"72% of token consumption accounts for only 27% of expenditure"—The Key Issue in the AI Bull-Bear Debate

AI demand data is creating a rare market divergence: the same set of numbers is being interpreted by both bulls and bears to claim victory. At the core of this debate is a disturbing structural contradiction—Token consumption is exploding, but the vast majority of incremental demand is flowing to models with the lowest fees, while GPU costs to service these Tokens are rising in tandem.

According to the latest report from Bobby Molavi of Goldman Sachs, OpenRouter data shows that open-weight models handled 72% of Token traffic in the second week of September, yet they only accounted for 27% of the platform’s total spending; at the same time, closed-weight models, while their share dropped from 89% in early January to 73%, still firmly control the majority of monetized revenue.

More notably, about 69% of Tokens on the OpenRouter platform in mid-September came from Chinese AI labs. Molavi characterizes the current market as “unplayable,” and points out that AI productivity gains have yet to leave any measurable trace outside the tech sector.

This structural contradiction is reshaping market expectations for AI investment returns. GPU rental prices continue to rise, Token prices keep dropping, and both ends are under simultaneous pressure. This means compute providers are pocketing rental dividends, but enterprise clients and ultimate shareholders who foot the bill are facing a predicament where ROI is hard to realize. Meta’s recent fundraising of up to $95 billion from shareholders for AI capital expenditure—with its chip suppliers buying in at the same time—marks the shift of “vendor financing” from a debt to an equity model, making the question of capital cycle sustainability even more urgent.

Explosion of Token Demand: The Real Picture Behind a 250x Growth

The growth in demand itself is unquestionable. Citing OpenRouter data, Molavi notes that Token demand has grown 27-fold since last November, with open-weight models now taking up 72% of Token share. Looking over a longer cycle, OpenRouter’s weekly Token volume soared from 0.5 trillion in mid-January 2025 to 128.4 trillion in the week of September 14, about a 250-fold increase in 20 months.

GPU pricing also confirms the authenticity of this demand. H200 rental prices are up 160% this year, B200 up 120%, and H100 up 35%. Even more compelling, the current rental price of the A100 chip produced in 2020 is already higher than at the start of the year. This directly challenges the previous core assumption about GPU depreciation cycles—if a chip can generate higher returns after six years, then the pessimistic three-year depreciation outlook no longer stands, and the long-term profitability of hyperscalers needs to be reassessed.

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

Token Price Free-Fall: The Fastest Deflation in History

However, the other side of this demand explosion is that Token prices are collapsing at an unprecedented pace. According to Epoch AI’s September 22 research report, “The Plunging Price of Thought,” since 2023, the cost of reaching the same level of AI performance has fallen by about 47% per quarter, or about 13x per year; in frontier models, this is even faster, averaging a 66% drop per quarter, or about 75x per year.

Epoch AI provides the most straightforward case:

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

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

However, the actual prices paid by clients have declined much more slowly than the list price. YipitData’s tracking of nearly 20 quadrillion annualized Tokens shows that effective API Token pricing dropped by 39% in the second half of 2025 (44% on OpenRouter), but up to May 2026, it has declined by only 6% year-to-date.

The reason is: developers continuously chase cheaper models, while enterprise clients are increasing their use of high-end models. Frontier labs are temporarily supporting the revenue side with premium pricing for the latest models—but Epoch data shows the half-life of this premium is measured in quarters, as open-source clones quickly replicate tasks at a fraction of the cost.

The Core of the Bull-Bear Dispute: Who Makes Money and Who Pays

This is precisely the core of Molavi’s “two interpretations of the same data” dilemma.

The bull case: demand keeps expanding, the Total Addressable Market (TAM) is still growing, and the explosive increase in Token consumption proves AI is being widely adopted.

The bear case: per-Token costs are rising, per-Token revenue is dropping, and incremental traffic is dominated by the lowest-fee open-weight models. OpenRouter data illustrates this clearly: open-weight models process 72% of Tokens but only take 27% of spending; Chinese AI labs contribute nearly 70% of the platform’s Token traffic. The Goldman momo basket notably peaked at the end of June, syncing with the large-scale entry of Chinese open-weight models into the AI market.

Molavi sums up the contradiction as: rising GPU rental prices and falling Token prices are the two ends of the same squeeze. It’s positive for GPU holders; but for companies paying rent, it’s a different story.

Productivity Gains Missing: The Silence of the S&P 493

The ultimate logic of AI capital expenditure is that productivity increases lift returns for society as a whole. So far, this logic hasn’t been validated by data outside of the tech sector.

Molavi cites a strategist report pointing out that AI “has yet to show measurable productivity gains in key data outside the tech universe”: revenue per capita and profit margins for the Mag 7 (the seven tech giants) have improved, but comparable indicators for the S&P 493 (the other 493 S&P 500 companies) are almost flat.

This reality is changing the logic of corporate AI spending decisions. Molavi observes that many firms are shifting from “correlation between Token spending and productivity/growth” to assessing “Token spending’s impact on costs and profit margins.” CFOs have identified the AI line item in their budgets and want to know what they’re getting for their money.

The ‘Boom Cycle’ and the Question of Capital Sustainability

What supports the current AI capital spending craze is what Molavi describes as a “boom cycle”: the more hyperscalers spend, the more revenue flows into the AI ecosystem, attracting venture capital and growth equity inflows, creating a two-way capital stream—all anchored by the initial AI capital spending itself.

The fatal weak point of this cycle is its sustainability: once capital expenditure slows, can the ecosystem’s revenue self-sustain?

The private market is no exception. Molavi notes that current exit cases are “numerous but extremely concentrated” (SpaceX, Cerebras, Cursor, Hugging Face), and some project revaluations “have no catalyst other than a letter of intent,” with private market deal logic appearing to assume “the Token era will never end, and everything will be disrupted.”

On the funding side, as debt windows tighten, Meta has recently announced it will raise up to $95 billion from shareholders for AI capex, with its chip suppliers simultaneously becoming equity holders. Vendor financing has evolved from a debt into an equity model, blurring the lines between provider and user in terms of revenue and financing.

Molavi distills the nature of the entire debate into four questions: What is the cost to build, who pays, is the pain worth it, and how long must we wait for returns?

The current data landscape is: Token demand grows 250-fold, but the incremental revenue goes to the cheapest models; GPU rentals rise, but the bill ultimately lands with shareholders; productivity gains outside the tech sector remain unseen.

Molavi’s conclusion is that demand is undoubtedly real, but “even as demand rises 250-fold, funds flow to the lowest-price models, rent flows to GPU holders, and the bill is passed to shareholders”—this is not the ROI story that was promised at the outset of capital spending. In this bull-bear tug-of-war, the bears may be holding the stronger end of the rope.