12 Key Takeaways From Ben Evans AI Eats The World May 2026
$700B hyperscaler capex, offloaded funding, power constraints, AI model commoditization, Anthropic’s revenue surge, and the mobile-network analogy challenging conventional AI narratives.
Welcome to Global Data Center Hub. Join investors, operators, and innovators reading to stay ahead of the latest trends in the data center sector in developed and emerging markets globally.
Ben Evans does not call platform shifts lightly. For two decades, he has explained how the technology industry moves PCs to web, web to mobile, mobile to cloud and what gets built, captured, and destroyed at each step.
His May 2026 AI Eats The World update is the sharpest read yet on the current cycle.
The headline figure: the big four hyperscalers plan to spend roughly $700 billion on AI infrastructure in 2026, more than twice global telecoms capex and approaching the scale of global oil and gas.
The framing is sharper than before: Evans now asks whether AI infrastructure will mirror mobile networks a trillion-dollar capex industry where little equity value accrued to the operators.
Evans does not run an AI fund. He is not long the trade.
His job is pattern recognition across cycles.
That makes his framework more useful than most for investors underwriting digital infrastructure exposure now.
Here are the twelve takeaways that matter most for anyone allocating to data center, power, or AI infrastructure capacity.
The Capex Question
1. The big four will spend roughly $700 billion on AI infrastructure in 2026 — more than twice global telecom capex.
The 2026 capex guidance for Microsoft, Alphabet, Amazon, and Meta approaches $700 billion combined. For context, global telecom capex is roughly $300 billion annually, while oil and gas spending is around $1 trillion. AI infrastructure is now scaling alongside the world’s largest industries, with spending up roughly 7x since 2022 as tech leaders prioritize the risk of underinvesting over the possibility of excess capacity.
2. Capex-to-sales ratios are at all-time highs across the largest software companies.
Meta now spends about 55% of revenue on capex, Microsoft 54%, Alphabet 45%, and Amazon 25%. In just three years, three of the world’s largest software companies have transformed into infrastructure-heavy businesses at a scale that rivals or exceeds traditional telecom operators. As Evans puts it, they are “challenging financial gravity” these are no longer asset-light companies.
3. US data center construction has overtaken US office construction.
US data center construction spending now runs around $50 billion at a seasonally adjusted annual rate, matching or exceeding US office construction for the first time on record. Office construction has been declining since 2020. Data center construction has been rising on a near-vertical curve since 2022. The buildings of the American economy have shifted from desks to servers in less than four years. The figure excludes the compute inside the buildings.
4. The funding mechanism has migrated from cash flow to “structure.”
Evans’s own term, lifted from the deck: “Here comes the structure.” Meta has formed a $27 billion joint venture with Blue Owl Capital to help fund its Hyperion data center, while private capital is increasingly being drawn into what Evans calls the “$3 trillion AI building boom.” After three years funded largely by operating cash flow, the next phase is increasingly being financed through external balance sheets.
5. OpenAI is the test case: $1 trillion per year of aspirational infrastructure deployment.
OpenAI’s stated commitments imply 30 GW+ of new capacity, ranging from roughly $600 billion to $1.4 trillion by 2030. The internal aspiration referenced in the deck targets 1 GW per week, or about $1 trillion annualized at ~$20B per GW. OpenAI cannot fund this directly. Evans describes the setup as “Other People’s Balance Sheets, circular revenue and a lot of plate-spinning.” The plate-spinning is the warning the structure is the test.
The Bottleneck Map
6. Three bottlenecks now constitute a “pig in a python.”
Evans identifies three stacked constraints: GPU and memory supply, multi-year power bottlenecks (except China), and rising political and community resistance to data center siting. Together, these sharply reduce announced capex before it becomes operational capacity. The key geopolitical signal is that China is the only major economy able to add power at AI speed, meaning power-ready regions elsewhere may have structural pricing power not yet reflected in land or assets.
