Nvidia's data centre business took in $75.2 billion last quarter, up 92% on the year. Twelve years ago the same segment made $57 million. That is roughly a 1,300-fold increase, and since ChatGPT arrived in late 2022 the line has been doubling about every eleven months. Morgan Stanley, which calls Nvidia "the most compelling story in semiconductors," now reckons its AI revenue is heading for a $100 billion annual run rate. Meanwhile the business that built the company, gaming and consumer graphics, sits flat at $5 to $6 billion a quarter. It has not shrunk. It has simply become a rounding error.
Read that chart and one thing follows immediately: every company writing those cheques has an overwhelming financial reason to stop. As one analysis of the numbers put it, the incentive to route around a 1,300x toll booth is not philosophical, it is arithmetic. And in the past few weeks the bypass projects have all surfaced at once.
OpenAI unveiled its first in-house chip, Jalapeño, on 24 June. It is an inference chip, not a training chip, which is the tell. Training is where Nvidia's general-purpose GPUs and the CUDA software ecosystem are hardest to dislodge. Inference is the daily grind of answering user queries, a cost that never stops, and a workload specialised silicon can attack. Broadcom's chief executive Hock Tan says Jalapeño's inference costs could run about 50% below mainstream AI GPUs at performance roughly comparable to Blackwell. The chip went from design to tape-out in nine months, against the usual two to three years, partly because OpenAI used its own models to help design it. Richard Ho, who runs hardware at OpenAI, frames the goal as control of the whole stack from model to chip to data centre. The roadmap targets 10 gigawatts of custom compute by 2029.
Meta is further along than most people realise. Reuters reports the next generation of its MTIA accelerators could enter production as early as September, with one chip already through roughly six weeks of testing. Anthropic is reportedly exploring its own silicon with Samsung. Google and Amazon have been shipping homegrown accelerators for years.
Here is the part nobody says out loud. OpenAI's escape chip is designed with Broadcom. Meta's escape chip is designed with Broadcom. The great flight from a single point of dependence is being organised, to a striking degree, through a single supplier. And the fabrication runs through TSMC either way. The bypass has a toll booth of its own; it is just newer, and cheaper for now.
That concentration is not theoretical. In early June, Broadcom beat expectations on revenue and earnings but guided Q3 AI chip sales to $16 billion against the $17.2 billion analysts wanted, and declined to raise its full-year AI forecast. The stock fell 14%. AMD dropped almost 11% and Intel more than 11% the following day. One cautious sentence from the company building everyone's alternative was enough to shake the entire sector.
The people actually selling compute insist none of this reflects weakening demand. Pat Gelsinger, the former Intel chief now at Playground Global, told CNBC he thinks of AI demand as "almost unlimited," with energy availability "the only real limiter." Cerebras chief executive Andrew Feldman says industry demand "far outstrips available capacity." Lumentum, which makes optical components for data centres, says it is sold out for five years. What has changed is the mood inside customer organisations. The era of "tokenmaxxing," where staff were encouraged to burn as many tokens as possible on principle, is giving way to finance departments asking what the tokens bought, particularly when open-weight models from DeepSeek and Alibaba come close for far less.
The uncomfortable footnote is who pays for the boom. Memory makers such as Micron have redirected output to high-margin data centre buyers, and IDC now expects the global smartphone market to post its steepest annual fall on record in 2026, down 13% to a decade low, in what it calls a shock originating in the memory supply chain. The AI build-out is not just reshaping who sells chips. It is quietly deciding who does not get them.