Meta just locked in a multi-generation partnership with Broadcom to build custom AI accelerators. Not a small bet - they're committing to over 1 gigawatt of custom silicon. That's the power consumption of a small city, dedicated entirely to running inference, recommendations, and generative AI across Facebook, Instagram, WhatsApp, and the rest of Meta's platform.
The chips are called MTIA - Meta Training and Inference Accelerators. This isn't Meta's first swing at custom silicon, but it's the first time they've publicly committed to multiple generations with a manufacturing partner. That signals intent. They're not experimenting anymore. They're building infrastructure for the long run.
Why Build Your Own Chips?
The obvious question: why not just buy GPUs like everyone else?
Because Meta's workloads aren't like everyone else's. Most of their AI isn't training massive models - it's running billions of inference requests per day. Recommending posts. Ranking feeds. Translating messages. Moderating content. These tasks don't need the raw horsepower of an H100. They need efficiency at scale.
Custom silicon lets you optimise for exactly that. You strip out everything you don't need and double down on what you do. Broadcom brings the manufacturing expertise. Meta brings the workload data. Together, they're building chips that do one thing extremely well: run Meta's specific models, faster and cheaper than general-purpose hardware could.
The economics matter here. When you're running inference at Meta's scale, even small efficiency gains compound into massive cost savings. A 10% improvement in power efficiency across 1GW of compute saves tens of millions of dollars a year. That's not theoretical - that's why they're doing this.
The Bigger Pattern
Meta isn't alone in this. Google built TPUs. Amazon built Trainium and Inferentia. Microsoft is working with AMD on custom chips. Apple's been doing custom silicon for years. The hyperscalers have all reached the same conclusion: when you're operating at this scale, buying off-the-shelf hardware is leaving money on the table.
What's interesting is the timeline. Meta's MTIA project started quietly a few years ago. Now they're talking about multiple generations, public partnerships, and gigawatt-scale deployments. That suggests they've crossed a threshold. The chips work. The economics make sense. Now it's about scaling production.
For NVIDIA, this is the long-term headwind nobody's talking about loudly yet. Yes, demand for GPUs is still insane. But the biggest customers - the ones who can afford to build their own chips - are systematically moving workloads off general-purpose hardware and onto custom accelerators. Not overnight. But steadily. Generation by generation.
What This Means for Builders
If you're running a startup, you're not building custom chips anytime soon. But the principle still applies: optimise for your workload. Don't assume the default tool is the right tool just because everyone else uses it.
Meta's bet on custom silicon is about control. Control over cost, performance, and supply chain. For most businesses, that control comes from choosing the right cloud provider, the right model size, the right deployment strategy. The companies winning in AI aren't necessarily the ones with the biggest models - they're the ones with the tightest fit between their tools and their actual needs.
Meta's partnership with Broadcom is a signal. The AI infrastructure landscape is fragmenting. General-purpose chips will still exist, but the most demanding workloads are migrating to purpose-built hardware. That changes who wins, who loses, and where the money flows over the next decade.
1GW of custom accelerators. That's not a side project. That's a new category of infrastructure. And Meta's not building it to rent it out - they're building it because at their scale, it's the only thing that makes sense.