The price of intelligence fell 43% in ten weeks. Almost none of it was a price cut.

I.Ten weeks, forty-three percent

On 10 August, Jefferies circulated a note flagging that enterprise inference costs had hit their 2026 low, citing the Silicon Data LLM Token Expenditure Index — a daily, expenditure-weighted measure of what the market pays per million tokens across frontier APIs, open-weight platforms, brokered instances and self-hosted deployments. The series read $2.04 on 31 May, $1.45 in late July, and $1.16–$1.18 between 6 and 8 August. That is a 43% decline in ten weeks in the realised price of the core input to every AI product on the market.

Two dated events sit inside the steepest part of the curve. On 30 July, OpenAI cut GPT-5.6 Luna from $1/$6 to $0.20/$1.20 per million input/output tokens — roughly 80% off — and Terra by 20% to $2/$12, leaving the flagship Sol tier untouched at $5/$30. The company attributed the cut to rewritten production GPU kernels and a redesigned speculative-decoding draft model, framing it as an efficiency dividend rather than a loss leader. On 31 July, DeepSeek shipped V4-Flash-0731, a 284B-parameter mixture-of-experts model with 13B active parameters under an MIT licence, and held its price at $0.14/$0.28 while improving agentic benchmark scores.

II.The buyers moved before the sellers did

An expenditure-weighted index does not measure list prices. It measures what buyers chose. And by mid-2026 buyers had already moved: the State of Open Source AI report (v1.0.1, July 2026) finds the seven highest-volume models on OpenRouter all ship open weights, with open models taking 72.4% of top-10 routed token volume by late July — up from roughly a third in late 2025. Chinese open-weight providers served about 46% of routed tokens against 35% from US providers.

The composition matters more than the headline. A separate Q2 2026 breakdown puts Xiaomi at 21.1% of weekly OpenRouter tokens, Alibaba's Qwen at 13.9%, and MiniMax, Zhipu, DeepSeek and StepFun together at 24.6% — a Chinese share that stood below 2% a year earlier. On coding volume the capability gap has effectively closed; on reasoning benchmarks closed models still lead by three to eight points, and on multimodal and computer-use tasks the lead is wider. Buyers are not abandoning frontier models. They are routing the cheap 90% of their traffic away from them, which drags a spend-weighted average down without anyone cutting a price.

III.What an expenditure-weighted index cannot tell you

The honest read is that this index is being asked to carry more than it can. CIO reported on 3 July that the index had fallen 20% from its May peak and that the weighting methodology makes the cause genuinely unidentifiable: enterprise price pressure, a shift to less token-hungry models, and outright demand weakness all produce the same line. Silicon Data's own reading, reported on 8 June, was that the softness reflected slowing migration toward premium closed models rather than a reversal — while macro strategist Andreas Steno Larsen warned in the same piece that sustained weakness would end the memory, hardware and data-centre trades for this cycle. Both cannot be right, and the index alone cannot adjudicate between them, because Silicon Data publishes no matching volume series.

There is a harder caveat: not every price is falling. Anthropic's Claude Sonnet 5 is running a promotional $2/$10 through 31 August that reverts to $3/$15 on 1 September — a 50% increase. DeepSeek has announced a peak-hours surcharge doubling all billing items during two daily Beijing windows, with no effective date set, and has signalled a further increase it calls significant. Part of the August low is therefore a promotional artefact with a known expiry.

IV.The capex math now runs on elasticity

Microsoft, Google, Amazon and Meta have guided to a combined $725–785bn of 2026 capital expenditure, roughly 77% above 2025 and an estimated 93% of operating cash flow, against 33% in 2023. That analysis puts the capex-to-revenue divergence near 46%, already wider than the 32% peak of the 2001 telecom overbuild. An earlier February tally at $660–690bn flagged the same structural risk: infrastructure built today may take 18–36 months to earn proportionally, and Amazon's stock fell 8–10% on its announcement.

That investment is underwritten by revenue billed per token. If realised price per token falls 43% while token volume rises faster, the buildout is early rather than wrong. The volume evidence is real: OpenRouter alone reached 25 trillion tokens a week, about 100 trillion a month, a fivefold rise in six months as of 26 May. Jefferies made the Jevons argument explicitly on 26 June, seeing zero sign of capex slowing and rotating into memory — SK Hynix, Kioxia, Samsung — while trimming internet names. Judgment, not fact: elasticity above one is the load-bearing assumption of the entire 2026 capex cycle, and it is currently supported by platform-level volume data rather than audited industry figures.

V.Cheap tokens, expensive bills, invisible margins

For buyers, the 43% decline has not shown up as savings. The arithmetic is that cost equals price per token multiplied by tokens per task, and the second term has grown faster: one analysis puts the fall in token prices since 2023 at roughly a thousandfold against a consumption rise several orders larger, driven by orchestration overhead, context accumulation, retry loops and tool descriptions. Worked examples show a fraud-check agent at 13,500 tokens against 800 for a chat turn, and one fintech tripling its bill while growing users tenfold — cheaper per unit, more expensive in total.

The supply side has its own blind spot. Fireworks, Baseten and Together AI raised roughly $3.8bn in four weeks selling managed inference on customer-tuned open weights, yet as Forbes noted on 18 July, none disclose gross margins — the one number that would reveal whether falling token prices are an efficiency dividend or a subsidy. Long GPU commitments against short customer contracts leave little room to hold a price through a supply shock. Morgan Stanley's 3 August framing is the useful one: enterprise AI tasks average around $55 of revenue against $2–5 of direct cost, 63% of surveyed enterprises already run open models, and moving a workload from closed to open cuts price roughly 70%. At those spreads, substitution is close to free money for the buyer and a direct revenue transfer away from the frontier labs.

VI.How it plays out: three scenarios

Horizons are two to five years; probabilities are analytical judgment, not model output, and reflect the balance of evidence as of 12 August 2026.