Meta's homegrown AI chips head to production this September
Meta will start manufacturing its own modular AI accelerators in September, chipping away at Nvidia's grip. For Indian developers renting GPUs, the ripples could reach INR bills.
The News
Meta will begin manufacturing its own custom artificial-intelligence accelerators in September, moving a programme first detailed in March 2026 from the drawing board to the factory floor. The chips belong to the company's Meta Training and Inference Accelerator (MTIA) line, built to run the ranking and recommendation models that power Facebook and Instagram, alongside heavier inference workloads.
According to TechCrunch, which reported the timeline, at least one of the new chips has been through roughly six weeks of testing. The defining design choice is modularity: each MTIA generation is assembled from reusable chiplets, so Meta can reshuffle components as model architectures shift rather than redesigning silicon from scratch. Meta declined to comment.
The supply chain reads like a who's who of the semiconductor world. Broadcom is handling chip design, TSMC the manufacturing, with Samsung supplying memory, Sandisk providing storage and Sumitomo Electric contributing fibre-optic equipment. The effort sits inside a colossal spending plan: Meta has guided 2026 capital expenditure to between 125 billion dollars and 145 billion dollars, and intends to bring 7 gigawatts of compute online this year, doubling that next year.
Why It Matters
The pitch is straightforward. Every accelerator Meta builds itself is one it need not buy at a premium from Nvidia or AMD during a chronic shortage, even as it keeps those relationships alive as a hedge. Owning the silicon also lets Meta tune performance-per-watt for its own narrow set of jobs, which is exactly where recommendation engines and inference reward specialisation.
This is not a lone experiment. TechCrunch notes that OpenAI is developing an inference chip with Broadcom, Anthropic has been in discussions with Samsung, and Amazon and Google already run proprietary processors at scale. The clearest precedent is Google's Tensor Processing Unit, unveiled back in 2016, which quietly grew into a genuine alternative to buying merchant GPUs. The lesson from that decade-long bet is that custom silicon rarely displaces Nvidia outright; it caps the bill and buys leverage. Meta is now making the same wager with a far larger chequebook.
The modular-chiplet approach is the interesting wrinkle. Rather than committing to one monolithic design that could be obsolete before it ships, Meta is treating hardware more like software, swapping blocks as the workload evolves. If it works, it shortens the punishing gap between chip conception and deployment.
Indian Angle
For India, the story is really about the price of compute. Homegrown model builders such as Sarvam and Ola-backed Krutrim rent almost all their training and inference capacity from hyperscale clouds, and those bills are quoted in dollars. When Meta, Google and Amazon shave their own silicon costs, the savings eventually filter into cloud pricing, easing a burden that lands hard on rupee-denominated startups competing on thin margins.
There is also a strategic signal for policymakers. Krutrim has already talked up its own Bodhi chip ambitions, and the India Semiconductor Mission under MeitY has staked billions on domestic fabrication. Meta's chiplet route is a reminder that the near-term win is not building a leading-edge fab overnight but mastering design and packaging, an area where Indian engineering talent already runs deep inside firms like Broadcom and Nvidia.
Finally, the capex numbers reset expectations for Indian enterprises budgeting AI projects. If a single company can commit up to 145 billion dollars a year, Indian CIOs planning inference-heavy deployments should assume compute costs stay volatile and design for portability across clouds rather than locking into one vendor.
FAQ
When does production begin?
Meta plans to start manufacturing the new MTIA chips in September 2026, following roughly six weeks of testing on at least one design. The programme was first detailed in March 2026.
Is Meta dropping Nvidia?
No. The goal is to reduce reliance on Nvidia and AMD amid a GPU shortage, not replace them. Meta is expected to keep buying merchant GPUs while adding its own chips for specific workloads.
What does this mean for Indian AI startups?
Cheaper in-house silicon at the hyperscalers should, over time, translate into lower cloud pricing, easing dollar-denominated compute bills for Indian model builders such as Sarvam and Krutrim.
Where can I read the original coverage?
The reporting was published by TechCrunch on 9 July 2026, with the full detail linked below.
This story was reported by TechCrunch. Read the full original coverage at TechCrunch.