onto flash in the data centers running the models. So the storage shock is real, and it lands wherever inference runs: mostly in hyperscale and neocloud fleets, but also in the enterprises that run inference on their own infrastructure. For the enterprise CIO, the bill is for access, governance and readiness, and the question for investors is who gets to collect the toll.

For subscribers

The full report lays out the framework we now use for this market. Every enterprise reaches agent-ready data by one of three roads: it pays the system-of-record vendor’s toll, it federates and queries data in place, or it copies data out into a lakehouse it controls. The mix of those roads decides who captures the spend far more than it decides the size of the spend. We size all three roads, show where read tolls are eroding and write tolls are holding, and map the vendors across the Connect, Manage and Use layers.

Subscribers also get the capture math by scenario. Systems-of-record vendors keep anywhere from 18% to 45% of recurring agent data spend in 2030 depending on the road mix, a range we think investors have not yet modeled. The report closes with the gross-versus-net budget view, the indicators we are tracking quarter by quarter, and what would change our view. Eight exhibits and the model behind them are included.

Inside the full report:

• The three roads to warm data, and why the road mix decides who gets paid

• Read tolls versus write tolls across Salesforce, ServiceNow, Microsoft, SAP and Workday

• The readiness bill: roughly $18 billion to $40 billion this year, built per functional area and checked against the services market

• Connect, Manage, Use: where $25 billion to $37 billion of recurring spend lands by 2030

• Capture by scenario: why systems of record keep anywhere from 18% to 45%

• What it means for Salesforce, ServiceNow, Snowflake, Databricks, Microsoft, MongoDB and the systems integrators

• Gross versus net budgets, and the three tests we are tracking each quarter

The report also has a deeper backend. Atlas, the Creative Strategies research portal, holds the companion packet for this report: the scenario outputs behind every exhibit, the toll calibration and the readiness build, with an agent you can question directly, for example on how capture changes if read access stays metered. Subscriptions with Atlas access include it.

Read more

Every note we publish rests on more work than fits on the page: the models behind the forecasts, the scenario outputs, the assumptions and how confident we are in each, and the data behind every exhibit. Atlas, the Creative Strategies research portal, is where that work lives. Each major report now comes with a companion research packet in Atlas, and our research agent is built to work through it with you. You can test our assumptions against your own, rerun the analysis under different inputs, and get to the specific numbers that matter for your work, without waiting on a call.

If you have Atlas access, the companion packet for this note is live now. Open it, ask the agent your hardest questions, and push on the parts of the argument you find least convincing. That is what it is there for. If you would like access for yourself or your team, reach out to discuss.

Additional research coverage in CS Atlas:
Micron Q4 FY26: The Memory Cycle Is Carrying a Larger Bit Base
HPE Networking Day Takeaways on Open AI Infrastructure
Synopsys Investor Day Connects Agentic EDA to Commercial Value

We understand and appreciate how much the historical lens on memory is still influencing how the market is viewing the category. Seasoned investors have spent decades learning to value memory suppliers on where they sit in the price cycle, and for most of that history it was the right approach. We think it now understates what this industry earns through a full cycle driven by AI—the workload that changed it all. In it’s simplest form, memory revenue is bits times price. In the two decades of covering memory and with each cycle we have observed, price did the swinging and bit demand kept compounding underneath it, so the floor of any downturn comes down to two things: how many bits ship at the bottom, and what they sell for when they do. The first of those has changed quite a bit. Our Base case has the industry shipping about 2.5 times as many conventional DRAM bits in 2030 as it did in 2026. HBM, which was a negligible share of revenue in the 2019 trough, reaches close to two-fifths of sector revenue by 2030.

The way we think about it is that the cycle is the waves and the bits are the tide. Every downturn is measured against a shoreline, and the tide in memory is coming in faster than at any point we have seen. Our thesis, and what we have modeled, is that the TAM is still rising, that there will be real downturns along the way, and that those downturns will play out from a revenue base several times larger than anything the industry earned before 2025. We call this betting on bits instead of history. The approach accepts that memory, at least parts of it, stays cyclical, and the one question it asks is whether the bits keep coming. Our model says they do, simply because the workload demands it, and by the end of the decade they arrive faster than the suppliers can build for them.

What history does and does not show

The usual argument for a higher floor points at the history and says each trough sits above the last one. The record does not cooperate, and any investor who knows the data will raise it early. By our estimates, memory revenue bottomed at $105 to $110 billion in 2019. After peaking at $150 to $155 billion in 2021, it fell to $85 to $90 billion in 2023, below the prior trough. The drawdowns were not gentle either, about one-third from 2018 to 2019 and close to 40% from the 2021 peak to the 2023 bottom.

That is a fair objection, and it is the reason we do not build this case on the revenue line. Revenue history blends two series that behave very differently. Bit demand grew through every one of those cycles. The damage came from realized pricing, which fell by about half in each downturn as new supply and customer inventory corrections landed at the same time. So a floor drawn from past revenue troughs inherits every accident of timing in that pricing. A floor drawn from bits asks something narrower. How much memory does the world need at the bottom of the next cycle, and what is the lowest price the suppliers have ever accepted for it?

Exhibit 1 puts the two regimes side by side on a log scale. Nothing the industry earned from 2014 through 2025 reaches our Low case in any forecast year, and that includes the 2018 peak of $155 to $160 billion. The Low case is where accelerator demand and pricing soften together, and even there 2027 revenue comes in near $712 billion, about 4.5 times that peak. We do not treat the Low case as a floor, and the floor itself comes in the next section. Even our downside case, in other words, lives in a different revenue regime from every cycle in the historical record.

