This report is one piece of a bigger picture we track every day in Atlas. Our Atlas members get every note we publish, the models behind them, and an agent that can answer questions against all of it.
Here is what went up in Atlas since Friday:
Cerebras Q3 2026: More Customers, Not Enough Data Centers (Friday). Cerebras now sells into three hyperscalers, but it can’t get data center space fast enough, and that has it behind on OpenAI.
OpenAI’s Cyber Event: Daybreak Moves From Model Access to Fixing Code (Friday). OpenAI now sells two security products, and we look at which cyber vendors are most exposed and which gain.
Discussion With Bloom Energy Management: Power Quality and Native DC Are the Next Selling Points (today). Utilities are starting to ask AI customers to smooth their load, and Bloom thinks its fuel cells feeding an 800 VDC bus are part of the answer.
Want to go further on this report? Ask Atlas Agent about the 2030 capacity gap
What we do not think is changing
There are multiple debates shaping the conversation as we close out 2026. One is the increased concern degree industry is overbuilding compute for AI. This argument is largely built on prior cycles, which we address below. Underneath that is a bigger question, which is whether this buildout lasts, or whether demand fades once the first wave of spending is done. Both are fair questions, and important ones to keep wrestling with. We are sensitive to that sentiment driving the market and that sentiment is often reality until it changes. We also know, and many prior cycles have demonstrated this, that the market can be disconnected from technology industry fundamentals.
From a demand perspective, think about the compute needs in the pre AI era. Before AI, the world ran the entire internet, every website, search engine, streaming service and cloud application, on roughly 85 to 90 GW of data center capacity. That is our estimate of the current compute installed base just to serve the modern web and enterprise software. Going forward, our models currently have AI alone needing about 175 GW by 2028 and about 375 GW by 2030, roughly two to four times what it took to run everything else, and the workload math in this report will explain why this workload necessitates more compute just to serve the same internet population using agentic AI regularly.
From a market belief standpoint, it also helps to remember that markets can take a long time to believe in a new platform shift or paradigm. Apple is one of our favorite historical examples which traded below 20 times earnings at the end of every year from 2010 through 2018 while investors debated whether smartphones would become a commodity and whether Apple was about to be disrupted. It took close to a decade for that view to change. With that in mind, there are a few fundamentals we do not believe are changing, and they shape how we think about where we are in this cycle.
This buildout is not waiting on a new user base. Every prior buildout we have studied was built ahead of users who still had to buy something new to take part. The PC needed buyers to bring a computer home, the fiber build of 2000 needed households to sign up for broadband, and the mobile internet needed people to buy smartphones. AI arrives on hardware people already own. About 6 billion people used the internet in 2025, roughly three quarters of the world’s population, according to the ITU, and the GSMA counts about 4.7 billion people using mobile internet on their own device. Anyone with a modern PC or smartphone can use an AI assistant today, and increasingly an agent, without buying anything. Our Base case already has about 750 million daily consumer users and 120 million enterprise seats in 2026. The base that can adopt agentic AI is measured in billions, and it is already installed.
Enterprises are still at the early-adopter stage. As in every adoption cycle we study, we break the enterprise market out by adopter profile. The early adopters, roughly 15% to 20% of enterprises, are technically strong, comfortable with risk and willing to move fast, break things and learn as they go. They work out the problems in a new technology and feed what they learn back to the companies building it, and that feedback loop is how a technology gets refined into something the mainstream will buy. Enterprise AI is in that stage today. Between 60% and 70% of the organizations in our research have AI in production somewhere, but most of the leaders have reached governed production, where a workflow is live with permissions and monitoring, and few have reached company-wide scale.
Governance and security are the gates to scale. Our CIO, CTO and CSO research tells us the enterprise is early in adoption at scale, and that governance and security have to be cleared first. Agents that hold credentials, call tools and change systems have to be governed like any other identity, and security has become the approval step before an agent goes into production. Measurement is early too. In our September survey of 200 CIOs and CTOs, only 33 had a finance-reviewed result against the full cost of AI, and 113 had not yet measured full cost at all. Getting to a company where every employee uses agents every day, in a company-approved way, depends on how much on-premises infrastructure, data and software has to be rebuilt first. That could move quickly or slowly, and it will vary a lot by company. We have conviction on the direction and less certainty on the timing.
We remain compute constrained, and the enterprise has barely started. Even with enterprise adoption this early, the industry is short of compute. SK hynix has said its DRAM and HBM are effectively sold out through 2029, Alphabet reported a cloud backlog of $561 billion last quarter, and our Base case has serving capacity short of demand from 2027 onward, about 10% short by 2030. The insight behind this is purely the computational complexity of the workload.
One person using an agent puts far more load on infrastructure than one person browsing the web or using a SaaS application, and electricity is the simplest way to compare them, because a CPU cycle and an AI operation do not convert cleanly. Google put a 2009 search at about 0.3 Wh including indexing, and with roughly 5 trillion searches a year across about 6 billion internet users, a typical person searches two or three times a day, which comes to under 1 Wh. The IEA models a single agent task at about 1 Wh of GPU electricity without reasoning and about 50 Wh with it. At the six tasks a day our Base case assumes for one active agent in 2026, one agent uses roughly 6 to 300 Wh a day, or about 10 to 400 times a person’s daily search. A single reasoning-heavy agent task uses about as much electricity as 200 median Gemini chat prompts, and more than the data center share of 10 hours of streamed video. Both comparisons lean conservative, since the search figure includes overhead and indexing while the agent figure counts only GPU electricity.
Our focus for this report is to articulate how computationally complex AI is as a workload compared to any other software in history—hint its much more like graphics than web or apps. While we know we need more GW of compute nobody knows exactly how much compute it will take to serve AI to every technology user. We are confident it is many times what is installed today, and that the enterprise, the part of demand that pays the most per user, has not arrived at scale yet.
In the full report, for paid subscribers:
What goes into one token. The math and memory behind a single token, and why memory, not compute, is the bottleneck once agents run.
Why the same agent task can cost 12 times more. We build one agent task from the ground up and show how model choice and cache reuse swing the compute by more than an order of magnitude.
Why request counts mislead. On one major AI gateway, a quarter of requests accounted for about half a percent of spend. We show what to count instead.
The software and architecture levers that stretch the fleet, and why efficiency gains add to demand rather than cutting it.
What the energy numbers actually measure, so you can tell a fair comparison from a misleading one.
The 2030 capacity gap. How 158 GW of installed accelerator power becomes only about 27 GW of useful serving capacity, against about 30 GW of demand.
Who benefits. The companies best placed, graded by position: memory first, then accelerators and commissioned power.
Companies covered in this report
NVIDIA, AMD, SK hynix, Samsung, Micron, Alphabet (Google), CoreWeave, DeepSeek, Alibaba (Qwen), Moonshot AI, Anthropic, Vercel and Microsoft (GitHub).




