Why falling AI token prices can increase compute demand, what the H100 rental chart shows, and how NVIDIA's business shifted from gaming to data centers.
More Macroplane supply-chain guides · Baskets · Investment theses · Macro trends
AI can get cheaper to use while the infrastructure behind it stays expensive. Lower token prices make more tasks worth automating. If usage expands faster than efficiency improves, total compute demand rises. That is the Jevons paradox argument behind the AI buildout—and the reason a falling token price does not automatically undermine NVIDIA ($NVDA).
The chart below captures the tension. Through the summer of 2026, a token-price benchmark falls while H100 GPU rental prices recover. Software buyers get cheaper access to intelligence; infrastructure buyers still compete for capacity.
There is a second signal in NVIDIA's ($NVDA) accounts. Gaming has disappeared as a separately reported revenue category, and data centers generate roughly 93% of sales. That does not mean the company has stopped selling gaming GPUs. It means AI infrastructure has become the business around which the financial story revolves.
*Reader-supplied a16z Growth chart. Its footer credits Ornn, Silicon Data and Bloomberg, compiled by Citadel Securities, with data as of August 2026. The same divergence is discussed in a16z's State of Markets II. Values discussed below are approximate visual readings, rather than a reconstruction of the underlying series.*
The chart compares dollars per H100 GPU-hour on the left axis with a token expenditure benchmark on the right. Around June, both decline. During July and August, the token series continues toward roughly $1 per million tokens, while H100 rental prices recover toward approximately $2.70 per hour.
Those are different products. A token is a unit of model input or output. A GPU-hour buys time on hardware. Their relative prices depend on how much useful output the hardware produces, how heavily it is used, and what the model provider charges.
The token benchmark also needs care. Silicon Data describes it as a usage-weighted price per million tokens across its observed market. It can move because of price changes, users switching models, or changes in the basket. It does not measure total AI spending or the price of a constant level of intelligence. Silicon Data methodology.
The divergence is therefore consistent with growing demand absorbing efficiency gains. It is not proof that cheaper tokens caused higher rental rates: the chart contains neither total GPU consumption nor a counterfactual showing what demand would have been without the price decline.
Compute cost discussions often collapse three separate measures into one: the price of hardware capacity, the price of model access, and the cost of getting a useful job done. An improvement in one does not guarantee an improvement in the others.
| Measure | What it buys | What can change it |
|---|---|---|
| GPU rental price | One GPU-hour on a specified accelerator | Supply, location, contract terms, connectivity and availability |
| Token price | A quantity of input or output from a particular model | Competition, serving efficiency, model choice, caching and commercial pricing |
| Cost per successful task | An acceptable answer or completed workflow | All of the above, plus context length, retries, tool calls and human review |
A model that costs less per token can still produce a larger invoice if it reads longer documents, generates more reasoning, or attempts a task repeatedly. A pricier model can be cheaper per completed job if it succeeds on the first attempt.
For an application builder, the useful question is: what does it cost to complete this task at the required quality and latency? For an infrastructure operator, it is: how much productive, paid work can this fleet deliver after operating expenses and financing costs?
A simplified serving equation explains the apparent contradiction:
GPU rental cost per million tokens = hourly rental price × 1,000,000 ÷ tokens served per hour.
This is a narrow hardware-cost calculation, not a complete API price model. It excludes CPU capacity, networking, storage, operations and provider margin. The throughput assumption must also specify the model, input/output mix, context length, quality and latency requirements.
Consider an illustrative example, not a measured industry benchmark. A GPU rented for $2 per hour serves one million comparable tokens per hour. Its rental cost is $2 per million tokens. Later, the hourly rent rises to $2.70, but improved serving doubles throughput. Rental cost per million tokens falls to $1.35.
The hardware-hour became 35% more expensive. The hardware cost per token fell 32.5%. Both statements can be true.
Better batching, lower-precision computation, caching and smaller models can improve the economics of serving. Competitive providers may pass savings to customers. That creates room for token prices to fall even when the underlying hardware remains scarce.
Fixed-capability inference has already become much cheaper. Stanford's 2025 AI Index reports that the price of querying a model meeting a GPT-3.5-level MMLU threshold fell from $20 per million tokens in November 2022 to $0.07 in October 2024—more than a 280-fold reduction. That comparison is a benchmark-specific capability threshold, not evidence that every frontier task became equally cheap. Stanford AI Index, Chapter 1.
Jevons paradox occurs when an efficiency improvement makes a resource cheaper to use and the resulting expansion in activity more than offsets the savings per unit. William Stanley Jevons developed the argument around coal in his 1865 book: more efficient engines made coal-powered industry more economical and encouraged its expansion. The Coal Question.
For AI, the mechanism is easy to imagine. An expensive system is reserved for high-value questions. A cheaper system can read every support conversation, examine every code change, or classify every document. The savings expand the set of tasks that clear an economic threshold.
