Our view is that the world is entering the Intelligence Age, which continues the Information Age, and that this is still the buildout. That is the subject of this full report, however, before we dig into our views on prior cycles and how they inform today, we want to highlight feedback we heard from investors over the last few days.
We spent the last few days at TomorrowX, a large investor conference, and had dozens of conversations with institutions of all sizes and positions, as well as many retail investors and hedge funds. We came away feeling that overall sentiment is not what we would call bearish, in the sense of disbelief in AI or doubt that AI will lead to value creation we have not yet anticipated. It is more that there is uncertainty around the category. If we had to put a finger on it, we would say there is a lack of conviction on almost any topic around the AI buildout right now, which is the opposite of what we saw at the end of last year and through most of 2026.
A few views came up more than any others among the investors we talked to, and we want to share them along with where we agree and disagree.
On compute demand. Many investors are wrestling with whether we need as much compute as is forecast. They point to models getting more efficient generation on generation and getting more out of the same compute, and to new hardware adding to those gains. The argument goes like this. If we double the roughly 135 GW of data center capacity installed globally today, and much of that gets built with compute capable of serving larger models to the masses, why would the GW we need keep climbing well past 300 GW by 2030, which is where most AI semiconductor revenue and unit forecasts are built.
We think this is a fair concern, and we agree there is likely a time coming when compute demand in GW comes into parity with what it takes to serve a global workforce and consumer base using agentic AI. We highlighted this in our report on compute demand on Tuesday, We Need More Compute.
On memory. This was perhaps the spiciest set of conversations we had. Most of the sentiment is that every bit of memory is inevitably going back to commodity pricing. The view is that current margins are only what they are because supply is short of demand, so they cannot last, and the whole of memory goes back to historical margins, which ranged from negative to about 20% on average.
We agree that 75% to 85% margins are not sustainable long term, because some parts of the memory vendors’ portfolios will still follow standardized component pricing. That means segments like consumer devices and other electronics, where it is very easy to swap one vendor for another and buyers choose on price.
We disagree, however, that all of memory goes back there. We see a growing segment of memory bits moving from commodity pricing to value pricing, because that memory is more mission critical and is co-designed for AI workloads. We think certain memory vendors have a better chance of shifting more of their annual sales to value pricing, and for them margins will hold. This is already playing out in the segments where co-design is happening today. If we are right, margins for these tiers of memory will look much more like logic, where margins have historically averaged in the 50% range. We tease this out, along with what the market is and is not pricing, in our last memory report, Bet on Bits, Part 2.
On AI diffusion. We were happy to find that the consensus view is that we are just in the adoption phase of AI within enterprises, and that the timing and economics depend purely on how fast enterprises adopt and deploy at scale. We agree strongly. Where we disagree is with the idea that this adoption curve will not keep demanding significantly more GW than we have today, per our points above. Still, there is at least conviction that AI has value, that it will diffuse to the enterprise, and that enterprises will put budget to it. The variations inside those assumptions change what it takes to fund AI, but we all agree AI will diffuse and create economies of scale for compute.
On neoclouds. This was also spicy, and it follows from what we wrote in our NVIDIA note after Jensen Huang’s departure. We look at it two ways. First, the neoclouds exist primarily as proxies for hyperscaler demand and lab training. Most of their demand comes from hyperscalers that need capacity for enterprise AI workloads, and from labs and startups training models. The capital to fund buildouts ahead of established customer contracts comes from knowing that hyperscalers cannot meet their own demand, and from consistent demand for training. NVIDIA was incentivized to keep using its balance sheet to help these GPU-as-a-service companies fund the buildout, because NVIDIA hardware is well positioned on tokens per dollar per watt and on the fungibility a diverse set of demand needs. The GPU is more flexible than a custom ASIC for this kind of customer.
Second, should a slowdown come from the neoclouds and GPU-as-a-service providers, we do not believe it changes the demand profile in GW. We do think it changes who ends up buying that power in any consolidation across the GPU services business, and in that scenario it benefits the hyperscalers, in our opinion.
We understand sentiment sitting in a pocket of uncertainty, and finding conviction among the wider investor pool was very difficult. We think folks are in sit-and-wait mode, developing new theses and figuring out what bets to make in 2027 (hint: cybersecurity should be one of them). We also think there is a great deal of disconnect between investors and what we are seeing in the fundamentals at a technology level. We still believe we are early in both the adoption of AI at scale and the compute demand needed to satisfy that adoption economically. Our view is that the world is entering the Intelligence Age, which continues the Information Age, and that this is still the buildout.
The rest of this report looks at:
How the Information Age set up the Intelligence Age, and the test that separates a speculative buildout from a productive one
Why stock prices can fall while the technology is working
What Apple and Amazon in the 2010s show about how the market prices a long buildout
Why the cost of doing business is so high right now, and how that compares with 2000
Our survey data on enterprise AI returns, and why security has become the approval step for AI at scale
Where the speculative risk sits, and the six signals that would change our view
Companies covered in this report: Microsoft (MSFT), Alphabet (GOOGL), Amazon (AMZN), Apple (AAPL), NVIDIA (NVDA), SK hynix (000660.KS)
Subscribe to read the full analysis, including our survey data on enterprise AI returns and the signals we are tracking.
Go deeper in Atlas. The companion packet behind this report is in CS Atlas. It has the history of the five technology surges, the Apple and Amazon valuation history year by year, our enterprise adoption ladder, the survey tables on AI returns and the investor views we heard this week with our response to each. Paid subscribers can ask the Atlas agent questions grounded in it. A few to start with:
How does today’s buildout compare with 2000 on capacity, funding and contracts?
What did the market pay for Apple and Amazon each year from 2010 to 2023?
How many enterprises measure the full cost of AI, and how many are at or above break-even?
Which six signals would change the view that this is still the buildout?
Further Coverage in Atlas since Tuesday:
Everpure Q3 FY27: Cutting Prices While Flash Is Still Expensive. Everpure beat its revenue guide by about 10%, raised the year and cut prices with flash costs near decade highs.
Palo Alto Networks Q1 FY27: A Strong Quarter, and a Bet on Training Its Own Models. Every metric beat, the full-year guide rose by more than the beat, and the company is moving into training its own models.
NVIDIA Q3 Fiscal 2027: A Strong Quarter, but We Are Moving Our View Toward Risk. NVIDIA raised its fiscal 2028 growth outlook to about 75%, and we look at what Jensen Huang’s exit means for talent, supply and the neocloud customer base.
Discussion With Bloom Energy Management: Power Quality and Native DC Are the Next Selling Points. Bloom says utilities are starting to ask AI customers to smooth their load, and its fuel cells can feed an 800 VDC bus with no inverter.



