Only 3% of U.S. consumers pay for AI — and the 'Super Intelligence' rebrand can't change that math

Menlo Ventures reported in September 2026 that 55% of U.S. AI users pay for at least one AI product — a number that sounds healthy until you notice that 55% of 64% is roughly 35% of all adults, and most of those bills are below $20 a month. Card-data surveys tell the stricter story: only around 3% of U.S. consumers actually spent money on AI in Q1 2026, per a16z's Consumer Edge dataset, and the Bank of America Institute pegs paying households at the same 3%.

Hyperscaler capex is now on track for roughly $800 billion in 2026 — about twenty times what consumers are paying for AI globally. Goldman Sachs Research forecasts global AI investment will exceed $1 trillion this year, with U.S. AI spend alone reaching 1.8% of GDP and climbing to 2.8% by 2028 — a level it calls 'consistent with 2–5% of GDP peak investment impulses observed in prior general-purpose technology buildouts.'

Three of the five hyperscalers above (Alphabet, Amazon, Oracle) printed negative free cash flow in their most recent quarter. That is the financial signature of a buildout being funded by debt and balance-sheet cash rather than the products it is supposed to enable. The $1 trillion-in-2026 macro estimate implies the U.S. AI investment share of GDP nearly doubles between 2026 and 2028 — historically a ratio only seen at the height of prior technology super-cycles, none of which ended without a digestion phase.

Executive Order 14434 — 'Inaugurating the Era of Super Intelligence,' signed September 29 and published in the Federal Register on October 2 — directs executive agencies to substitute 'Super Intelligence' / 'SI' for 'Artificial Intelligence' / 'AI' in correspondence, public communications, policy documents, websites and reports. It defines SI as 'the technologies and systems encompassed by the term artificial intelligence' — the same scope, renamed. It does not create demand, authorize appropriations, alter procurement, or relax any export controls.

If the consumer is barely paying, the bill is being footed by hyperscaler balance sheets and routed through NVIDIA and a handful of foundries. NVIDIA's Q2 FY2027 data-center revenue of $89.0 billion (+117% YoY) breaks into $49 billion from hyperscalers and roughly $40 billion from 'AI cloud,' enterprise and sovereign customers — a more diversified base than the consensus narrative implies. The 33% non-hyperscaler slice is the segment most resilient to a consumer-AI disappointment, and the one most exposed to sovereign-AI contracts and neocloud leasing demand.

Two private labs are doing the opposite of the consumer data. [OpenAI] is on an annualized run rate approaching $70 billion as of late September 2026, and [Anthropic] surpassed a $65 billion annualized rate by the end of July — both driven primarily by enterprise and API contracts, not consumer subscriptions. Upstream, TSMC fabricates the leading-edge AI accelerators and SK Hynix supplies the HBM memory that every NVIDIA and AMD AI GPU depends on; both have raised prices and locked multi-year supply commitments as a result.