Meta is preparing to deploy a new generation of internally designed artificial-intelligence processors as it seeks tighter control over the soaring cost and energy demands of its AI infrastructure. The company is testing its MTIA 450 chip, code-named Arke, with deployment planned for the first half of 2027, while the more broadly deployed MTIA 500, code-named Astrid, is expected to enter data centers by the end of 2027. Working with Broadcom on design and TSMC on manufacturing, Meta is betting that purpose-built processors can deliver better performance per watt and per dollar than general-purpose AI hardware. The strategy also reduces dependence on Nvidia while giving Meta greater control over an increasingly critical component of its business. With more than a gigawatt of custom-chip capacity planned over a 12-month period, the effort represents a significant shift toward vertical integration as AI infrastructure becomes one of technology’s largest capital expenses.
Key Takeaways
- Meta plans to begin deploying MTIA 450, or Arke, during the first half of 2027, followed by the more widely used MTIA 500, or Astrid, by the end of that year.
- The company is emphasizing specialized inference processors because designing chips specifically around its workloads can improve performance per watt and per dollar while avoiding the additional cost of hardware engineered simultaneously for training and inference.
- Meta’s partnership with Broadcom and TSMC gives the company an alternative to relying exclusively on Nvidia and AMD, with more than one gigawatt of custom silicon capacity already planned and additional expansion expected.
In-Depth
Meta’s latest custom AI processors illustrate an increasingly important reality of the artificial-intelligence boom: computing economics may ultimately matter almost as much as raw processing power. As AI deployment expands across billions of daily interactions, even modest improvements in electricity consumption and hardware efficiency can translate into enormous savings.
The MTIA 450, code-named Arke, is being tested for deployment during the first half of 2027. Its successor, MTIA 500, or Astrid, is expected by the end of 2027. Meta says early Arke hardware performed within roughly 2% to 3% of engineering simulations, an encouraging result as production moves toward scale.
The strategy is deliberately focused on inference—the process of running already-developed AI models—rather than attempting to make every processor handle every possible workload. Meta canceled a planned chip called Olympus, which would have combined training and inference capabilities, after concluding that such versatility could raise costs substantially.
That decision reflects some old-fashioned economic discipline inside an extraordinarily expensive technological race. When infrastructure is measured in gigawatts, unnecessary capability becomes expensive overhead.
Meta is not abandoning outside suppliers. Nvidia and AMD remain important components of its infrastructure. Instead, the company is building an alternative alongside them. Working with Broadcom on chip development and TSMC on manufacturing allows Meta to specialize hardware around its own software requirements.
If the approach scales successfully, custom silicon could give major technology companies greater bargaining power, lower operating expenses, and reduce dependence on any single semiconductor supplier.
Sources
- https://www.reuters.com/world/asia-pacific/meta-unveils-plans-batch-in-house-ai-chips-2026-03-11/
- https://www.reuters.com/world/asia-pacific/meta-put-ai-chip-into-production-september-it-looks-double-computing-capacity-2026-07-09/
- https://about.fb.com/news/2026/03/expanding-metas-custom-silicon-to-power-our-ai-workloads/
- https://investors.broadcom.com/news-releases/news-release-details/broadcom-announces-extended-partnership-meta-deploy-technology