ReadingResearchRadarInvestment framework
Sign in / Sign up
← Back to reading
NewsMaterial4 min readDisconfirmAI Editorial

Theory Ventures Partner Tomasz Tunguz: The AI Industry Chain Gross Margin Waterfall — Applications Capture 40%, Neocloud Post-Depreciation Margin Is Only $0 to $2 per Watt-Year

Theory Ventures partner Tomasz Tunguz normalizes the five layers of the AI value chain into dollars per watt-year ($/w/yr), constructing a gross margin waterfall based on CoreWeave’s Q3 2026 financial report and Anthropic’s S-1/A filing. The analysis shows that applications capture 40% of the total pool, while Neoclouds show a pre-depreciation margin of 28% but only $0 to $2 per watt-year after depreciation. Inference vendors capture 24%. Three key implications emerge: rising compute prices test the ability to pass on pricing power; differing application ROI determines differences in pricing power; and vertical integration is encroaching on adjacent layer margins. The author acknowledges that the model relies on approximate assumptions and contains significant error.

2026-11-11 08:00 GMT+8

The AI industry chain faces a fundamental measurement difficulty: utilities sell kilowatt-hours, neoclouds sell GPU hours, inference providers bill by token, and application layers price based on completed tasks. Each of the five layers quotes in different units, making it impossible to directly compare who earns more. In an analysis published on November 11, Tomasz Tunguz, partner at Theory Ventures and former Google product manager, proposed a normalization solution: convert all layers to dollars per watt-year ($/w/yr), which represents how many dollars of revenue are generated annually for every watt of continuous power. This denominator brings the entire chain onto the same ruler.

After normalization, Tomasz Tunguz assembled a gross margin waterfall using CoreWeave’s Q3 2026 earnings report, Anthropic’s S-1/A filing, disclosures from Core Scientific with the U.S. Securities and Exchange Commission (SEC), and CBRE data center leasing data. Assuming an application layer gross margin of 40%, the distribution of the total gross profit pool is as follows: Power 3%, Data Center 5%, Neocloud 28% (before GPU depreciation), Inference 24%, and Applications 40%. The three layers of Applications (40%), Neocloud (28%), and Inference Providers (24%) collectively capture 92% of the total gross margin—this figure is before GPU depreciation.

Sign in / Sign up
ReadingResearchRadarInvestment framework

However, this waterfall has a critical accounting boundary. The Neocloud figure of 28% is pre-depreciation. The capital expenditure for GPU hardware—$35 to $50 per watt, depreciated over six years—adds $6 to $8 per watt in annual capital costs, a portion of which effectively belongs to Nvidia (which has a hardware gross margin of approximately 75%). CoreWeave’s depreciation and amortization in Q3 amounted to $1.89 billion, representing 52% of its revenue. After deducting GPU depreciation, Neocloud’s gross margin drops to $0 to $2 per watt-year under the current blended rate of $9 per watt-year. In other words, while Neocloud appears to earn significantly, most of that gross margin is consumed by GPU hardware costs.

From this, Tomasz Tunguz draws the first implication: rising compute prices are testing who can transmit costs to end users. Anthropic’s S-1/A filing discloses that short-term burst compute has reached $13 per GPU-hour. On CoreWeave’s Q3 earnings call, CEO Michael Intrator confirmed that prices for new four-to-five-year contracts are shifting from $40 per megawatt-year toward $20 to $25 per megawatt-year—more than double CoreWeave’s current blended fleet rate of $9 per watt-year. If application layers can charge end users higher prices due to improved effectiveness, cost increases can cascade through the chain; if not, Neoclouds, inference providers, and applications must compete for a fixed pool of gross margin.

The second implication is that pricing power for different applications depends on their return on investment (ROI). Tomasz Tunguz points out that an AI that saves 30% in costs does not possess the same pricing power as one that saves 70%. Small models, fine-tuning, and routing techniques can help maintain margins, but unlike the SaaS era, different vertical software markets will support varying sizes of gross profit pools. This means the 40% gross margin in the AI application layer is not evenly distributed but highly differentiated.

The third implication is that vertical integration is already absorbing adjacent layer margins. In 2025, CoreWeave acquired Weights & Biases ($1.03 billion), OpenPipe, Marimo, and Monolith AI ($185 million), moving into training, evaluation, and enterprise workflows. Nebius spent over $1 billion in 2026 acquiring four companies spanning search, model optimization, cold-start scaling, and orchestration to build its Token Factory. Owning multiple layers allows integrators to absorb price fluctuations that squeeze independent suppliers.

Tomasz Tunguz himself marked the boundaries of this model: the conversion to $/w/yr relies on modeling approximations rather than audited disclosures, involving parameters such as quarterly starting active power, facility PUE (1.2 to 1.45), rack power draw (150 to 180 kW), billing utilization (65%), and cache hit rates (87% to 90%), leading to significant error across workloads. The inference layer modeling is based on open-weight models rather than proprietary frontier lab APIs. The 40% application gross margin is an assumed input, not a measured result. This waterfall describes a static snapshot and does not predict which layer will collapse first or when.

Sources

  • tomtunguz.comTomasz Tunguz↗
Continue researching →