7. Microsoft’s CTO has stated directly that capacity has been “almost impossible to build fast enough since ChatGPT launched.”
Kevin Scott’s quote is the strongest first-person admission yet that the supply side of AI infrastructure is structurally behind demand. The constraint is not pricing. It is delivery. Three years into the cycle, the largest hyperscale operator in the world is on record that the build has not kept up with demand and may not for some time.
The Commodity Problem
8. Frontier LLMs are converging within 10 to 15 points on benchmarks.
OpenAI, Anthropic, Google, Meta, and Chinese labs now cluster within 10 to 15 points of each other on aggregate benchmark scores. A new leader emerges every few weeks. Evans’s verdict is direct: “For most general and consumer use, models are very similar, and crucially there are no network effects.” The second clause is the heavier one. The models can be excellent. They cannot be defensible at the model layer.
9. Anthropic has crossed OpenAI on monthly revenue.
Anthropic’s gross monthly revenue is around $3.5B, surpassing OpenAI’s roughly $2B net monthly revenue, with the crossover occurring in early 2026. Private valuations now sit near $850B for OpenAI and $900B for Anthropic a combined scale that exceeds the total value of all U.S. venture-backed IPOs from 1995–2000. The model-layer race is proving not to be winner-takes-all, and much of the capital assumes a structure that may be misread.
10. Commodity infrastructure historically does not capture value up the stack.
This is the central counter-thesis in the May 2026 deck: global mobile data traffic is up ~20x since 2010, while the MSCI global telco index has stayed largely flat. Despite massive capex, value accrued mainly to platforms like Meta, Apple, Google, and Amazon, not operators. Evans asks whether AI infrastructure will repeat this pattern. The scarcity thesis around power, land, and interconnection remains unproven, leaving telecom as the key bear case for data center investors.
11. The provisional thesis: “Models will just be infra. Innovation will move up the stack.”
Evans’s stated thesis is five clauses, each a direct claim: chat is a terrible user experience; general use needs apps; the labs cannot build all the apps; models are commodities with no network effects; models will end as infrastructure. The conclusion is that innovation, and therefore equity value capture, will move up the stack to the products and workflows that consume the infrastructure rather than the infrastructure itself.
The Deployment Gap
12. Consumer use is “a mile wide and an inch deep.”
ChatGPT has ~900M weekly active users, but only ~5% are paying. Usage remains relatively light, with under ~5,000 prompts per year for top users and ~1,200–1,500 for the top 20%. Daily penetration is ~14% of US adults, 26% of teens, and 22% of ages 14–29. Even so, this doesn’t yet justify a ~$700B capex buildout on monetized demand alone. Enterprise adoption shows a similar gap: Accenture books ~$2.2B per quarter in GenAI contracts, while most firms still report pilots far exceeding production use outside software development.
Strategic Implications for Digital Infrastructure
Three implications survive for anyone underwriting digital infrastructure exposure now.
First, the mobile network comparison is the key bear case the bull side hasn’t answered: history suggests commodity infrastructure rarely captures long-term value, meaning much of the equity upside may flow to application and service layers rather than data center infrastructure itself.
Underwriting should price both scenarios the scarce-input case and the commodity-infra case and discount accordingly.
Second, the funding shift is the cycle’s leading indicator of stress. Cash-flow-funded capex is a different asset class from joint-venture or debt-structured capex. The Meta-Blue Owl JV, the Oracle debt repricing, and the OpenAI plate-spinning structure are early signals of how the funding migrates.
The rate of that migration matters more than any single deal. Investors who model that migration explicitly will hold a different view of cycle duration than those who do not.
Third, power scarcity remains the key structural moat. Markets that can deliver reliable, dispatchable power within 24–36 months command a premium, especially where permitting is fast and grid capacity already exists. Most emerging markets, however, still lag on both fronts.
The ones that do are the addressable universe for the next phase of digital infrastructure capital formation.