Full Research Note for Subscribers

• The trough-price floor: every bit in our 2026 to 2030 forecast priced at the worst realized prices of the last cycle, showing where the floor sits with pricing optimism taken out entirely

• The downturn already in our Base case: the price cycle our forecast already carries, and why revenue holds through it

• The supply balance: why the true shape of the TAM remains unobserved

• What Micron’s latest print adds: the new evidence for the floor argument, and where it pushes back

• What would break the view: the four conditions we are tracking

• What it means for the suppliers: who the floor favors, and how investors should size it

• The Atlas companion packet: rerun the floor with your own trough prices and bit paths, and take your questions to our research agent

Read more

Agents create more work for the rest of the computer

Late last year, when most industry discussion centered on GPUs, we began making the case that the CPU was about to regain strategic importance. We formalized that argument in March with Secret Agent CPU, which set out our preliminary TAM forecast and explained why agentic AI changes the CPU-to-GPU ratio in the data center, and we raised the model in Secret Agent CPU, Revisited. The view came from close study of the workload, which is how we approach every compute cycle.

The simplest framing is that the GPU does the thinking and the CPU carries out the actions. An agent runs tools, modifies files, queries databases, and evaluates results before deciding on its next step, and that execution runs on CPUs. A coding agent illustrates the pattern: it writes a change, runs the tests, reads the failure, and iterates. Faster inference shortens the reasoning steps, faster CPUs return test results sooner, and additional cores allow more of those jobs to run concurrently rather than queue.

We think it is also important to note that this work extends beyond new AI infrastructure. An agent calling a business service creates work in that service’s software and databases, so part of this demand lands on server CPU infrastructure of every kind, including general-purpose fleets running workloads that have nothing to do with AI. Agents also operate at machine speed, issuing requests far faster than a person can click through an application, and we think that pace means general-purpose infrastructure will need to be upgraded as well as expanded to keep up. Some of the new work fits within existing headroom. The remainder, together with that upgrade cycle, is where the structural expansion of this market begins, and it is the portion this guidebook is designed to measure.

More workers and faster workers

In the AI factory, GPUs run the model's reasoning, CPUs carry out the resulting work, and memory holds work in progress. We describe CPU design along three dimensions. Wide is the ability to run more independent jobs, fast is the speed of an individual worker, and fed is how much memory bandwidth, cache, and capacity each active worker can draw on. More cores serve concurrent users and parallel tools, faster cores shorten dependent steps, and a design that cannot keep its workers fed delivers the full benefit of neither.

Most server CPUs in service today were tuned for a different workload, with cloud parts optimized for core density and cost per virtual machine and accelerator hosts optimized for feeding GPUs. These we have defined in the past as cloud native CPUs. Agentic work asks for all three at once, and because every design trades width, speed, and memory against a fixed budget of power, area, and cost, no single balance point fits every agentic workload. That is why we expect CPU designs built specifically for agentic execution (agentic native CPUs), and why the vendor approaches in this guidebook differ as much as they do.

Exhibit 1. CPU work extends from the agent environment into existing business systems

Conceptual workflow; connections do not imply one coherent hardware domain. Source: Creative Strategies analysis.

Public runtimes already separate model serving from execution, and the always-on designs differ in where that execution happens. Meta describes an isolated environment for Muse’s browser and tools, and Anthropic separates its harness, execution containers, and durable session state (Meta, Anthropic). In our testing of OpenAI’s dots, the agent keeps its own cloud Linux environment available and can also delegate work to a user’s linked computer, which moves part of the execution load out of the data center. None of these designs implies a physical core permanently reserved for each user. Where the work runs, however, changes what each operator needs from cores, memory, storage, and networking, and we examine those differences in the full report.

More activity does not automatically mean more purchases

The CPU opportunity depends on the amount of useful work growing faster than efficiency gains and existing capacity can absorb it. Sharing and suspending environments can support a large agent population with modest purchases, while memory limits can require more servers yet favor lower-cost CPUs. Competition divides along similar lines. A host that exchanges data constantly with an accelerator benefits from close integration, an independent tool pool rewards execution performance, memory, and software support, and different suppliers can win each within the same deployment. We also expect every implementation of personal agents to vary by provider, shaped by the strengths of its models and the infrastructure it can access, as the contrast between Muse and dots already shows. There is no standard way to build an agent computer, and that variety is why CPU vendors are taking such different approaches to cover the breadth of use cases.

Our measure is completed tasks at consistent quality and response time, relative to full-system cost and energy. The full report applies that measure to compare vendors, then separates the capacity customers require from the CPUs they purchase each year.

Inside the Full Report

• Our server CPU forecast through 2030, with value nearly tripling to about $220 billion, merchant sales separated from captive silicon, and the split across general-purpose servers, accelerator hosts, and agent execution.

• Where NVIDIA, AMD, Arm, Intel, Qualcomm, and the hyperscalers are strongest, architecturally and commercially, and which agent workloads each design fits best.

• Why dedicated CPU racks are moving into the scale-up domain, the CPU-to-GPU ratios they imply, and where mixed-vendor configurations still need proof.

• How Muse and Dots place agent work in different parts of the stack, from our own testing, and what each design demands of cores, memory, storage, and networking.

• One agent workflow traced step by step, showing where CPU speed reaches the user and when it raises GPU utilization.

• A sizing method that links task activity to cores, environments, and racks, with a worked comparison of wide, fast, and fed designs under queues and memory limits.

• Atlas integration companion packet for subscribers to go deeper with the research.

Read more

We have spent the last several weeks working through the optical supply chain and the manufacturing steps that will set the pace of optical content inside AI systems. Demand is the easy part of the forecast everyone would adopt optical today if could be manufactured and was reliable at scale. The industry needs more bandwidth between accelerators, and copper hits a physical limit as link speeds go up and scale-up domains get bigger. The harder question, and the one we care most about, is how fast the supply chain can turn that demand into qualified optical engines that ship at repeatable yield.

