NVIDIA Takes 79% of AI Profit. Applications Get 7%. The Value Chain Won't Self-Correct.

In a healthy value chain, profit distributes roughly proportional to value created. The company that builds the thing the end user buys captures the largest share. The company supplying raw materials captures less. This is economics 101.

The AI value chain does not work this way.

Two years after the generative AI explosion, Apoorv Agrawal at Altimeter Capital published an updated breakdown of where the money actually goes. The total market grew from $90 billion to $435 billion. Nearly five times larger. And the structural inversion that was visible in 2024 has not corrected. It has calcified.

The Numbers

The AI economy now has three layers, and their profit shares look nothing like what you would expect from a mature market.

Semiconductors: approximately $300 billion in revenue, 73% gross margin, capturing 79% of all AI gross profit. This layer is, functionally, one company. NVIDIA’s data center business alone runs at roughly $250 billion annualized ($62 billion per quarter). In two years, NVIDIA added $175 billion in revenue. That single increment is three times the entire application layer.

Infrastructure: approximately $75 billion in revenue, 55% gross margin, 14% of gross profit. This is the hyperscaler layer. Amazon, Microsoft, Google, and Oracle compete aggressively here. Their combined AI capex reached $443 billion in 2025 (up 73% year-over-year, per Epoch AI) and is projected to hit $450 billion of $600 billion total capex in 2026.

Applications: approximately $60 billion in revenue, 33% gross margin, 7% of gross profit. This is where the products live. The chatbots, the coding assistants, the enterprise tools. And this layer is itself a duopoly: OpenAI and Anthropic control roughly $45 billion of that $60 billion. About 75%.

Read those numbers again. The layer closest to the end user, the layer building the products that justify the entire investment, captures seven cents of every dollar of profit.

Why This Matters for Governance

When we analyzed AI’s economic structure a month ago, the 10:1 spend-to-revenue ratio was the headline. That ratio remains, but the Altimeter data reveals something more structural. The problem is not just that spending outpaces revenue. The problem is where profit concentrates.

A value chain where one company captures 79% of profit is not a market. It is a dependency.

Every enterprise AI strategy depends on compute. That compute depends on NVIDIA GPUs. NVIDIA’s gross margin of 73% on AI hardware tells you exactly how much pricing power they have and how little leverage their customers possess. Jensen Huang called custom ASICs “noncompetitive” and noted that “a lot of ASICs get canceled.” He can afford to say that when you hold 79% of the profit pool.

Amazon is trying to change this. Their custom Trainium2 chips have a run rate above $10 billion, with 1.4 million chips deployed. Google ordered up to 1 million TPU chips (many for Anthropic’s workloads). These are serious efforts. But as Andy Jassy put it with remarkable candor: “As fast as we’re adding capacity right now, we’re monetizing it.” The hyperscalers are building alternatives to NVIDIA while simultaneously buying everything NVIDIA can produce.

This is not a market correcting itself. It is a market deepening its dependency while talking about reducing it.

Concentration Risk Is Governance Risk

Most enterprise governance frameworks treat vendor risk as a procurement problem. Evaluate the vendor, negotiate the contract, monitor the SLA. Standard practice.

But the AI value chain introduces a different kind of vendor risk. Your direct vendor (say, Microsoft Azure or AWS) is itself dependent on a single supplier that captures nearly four-fifths of the profit in the entire chain. Your governance framework evaluates your vendor. It rarely evaluates your vendor’s vendor. And it almost never asks: what happens if the company that captures 79% of profit decides to change pricing, allocation, or terms?

This is not theoretical. NVIDIA’s allocations already determine which cloud providers can offer what capacity, and when. An enterprise that has not modeled this dependency into its AI governance is operating with a blind spot the size of $300 billion.

Three specific governance questions follow from this data:

Single-supplier exposure at the infrastructure layer. If your AI workloads run on one cloud provider, and that provider depends on NVIDIA for 80%+ of its AI compute, your “multi-cloud strategy” is cosmetic. Governance should model the actual supply chain, not the contractual surface.

