Meta's MTIA 450 chip goes to data centers in the first half of 2027
Meta says its own AI silicon is ready — and a Chinese driving-chip maker says the same about its scale.
Meta plans to put MTIA 450 — its third-generation custom accelerator, codenamed Arke — into data centers in the first half of 2027, with MTIA 500 (Astrid) finishing design work in about a month and reaching production data centers by the end of 2027, according to Bloomberg. Yee Jiun Song, the engineering vice president who runs Meta's custom silicon program, says each generation delivers better performance per watt and per dollar than buying vendor hardware, and that the chips run Meta's models more efficiently than current Nvidia offerings.
The proof point is a delivery, not a benchmark. Twelve MTIA 450 chips arrived from TSMC on September 1, with measured performance within 2% to 3% of Meta's simulations — and on day one the team was already running Meta's own models alongside DeepSeek's and Alibaba's. Meta has committed to more than a gigawatt of these chips over a 12-month window. The roadmap is aggressive by silicon standards: Meta's own engineering blog describes shipping a new chip roughly every six months, built from reusable chiplets that can be upgraded separately and dropped into the same chassis, rack, and network fabric.
The design choice worth understanding is what Meta optimized for. MTIA 450 doubles high-bandwidth memory throughput to 18.4 TB/s over the previous generation and adds hardware acceleration for attention operations — the parts of transformer inference that get brutally expensive as context windows grow. Meta is not chasing peak training FLOPS; it is buying down the cost of answering inference requests for billions of daily users. Song told Bloomberg that when you are building gigawatts of capacity, a potential 30% cost increase is unacceptable — the reason Meta killed its planned Olympus chip for 2028 or 2029 and focused on inference instead.
That framing lands in the middle of the industry's loudest argument this month. As frontier labs talk about pacing and the market debates a slowdown, Meta is quietly converting its Nvidia bill into owned capacity — the same vertical-integration move OpenAI made with Broadcom and Google made with TPUs years ago. Custom silicon is slow, expensive, and unforgiving; but at Meta's scale the savings compound, and the leverage over suppliers is the real prize.
Horizon Robotics put its 15 millionth driving chip in a Volkswagen. The Chinese chipmaker said cumulative Journey-series production passed 15 million units on September 3 — the first 10 million took five years, the next 5 million took 12 months. The milestone vehicle is the FAW-Volkswagen ID. AURA T6, launching September 20 with the Journey 6H chip and Horizon's HSD driving model, with HSD V2.1 due next; Horizon CEO Yu Kai was handed car 001.
The momentum is real; the profit is not yet. Journey shipments hit 2.218 million units in the first half of 2026, up 12.1% year over year, on revenue of RMB 2.055 billion (about $306 million), up 32.9%, at a 66% gross margin. But adjusted net loss widened 25.4% to RMB 1.671 billion (about $249 million), research spending swallowed more than 70% of revenue, and the company does not expect break-even until around 2028. Horizon claims more than 50% of China's self-brand ADAS chip market and 22.8% of urban NOA chips — a lead that in-house programs at BYD, Li Auto, XPeng, and NIO's spun-out Shenji are all built to erode.
What to watch: whether Meta's MTIA 450 benchmarks hold up once real traffic hits it in 2027 — and whether the second half of Meta's gigawatt commitment lands on schedule.
Should chip buyers celebrate when a hyperscaler designs its way out of Nvidia's pricing? Tell us in the comments.
Sources: Bloomberg · Meta AI Engineering · Investing.com · QbitAI · ChinaEVHome