In our inter interview Lightmatter CEO Nick Harris gave us a closer look at one of the more differentiated attempts to solve that problem. Most commercial lasers are still made on small indium phosphide wafers by a fairly small group of suppliers, and Nick makes the case that Lightmatter can move more of that work onto the 300-millimeter silicon base mature CMOS foundries already run. If that holds up at volume, it opens a different path to laser capacity. He walks through how the approach works and how Lightmatter is designing around the reliability problem that comes with thousands of optical links in one system.

He also gives his own timeline for when near-package and co-packaged optics ramp, which we found useful to hold up against our optical attach ladder, along with the reliability data Lightmatter has built up so far. As always, we would separate a product shipping from broad attach across deployed racks, and that gap is a big part of what we dig into with Nick Harris. The full video, audio and edited transcript are available to subscribers below.

Subscribers with CS Atlas access can deeper on Lightmatter and the wider optical transition. The Lightmatter packet breaks down Passage and Guide and places the company on our timing ladder from NPO through CPO and the longer-term photonic-interposer roadmap. A separate manufacturing section tests the 300-millimeter claim against the yield and qualification steps that still have to be proven.

Subscribers can then use our optical-solutions comparison and interconnect SWOT to see which companies control each part of the stack, how their products differ, and where they are focused across scale-up, scale-out, and scale-across networks. Each company entry maps the manufacturing unit it must scale and the evidence we need before moving its timing or competitive position higher.

For Subscribers

• Full video interview

• Why Lightmatter views interconnect as the problem that now sets the pace of compute.

• How Passage and Guide address optical links from NPO through a photonic interposer.

• How bonding indium phosphide onto 300-millimeter silicon photonics wafers changes the manufacturing approach.

• Why reliability, optical test capacity, and foundry process control still govern scale.

• Where Nick expects NPO, CPO, and photonic interposers to enter the deployment curve.

Read more

In our Bloom Energy report, we looked at behind-the-meter power as a way to get data centers running sooner while customers wait on utility service. We wanted to take that question to Billy and get a better sense of what it actually takes to make these projects work. We’ve made the point in prior reports that customers may be willing to pay more for earlier access to power, and we still think that’s true. But the developer still needs an agreement that covers the cost of building and operating the plant.

Billy spent a good chunk of our conversation on a problem we had given less attention to. The power plant is a small share of the total investment at a compute site, but an outage can put a much larger investment at risk. So the question becomes how much of that risk the power developer can afford to take on. Adding spare generation or batteries costs money, and the customer’s contract has to run long enough for the developer to earn it back. That gives us another way to look at the BTM opportunity from the Bloom report: what reliability the customer needs, what it costs to provide, and whether the two sides can agree on who pays when something goes wrong. It also adds a piece we were missing in our TCO analysis and in our Gigawattonomics model, as we track AI capex ROI all the way down to the power shell provider.

The conversation also helped us understand another challenge with running data centers on natural gas, which is getting enough gas to the site. The US has plenty of gas, but a data center may still need a pipeline extension or more compression before it can use that supply. The companies building that infrastructure need customer commitments to justify the investment, so it becomes one more project that has to be funded and built on a schedule that lines up with when the data center needs power. It’s the same pattern we keep seeing, with the constraint moving further down the stack.

Below, we get into how these projects come together, why the contracts are hard to structure, and where Billy sees the next constraints showing up.

Our companion report develops the investment implications through project examples, gas-demand scenarios, and the economics of dedicated generation and who is best positioned. Subscribers, with Atlas access can explore our full power landscape research, and BTM / natural-gas and compute research alongside the interview.

In the conversation:

• Why generation, power delivery, and the customer’s opening date are difficult to align.

• What happens when a power supplier cannot buy replacement electricity during an outage.

• How reliability requirements and contract length affect the price of power.

• Why pipeline companies need credible customer commitments before construction begins.

Read more

In Secret Agent CPU, Revisited, published May 26, we raised that forecast. Our conversations across the industry were strengthening our conviction that dedicated CPU capacity would become an essential part of running agentic systems at scale. We remain bullish on CPUs because of the clarity in inference workload demand and the clear role the CPU plays becoming a larger infrastructure requirement.

But while we talk a lot about the CPU-to-GPU ratio, it is very hard to isolate that number and turn it into a demand forecast. The mix depends on the infrastructure operator and the decisions it makes about serving AI. Key point here, each hyperscalers ratio will be different. There will be no standard CPU/GPU ratio configuration. The main key here is to understand those decisions, we need to understand the scale-up domain and why this is priority number one for serving inference at scale and monetizing compute infrastructure.

Why we add “domain” to scale-up

A lot gets talked about scale-up, which involves connecting accelerators so they can work closely together in a single compute fabric. We add “domain” firstly because its a helpful mental model to understand networked infrastructure topology and because we want to “connect” how much networked compute and memory can cooperate efficiently on the same workload.

The goal is to make more compute and memory work tightly together as one domain. Today, one way to think about that is a rack whose accelerators cooperate to serve a model to many customers at the same time. Depending on the model and workload, that can mean thousands of simultaneous users, although there is no fixed number of users per rack. Longer conversations require more memory, while more demanding reasoning can occupy the processors for longer. Operators want to expand that domain where doing so lets them serve larger models or more users economically. Across the broader datacenter, they also connect multiple domains, without requiring every processor to participate in one tightly coupled system.

There are two ways this creates demand for more racks. An operator can run additional copies of the model and distribute customers across them. Or a model can require more resources than one rack provides efficiently, which means multiple racks must cooperate on its workload execution.