Application layer concentration. OpenAI and Anthropic controlling 75% of the application layer means that most enterprises building on foundation models are building on a duopoly. Agrawal’s analysis notes that at the current pace, the application layer would need more than ten years to reach the same revenue share that cloud computing achieved. The competitive dynamics that drive down prices and improve terms depend on more than two players reaching scale.

Capex-to-revenue mismatch as systemic risk. Hyperscalers spending $443 billion while AI applications generate $60 billion in revenue creates a structural fragility. If the revenue doesn’t materialize on a reasonable timeline, the capacity investments will need to be rationalized. Your governance framework should include scenario planning for what happens when your cloud provider starts making different allocation decisions under margin pressure.

The Only Competitive Layer

Here is the structural irony. Of the three layers, infrastructure is the only one with genuine competition. Semiconductors are a one-player game. Applications are a two-player game. Infrastructure, with Amazon, Microsoft, Google, Oracle, and a growing set of specialized providers, is the only layer where market dynamics produce the kind of competition that benefits buyers.

This means enterprise governance should weight infrastructure decisions more heavily than most organizations currently do. Not because infrastructure is the largest cost (it isn’t, for most enterprises), but because it is the only layer where your procurement choices actually create leverage. You cannot meaningfully negotiate with NVIDIA. You have limited leverage against an OpenAI/Anthropic duopoly. But you can create real competitive pressure between cloud providers, and you should structure your governance to do exactly that.

What the Data Suggests

Two years of data have confirmed three things.

First, the value chain inversion is not a phase. It is a structure. NVIDIA’s profit share dropped from 87% to 79%, but that decline came from the infrastructure layer growing, not from applications catching up. The application layer’s 7% share is a structural condition, not a starting point.

Second, the hyperscaler response (custom silicon) is real but insufficient. Amazon’s $10 billion Trainium run rate is meaningful. It is also 4% of NVIDIA’s AI data center revenue. Google’s TPU deployment is serious. It is also primarily serving one customer (Anthropic). Custom chips are narrowing the dependency; they are not eliminating it.

Third, the application layer’s consolidation is accelerating, not dispersing. Two companies controlling 75% of a $60 billion market means the remaining 25% is split among hundreds of startups and incumbents. The expected pattern of a maturing market (where competition drives profit share toward equilibrium) is not occurring. The leaders are pulling away.

For enterprise AI governance, this means the standard assumption that “the market will sort itself out” is not supported by the data. Two years in, the market is sorting itself into a monopoly, a duopoly, and a competitive middle that subsidizes both.

The Governance Response

If you are building an enterprise AI strategy on the assumption that you will have meaningful vendor choice in eighteen months, the data does not support that assumption. Governance frameworks need to account for structural concentration, not just individual vendor risk.

Practical steps:

Map your actual supply chain, including your vendors’ dependencies, not just your direct contracts. If three of your AI initiatives depend on GPT-4 running on Azure backed by NVIDIA H100s, you have one point of failure with three labels.

Build optionality into your architecture now, while switching costs are still low. The application layer is young enough that migrating between foundation models is feasible. In two years, accumulated fine-tuning, custom tooling, and workflow integration will make it much harder.

Model concentration scenarios in your risk framework. What happens if NVIDIA raises prices 20%? What happens if OpenAI’s pricing changes after their next funding round? What happens if your cloud provider prioritizes its own AI products over third-party workloads? These are not edge cases. They are the natural behavior of concentrated markets.

Govern the economics, not just the technology. Most AI governance frameworks focus on model risk, data privacy, and compliance. Those matter. But the economic structure of the value chain is a governance concern of equal weight. A technology that works perfectly but costs three times your projection because of supply chain concentration is still a governance failure.

The AI value chain will not self-correct into competitive equilibrium. Two years of data confirm that. The question is whether your governance accounts for the market as it actually is, or the market as you hope it will become.

This analysis synthesizes The Economics of Generative AI: Two Years Later by Apoorv Agrawal (April 2026), with data from Epoch AI, Amazon, and NVIDIA earnings reports.

Victorino Group helps organizations navigate AI vendor concentration risk and build governance for multi-provider strategies. Let’s talk.