Tracking the AI buildout

A cut you won't find elsewhere — we built it by joining these four companies' capex and R&D straight from their 10-Ks. For decades big tech were research businesses: in 2016 they spent far more inventing ($38B on R&D) than on physical capacity ($24B of capex). AI inverted that. Capex stayed below R&D for ~20 years, then passed it for good in 2024 and went vertical — by 2025 they poured $247B into capex versus $161B into R&D (1.5×). The most valuable software companies on earth now spend half-again as much pouring concrete and buying chips as they do inventing — and the gap is widening fast. They've quietly become industrial companies.

Where the money to fund the buildout comes from. Amazon's mix is tilting from low-margin retail toward high-margin services — AWS, advertising, and subscriptions are now ~34% of revenue (advertising alone more than tripled since 2020). Those are the profit engines that pay for the capex.

In June 2024, Leopold Aschenbrenner published dated forecasts for the AI buildout that read as aggressive: ~$500B of annual AI investment by 2026, ~$2T by 2028, ~$8T by 2030. Two years in, the measured floor is tracking his curve: the five hyperscalers' total capex alone hit $482B in the trailing twelve months through 2026Q1— and that excludes every non-hyperscaler dollar (frontier labs, neoclouds, sovereigns, power). The open bets sit further out: 10 GW-class clusters by 2028 and AGI “strikingly plausible” by ~2027.

This is what the buildout is for. AWS has roughly doubled in four years to ~$129B; Google Cloud has tripled; Microsoft's Intelligent Cloud cleared $100B; Oracle is the smallest, though its infrastructure line inside it is growing fastest (see Revenue by segment above). The big three alone now book ≈$294B/year — the revenue engine that makes $594B of annual capex rational rather than reckless. When this curve bends, the capex curve will too.

The whole story of U.S. industry in one frame. Total industrial production has crept up just +2.4% since 2022, and manufacturing outside high-tech is actually below its 2017 level (96) — the industrial base is flat. Semiconductors are the lone exception: +44.6% since 2022, +11.5% year-over-year, now at 182. Effectively all the growth in American industry is chips — the AI buildout showing up in the raw macro data.

The other side of the border — and the surge behind the headlines. For years China assembledthe world's computers and parts (blue) while importing the chips inside them. No longer: its integrated-circuit exports (orange) are up 12× since 2010 and have gone near-vertical — about $38B in a single month, +96.1% year-on-year (Jan–Jun 2026), pulling clear of computers. This is China's own customs office confirming the same chip boom the U.S. production chart shows from the other side — a global race, both governments' numbers pointing the same way.

South Korea is Samsung + SK hynix — most of the world's DRAM and HBM — so its memory exports read the whole cycle. The 2023 glut is unmistakable (value fell to ~$43B, implied price cratered), then the AI upcycle drove it to a record $94.613B in 2025 as HBM demand exploded. The orange line — implied price per kg — is a free read on ASPs that usually costs real money.

A cut you won't find elsewhere. Memory pricing data is sold for real money — but every customs record carries value andweight, so value ÷ weight is a free implied-ASP proxy. Does it work? Against what the makers actually disclose, the proxy's quarter-on-quarter moves track SK hynix's DRAM ASP at r = 0.747 (n = 14) and Micron's at r =  0.709— it tracks the Korean makers best, as you'd expect from Korea-origin customs. The gap from 1.0 is the HS-6 code blending DRAM with NAND; the KCS HS-10 split (DRAM 8542.32.1010 / NAND .1030) fixes it next.

The new bottleneck. U.S. utility output sat essentially flat for ~15 years — then it bent upward, now +10% since 2020and climbing, as data centers pile gigawatts of load onto a grid that wasn't growing. Power, not chips, is increasingly the limiting factor on the buildout. (EIA generation-by-source in TWh is the natural next layer — a free key away.)

The frontier is crowded and close. Two dozen models sit within a fairly narrow band, and they come from a dozen labs across the US and China — not one or two. Frontier-level capability has largely commoditized at the top; the interesting differences are now in price and availability, not raw smarts.

Capability keeps getting cheaper. The attractive quadrant (top-left — smart and cheap) is filling in, with open and Chinese models delivering near-frontier intelligence at a fraction of the price. Together they drag the whole price-performance frontier down and to the left — which is exactly the demand that justifies the capex on the other charts.