Extending the scale-up domain across racks gives the operator more tightly connected resources to work with. Models can also run across separate domains through scale-out networking, although the communication cost changes. Connecting racks alone does not establish that they form one scale-up domain.

This is where model size and model context enter the conversation. The model’s weights take up memory, and serving active users requires additional working memory. For many models, the KV cache holds information from the tokens already processed so the model can continue generating its response. Longer context and more simultaneous sessions increase that requirement.

A larger domain gives an operator more room to place the model and its active state. Depending on the workload, that can make a larger model practical to serve or improve throughput at the response time customers expect. The operator still has to balance those benefits: a larger model or longer context can consume capacity that otherwise serves additional users.

And shared memory needs a qualification. Memory remains distributed across processors, with a cost to accessing it remotely.More connected compute does not, by itself, make the same model smarter or more capable. It gives operators more capacity to deliver the capabilities of the models they choose.

Interconnect determines how far the domain can grow

Our view is that hyperscalers and neoclouds have a strong incentive to expand these domains where the additional scale improves serving economics. Some enterprises will face the same decisions as they deploy larger AI systems on-premises.

The limit is how many accelerators can exchange data fast enough to work efficiently together. When a model runs across multiple accelerators, each needs results from the others to continue its work. If those results arrive too slowly, adding accelerators can leave more expensive hardware waiting for data.

Copper has limits around reach and signal integrity as speeds increase. The underlying boundary depends on the architecture, including its physical layout and power budget.

Optical interconnects expand the possibilities for reach and bandwidth density, which is why we have spent so much time on this part of the market. But the optical system still has to be manufactured, qualified, and serviced. In Optics Won’t Scale as Fast as the Market Expects, we explained why demand for optical connectivity can move faster than qualified supply.

The race is to build larger scale-up domains that can serve larger models and more users concurrently and economically. If the processors spend too much time waiting for data, operators are paying for compute they cannot fully use. How well each company solves that problem will help determine its margins from serving AI at scale.

Where the CPU changes the equation

The CPUs already present in GPU systems give us a baseline ratio. For example, NVIDIA’s GB200 NVL72 contains 36 Grace CPUs and 72 GPU packages, or one CPU for every two GPU packages. Counting the dies inside those packages would produce a different number, so we need to keep the denominator consistent.

Where the ratio conversation becomes more interesting is when we add dedicated agentic CPU capacity which will manifest itself as dedicated racks of CPUs in the scale up domain of GPU compute racks.

This is why we separate head-node CPUs from dedicated agentic CPUs in our models. While each customer building their own compute racks, like AWS, Azure, and Google, can optimize the number agentic CPUs for their own unique needs the template we have today is of NVIDIA’s Vera CPU Rack which supports up to 256 CPUs and is positioned alongside its accelerator systems. Arm’s AGI CPU also targets agentic infrastructure with AMD, Qualcomm, and Intel following suit.

We do not yet know how much of that capacity every operator will deploy or what their scale up domain mix of CPU/GPU/premium accelerators will be. It could be one CPU rack supporting several GPU racks, or more, with the allocation changing as the workload, or economics, changes.

Consider a useful, but hypothetical, example deployment of a scale up domain of ten GPU racks, each with 72 GPU packages and 36 host CPUs. That gives us 720 GPU packages and 360 CPUs. Adding a dedicated rack containing 256 CPUs takes the total to 616 CPUs, or about 0.86 CPUs per GPU package. Two dedicated CPU racks take it to 872 CPUs, or about 1.21 to one.

That is math using stated configuration assumptions, not a deployed-system claim or a forecast. We highlight this use case to show how varied potential configurations can be and why trying to estimate any CPU:GPU ratio is not the right analysis or measurement of the market.

Demand still has to fit what can be manufactured

Our CPU forecast (available in institutional tier of CS Atlas) separates general-purpose servers, accelerator head nodes, and standalone agentic systems along with key vendor share assumptions. We then constrain shipments using our manufacturing assumptions since our forecasts are grounded also in what’s manufacturable.

Our most current base case implies the following growth from a 2026 baseline:

• 2027: CPU silicon-value growth of 29%; CPU package shipment growth of 20%.

• 2028: CPU silicon-value growth of 33%; CPU package shipment growth of 22%.

• 2029: CPU silicon-value growth of 25%; CPU package shipment growth of 11%.

• 2030: CPU silicon-value growth of 35%; CPU package shipment growth of 29%.

These are conditional, and continually updated, Creative Strategies estimates, grounded in our most current assumptions and industry checks. Silicon value includes an imputed value for captive CPUs, so it is broader than merchant CPU revenue. Pricing and product mix explain why dollar growth can exceed shipment growth given a core assumption we are making on ASP for agentic CPUs specifically.

The model constrains CPU output by compatible foundry-node capacity and allocation after competing demand, then applies yield, assembly, and substrate assumptions. Its 2029 slowdown and 2030 acceleration depend heavily on node transitions and the capacity those transitions make available. For customer/clients with access to our data models we break down the full assumptions in base/bull/stretch scenarios.

We remain aggressive in our forecasts but also grounded in manufacturing reality. Wafer mix assumptions matter and we cannot assume every wafer is available to CPUs while GPUs and custom accelerators are competing for capacity. Nor does the model establish the industry’s absolute manufacturing ceiling. We will continue refining these assumptions as new information becomes available on customer allocations and system-level deployment constraints.

Any ratio estimate has to be grounded. We would therefore be careful about declaring that a two-to-one CPU-to-GPU ratio, or more, is either inevitable or impossible. It first needs a defined workload and counting convention, then a test against the supply model aligned with tangible tracking of hyper scaler scale up domain mix. Demand above our forecast and demand above feasible supply are different conclusions.

