The AI Capex-Adoption Gap: Big Tech's $725 Billion 2026 Buildout vs the ~10% of US Firms Using AI

Amazon, Microsoft, Alphabet and Meta, the four biggest hyperscalers, have committed about $725 billion to capital projects in 2026, 77 percent more than their record 2025 spending, while the share of US businesses using AI to produce goods or services stood at 10 percent in September 2025. Divide the first figure by the second and the AI investment debate compresses into one number: roughly $1.3 million of planned buildout spending for every US firm actually using AI in production, with the full method below 12.

That quotient is Capex per Adopter, the figure this continuing ProvenBrief series maintains. Both halves are tracked obsessively on their own: hyperscaler guidance in earnings coverage, and adoption in the Census Bureau's Business Trends and Outlook Survey (BTOS), which updates every two weeks 3. A July 17 Federal Reserve note by Paul E. Soto, Mason Thieu and Jeffrey S. Allen compiles these public indicators, attaches the data file, and invites researchers to "use and extend this framework to monitor developments in the AI buildout" 4. The note charts capex and adoption separately. The division of one by the other is the extension, and we carry it on both of Census's AI definitions, because the gap between them is itself a finding.

The numerator: $725 billion was the April reading, and the ledger has moved since

• Amazon expects capital expenditures to reach $220 billion this year, CEO Andy Jassy said July 30, up from the $200 billion forecast set in February; he cited rising memory prices and said even at that level capacity will not meet 2026 demand 5.
• Microsoft set calendar-year 2026 capex at $190 billion, with CFO Amy Hood attributing $25 billion of it to rising memory chip and component costs 1.
• Alphabet lifted guidance to as high as $205 billion in July, from as much as $190 billion in April 51.
• Meta raised its full-year projection by $10 billion, to a range topping $145 billion, citing component pricing and competition for land, power and skilled workers 1.

The $725 billion total was compiled by the Financial Times from first-quarter earnings and marked a 77 percent jump from 2025's record $410 billion 1. After the summer raises at Amazon and Alphabet, the four companies' latest guidance sums to roughly $760 billion, our tally. Two cautions from the Fed note travel with any of these numbers: hyperscaler capex includes non-AI spending, and because the companies increasingly lease data center capacity, headline capex could progressively understate the true buildout 4. The builders' demand evidence is contracted rather than realized: AWS's backlog stood at $496 billion last quarter and Google Cloud's at $460 billion in its first-quarter report 51. Jefferies analyst Brent Thill put the bull case in five words: "The bear thesis is garbage" 1.

The denominator: 10 percent of firms, or 22, depending on the question

Adoption moves in percentage points, not tens of percent. On Census's original BTOS question, whether a firm uses AI in producing goods or services, usage crept from 4.6 percent of firms at the start of 2024 to 10 percent in September 2025, a 5.4-point climb in about 21 months 2. In November 2025 Census re-worded the question to ask about AI use in any business function; that series opened at 17.3 percent 2, stood near 18 percent at year-end 6, and reached 22 percent in the August 23, 2026 release 7. Even under the broad definition, in the Information sector, where use is highest, more than six in 10 businesses reported no AI use at all 2.

The spread across surveys is wider than the spread within BTOS. Work-related generative AI use stood at about 41 percent of the workforce in the Real-Time Population Survey in November 2025, and the Atlanta Fed's Survey of Business Uncertainty found 78 percent of the labor force works at firms that have adopted AI; a Fed review of 16 adoption surveys found mid-2024 estimates spanning 5 to 40 percent 6. The 10-versus-22 gap is not noise to average away. It is the difference between AI as a production technology and AI as an office tool, and we report the ratio on both definitions.

Capex per Adopter: about $1.3 million per producing firm, about $590,000 on the broad definition

The method, in full: the numerator is the four companies' combined 2026 capex guidance; the denominator multiplies the BTOS adoption share by the Census Bureau's 2023 count of US employer firms, 5.58 million with one to 499 employees, the employer-business universe BTOS represents 83. On the strict production definition, 10 percent of 5.58 million firms is about 558,000 adopters, and $725 billion across them works out to roughly $1.3 million per firm. On the any-use definition, 22 percent implies about 1.23 million firms, or roughly $590,000 each. On the updated roughly $760 billion ledger, the readings rise to about $1.36 million and $620,000, our arithmetic 51. One asymmetry is permanent: the numerator is global spending by four companies, the denominator US firms only.

The spend side: a median $11.95 per employee, and a slide at the top

Buyer-side data shows the same shallowness. Among the roughly 70,000 companies on the payments company Ramp, the median AI-spending firm paid $11.95 per employee in July; the top 10 percent of firms spent a median $650 per employee, and the top 1 percent $7,400 97. In August, 56 percent of Ramp customers paid for AI products, rising 0.4 percent from the month before, and the top 1 percent slipped to $7,205 7. Average token costs fell to $0.68 per million tokens from a March peak of $1.15 as OpenAI and Anthropic cut prices, and Ramp economist Ara Kharazian says the competition is driving down both prices and spend at the top 1 percent of companies 7. Ramp's clientele skews technical, so its levels sit above the wider economy, where Census puts use at 22 percent 7.

What would close the gap

The two sides move on different clocks. Capex guidance resets quarterly and has drifted up roughly $35 billion since April 5; BTOS updates every two weeks, and realized adoption on the broad definition added about five points in its first nine months 27, and in November 2025 businesses expected roughly four more points within six months 2. The Fed note names its own tripwires for the shift from buildout to payoff: capex associated with productivity gains, a sustained change in labor market dynamics among AI-exposed groups, or a widening productivity gap between high- and low-exposure sectors 4. Until one of them trips, the series' headline reading stands: about $1.3 million of committed buildout for every US firm using AI in production.