ReadingResearchRadarInvestment frameworkAI Company Primary Sources
Sign in / Sign up
← Back to reading
NewsMaterial3 min readDisconfirmAI Editorial

Tomasz Tunguz: How GPU Price Hikes Elevate Harness Value Through Token Structure — The Transmission and Boundaries of AI's Four-Layer Coupling

Theory Ventures partner Tomasz Tunguz proposes a four-layer coupling framework for the AI industry, in which GPU cost inflation transmits through the input token share (up to 98%) to amplify the value of token optimization at the harness layer. This article focuses on this specific transmission mechanism: as the Ornn B300 index rises from $5.53 to $12.29 per hour, the economic value of reducing invalid context injection by harnesses is magnified. It also examines the boundaries of the evidence: key figures come from different experiments and models; vendor-reported data and current routing behaviors do not equal proven trends; and the framework describes a coupled structure but does not provide predictions.

2026-11-02 08:00 GMT+8

Tomasz Tunguz is a partner at Theory Ventures and a former Google product manager. In an article published in November 2026, he proposed that the AI industry consists of four layers: applications, harness & context, models, and data centers. Each layer faces an unresolved economic question, and the four layers form a coupled system. This article focuses on one specific transmission chain: how GPU cost inflation elevates the economic value of the harness layer, and where the evidentiary boundaries of this logic lie.

Tunguz’s transmission mechanism relies on a structural fact: in agent workloads, input context accounts for up to 98% of all tokens (he cites OpenRouter’s token composition data, ranging from 85.7% to 98.2%). This means that the vast majority of token consumption in model inference is not 'answering,' but rather 'feeding context to the model.' When GPU prices rise—the Ornn B300 index increased from approximately $5.53 per hour in mid-June to $12.29 per hour by the end of October, with a compound monthly growth rate of about 20%—the computational cost per token increases accordingly. At this point, the economic value of a harness that can reduce invalid context injection is magnified: what is saved are not output tokens, but the input tokens that make up 98% of the total volume.

Sign in / Sign up
ReadingResearchRadarInvestment frameworkAI Company Primary Sources

Two key figures support this logic. First, UC Berkeley’s HarnessTax study found in a comparison of 42 SWE-bench tasks that optimizing the harness can reduce the cost of solving the same task by 71% without producing a statistically significant difference in quality. Second, Tunguz’s own contract modeling shows that on a $250,000 AI assistant contract, harness optimization can lift the gross margin from 38% to 75%. These two figures come from different evidence families: the former from independent academic experiments and the latter from commercial contract projections, with different metrics: the former measures task cost, while the latter measures contract gross margin.

However, the evidentiary boundaries of this transmission chain need to be clearly marked. First, the Anthropic S-1 disclosure that Opus 5.5 reduced customer cost per task by 40%, and that quarterly revenue growth accelerated from 48% to 132% after the release of Opus 4.5, are vendor-reported data, not independently verified figures. Second, Tunguz observed that about 83% of agent tasks are now routed to smaller or domain-specific models, but this is a description of current behavior; he did not prove this is an irreversible trend—the same article also recorded secondary market GPU quotes dropping from $7.90 per hour to $4.60 per hour within a month, indicating that underlying costs themselves fluctuate violently. Third, the 98% input token share comes from token composition statistics of the OpenRouter public gateway, whose scope of application may not cover all deployment scenarios.

Tunguz uses a 'double pendulum' to metaphorize the four-layer coupling: each component follows simple deterministic laws, but the entire system exhibits chaotic motion, where tiny changes produce unpredictable large swings. This metaphor itself acknowledges the framework’s limitations—it describes a coupled structure but does not provide predictions. Readers can use it as an analytical map to locate where cost pressures land: when GPU prices rise, focus on the harness layer’s token optimization capabilities; when model efficiency improves, focus on pricing space at the application layer. But the framework does not tell you when these changes will occur, how large they will be, or which layer will collapse first.

Sources

  • tomtunguz.comTomasz Tunguz↗
Continue researching →