Each operator is making a hardware bet

Hyperscalers make distinct decisions about the infrastructure they want to build, working with system suppliers and manufacturing partners. They choose the balance of merchant and custom silicon, then determine the memory and networking required to make those processors productive within a TCO profile.

It is important to note how one operator may prioritize flexibility across many customer models. Another may have enough predictable internal demand to justify a more specialized accelerator. They can also assign different stages of inference to different systems when the benefit outweighs the cost of moving state between them. Every implementation is unique and many times bespoke which is why generalities harm more than hurt the analysis.

Mixing merchant and custom infrastructure across a fleet is different from making those processors cooperate inside one tightly connected domain. That requires compatible hardware and software. These decisions commit capital to assumptions about workloads that are still evolving, and those assumptions help determine the margin profile of serving AI at scale.

Our conviction is that enterprise and personal agents will create much more work around the model layer in the scale up domain. We will have many simultaneous sessions calling on CPUs and then returning to accelerators, with memory retaining the state needed to continue. The amount of dedicated new CPU infrastructure depends on how much existing capacity absorbs and how efficiently the software schedules that work. Another reason, we are bullish on the growth profile of AWS, GCP, and Azure.

For who benefits here in silicon, we have high conviction on both AMD and Intel, with Intel standing to doubly benefit with Intel Foundry well positioned to also manufacture CPUs as additional capacity is brought online. We maintain NVIDIA is also a primary beneficiary with their agentic CPU roadmap, and Arm and Qualcomm also coming online with solutions. This is not the server CPU TAM of old where share taking was the fundamental analytical baseline. This is a significant TAM expansion and the analysis now who is are the unit volume and price share gainers in a market that is growing in dollar value ~30% YoY in our supply constrained forecast.

What would make us less bullish is evidence that CPU work per completed task falls enough to offset usage growth, or that existing infrastructure absorbs much more demand than we expect. For now, we believe the expansion of agentic workloads supports more CPU capacity, with operators making different decisions about how much belongs around each accelerator deployment.

That is how we think about the CPU-to-GPU relationship: through the complete inference system and how its scale-up domains shape what operators can manufacture and deploy economically. Keeping that view grounded is our ongoing research goal as hardware choices evolve and we learn more about how these systems perform in practice.

Continue the research in Atlas

Subscribers with CS Atlas access can use Atlas to connect our CPU reports with the research on networking and optical supply. Ask how dedicated agentic CPU racks change the processor mix, which assumptions constrain growth, and what evidence would change the forecast.

Explore this research in Atlas

The link opens an editable question after sign-in. Answers depend on the research and model versions available with your access.

• CPO Summit Takeaways

• 800 VDC Transition and Timelines

• Learnings from AWS Meetings on Compute and Model Trends

We have spent time over the last month in conversations with stakeholders at different hyperscalers and have landed on a conviction about how infrastructure decisions are likely to play out during this AI buildout. We believe the future of compute is fungible. While we appreciate NVIDIA’s framing that the GPU is fungible, and that is true, it is clear to us that hyperscalers have a vested interest in building their data centers with workload fungibility in mind.

This viewpoint is an important baseline observation on the diversity of solutions they will use, spanning merchant compute, custom compute, copper, optical, liquid and air cooling, and power delivery from 800 VDC to older generations. We expect these technologies to coexist across their infrastructure, with software coordinating the work across that diversity as one compute fabric. Note, this is a particular advantage unique the hyperscalers and we believe will be a key part of their continued competitive advantage in designing AI infrastructure at scale and competing for enterprise workloads. Having this perspective raises a number of implications and questions that our report explores.

The value of this viewing infrastructure choices this way is in how it changes our assessment of the companies supplying this buildout. More choice gives customers another way to buy and build capacity, but the savings and supplier profits depend on which workloads move, what it costs to support them and who continues to supply the rest of the system. We work through six questions to understand where those economics are likely to improve and where the apparent benefit of greater choice breaks down.

1. How does greater choice change the customer’s purchasing power?

We expect hyperscalers to start by using a wider mix of systems for new workloads while keeping existing work on the hardware already running it. That includes buying from other suppliers and developing their own silicon and rack-scale designs, giving them more control over how the systems are built and which parts they buy. At first, these efforts may help them add capacity and better match hardware to their workloads. As the alternatives reach production at scale, customers gain more room to negotiate prices. We expect those benefits to develop before broad pricing pressure affects suppliers’ growth and margins.

2. Where does specialization offer better economics, and where does GPU flexibility remain more valuable?

Workload fungibility gives operators more choice over where they run their work. We expect custom silicon to gain ground where steady demand justifies the cost of designing and supporting it, while GPUs remain valuable for a changing mix of workloads. The benefit comes from matching work to the systems that can run it most economically. That requires enough software support to make those choices practical. A cheaper chip offers limited savings when it cannot run enough of the customer’s work or takes too long to get into service.

3. How much influence moves to the software that decides where the work runs?

The software deciding where work runs also influences which hardware the operator needs to buy or they design custom compute. We expect hyperscalers to keep considerable control over those decisions because better utilization and more purchasing options directly benefit their own businesses. Suppliers still have an opportunity where their software reduces the work required to deploy and operate a system. The distinction is who collects the economic benefit. More important orchestration software does not necessarily mean a larger revenue opportunity for an independent software company.

4. How does serving inference across larger compute domains change the network?

As hyperscalers connect more systems to serve inference at scale, the network helps determine how much of that capacity they can use together. Workload fungibility depends on being able to send work to available compute without adding too much delay or cost. This raises questions about where copper remains sufficient, where optical connections become necessary, and how much flexibility operators can gain while keeping tightly connected systems running efficiently. We examine what those choices mean for networking and optical suppliers as operators build larger compute domains.

