The biggest names in AI are racing to design their own chips, and the collective effort is turning up the heat on Nvidia, whose GPUs have been the near-mandatory currency of the AI boom. OpenAI recently unveiled its first custom chip, developed with Broadcom, joining a list of hyperscalers and frontier companies -- from Google and Amazon to Meta and now SpaceX-adjacent efforts -- that have concluded buying all their compute from a single vendor is both expensive and strategically dangerous.
The logic is straightforward economics. Nvidia commands gross margins that have run around 75%, meaning every dollar of GPU spend carries a large premium that flows straight to Nvidia's bottom line. For companies spending tens of billions a year on compute, even partial substitution with in-house accelerators tuned to their specific workloads can save enormous sums and, just as importantly, free them from the supply queues and allocation politics that have defined the GPU shortage.
The pattern is now industry-wide. Google has its TPUs, Amazon its Trainium and Inferentia, Meta its MTIA line, and OpenAI's Broadcom collaboration brings the most prominent model lab into the custom-silicon club. The enabling players are the merchant chip designers -- Broadcom and Marvell chief among them -- that provide the design IP and packaging expertise to turn a software company's requirements into working silicon. They, alongside TSMC's manufacturing, are quietly becoming the indispensable layer beneath the AI economy.
“Nvidia commands gross margins that have run around 75%, meaning every dollar of GPU spend carries a large premium that flows straight to Nvidia's bottom line.”
This is not Nvidia's death notice -- far from it. Its CUDA software ecosystem remains a deep moat, its newest GPUs still set the performance bar, and demand continues to outstrip supply. Custom chips tend to target specific, high-volume workloads like inference rather than the full breadth of training, so the near-term effect is margin pressure and diversification rather than displacement. But the direction of travel is unmistakable: the largest buyers are systematically reducing their dependence.
The strategic read connects to a broader theme running across the AI landscape -- that durable advantage is migrating to whoever controls the physical layer. Just as Amazon and rivals are racing to lock up data-center capacity and power, controlling your own silicon is a way to own your cost structure and your destiny rather than renting them. It's the same instinct that has Musk's world buying optical-interconnect startups and building data centers: vertical integration of compute.
The bear case for the challengers is that designing competitive chips is brutally hard, expensive, and slow, and many in-house efforts underperform Nvidia's roadmap or arrive late. CUDA lock-in is real, and software portability remains a genuine obstacle. What to watch: how much of their own inference the hyperscalers actually move onto custom silicon, whether OpenAI's Broadcom chip ships at scale, and whether Nvidia's margins finally compress as the buyers it depends on become its competitors.