Nvidia’s AI crown may be slipping as Meta and Broadcom race to build custom silicon that could make GPUs yesterday’s news.
The first tangible blow lands in early 2027, when Meta begins deploying three new inference chips—the MTIA 400, 450, and 500—each engineered to run generative-AI models inside its data centers. A fourth chip, the MTIA 300, is already live and handling ranking and recommendation workloads for Facebook and Instagram feeds.
What makes the timetable dangerous for Nvidia is cadence, not raw transistor count. Meta’s silicon team uses modular chiplets that can be swapped, tested, and relaunched every six months, a pace that compresses the traditional two-year foundry cycle by 75 percent.
"Rather than placing a bet and waiting for a long period of time, we deliberately take an iterative approach: 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. This tighter loop keeps our hardware better aligned with evolving models while enabling faster adoption of new technology."
Broadcom handles manufacturing and advanced packaging for portions of the lineup and counts Meta among its five major XPU customers, confirming that the social-media giant is now a serious semiconductor buyer with foundry pull.
The strategic shift is subtle but seismic. Training large language models still benefits from Nvidia’s GPU scale, yet most day-to-day AI compute inside Meta is inference—lighter matrix math that custom XPUs can execute at lower power and, crucially, without paying Nvidia’s margin.
Look, if Meta can prove that six-month refresh cycles translate into measurable CapEx savings, every hyperscaler with internal chip teams will accelerate their own roadmaps, and Nvidia’s data-center growth story suddenly has a timing problem, not just a pricing problem.
⚠️ LEGAL DISCLAIMER: This article is for informational purposes only and does not constitute financial or investment advice.
Source: Yahoo