This builds on our ongoing optical and networking research and our recent conversation with a senior AWS executive, covered in Atlas. Larger, sparse models need fast connections between the accelerators holding their different parts. The race is on for the larger scale up domain where optical links are necessary extend those connections across more racks (bigger compute clusters), giving operators more room to spread power and cooling demands. The return depends on whether the larger system serves more work at the response times customers need, after accounting for the cost of connecting it.

5. What is infrastructure flexibility worth over the life of a data center?

Power and cooling decisions made today will affect what an operator can install several equipment generations from now. We see a stronger case for spending against a defined upgrade path than for preparing every part of a facility for the highest possible density. The extra investment has to earn its way back through lower modification costs, less downtime or getting paid for new capacity sooner. Our facility analysis shows why the return changes substantially when earlier installation does not produce earlier income, and why the lease matters as much as the engineering.

6. Which companies retain the most value as operators gain more options?

We keep coming back to the work a supplier retains as the customer gains more options. Another company may supply the accelerator while NVIDIA continues to sell valuable infrastructure around it. A custom-silicon partner may win a design but retain less of the profitable work when it reaches production. This is why we assess implementation, proprietary IP, networking and support separately. It also explains how greater customer choice can pressure pricing and still support profit growth for suppliers that remain difficult to replace.

In the full report, we examine:

• When cheaper compute costs more. We test how a lower-cost system’s apparent advantage holds up once utilization, software compatibility and deployment delays enter the calculation—and where paying more for GPU flexibility or integration produces a better return.

• What supports our company preferences. Why Broadcom remains our largest modeled dollar opportunity in custom silicon, why MediaTek leads our incremental conviction ahead of Marvell, and what would change that order.

• How NVIDIA participates alongside custom compute. Which parts of the system can remain valuable to NVIDIA even when a customer chooses another accelerator.

• Which developments would change our view. The production orders, retained supplier work and completed facility upgrades we will assess over the coming quarters.

Take the analysis further in CS Atlas. Explore seven company assessments and three illustrative models, then connect the findings with our broader industry research. Ask which suppliers benefit as hyperscalers design more of their own infrastructure, compare company opportunities, and test what would change our views. We built Atlas so subscribers and clients can work through our research on their own terms and apply it to the decisions they face.

Not yet a paid subscriber? Join The Diligence Stack to access the full report and go deeper with our research in CS Atlas.

Want to engage more with our firm and our analysts, reach out here.

Read more

Our interconnect SWOT is also available in CS Atlas, where subscribers can explore how we assess Credo alongside its competitors. In our July scorecard, Credo received our highest score for current-cycle strength, its role in addressing key connectivity bottlenecks, and its opportunity in scale-up systems. This report builds on that work, including why we think owning more of the copper and optical design gives Credo room to grow and what we still need to see as those products reach customers.

The companion research is available to subscribers with CS Atlas access.

We will have full coverage of AI Infra Summit in Atlas as well this week so check CS Atlas later in the week for our event notes.

Why we are positive on Credo

We are big believers that vertically integrated companies have more sustainable competitive advantages than their less integrated peers. This is a key reason we have been positive on Credo from the start, and while management consistently articulates this, we still think it is not given enough weight in the analysis of Credo’s competitive position. While Credo still has a lot of runway in its AEC and ALC businesses, we think the market is undervaluing what DustPhotonics adds as interconnect designs move toward NPO, CPO and scale-in.

DustPhotonics gives Credo control of the photonic chip alongside its own SerDes, optical DSP and firmware. We think owning the SerDes is an important part of this advantage because Credo can optimize how the electrical signal is sent and received as it works through the optical design. When a customer needs a connection to use less power, its engineers have more choices about where to make that improvement and how to balance it against reach and reliability. They can work across more parts of the electrical and optical parts together, including how a change in one affects the other. We think that flexibility becomes more valuable as customers move to new architectures and look to integrate heterogenous compute diversification, giving Credo more ways to meet their requirements and compete for the next design.

We also think Credo’s position as a connectivity-focused company with both copper and optical products deserves more weight than is given. Several competitors are stronger on one side of that market, while broader platform suppliers approach connectivity as part of a larger system. Credo can work with customers on where copper still makes sense and where optics is needed, including deployments that use both. That gives it more ways to stay involved as the customer’s architecture changes, without needing every connection to move to optics for the business to grow.

The same flexibility applies when the customer needs a more reliable connection or encounters a problem during qualification. Credo has more of the design available to investigate and adjust, including how it works with the customer’s switches, network interface cards and software. That also gives it more options when working through a production issue with manufacturing partners. If the solution requires changing both the electrical and optical design, the work can be coordinated inside the company rather than across separate suppliers and development schedules.

Helping a customer get to market faster is one benefit we expect from that control and one regularly highlighted my management. The customer may also get a product better suited to its power and thermal requirements, or have less work to do resolving problems in qualification and operation. What the May DustPhotonics acquisition gives Credo is more ability to make those tradeoffs across the full connection, and we think that will matter as customer requirements change from one design to the next and vary across many bespoke rack scale designs, as is the industry trend. That is the broader competitive advantage we see in its vertical integration. Credo acquisition announcement

Winning more of the customer’s business

Credo’s current electrical/connectivity relationships give it a way to bring this broader set of capabilities to customers whose systems are already familiar. This is another element that favors an integrated approach in that it creates the environment for more customer co-design and less likely to be designed out. By covering many parts of the complete solution Credo can extend from the cable to optical components and complete transceivers, depending on what the customer needs help with.

