If your strategic AI plan treats frontier model providers as well-capitalised, fast-scaling, and on a path to comfortable profitability, the 2025 unit economics tell a different story. OpenAI reached an estimated $25 billion in annualised revenue in 2025. The same company's compute spend in 2025 was $16.3 billion, roughly 65% of revenue. Anthropic reached an estimated $19 billion annualised. Hyperscaler capex has doubled since ChatGPT launched. The compute layer is consuming more of the revenue, not less, as the providers scale.

Frontier lab compute spend, 2022–25:

The data points to a structural issue. In most scaling businesses, unit economics improve as revenue grows: fixed costs amortise across larger volumes, suppliers offer volume discounts, operational efficiency improves. The frontier AI lab data shows the opposite pattern for compute. As OpenAI grew revenue from a few billion in 2023 to $25 billion annualised in 2025, compute spend grew faster. The ratio of compute to revenue did not decline. It increased.

The mechanism is visible in the broader sector data. Capability is currently measured by training compute scale (larger training runs produce more capable models). Customer expectations are calibrated to continuous capability improvement. Providers that fall behind on capability lose share rapidly. The combination produces a structural incentive to spend more on compute for each successive model generation than the prior model required to maintain or expand market position.

[CHART fig_4218_2026 — OpenAI vs Anthropic annualised revenue, 2024–25]

The inference cost adds a second pressure. As model capability rises, customers use models for more complex tasks. The longer reasoning chains, larger context windows, and richer multi-step interactions of 2025-era models cost more per query to serve than the 2023-era equivalents did. Inference revenue per query has grown, but inference cost per query has grown faster.

Citi Research's analysis of hyperscaler capex provides the third angle. Hyperscaler capex (the major cloud providers' infrastructure spend) doubled in the three years since ChatGPT launched, driven largely by AI infrastructure buildout. The capex commitments through 2027–2028 (Stargate at $100–500B, OpenAI-Oracle's $300B contract, Google's $40B Texas data centre, Microsoft's $17.5B India expansion) extend the trajectory rather than slow it.

For executive teams making AI investment decisions, the unit economics raise several strategic questions.

The first: vendor pricing pressure may rise. Frontier model providers operating at 60%+ compute-to-revenue ratios cannot sustain current pricing while scaling indefinitely. The two paths forward are price increases (which providers have signalled they will pursue as enterprise contracts come up for renewal) or substantial cost reductions on the compute side (which depend on hardware efficiency improvements, custom chips, and architecture innovations that are not guaranteed). Strategic plans that assume pricing stability for the next 24 months may be working from an assumption that the unit economics do not support.

The second: provider sustainability is now a procurement consideration. The leading frontier providers are loss-making at the operational level, with their losses funded by continuing rounds of investment. The funding environment may continue at 2025 scale (see #69 The investment ramp doubled) or may tighten. Strategic plans that depend on a specific provider being available 36–60 months from now should evaluate that provider's path to operational sustainability, not just their current capability, as part of procurement diligence. The investment ramp gives short-term confidence; the unit economics suggest the long-term picture is more contingent.

The third: the cost structure favours scale concentration. Providers with $20B+ in annual revenue can absorb $15B+ in annual compute costs and still operate. Providers with $1–5B in revenue cannot. The combination of high capability requirements (large training runs) and high inference costs (sophisticated serving infrastructure) means the long tail of smaller frontier model providers is structurally disadvantaged. The "Cambrian explosion of AI providers" narrative of 2023–2024 is giving way to a "two or three frontier providers per category" reality in 2025–2026. Strategic plans should account for this consolidation.

The trajectory through 2026–2027: compute costs as a share of revenue will probably moderate from the 2025 peaks as inference cost optimisations roll out (faster chips, more efficient serving, model distillation) but will likely remain at much higher ratios than would be sustainable in a normal SaaS business. The frontier model business is not converging on SaaS-like unit economics. It is converging on something more like utility or telecom economics: capital-intensive, low-margin at maturity, with structural concentration in a small number of providers per capability category.

For strategic AI planning, three implications follow.

The first: procurement frameworks should evaluate vendor unit economics alongside capability and pricing. A vendor with strong capability and competitive pricing today but unsustainable compute economics is a riskier procurement choice than a vendor with weaker headline metrics and stronger unit economics. The procurement evaluation should ask: "Can this vendor sustain its current pricing and capability roadmap on the unit economics it has?"

The second: long-term contracts should include re-pricing mechanisms. As vendors face pressure to raise prices, the contracts that protect customer pricing for longer windows (24–36 months) will be valuable. The contracts that allow vendor-side re-pricing every 6–12 months will become more expensive over time. Strategic plans that lock in current pricing for 24-month horizons may produce substantial value as the cost pressure forces broader pricing changes.

The third: build optionality across providers. The frontier provider landscape is consolidating but has not yet collapsed to a single dominant vendor. Maintaining optionality (multi-provider procurement, abstracted API integration, switching capability) provides protection against any single provider's economics or strategy shifting. The cost of building this optionality is rising as integration complexity grows, but the value of having it rises faster.

The trajectory: AI compute economics will remain the structural feature shaping vendor behaviour, pricing strategy, and provider concentration through 2026–2028. Strategic plans built for an environment where compute is the dominant cost variable and provider consolidation is ongoing will produce better outcomes than plans built on the assumption that the vendor landscape stabilises and unit economics normalise. The trajectory of the data does not support that assumption. Neither does the trajectory of capital deployment. Strategic plans should anchor to the structural reality the data describes.

Sources

  • Primary: Stanford AI Index 2026, Chapter 4 (Economy) 4.2 — hai.stanford.edu/ai-index/2026
  • Frontier-lab unit economics: Epoch AI, 2026 — frontier-lab revenue and compute spend estimates (2022–2025)
  • Hyperscaler capex: Citi Research, 2025 — hyperscaler AI capex tracking and forward commitments
  • Infrastructure deals context: Stanford AI Index 2026 year-in-review (Stargate, OpenAI–Oracle, Google Texas, Microsoft India)