Meta is set to start making the newest versions of its custom AI chips in September, a direct push to lower its dependence on expensive GPUs during a brutal component shortage. That’s according to TechCrunch AI, citing a Reuters report based on an internal Meta memo. At least one of the chips cleared testing in roughly six weeks, the memo said, which is fast for silicon of this kind.
Here’s what’s happening and why it lands hard on the AI industry.
What Meta is building
These chips come out of Meta’s MTIA program (Meta Training and Inference Accelerator), first detailed publicly in March. Meta designed them with Broadcom, but the actual manufacturing goes to TSMC. The company is also sourcing parts from across the supply chain:
- RAM from Samsung
- Storage from Sandisk
- Fiber-optic equipment from Sumitomo Electric
Meta is taking a modular approach, building each generation from chiplets so it can swap in new designs as AI workloads shift. In its own words, quoted by TechCrunch AI: “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.”
The plan is to point these chips at ranking and recommendation algorithms, broader AI workloads, and inference for Meta’s apps. Meta has quietly been making its own AI chips since 2023, so this isn’t a first attempt. It’s the next step in a program that’s finally hitting scale.
Why this matters
Compute is the single biggest cost in modern AI, and Nvidia has been the toll booth everyone pays. Meta wants to route around part of that toll. Custom chips let it handle specific jobs, like recommendation systems, without buying a premium GPU for every task.
The money involved is staggering. Meta said in April it expects capital expenditures between $125 billion and $145 billion this year, much of it aimed at AI. It plans to deploy 7 gigawatts of compute this year and double that next year, per the memo TechCrunch AI cites. To feed its Muse Spark model series, Meta has been signing data center and power deals worldwide, plus:
- A deal with ARM for compute on its recommendation systems
- A multibillion-dollar deal with AMD for Instinct GPUs
- A multibillion-dollar deal with Amazon to use its homegrown CPUs
What stands out here is that custom silicon doesn’t mean Meta stops buying from Nvidia and AMD. It expects to keep spending big with them too. The chips trim the bill at the margins and give Meta leverage. They don’t replace the GPU fleet.
The bigger shift
Meta is far from alone in trying to slow the flood of cash heading to Nvidia. The pattern is now industry-wide:
- OpenAI last month unveiled an inference processor it’s building with Broadcom
- Anthropic is reportedly weighing its own chips with Samsung
- Amazon and Google already build their own training and inference silicon
- A wave of startups is chasing the same demand
This is significant because it marks a real change in the status quo. For years, the answer to “how do we get more AI compute” was “buy more Nvidia.” Now every major lab and platform is hedging with custom silicon. Nvidia still sits at the center, but the biggest customers are quietly building exits.
What to watch
If production holds to the September timeline, expect Meta to lean harder on MTIA for internal workloads first, especially recommendations and inference, where custom chips pay off fastest. The six-week test cycle is worth noting on its own. Fast iteration is exactly what a modular, chiplet-based design is supposed to deliver, and it hints Meta can push new generations out quicker than traditional chip cycles allow.
For practitioners, the takeaway is simple. The compute layer is fragmenting. More custom accelerators means more hardware targets, more optimization work, and more reasons to avoid locking your stack to a single vendor’s tooling.
Meta declined to comment. More detail is available at the original TechCrunch AI report.