AI demand data is creating a rare market divergence: the same set of figures is being selectively used by bulls and bears, each side claiming victory. At the heart of this debate lies a troubling structural contradiction—token consumption is exploding, but the vast majority of this growth is flowing to the cheapest models, while the GPU costs required to serve these tokens are rising in tandem.
According to Bobby Molavi of Goldman Sachs in his latest report, OpenRouter data shows that open-weight models accounted for 72% of token traffic in the second week of September but received only 27% of total platform spending; meanwhile, closed-source models, although their share declined from 89% in early January to 73%, still dominate the vast majority of monetization revenue.
More notably, in mid-September, approximately 69% of tokens on the OpenRouter platform originated from Chinese AI labs. Molavi characterizes the current market as “unplayable,” noting that the productivity gains from AI have so far left no measurable impact outside the tech industry.
This structural imbalance is reshaping market expectations for AI investment returns. As GPU rental rates continue to rise and token prices keep falling, both sides are under pressure—power suppliers reaping rental profits, while corporate clients and end shareholders bear the cost and face mounting challenges in realizing ROI. Meta’s recent fundraising of up to $95 billion from shareholders for AI capital expenditures, coupled with simultaneous equity stakes taken by its chip suppliers, signals that supplier financing has evolved from debt-based to equity-based models, making the sustainability of this capital cycle an increasingly urgent question.
Token Demand Explosion: The Real Picture Behind 250x Growth
The growth in demand is undeniable. Molavi cited OpenRouter data showing that Token demand has increased 27-fold since November last year, with open-weight models currently accounting for 72% of Token usage. Over a longer timeframe, the weekly Token processing volume on the OpenRouter platform surged from 0.5 trillion tokens in mid-January 2025 to 128.4 trillion tokens for the week of September 14, representing an approximate 250-fold increase over 20 months.
GPU pricing further confirms the authenticity of demand. Rental prices for the H200 have risen 160% this year, the B200 by 120%, and the H100 by 35%. More compellingly, the current rental price of A100 chips manufactured in 2020 has surpassed their levels at the beginning of the year, directly challenging the core assumption regarding GPU depreciation cycles—if a chip can still generate higher returns six years after production, then the pessimistic three-year depreciation forecast becomes untenable, necessitating a reassessment of the long-term profitability potential of hyperscalers.
For bulls, this is the strongest argument: demand is real, and the economic lifespan of mining assets exceeds what bears expect.
Token price freefall: The fastest deflation in history
However, the other side of this surge in demand is a collapse in token prices at an unprecedented rate. According to the September 22, 2023, research report “The Plunging Price of Thought” by Epoch AI, since 2023, the cost to achieve equivalent AI performance has decreased by approximately 47% per quarter, or about 13-fold annually; in the realm of frontier models, this decline is even faster, averaging a 66% quarterly reduction, or roughly 75-fold annually.
The most intuitive example provided by Epoch AI:
In January 2025, OpenAI’s o3 model answered 75% of the GPQA Diamond PhD-level science questions correctly, at a cost of approximately $0.30 per question; today, GPT-5.6 Luna achieves the same score at just $0.0004 per question—a 725-fold reduction in cost over 18 months, equivalent to a $50,000 car dropping in price to $69.
Compared historically on a横向 basis, the rate of decline in large language model inference costs is four times that of DNA sequencing, six times that of computing power improvement between 1940 and 2001, eighteen times that of lithium-ion batteries, and fifty-four times that of electricity.
However, the actual price paid by customers has decreased much more slowly than the listed price. Data from YipitData, tracking nearly two quadrillion annualized Token instances, shows that the effective Token pricing for API services fell by 39% (44% on OpenRouter) in the second half of 2025, but has only declined by 6% year-to-date as of May 2026.
The reason is that developers continue to pursue cheaper models, while enterprise customers are increasing their use of high-end models; cutting-edge labs have temporarily sustained revenue through premium pricing on the latest models—but Epoch’s data shows that this premium has a quarterly half-life, as open-source clones can soon perform the same tasks at a fraction of the cost.
The core of long-short divergence: who is making money and who is paying the cost
This is precisely the core of the "same data, two interpretations" dilemma described by Molavi.
The bullish thesis is: demand continues to expand, the total addressable market (TAM) is still growing, and the explosive growth in token consumption demonstrates widespread adoption of AI.
The bearish logic is this: the cost per token is rising, while revenue per token is falling, and the primary source of new traffic consists of open-source models with the lowest pricing. OpenRouter’s data clearly illustrates this trend: open-source weighted models processed 72% of tokens but accounted for only 27% of spending; Chinese AI labs contributed nearly 70% of the platform’s token traffic. The Goldman Momo Basket peaked at the end of June, a timing that closely aligns with the large-scale entry of Chinese open-source weighted models into the AI market.
Molavi summarized this contradiction as: the rise in GPU rental prices and the decline in token prices are two sides of the same squeezing effect. This is favorable for GPU holders, but presents a different picture for businesses paying rental fees.
Missing Productivity Gains: The Silence of the S&P 493
The ultimate logic behind AI capital expenditures is to drive widespread societal gains through increased productivity. However, this logic has not yet been empirically validated outside the technology sector.
Molavi cited a strategist report noting that AI "has not yet demonstrated measurable productivity gains in critical data outside the tech universe": revenue per employee and profit margins for the Mag 7 (the seven major tech giants) have increased, while the corresponding metrics for the S&P 493 (the remaining 493 companies in the S&P 500) have remained nearly flat.
This reality is changing the logic of AI decision-making in the business world. Molavi observes that many companies are shifting from a framework that links token spending to growth and productivity to one that examines token spending in terms of its impact on costs and profit margins. CFOs have identified AI as a budget line item and want to know what they are getting for their money.
The "Virtuous Cycle" and the Question of Capital Sustainability
The current surge in AI capital expenditures is fueled by what Molavi describes as a "virtuous cycle": the more capital large cloud providers invest, the more revenue flows into the AI ecosystem, attracting further investment from venture capital and growth equity firms, creating a two-way flow of capital—all anchored by the initial AI capital expenditures.
The fatal flaw in this cycle lies in its sustainability: once capital expenditures slow down, can the ecosystem’s revenue sustain itself?
The private market has also not been immune. Molavi noted that current exit cases are “substantial in number but highly concentrated” (SpaceX, Cerebras, Cursor, Hugging Face), with valuations for some projects being reassessed “without any catalyst beyond a term sheet,” and the entire private market’s transaction logic appearing to assume that “the token era will never end and everything will be disrupted.”
On the financing side, as debt financing windows narrow, Meta recently announced it will raise up to $95 billion from shareholders to fund AI capital expenditures, with its chip suppliers simultaneously participating as equity investors. Supplier financing has evolved from debt to equity, blurring the lines between revenue and financing for providers and users alike.
Molavi distilled the essence of the debate into four questions: What is the construction cost? Who will pay for it? Is the suffering worth it? And how long must we wait to see a return on investment?
The current data landscape is as follows: Token demand has grown 250-fold, but incremental profits flow to the lowest-cost models; GPU rental costs are rising, yet the bills are ultimately passed on to shareholders; productivity gains remain absent outside the technology sector.
Molavi concluded that the authenticity of demand is beyond doubt, but the claim that demand has grown 250-fold, with capital flowing to the cheapest models, rents going to GPU holders, and bills passed on to shareholders, was not the ROI story originally promised by capital expenditures. In this tug-of-war between bulls and bears, the bears may hold the stronger end of the rope.