Meta takes its silicon in-house
The day's biggest story comes from an internal Meta memo reviewed by Reuters: the company plans to begin manufacturing its own AI-specific chip, codenamed Iris, in September. It moves Meta's custom silicon from the drawing board to the assembly line, and it arrives with ambition attached — a separate report on the memo says Meta aims to roll out a new AI chip roughly every six months as part of one of its largest AI infrastructure upgrades yet. Coverage from MLQ.ai and others echoes the September timeline, framing Iris as Meta building more of its own hardware rather than buying it.
The subtext is Nvidia. Multiple outlets, including MSN, read Meta's move as a bid to loosen the tech industry's heavy reliance on Nvidia's GPUs — and the kind of shift that, if it spreads, cools demand for the chips that have powered the boom.
China's developers want their own chips too
Meta isn't alone in wanting off Nvidia's supply. Reports from finance.biggo.com and others say two of China's leading AI developers, DeepSeek and Zhipu AI, are building custom chips rather than relying entirely on Nvidia GPUs. It's the same instinct playing out on the other side of the Pacific: reduce dependence on a single outside supplier, and take control of the hardware roadmap.
But Nvidia isn't going anywhere
For all the talk of loosening its grip, Nvidia's dominance was on display today. A new model called Grok 4.5, released by SpaceXAI, was trained on 'tens of thousands' of Nvidia chips, according to TechCrunch — a scale of compute that caught the attention of both the tech world and Wall Street. And the addressable market may be about to widen: Beijing is set to let Chinese AI companies buy Nvidia's H200 chips, per The Japan Times, Finimize, GuruFocus and finance.biggo.com, in what's described as a limited rather than wide-open approval. The custom-silicon projects are long games; Nvidia is still where the training happens now.
A cheaper path to the leading edge
The manufacturing map is shifting too. Japan's Rapidus is staking its challenge to TSMC on price, disclosing plans (via Tom's Hardware) to sell its most advanced 2nm-class wafers by undercutting the industry leader. Meanwhile, Apple has committed $30 billion to Broadcom for chips to be made in Colorado, according to Startup Fortune — a renewal of the two companies' existing relationship, and another vote for domestic production.
The number to watch
All of this sets up July 16, when TSMC reports second-quarter results. Wall Street is treating the report as more than one company's profit line — analysts see it as a referendum on whether the massive, industry-wide AI buildout is holding up. With Meta building its own chips, China's developers following suit, and Nvidia's reach still expanding, TSMC's numbers will be the clearest read yet on how much of this momentum is real. Six days out, the day's news gives Wall Street plenty to chew on.