There are good alternatives in both cases. Broadcom and Marvell have substantial electrical and optical capabilities, and established module makers have experience getting products through manufacturing and into customer deployments. Some buyers will prefer separate suppliers because that makes parts easier to change. Others may benefit from having more of the design and support with one company, particularly when power, reliability and the deployment schedule have to be worked through together. While we recognize the option of flexibility, our extensive work analyzing the supply chain consistently lands on the observation that customers prefer a simpler supply chain when possible and “one neck to choke,” meaning a single source with more of a total solution is preferred more often than not.

Exhibit 1. Credo can do more of the work inside the connection

Source: Creative Strategies analysis of Credo’s public product portfolio and optical architectures. The map describes product responsibility, not measured competitive performance. Platform participation remains subject to qualification.

Our expectation is that having more ways to meet the customer’s requirements makes it a stronger competitor for the connections that remain open to outside suppliers. Time to market is one reason a customer might choose it; the ability to keep improving the product with that customer is why we think the advantage can last.

How much growth follows?

Our proprietary accelerator and networking models help us better estimate this dynamic. More accelerators, and in particular higher density of accelerators, create more demand for connections, but the number of units is only part of the equation. As systems spread within the rack and across racks, the distance between parts of the system and the bandwidth they need also change. Copper may still make sense for the short links, or dedicated ones once optics intra rack is reliable enough. And its already established optics serves the longer distances. We expect opportunities for both, rather than one date when the industry moves everything to optics. More in our report on optical timing scenarios below.

We envision a scenario where Credo could grow faster than accelerator deployments as the totality of their solutions get higher attach in rack scale builds. Our models let us test how much those assumptions matter, while keeping bandwidth growth from being counted twice. We also have to distinguish selling a component from selling a complete module. The module brings more revenue, but Credo has to pay for more hardware, inventory and support, so a bigger optical business does not automatically mean a higher company margin.

We are less certain about the pace of that ramp than we are about Credo’s position going into it. Qualifications can take longer than expected, manufacturing output can be uneven and customer schedules can move. Over the next year, we want to see repeat orders turn into gross profit and cash. Our view is that Credo is well positioned to win more business, with a path to growth that could justify its valuation. How quickly it gets there, and how much it earns along the way, are the parts we outline in our house view of Credo.

Inside the full report

• What DustPhotonics adds. How joint electrical and optical design could improve power, reliability, qualification and delivery—and where manufacturing dependencies remain.

• Where Credo can win. Why customers might choose Credo over Broadcom, Marvell or established module suppliers, depending on what they’re buying.

• How optical changes affect the business. What Credo could gain or lose as customers adopt linear optics, NPO and CPO.

• The case for outgrowing deployments. Our accelerator and networking models test how content, share and delivery assumptions affect calendar-2028 optical revenue and gross profit.

• What would justify the valuation. Growth and cash-margin scenarios through FY32 show what could support the September 11 valuation—and what falls short.

• What changes our conviction. The orders, design wins, margins and cash conversion that distinguish a lumpy ramp from a weaker thesis.

• Go deeper in CS Atlas. The companion, Credo and AI Interconnects: Technology, Competition and Growth, includes technical detail, competitive comparisons, full optical assumptions and 20 investor questions. Readers with Atlas access can explore those assumptions and ask their own follow-ups.

Read more

Using agentic AI to defend the enterprise

The first leg of AI adoption in the enterprise centered on basic LLMs that employees could ask questions or use to draft content. Those early pilots did not raise anywhere near the security concerns enterprises face today as agents run wild inside their four walls. In less than a year, AI went from those early pilots to software that can interpret a goal, decide what to do next, and keep working around the clock. Enterprises have already recognized the potential, but agent use has moved ahead of governed production. They are now working through how much authority to give systems that can change files, operate software, and access company data. That means setting clear limits, requiring approval for sensitive work, and having a way to recover when something goes wrong. Our first cybersecurity report examined the controls enterprises need as AI gains access to their systems. This report looks at the next phase: using agentic AI to defend those systems. Agents can help investigate threats, find vulnerabilities, and prepare fixes. The next challenge is how much of that response the enterprise can safely let them carry out.

One example from the practitioner interviews helps illustrate the defensive use case: a security agent detects suspicious activity and disables an employee’s account. That could stop an attacker, but a mistaken decision could also interrupt legitimate work. The company needs to know why the agent acted, what it was allowed to do, and how to restore access safely. Those requirements determine how much of the response it can automate.

We measure progress in remediation through verified closure: confirming that a security problem has been resolved and the affected system still works. Disabling an account may contain an attack, but the team still needs to check whether the attacker retains another way in. The opportunity is to help security teams complete more defensive work safely. This report examines how that work divides among frontier AI developers, security platforms, and enterprise teams, and where it could create new revenue.

Exhibit 1. AI Makes Discovery Faster; Enterprise Change Control Becomes the Limiting Step

Frontier labs are becoming part of the defense stack

We believe an under appreciated aspect of frontier AI models is how frontier model labs are becoming suppliers of cyber-defense capability. OpenAI’s Daybreak program gives approved defenders access to models with fewer restrictions on authorized security work, with separate approval for its purpose-trained cybersecurity models. Anthropic’s Project Glasswing gives selected partners access to an unreleased frontier model for defensive work. Both programs make advanced cyber capabilities available through controlled access rather than unrestricted public use.

Cybersecurity creates a distribution problem because the same model can help a defender validate an exploit or help an attacker develop one. Public services therefore block some advanced requests, including legitimate defensive work. A trusted enterprise program can allow more capability because access is tied to a verified user and a defined defensive purpose, with the provider monitoring each session. OpenAI’s Daybreak program is the clearest current example, while Anthropic’s Project Glasswing uses the same basic approach with controlled access to an unreleased model.

