The AI Stack Isn't Flipping

Two years ago, Apoorv Agrawal, an investor at Altimeter Capital, published an analysis showing that the generative AI value chain was inverted: semiconductors captured 83% of revenue and 87% of gross profit, while applications—closest to end customers—earned almost nothing. He predicted this would eventually flip, following the pattern of every prior platform shift.

His update this week shows the AI ecosystem has grown 5x, from ~$90B to ~$435B in annualized revenue. But the shape hasn't changed. Semis still capture 70% of revenue. Apps are stuck at 14%. NVIDIA alone added $175B in incremental revenue—three times the size of the entire app layer today.

Apoorv still believes the stack will flip eventually. He's just less sure about when.

Here's the problem: the frame is wrong. AI isn't cloud.

Cloud commoditized compute. AI is building moats around intelligence.

Cloud economics flipped because infrastructure became abundant and cheap. Apps won because building on AWS was eventually cheaper than owning data centers. The value migrated up the stack as compute commoditized.

AI economics don't work the same way. Here's why:

1. Training compute is deflationary, but frontier intelligence is not

Custom ASICs—Google's TPUs, Amazon's Trainium, Microsoft's Maia, Meta's MTIA—will compress NVIDIA's margins over time. That part's real. Google's TPU v7 reportedly forced NVIDIA to cut pricing ~30% for some customers. Amazon's Trainium business crossed $10B annual run rate. OpenAI signed a multiyear deal with Broadcom for 10GW of custom accelerators.

But here's what doesn't happen: those savings don't get passed to the app layer.

The hyperscalers aren't buying GPUs to resell compute. They're buying GPUs to train proprietary models that become the infrastructure. When Google trains Gemini on TPUs or Amazon trains Olympus on Trainium, they're not just saving money on training—they're internalizing the entire margin structure.

OpenAI and Anthropic aren't app companies in the SaaS sense. They're selling access to intelligence as a service. The line between infrastructure and application collapses. The semiconductor margin doesn't migrate to apps—it gets absorbed by model providers and hyperscalers building their own silicon.

2. Vertical integration breaks the cloud analogy

In cloud, AWS was a neutral platform. Amazon didn't compete with most of its customers. The infrastructure layer stayed separate from the application layer.

In AI, the biggest semiconductor buyers—Google, Meta, Amazon, Microsoft—are also building the mines. They're vertically integrated. They're not just providing picks and shovels to prospectors. They're prospecting themselves.

NVIDIA's dominance is temporary not because apps will capture value, but because hyperscalers will internalize semiconductor margin through custom silicon and then keep it at the infrastructure layer. Jensen Huang can dismiss ASICs as "noncompetitive" all he wants. The custom silicon programs aren't trying to beat NVIDIA on performance—they're trying to eliminate the rent.

3. The app layer is structurally trapped

Apoorv's data shows apps grew 12x in two years but are still stuck at 33% gross margins. Why?

Because they're paying rent to two landlords: the model providers (OpenAI, Anthropic, Google) and the cloud providers (Azure, AWS, GCP). Most "AI apps" are thin wrappers around frontier models with no defensible moat.

The few that work—Cursor, Glean, Harvey, Abridge—succeed because they have proprietary data flywheels or deep workflow lock-in. But those are exceptions. The modal AI app is a feature, not a company.

The app layer can't capture value the way SaaS did because SaaS companies owned their code and their customer relationships. AI apps rent the intelligence and rent the infrastructure. There's no margin left to capture.

The stack isn't flipping. It's collapsing into a barbell.

The winners in this stack are:

• Hyperscalers who control training compute and can subsidize inference to kill competition
• Frontier model labs with enough capital to keep training and enough distribution to avoid getting OEM'd
• Vertical AI companies that own proprietary workflows or data—Harvey in law, Abridge in healthcare, Glean in enterprise search

The losers are undifferentiated SaaS wrappers and anyone trying to arbitrage model pricing without building a data moat.

Everything in between—the middleware, the orchestration layers, the "AI platforms"—is getting compressed. The stack isn't flipping. It's collapsing into a barbell: semis + hyperscalers on one end, vertical AI applications with deep workflow integration on the other.

What this means if you're building or buying AI

Stop waiting for AI to get cheaper.

If you're advising companies on AI adoption, the takeaway isn't "wait for apps to mature." It's "build your AI capability in-house or accept that you're renting infrastructure forever."

The cloud playbook was: let someone else own the infrastructure, focus on your application layer. The AI playbook is different. If intelligence itself is infrastructure, and you don't control it, you don't control your business.

The companies that win are the ones that recognize this early and act accordingly.

For strategic AI advisory, orchestration design, or helping your organization navigate this transition: andmaverick.com