In early 2025, Anthropic was fighting for survival. Its gross margin on inference sat at a miserable 38%. By mid-2026, that number has nearly doubled to over 70%. And here’s the kicker — they did it while slashing prices by 3x.
That’s not a miracle. That’s a structural shift in the AI value chain. And most people are looking in the wrong direction.
The biggest lie in AI is that the hardware companies are the only winners.
For three years, the narrative was simple: Nvidia, TSMC, and the storage giants were the picks-and-shovels kings. The model labs were bleeding cash. But that story is dead. The profit center is moving — fast and quietly — from the infrastructure layer to the model layer. And the real chess game is being played by Nvidia, who is deliberately leaving money on the table to build a fortress that will last decades.
You’ve probably read that Nvidia has a monopoly. You’ve probably heard that they can charge whatever they want. But you haven’t heard why they choose not to.
The Day the Profit Center Shifted
In December 2025, something snapped. SemiAnalysis, a research firm that lives inside AI workflows, noticed a pattern: their analysts were using AI agents to do tasks that used to cost thousands of dollars — building financial models, analyzing earnings reports, creating dashboards. The cost per million tokens for these agentic workloads? $0.99. That’s one-fifth of the official Claude Opus 4.5 price.
How? The secret is in the structure: a 300-to-1 input-to-output ratio, over 90% prompt cache hits, and cache pricing at $0.50 per million tokens. The enterprise pays pennies for tokens but gets output worth tens of thousands of dollars. The gap between cost and value is the new profit engine.
Token production costs are collapsing. Token value creation is exploding. And the gap is now being pocketed by model companies, not chip makers.
For the first time in AI history, the model layer has pricing power. And they’re using it.
The Nvidia Paradox: Restraint as a Weapon
Here’s where it gets weird. Nvidia could double GPU prices tomorrow and still have customers. Their nearest competitor, AMD, is still years behind in software and scale. Yet Nvidia is underpricing its GPUs. Why?
Think of Nvidia as the world’s most powerful central bank. A central bank’s job isn’t to maximize profit — it’s to maintain the stability and growth of the entire monetary system. Nvidia is doing the same: suppressing GPU prices to nurture the AI ecosystem, ensuring long-term demand expansion. If they squeeze too hard now, they kill the golden goose.
But that’s only half the story. The other half is hidden in the fine print.
In 2025, Nvidia introduced SOCAMM — a proprietary memory module for the GB300 architecture. It’s a simple trick: take commodity DRAM, package it into a closed ecosystem, and charge a 60% gross margin. The GPU market is watched by regulators and analysts. The memory module market? No one is looking. Nvidia is quietly moving profits from the monitored business to the unmonitored one.
The same goes for networking. SemiAnalysis revealed that Nvidia charges neocloud providers twice as much for the same InfiniBand switch as it charges hyperscalers. Price discrimination in plain sight. The network business is smaller than GPU sales, but its margins are fatter — and it’s invisible to public scrutiny.
The One Chart That Explains Everything
SemiAnalysis built a framework they call “One Chart to Rule Them All.” It shows the floor and ceiling for GPU rental pricing. The floor is set by what neoclouds need to earn a 15.6% IRR. The ceiling is set by what customers are willing to pay before they switch to alternatives. Between them is a massive gap.
Right now, Nvidia is pricing near the floor. The gap to the ceiling is about 40%. That means Nvidia could raise GPU prices by 40% and still have customers feeling like they’re getting a deal. The neoclouds would still make money. The end users would still see lower costs than historical trends. Nvidia has a $40 billion price hike in its back pocket, and it’s not using it.
Why? Because the moment they raise prices, they trigger a chain reaction. Customers start looking harder at alternatives like AMD MI400, Intel Falcon Shores, and custom chips from Google and Amazon. Anthropic is already training its flagship models on Google TPUs and Amazon Trainium. The cord is not cut yet, but it’s fraying.
Nvidia’s restraint is a strategic bet: keep the ecosystem so dependent on CUDA and the Nvidia stack that when the day comes to finally raise prices, the switching costs are too high. But that day is not today.
What Happens Next
Anthropic crossing 70% gross margin is not an anomaly. It’s a signal. When agentic AI turns tokens from conversation tools into production engines, the willingness to pay skyrockets. The model layer is now the profit center of the AI industry.
Meanwhile, Nvidia is sitting on a powder keg. They have 40% pricing power in GPUs, 60% margins in SOCAMM, and 2x price discrimination in networking. When they decide to shift from “ecosystem nurturing” to “value capture,” the entire AI value chain will be reshuffled.
But the wolves are at the door. Anthropic, Google, and Amazon are building their own alternatives. The day Nvidia squeezes too hard is the day the hyperscalers accelerate their exodus. The tension between Nvidia’s power and its restraint is the most important story in AI right now.
Profit always flows to the path of least resistance. When the tide turns — and it will — everyone will have to recalculate their bets.
The only question is whether you’re betting on the central bank or the revolution.
FAQ
Q: If Nvidia could raise prices 40%, why don't they?
A: Because they're playing a multi-decade game. Raising prices now would accelerate customers' migration to alternatives like Google TPUs and Amazon Trainium. Nvidia is suppressing prices to keep the ecosystem dependent on CUDA, knowing that switching costs will be even higher later. It's a bet on long-term dominance over short-term profit.
Q: Does this mean I should invest in AI model companies instead of chip companies?
A: Not necessarily. The profit shift is real, but it's early. Model companies like Anthropic are capturing value now, but their margins are fragile — competition is fierce and open-source models are catching up. Nvidia's hidden profits (SOCAMM, networking) are less visible but more durable. The smart money is watching both layers and the tension between them.
Q: Isn't this just a temporary phase before commoditization kills margins?
A: That's the conventional wisdom, but it ignores the agentic AI effect. When tokens become production tools (not just chat), the value they create far exceeds their cost. The gap between cost and value is widening, not shrinking. That's why model companies can raise prices even as costs collapse. Commoditization might hit simple chatbots, but for high-stakes autonomous work, reliability and safety create a moat that pricing alone can't breach.