Frontier labs are likely to keep investing, making the case for the frontier strong, because gains in coding and reasoning already apply to cyber work. A model has the capabilities to stay with a long technical investigation or work across a large code base may create more defensive value than a public assistant that has to block advanced cyber tasks. That supports a restricted tier for vetted enterprises and governments, often through security-vendor partnerships. We are ready to treat trust-gated cyber access as an emerging product pattern. Private model weights and on-premises deployment remain unproven. A sovereign model is also a reasonable scenario for customers that require tighter isolation around sensitive telemetry, although provider and customer disclosures still have to show which delivery forms become real products.

Continue this question in Atlas: What evidence would move restricted frontier-model monetization from Medium to High conviction?

Model intelligence still needs enterprise authority

A defensive agent needs both model reasoning and the customer’s operating context. The model can help investigate a threat and prepare a response; identity and security platforms enforce what it is allowed to do. The enterprise owns the approval and recovery decisions. Partnerships are the practical near-term path because model developers and control platforms each supply part of the work needed to defend a live system.

The first financial proof we are watching for should appear in higher security-platform consumption or broader paid identity and recovery coverage as customers use AI for more defensive work. The real test is whether better reasoning helps teams resolve more exposure safely. Products that stop at recommendations leave the customer with the work of carrying out the response and the risk if it fails.

Inside the Full Report

• How agentic AI can help investigate threats, contain attacks, and complete fixes.

• The delivery paths and value-capture scenarios beyond today’s controlled-access programs.

• Why the model layer and the security control layer remain complementary.

• A threat-to-control map for the authority and tool paths agentic systems create.

• A safe-remediation framework built around skill-level authority and recovery.

• An auditable completeness scorecard for the leading positions.

• A 2030 market model that separates gross spend from incremental revenue.

• The pricing units and buyer evidence that would confirm or break the thesis.

• Integrated companion research in CS Atlas for subscribers to engage more with the research.

Read more

More suppliers can participate in the same program

The custom ASIC market has become more diverse and competitive as customers mix and match suppliers across their programs. Program wins once gave us a clearer basis for assigning beneficiaries. Customers now have a wider menu of compute, I/O, networking, and front-side and back-side services, and can divide that work among vendors. Supplier revenue and margins depend on the IP, products and design services retained as customers develop more of their own tooling.

Our April reports, Google’s TPU Strategy Offers a Clearer View of the Next AI Bottleneck and Custom ASIC Is No Longer One Market, connected workload specialization in a growing market with different supplier roles and margin profiles. Our September model more clearly separates compute from networking and allows for more sourcing combinations. As Google and Amazon continue to deepen their own tooling, they gain more control over which responsibilities they retain internally and which they assign to suppliers. We expect the wider availability of supplier IP to extend this mix-and-match approach across custom-silicon customers, including those with less in-house IP or tooling.

Our conviction in GPU flexibility remains intact because models are still changing, and inference workloads have very different memory and latency requirements. For enterprises managing that diversity, we continue to see GPUs as the strongest general-purpose choice on total cost of ownership (TCO): the ability to redeploy capacity as demand changes protects utilization. Specialized ASICs can retain an advantage where customers control a sufficiently large, predictable workload.

Broadcom leads in dollars, while share becomes a layer-by-layer question

Broadcom remains the largest modeled dollar beneficiary and the best-proven integrated supplier with a SerDes moat. The revenue and profit it retains depend on the mix of proprietary IP, products and design services customers continue to buy.

MediaTek is our highest-conviction incremental beneficiary. We see its TPU business expanding within a growing market, while its NVIDIA relationship gives it a route to customers developing custom accelerators for NVLink-connected infrastructure. Reaching that broader customer base still requires production volume and MediaTek retaining its implementation role through subsequent generations and widening customer base.

The broader vendor map includes Marvell, Qualcomm, Intel, AMD, GUC and Alchip. Their different combinations of owned IP, implementation experience and production responsibility make a single ranking of technical capabilities inadequate for judging the business opportunity.

How to read the visual: the customer assigns responsibilities across the program. Each box can go to a different supplier; one vendor may also fill several roles. This is a sourcing example, not a disclosed customer design or an allocation of market value.

A larger opportunity still requires program-level diligence

Our 2028 base case puts the value of custom chips and the networking used with them above $350 billion, with a stretch case above $400 billion. Those figures include the estimated value of chip development done inside companies such as Google and Amazon, so they are larger than the revenue outside suppliers would report. We model supplier revenue separately. We also exclude networking used with off-the-shelf GPU systems. The full report explains how different growth and networking assumptions affect these estimates.

Suppliers can earn more gross profit from a growing market even if they receive a smaller share of each customer’s spending. Profit could exceed our estimates if customers buy more of their higher-margin products. Revenue growth from buying components on a customer’s behalf adds less profit because much of that money goes to other suppliers.

As customers choose suppliers for their next designs, we will look at who supplies the compute, interfaces, packaging and networking. Accelerator sales can meet expectations while supplier earnings differ considerably, depending on which parts of the system each company provides.

Inside the Full Report

• The five wallet layers that show where front-side, compute, back-side, I/O, and network value can accrue.

• A vendor capability map covering Broadcom, Marvell, MediaTek, Qualcomm, Intel, AMD, GUC, and Alchip.

• The base, high-attach, and stretch cases, with the assumptions that bridge custom logic and networking.

• Broadcom’s modeled logic allocation and the sensitivity of its total-program share to networking participation.

• The proof milestones and customer sourcing behavior that would change our view.

• What has changed since April, why MediaTek leads our incremental conviction, and how program mix affects valuation.

• Full companion deep research pack in CS Atlas for subscribers to dive deeper with this report.

Read more