The AI Market: the dollar, not the token, is the true measure of power
Dragonfly Managing Partner Haseeb Qureshi sharply criticized the current approach to analyzing the AI model market. His main thesis: the share of tokens consumed is an extremely unreliable and misleading metric. Models should be compared solely by dollars spent, not by the volume of generated tokens. I consider this thesis fundamentally important for understanding the real dynamics of the industry.
Four Pitfalls of the Token Metric
The first problem is subsidies. Chinese labs regularly launch new models with aggressive discounts or even free access. This attracts users who migrate between free models, inflating token consumption without spending real money. In such cases, "market share" charts paint a false picture.
The second problem is different model sizes. A small model like Qwen 3.5-27B costs roughly a hundred times less per token than the flagship Claude Opus. Increased usage of Qwen might look like a sharp jump in market share, although economically it is a drop in the ocean. Analysis should be done within weight classes.
The third problem is multi-agent systems. The same amount can be spent on a complex system based on DeepSeek or GLM 5.2 and on a single advanced model like Opus or GPT-5.5 Pro. But a multi-agent configuration will "burn" far more tokens for the same money. As Qureshi accurately noted: if 5% of Opus usage shifts to such a system with four times the token consumption, the chart will show an ~18% loss of Opus share, although actual spending shifts by only 5%. Such charts exaggerate the importance of cheap tokens.
The fourth problem is the OpenRouter sample. Large companies, once committed to a single leading lab, prefer to work directly with Anthropic or OpenAI, avoiding OpenRouter's markup. On charts, this looks like a decline in the US share, even though the tokens simply move off the platform. Conclusion: OpenRouter is useful for assessing share within open models, but not for comparing open and closed ones.
Is the Future in Cheap Models?
SageRoad Research founder Trevor Noren developed a similar idea, linking it to pricing pressure on the industry. He cited JPMorgan's assessment: many tokens in the future will be consumed not by advanced models, but by small open models sufficient for specific tasks. Amazon already offers around fifty open models at a fraction of the cost of advanced ones, while Nvidia, along with Dell, Lenovo, and HP, is creating computers for AI agents.
The cost example is particularly illustrative. Running the Artificial Analysis Intelligence Index task set on Claude Opus 4.8 costs $3,700 for a score of 56, while DeepSeek V4 Pro scores 44 points for just $186 — roughly 20 times cheaper. Conclusion: advanced intelligence is not needed for everything, and where it is necessary, GLM 5.2 from Z.ai appears comparable to the top models from Anthropic and OpenAI.
Noren believes model commoditization will come not only from competition among leading labs, but also from companies seeking cost control through cheaper, specialized models. Both positions converge on one point: the AI market should be measured by money, not tokens, and under pricing pressure, the advantage is increasingly shifting towards cheap models.
My conclusion: the AI market is entering a phase of maturity where "raw volume" metrics give way to economic efficiency. Investors and analysts who fail to shift to dollar-based valuation risk missing a tectonic shift towards pragmatic model selection. Chinese labs have already captured the "efficient frontier" in the small model segment, and this is changing the rules of the game.