Meta is on track to begin mass production of its latest in-house AI chips this September, according to an internal memo cited by Reuters. At least one chip variant reportedly cleared its testing phase in roughly six weeks—which, if accurate, is a legitimately fast validation cycle for custom silicon. The company is designing the chips with Broadcom, fabbing them at TSMC, and sourcing RAM from Samsung, storage from Sandisk, and fiber-optic interconnects from Sumitomo Electric. It's basically a who's-who of the global semiconductor supply chain, which tells you something about how seriously Meta is taking this.
What Is MTIA, and Why Should You Care?
These chips fall under Meta's Meta Training and Inference Accelerator (MTIA) program, which the company has been running since 2023. In March, Meta publicly detailed four new chips under this umbrella—some already deployed, others rolling out this year or next. The design philosophy is modular: chiplets rather than monolithic dies, so Meta can swap in updated components as AI workloads evolve without scrapping the entire architecture. That's smart engineering, not just marketing. Monolithic chip design in a rapidly shifting AI landscape is like buying a fixed-rate mortgage right before rates drop.
"Each MTIA generation builds on the last, using modular chiplets, incorporating the latest AI workload insights and hardware technologies, and deploying on a shorter cadence." — Meta
The target workloads are Meta's ranking and recommendation algorithms (the stuff that decides what ends up in your feed), broader AI training runs, and inference for its consumer applications. These are high-throughput, latency-sensitive jobs that don't necessarily need the same floating-point horsepower as training a frontier language model—which is exactly where custom silicon starts making economic sense over general-purpose GPUs.
The Real Motivation: Nvidia Costs Are Eating Everybody Alive
Let's be direct about what's driving this. Meta has projected capital expenditures between $125 billion and $145 billion for 2026, a significant chunk of which is pointed squarely at AI infrastructure. When you're writing checks that large to Nvidia and AMD, the ROI calculus on building your own chips—even with the enormous upfront NRE (non-recurring engineering) costs—starts looking pretty compelling.
This doesn't mean Meta is walking away from Nvidia or AMD. The company still expects to spend heavily with both. Custom silicon and merchant GPUs aren't mutually exclusive; the strategy is to offload specific, well-understood workloads to cheaper in-house hardware while still leaning on Nvidia's H-series and AMD's Instinct GPUs for jobs that genuinely need that level of raw compute. It's a sensible hedge, not a declaration of war.
Meta Is Also Buying Compute From Basically Everyone Else
Meanwhile, Meta has been signing deals at a pace that suggests someone at Menlo Park has a very tired procurement team. The company struck an arrangement with ARM for recommendation system compute, cut a multibillion-dollar deal with AMD for Instinct GPUs, and—in one of the stranger moves of the past year—signed an agreement to use Amazon's homegrown CPUs for AI-related workloads. Meta also plans to deploy 7 gigawatts of compute capacity this year, with ambitions to double that figure in 2026, per the Reuters report. For context, 7 gigawatts is roughly the output of seven large nuclear reactors. The power demands of modern AI infrastructure are not subtle.
The Bigger Picture: Everyone Is Defecting from Nvidia's Monoculture
Meta isn't operating in isolation here. OpenAI recently unveiled its own inference processor, also co-designed with Broadcom. Anthropic is reportedly in discussions with Samsung about custom silicon. Amazon has its Trainium line. Google has its TPUs. The through-line is obvious: every major AI spender is trying to carve out workloads where they can escape Nvidia's pricing power.
The outcome isn't a world where Nvidia loses—Jensen Huang is not losing sleep over this. But it is a world where Nvidia gets squeezed on the more commoditized inference and recommendation workloads, while remaining dominant on frontier training runs where nothing else comes close. That's actually a more interesting competitive landscape than "Nvidia wins everything forever," even if it's also less dramatic than the "Nvidia killer" headlines suggest.
What to Watch For
September production start is the near-term milestone, but the real signal will come from Meta's subsequent earnings calls. Watch for any commentary on GPU procurement trends—if MTIA is genuinely moving the needle on unit economics, you'd expect to see external GPU spend flatten even as overall compute capacity grows. If both numbers keep climbing together, it means the custom chips are supplementing rather than substituting. Either outcome tells you something useful about where the economics of AI infrastructure are actually heading.
What is Meta's MTIA chip program?
MTIA stands for Meta Training and Inference Accelerator—a program Meta has run since 2023 to develop custom AI chips for its ranking, recommendation, and inference workloads, reducing reliance on third-party GPUs.
Who is manufacturing Meta's custom AI chips?
Meta is co-designing the chips with Broadcom and having them manufactured by TSMC (Taiwan Semiconductor Manufacturing Company).
Will Meta stop buying Nvidia GPUs?
No. Meta still expects to spend heavily with Nvidia and AMD. The MTIA chips are designed to handle specific workloads like recommendations and inference, not replace Nvidia for frontier model training.
How much is Meta spending on AI infrastructure in 2026?
Meta has projected capital expenditures of $125 billion to $145 billion in 2026, with a large portion directed at AI compute, data centers, and power infrastructure.
Dispatch desk
