Here’s the thing about Nvidia: they’ve been running the most profitable protection racket in tech history. A single H100 GPU costs $40,000, and Big Tech companies are buying millions of them. That’s roughly $30,000 in pure profit per chip for Nvidia—basically free money with a manufacturing cost of $3,000-$5,000.
But here’s where it gets interesting. Google, Amazon, Meta, and Microsoft aren’t the type to keep paying tribute forever.
For years, they had no choice. Nvidia owned 90% of the AI accelerator market in 2022. When you control 90% of something, you’re not just a player—you’re the game. But then something beautiful happened: the AI boom exploded faster than anyone could manufacture chips. Suddenly, the world’s most sophisticated tech companies were stuck in line, waiting for Nvidia to build more GPUs. Amazon literally told investors it couldn’t fulfill demand because of chip shortages. Anthropic couldn’t train Claude fast enough. The bottleneck became unbearable.
And that’s when Big Tech did what Big Tech always does: they built their way out.
Google started this years ago with its Tensor Processing Unit (TPU) back in 2015—a skunkworks project nobody paid attention to. Amazon followed with Inferentia and Trainium. Meta launched MTIA. Microsoft built Maia. By 2023, every major hyperscaler had custom AI chips in production. The financial press was too busy watching Nvidia’s stock to notice their biggest customers were quietly spending billions to escape Nvidia’s grip.
Then Amazon dropped Project Rainier—an internal AI cluster built entirely on its own Trainium2 chips. We’re talking 500,000 custom processors in a single system. For context, OpenAI used about 25,000 Nvidia GPUs to train GPT-4. Rainier is 20 times that, all on Amazon’s own silicon.
That was the signal. Six months later, in April 2026, the dam broke.
Broadcom signed a five-year TPU partnership with Google through 2031. Meta committed to $2 billion per year in design fees for custom chips. Google started talks with Marvell for two additional custom AI chips. Amazon announced a $25 billion investment in Anthropic, with Anthropic committing $100 billion to AWS over 10 years—centered on custom chips. Over $200 billion in long-term commitments in two weeks.
Nvidia’s data center share has already dropped from 90% to 75%. More than half of hyperscaler inference workloads now run on custom ASICs. Custom chip sales are growing at 45% annually—nearly three times GPU growth.
Here’s the real play: the “toll roads.” In the PC revolution, Intel didn’t pick sides between Apple, Dell, and IBM. It just supplied chips to everyone and collected its fee. Intel went up over 10,000% from 1981 to its peak.
The same thing is happening with AI infrastructure. The real money isn’t betting on which AI company wins—it’s owning the companies that get paid regardless. Arm (architecture), Synopsys and Cadence (design tools), Broadcom and Marvell (custom silicon builders), Taiwan Semiconductor (manufacturing), and networking players like Arista and Credo.
The difference between 2023’s AI speculation and 2026’s custom silicon trade? This one’s already in the numbers. The contracts are signed. The revenue is real. The only question is whether you’re positioned on the right side of it.