Agents can expand consumption within each task, too. A workflow may plan, retrieve information, invoke tools, check its work and revise its answer. Cheaper inference makes additional steps affordable. Whether those steps produce enough useful value is the commercial test.
The arithmetic distinguishes ordinary rebound from a full Jevons outcome:
| Illustrative scenario | Tasks completed | Compute per task | Total compute |
|---|---|---|---|
| Before efficiency gain | 100 | 1 unit | 100 units |
| Same activity, twice the efficiency | 100 | 0.5 units | 50 units |
| Partial rebound | 150 | 0.5 units | 75 units |
| Full offset | 200 | 0.5 units | 100 units |
| Jevons outcome | 300 | 0.5 units | 150 units |
If compute per comparable task halves, task volume must more than double for physical compute consumption to increase. More token usage by itself is insufficient evidence: a changing model mix can make tokens more or less computationally expensive.
Rising spending is a separate test. Spending equals price multiplied by quantity. A larger compute bill can reflect higher rental rates, greater consumption, or both. A rigorous Jevons claim needs evidence that efficiency-driven demand outweighs the physical resource savings.
The claim that NVIDIA ($NVDA) is “not even selling gaming GPUs anymore” is literally incorrect. Its GeForce RTX 50-series product range remains marketed with buying options. Official GeForce lineup.
What changed is disclosure. Starting in the first quarter of fiscal 2027, NVIDIA ($NVDA) changed its market-platform revenue presentation. The latest filing reports Data Center across Hyperscale and AI Clouds, Industrial, & Enterprise, alongside Edge Computing. It no longer presents Gaming as a standalone market-platform revenue line. Its accounting segments, Compute & Networking and Graphics, remain separate from that presentation. Q2 FY2027 Form 10-Q.
For the quarter ended July 26, 2026, NVIDIA ($NVDA) reported:
| Revenue measure | Q2 FY2027 |
|---|---|
| Total revenue | $96.2 billion |
| Data Center revenue | $89.0 billion |
| Data Center share | Approximately 92.5%, calculated from reported revenue |
| Edge Computing revenue | $7.2 billion |
Sources: earnings release and Form 10-Q.
Edge Computing is broader than gaming, so its revenue should not be relabeled as GeForce sales. The filing attributes growth to Blackwell workstations, partly offset by slower consumer PC sales tempered by elevated memory and systems prices.
The sharper argument is about economic weight. NVIDIA ($NVDA) still serves gamers, but its reported revenue overwhelmingly comes from data centers. Falling token prices can support that business if they unlock enough additional workloads. They do not guarantee continued GPU pricing power or profitability for every buyer of its hardware.
An AI workload depends on more than an accelerator. It needs memory, network capacity, electricity, cooling and a site that can actually operate. Software efficiency can improve quickly; expanding physical capacity involves construction, equipment delivery and power connections.
The IEA's 2026 Key Questions on Energy and AI report says global data-center electricity demand grew 17% in 2025, while consumption at AI-focused data centers rose 50%. Its updated central projection puts total data-center demand at around 950 TWh in 2030, versus 485 TWh in 2025. That forecast covers all data centers, not AI alone. IEA 2026 executive summary.
This is evidence of expanding physical demand alongside efficiency improvements. It still does not isolate how much growth was caused by lower token prices rather than better capabilities or other changes.
The supply-chain implication is conditional but important: if cheaper inference expands useful activity sufficiently, demand can spread from GPUs into memory, networking and power infrastructure. The supplier that captures value may be the one controlling the hardest capacity constraint.
For more context on those layers, see our AI supply-chain map and data-center infrastructure guide.
Jevons paradox describes a possible demand response. It is not a law requiring every AI infrastructure investment to work. Capacity can expand faster than useful demand; customers can refuse to pay for additional activity; and new hardware can undermine the value of older fleets.
The August chart is also a snapshot. Ornn's published H100 SXM benchmark was $2.67 per GPU-hour on October 10, 2026, down 6% over thirty days. Silicon Data displayed $0.98 per million tokens for October 9. These readings preserve the distinction between low token prices and meaningful hardware rental costs, but they do not show compute prices rising continuously. Ornn H100 benchmark, Silicon Data token benchmark.
An investor should track several measures together:
Strong demand can coexist with poor returns for an operator that overpaid for hardware, borrowed heavily or cannot energize its facility. Token commoditization can also squeeze a model provider even while its infrastructure suppliers benefit.
The central question is how much additional useful work cheaper intelligence creates. If that work grows faster than resource efficiency, the infrastructure market expands. If it does not, falling token prices eventually expose excess capacity. The price of a token is the starting point for that analysis, not the conclusion.
Research checked October 11, 2026. Hypothetical calculations are labeled as illustrations; forecasts and benchmark readings retain their source dates.