Meta just committed 1 gigawatt of custom AI chips with Broadcom through 2029

Meta and Broadcom announced a multibillion-dollar expansion of their custom AI chip partnership on April 14, extending through 2029. The initial commitment: over 1 gigawatt of computing capacity built on Broadcom-designed ASICs, with plans to scale to multiple gigawatts by 2027.

One gigawatt is enough to power roughly 750,000 average American homes. Meta is dedicating that to custom silicon alone, on top of its existing commitments to 6 GW of AMD GPUs, millions of Nvidia units, and Arm-designed processors. Mark Zuckerberg called it “the massive computing foundation we need to deliver personal superintelligence to billions of people.”

What we know

• Broadcom will co-develop multiple generations of Meta’s MTIA (Meta Training and Inference Accelerator) chips using its XPU platform. The scope covers chip design, advanced packaging, and networking.

• Meta has shipped four MTIA generations in two years, with two more in the pipeline:

  • MTIA 300 (in production): 200 GB/s HBM bandwidth, used for ranking and recommendation
  • MTIA 400 (deploying to data centers): 400% higher FP8 compute vs. the 300, 51% more HBM bandwidth
  • MTIA 450 (mass deployment early 2027): HBM bandwidth doubled from the 400, 75% more compute
  • MTIA 500 (mass deployment 2027): 50% more HBM bandwidth, 80% more capacity vs. the 450

• From the MTIA 300 to the 500, HBM bandwidth climbs 4.5x and compute FLOPS jump 25x.

• The next-generation MTIA chips will be the first custom AI silicon to use a 2-nanometer process (fabricated by TSMC), promising roughly 30% power reduction versus 3nm.

• Broadcom is also supplying Ethernet networking for Meta’s AI clusters: high-radix switches, optical connectivity, PCIe switches, and high-speed SerDes for rack-scale interconnects.

• The software stack is PyTorch-native (Torch FX IR, TorchInductor, Triton, MLIR, LLVM), with communication via Meta’s custom Hoot Collective Communications Library and a runtime built in Rust.

What we don’t know

• The exact dollar value of the deal. Multiple outlets describe it as “multibillion-dollar,” but Meta and Broadcom aren’t disclosing a specific figure.

• How MTIA performance compares head-to-head with Nvidia’s Blackwell or AMD’s MI400 on equivalent workloads.

• What happens to the Broadcom relationship if Meta’s parallel Nvidia commitments grow faster than expected.

The board shakeup

There’s a governance footnote worth reading. Broadcom CEO Hock Tan is stepping down from Meta’s board (where he’s sat since 2024) and transitioning to an advisory role focused on Meta’s custom silicon roadmap. The reason is straightforward: a sitting board seat alongside a multibillion-dollar silicon partnership creates a conflict-of-interest problem. He’ll still influence the chip strategy; he just won’t vote on it.

Tan is bullish. “This initial MTIA deployment is just the beginning of a sustained, multi-generation roadmap,” he said. This is Broadcom’s third major AI chip deal in three weeks, following agreements with Google and Anthropic.

What this means for you

If you’re a developer building on Meta’s infrastructure (PyTorch, LLAMA models, or Meta’s API ecosystem), this signals that Meta isn’t just buying GPU time anymore. It’s designing the hardware layer specifically for its workloads. The PyTorch-native software stack and Rust runtime suggest these chips are being optimized from the silicon up through the compiler to the framework you’re already using.

For the broader industry, the pattern is now hard to miss. Google has TPUs. Amazon has Trainium and Inferentia. Meta has MTIA. The era where Nvidia was the only game in town for AI compute is over. Nvidia’s GPUs will still dominate training for years, but inference (where most of these custom chips are aimed) is where the hyperscalers see the cost advantage in owning the full stack.