The CUDA moat springs a leak | Tuesday, June 30, 2026

On June 25 Qualcomm confirmed it is buying Modular, the startup behind the Mojo language and the MAX inference engine, for about $3.92 billion. The bet: the way to beat Nvidia is to make its software moat irrelevant. Days earlier, Amazon disclosed that its custom silicon business, Trainium and Graviton, crossed a $20 billion annual run rate.

Two different companies, one message. The lock-in that let Nvidia price like a monopoly is taking fire from several directions at once. For a CIO signing multi-year inference contracts, the question is shifting from which GPU to whether you are still required to care.

The deal does not arrive alone. Amazon told investors its custom-chip business now runs at a $20 billion annual rate, growing triple digits, with multi-gigawatt Trainium commitments from OpenAI and Anthropic and active talks to sell the chips to outside data centers. Google's TPUs, OpenAI's inference silicon, and now a hardware-agnostic software stack all point the same way. The industry is trying to turn Nvidia's moat into a commodity.

For enterprise buyers, the takeaway is not that Nvidia is in trouble. It is that the cost of being locked in is now negotiable. Inference, not training, is where production spend lives, and inference is exactly where a portability layer changes the math. The leverage in your next compute contract just improved, even if you never switch a single chip.

The tidy read on today is that Nvidia's moat is finally cracking. The harder read is that software portability and actual switching are very different things, and the gap between them is where lock-in lives.

Owning a hardware-agnostic compiler does not retire the switching costs that matter: retuned models, revalidated pipelines, retrained MLOps teams, and the simple fact that roughly 95% of enterprise AI usage still runs on frontier models served on Nvidia. Buyers tell surveys they want optionality, then keep shipping on the stack their engineers already know. Qualcomm's bet pays off only if enterprises do the unglamorous work of writing to an abstraction layer before the renewal, not after. Most will not, which is precisely why the moat has lasted this long. The credible threat to switch is worth more than the switch, but only if someone actually builds the off-ramp.

For three years the trade was simple. Buy Nvidia, rent Nvidia, build on Nvidia, and assume the moat only deepens. This month the assumption took fire from every direction at once. A chip company bought a software layer. A retailer turned its internal silicon into a $20 billion business and started shopping it around. A search giant raised $85 billion to out-build everyone, then watched its best researchers walk out the door.

None of this means the incumbents are in trouble. It means the structure is loosening, and loosening structures are where leverage lives. The enterprises we work with are not trying to predict which chip wins. They are buying optionality. They write to abstraction layers, keep a second model qualified, and make sure no single vendor can reprice them without a fight.