A six-month investigation into the largest concentrated capital reallocation since the dot-com bubble — who paid, who received, and what the bill looks like when it comes due.
Three hyperscalers now account for 71% of all AI infrastructure spending — a concentration ratio higher than oil, semiconductors, or cloud itself.
A note on why this issue reads less like a technology quarterly, and more like a sovereign-bond prospectus.
Six months into 2026, the numbers have moved past what any analyst model can comfortably call "forecastable." The four largest AI infrastructure spenders in the United States have together committed more capital in 180 days than the entire global semiconductor industry invested in the five years from 2017 to 2021 combined.
This issue is built around a single question: what does the capital know that the products do not? We have stripped out the speculation, the demo-stage valuations, and the seeding rounds, and focused instead on the harder money — committed capex, drawn-down facilities, hyperscaler offtake agreements, and the sovereign funds now writing directly into compute reserves.
What emerges is a picture of concentration. Three firms, three geographies, three use cases. The long tail of the AI startup market has not disappeared, but it has been quietly re-priced — downward — even as the headline numbers climb. The dispersion between the top decile and the median has rarely been wider.
Capital figures drawn from 10-K and 10-Q filings, sovereign wealth disclosures, and 412 verified term sheets dated 1 Jan – 30 Jun 2026. All values in nominal USD.
"Committed" = signed offtake. "Deployed" = cash settled. The gap matters.
Round counts fell, dollars rose, and the median quietly went the other way.
The shape is now familiar: fewer, larger rounds, with capital concentrating at the top of the funnel. Deal count fell 22% year-on-year even as dollars committed climbed 41.8%. The median round, stripped of mega-rounds, is down to $8.4M — its lowest level since 2019.
The concentration ratio is now higher than oil, semiconductors, or cloud itself.
For the first time, two firms — Microsoft and Alphabet — will each spend more on AI infrastructure in a single year than the entire global semiconductor capex cycle of 2018. Amazon and Meta are not far behind. Together, the four commit $293B in 2026.
Top-4 hyperscalers account for 71.2% of all AI infrastructure spending globally — surpassing the four-firm concentration of oil majors (62%), semis (58%), and cloud (67%).
| Firm | Capex | YoY | Share |
|---|---|---|---|
| Microsoft | $112B | +38% | 38.2% |
| Alphabet | $98B | +44% | 33.4% |
| Meta | $48B | +52% | 16.4% |
| Amazon | $35B | +27% | 11.9% |
| Others | $119B | +19% | — |
| Total | $412B | +34% | 100% |
Each dot is a funded AI company. The amber dots are the ones the market cannot afford to be wrong about.
| Company | Valuation | Rev mult |
|---|---|---|
| OpenAI | $340B | 182× |
| Anthropic | $180B | 156× |
| xAI | $120B | 210× |
| Databricks | $62B | 38× |
| Mistral | $14B | 88× |
| Perplexity | $9B | 42× |
Sovereign funds are now the single largest capital pool in AI — a structural change in 18 months.
Capital source → deployer → use case. The flows now diverge by region more than they converge.
The United States dominates foundation-model funding; China concentrates on compute sovereignty; Europe, late and smaller, skews toward application-layer capital. The three regions are no longer running the same race.
| Region | H1'26 | YoY | Share |
|---|---|---|---|
| United States | $198B | +42% | 59% |
| China | $84B | +58% | 25% |
| Europe | $52B | +31% | 15% |
| Rest of world | $18B | +22% | 4% |
Eleven events that redrew the map of who pays for intelligence.
Probability on one axis, severity on the other. The amber dot is the one keeping desks awake.
The 2026 surge is not, in our reading, a bubble in the conventional sense. The capital is real, the customers are real, and the unit economics — at the hyperscaler tier — are improving. What it is, instead, is a concentration event: a once-in-a-decade reordering in which a small number of firms, geographies, and use cases absorb the overwhelming share of available capital, leaving the long tail to re-price downward.
The question for the next eighteen months is not whether the capital continues. It will. The question is whether the returns on that capital — measured not in valuation marks but in deployable intelligence at sustainable unit cost — begin to converge with the commitment curve. On present evidence, they will not. The gap is the story.