Goldman Sachs Report: Open-Source Models Account for 72% of Token Consumption but Receive Only 27% of Spending

AI demand data is creating a rare market divergence: the same set of numbers is being used by both bulls and bears, and both sides claim victory. At the heart of this debate is an unsettling structural contradiction—token consumption is exploding, yet the vast majority of the incremental volume flows to the lowest-priced models, while the GPU costs of serving those tokens are climbing in tandem.

According to Bobby Molavi of Goldman Sachs in his latest report, OpenRouter data shows that open-weight models handled 72% of token traffic in the second week of September but captured only 27% of total platform spending; meanwhile, closed-source models, although their share fell from 89% in early January to 73%, still firmly control the vast majority of monetization revenue.

More noteworthy is that about 69% of tokens on the OpenRouter platform in mid-September came from Chinese AI labs. Molavi characterized the current market as "unplayable" and noted that the AI productivity dividend has so far left no quantifiable trace outside the tech industry.

This structural contradiction is reshaping market expectations for AI investment returns. GPU rents keep rising while token prices keep falling, with pressure from both ends at once, meaning compute suppliers enjoy the rent windfall, while the enterprise customers and ultimate shareholders footing the bill face a difficult-to-realize ROI. Meta recently raised up to $95 billion from shareholders for AI capital expenditure, with its chip suppliers taking stakes at the same time, marking that "vendor financing" has evolved from a debt form to an equity form, making questions about the sustainability of the capital cycle ever more urgent.

Token demand explosion: the real picture behind 250-fold growth

The growth in demand itself is beyond doubt. Molavi, citing OpenRouter data, noted that token demand has grown 27-fold since last November, with open-weight models now holding a 72% token share. Over a longer horizon, OpenRouter's weekly token processing volume soared from 0.5 trillion in mid-January 2025 to 128.4 trillion in the week of September 14, a roughly 250-fold increase in 20 months.

GPU pricing likewise confirms the authenticity of demand. H200 rental prices rose 160% during the year, B200 by 120%, and H100 by 35%. Even more telling, the A100 chip produced in 2020 now commands a rental price higher than at the start of the year, which directly strikes at the core assumption about GPU depreciation cycles—if a chip can still generate higher returns six years later, then pessimistic depreciation forecasts based on a three-year benchmark do not hold up, and the long-term profit potential of hyperscalers therefore needs to be reassessed.

For the bulls, this is the strongest argument: demand is real, and the economic life of compute assets is longer than the bears expected.

Free fall in token prices: the fastest deflation in history

Yet the other side of the demand explosion is that token prices are collapsing at an unprecedented pace. According to a research report titled "The Plunging Price of Thought" published by Epoch AI on September 22, since 2023 the cost of reaching an equivalent level of AI performance has fallen by about 47% per quarter, or about 13-fold per year; in the frontier model segment, the pace is even faster, falling by an average of 66% per quarter, or about 75-fold per year.

The most intuitive case given by Epoch AI:

In January 2025, OpenAI's o3 model answered 75% of questions correctly on the GPQA Diamond doctoral-level science benchmark, at a cost of about $0.30 per question; today GPT-5.6 Luna achieves the same score at a cost of only $0.0004 per question, a price decline of 725-fold in 18 months—equivalent to a car priced at $50,000 dropping to $69.

Compared across history, the decline in large language model inference prices is 4 times faster than DNA sequencing, 6 times faster than the improvement in computing power between 1940 and 2001, 18 times faster than lithium-ion batteries, and 54 times faster than electricity.

However, the prices customers actually pay are falling far more slowly than list prices. YipitData's tracking of nearly 2 quadrillion in annualized token data shows that effective token pricing for API services fell 39% in the second half of 2025 (44% on OpenRouter), but was down only 6% year-to-date as of May 2026.

The reason: developers keep chasing cheaper models, while enterprise customers are increasing the share of high-end models, and frontier labs' ability to command a premium for their latest models has temporarily supported the revenue side—but Epoch's data shows that the half-life of this premium is measured in quarters, because open-source clones can soon accomplish the same tasks at extremely low cost.

The core of the bull-bear divergence: who is making money, who is paying

This is precisely the core of the "same data, two interpretations" dilemma described by Molavi.

The bulls' logic is: demand keeps expanding, the addressable market (TAM) is still growing, and the explosive growth in token consumption proves that AI is being widely adopted.

The bears' logic is this: cost per token is rising, revenue per token is falling, and the main body of incremental traffic is precisely those open-source models that charge the lowest fees. OpenRouter data clearly presents this picture: open-weight models process 72% of tokens but capture only 27% of spending; Chinese AI labs contribute nearly 70% of the platform's token traffic. The Goldman momo basket peaked exactly at the end of June, a timing that coincides closely with the large-scale influx of Chinese open-weight models into the AI market.

Molavi summarizes this contradiction as follows: rising GPU rents and falling token prices are the result of the same squeeze seen from both ends. For GPU owners it is good news; for the enterprises paying the rent, it is a different picture.

The missing productivity dividend: the silence of the S&P 493

Molavi cites a strategist report pointing out that AI "has yet to show a quantifiable productivity improvement in key data outside the tech universe": revenue per employee and profit margins have both improved for the Mag 7 (the seven tech giants), but the corresponding metrics for the S&P 493 (the remaining 493 companies in the S&P 500) are essentially flat.

This reality is changing the AI decision-making logic of the corporate world. Molavi observes that a large number of enterprises are shifting from "the correlation between token spending and growth and productivity" to a review framework of "the impact of token spending on costs and profit margins." CFOs have already found this AI budget line, and they want to know what the money spent has bought.

The "boom loop" and the question of capital sustainability

What underpins the current AI capex frenzy is what Molavi describes as a "boom loop": the more the hyperscalers spend on capex, the more revenue flows into the AI ecosystem, which in turn attracts venture capital and growth equity capital, forming a two-way flow of funds—and the anchor of all this is the initial AI capex itself.

The fatal weakness of this loop lies in its sustainability: once capex slows, can the revenue of the entire ecosystem sustain itself?

The private market is likewise not immune. Molavi points out that current exit cases are "considerable in number but extremely concentrated" (SpaceX, Cerebras, Cursor, Hugging Face), that the valuation re-rating of some projects has "no catalyst other than a letter of intent," and that the deal logic of the entire private market is "as if the token era will never end and everything will be disrupted."

On the financing side, as the debt window narrows, Meta recently announced it would raise up to $95 billion from shareholders for AI capex, with its chip suppliers simultaneously participating as equity investors. Supplier financing has evolved from a debt form to an equity form, and the boundary between revenue and financing between providers and users is increasingly blurred.

Molavi boils the essence of the entire debate down to four questions: how much does it cost to build, who pays, is the pain worth it, and how long must we wait to see revenue returns.

The current data picture is: token demand has grown 250-fold, but the incremental revenue flows to the lowest-priced models; GPU rents are rising, but the bill is ultimately passed on to shareholders; and the productivity dividend is nowhere to be seen outside the tech industry.

Molavi's conclusion is that the authenticity of demand is beyond doubt, but "demand growing 250-fold, while the money flows to the cheapest models, the rent flows to GPU owners, and the bill is passed on to shareholders" is not the ROI story that the capex originally promised. In this tug-of-war between bulls and bears, the bears may be holding the stronger half of the rope. Risk disclosure and disclaimer

Markets carry risk, and investment requires caution. This article does not constitute personal investment advice and does not take into account the specific investment objectives, financial situation or needs of individual users. Users should consider whether any opinions, views or conclusions in this article are suitable for their particular circumstances. Investment based on this is at the user's own risk.