# Meta Bets on In-House AI Chips to Tame Soaring Compute Costs > Source: > Published: 2026-09-16 00:16:21+00:00 Meta Platforms (NASDAQ:META) is preparing to deploy a new generation of internally designed AI chips in its data centers during the first half of 2027, part of a broader push to rein in the ballooning energy and computing costs tied to its artificial intelligence expansion. The move matters now because AI infrastructure spending has become one of the biggest swing factors in Big Tech earnings, and Meta is signaling it wants more control over that cost line rather than depending entirely on outside chipmakers. The company is currently testing its third-generation processor, the MTIA 450, code-named Arke. Twelve units delivered by Taiwan Semiconductor Manufacturing Co. on September 1 performed within 2% to 3% of Meta’s internal simulations, an early sign that the design is tracking close to expectations. Meta’s next chip, the MTIA 500 or Astrid, is expected to finish design work within roughly a month and reach data centers by the end of 2027, suggesting the company plans to iterate on custom silicon on a fairly tight annual cycle rather than treating this as a one-time hardware refresh. ### Why Meta is building its own chips Meta’s custom silicon program, known as MTIA, is designed to reduce the company’s reliance on third-party AI accelerators while improving efficiency on two fronts that directly affect Meta’s bottom line: power draw and cost per unit of compute. Meta engineering vice president Yee Jiun Song said each generation “takes on a little bit more risk technologically and gets us better performance,” pointing specifically to gains in performance per watt and per dollar. #### Partnerships behind the hardware Meta is not building these chips alone. The company is working with Broadcom (AVGO) on chip design and with TSMC on manufacturing, a structure similar to how other hyperscalers have approached custom silicon rather than competing head-on with established GPU makers. #### Scale of the rollout The deployment plan is large by any measure. Meta has committed to installing more than a gigawatt of MTIA chips over a 12-month period, with the pace expected to pick up further if demand for AI compute stays strong. | Chip | Code Name | Status | Expected Milestone | |---|---|---|---| | MTIA 450 | Arke | In testing | Data center deployment starting H1 2027 | | MTIA 500 | Astrid | In design | Design completion in about a month; data centers by end of 2027 | ### A narrower focus: inference over training Meta has also scaled back the scope of the project. The company canceled its planned Olympus processor, which had been intended to handle both AI training and inference, partly due to cost concerns. Instead, Meta is concentrating its custom chip efforts on inference, the ongoing process of running trained AI models rather than building them. Song described the current generation as the workhorse chips Meta plans to use for general-purpose inference. The Arke chips have already run Meta’s own models alongside models from DeepSeek and Alibaba, an early indication of the hardware’s flexibility across different AI systems. ##### Why inference matters more right now Inference costs scale with usage, meaning every additional AI-powered feature Meta ships, whether in ads, recommendations or assistants, adds ongoing compute expense. Training a model is a large but finite cost, while inference is a recurring one, which is likely why Meta has prioritized it in this phase of the MTIA roadmap. ### What comes next for investors The central question for Meta investors is whether custom silicon can meaningfully bend the company’s AI cost curve. If MTIA deployment lowers power consumption and inference expense at scale, Meta could protect its margins while continuing to expand AI features across its products. Analysts and investors are likely to watch several signals in the coming quarters: how quickly Meta actually deploys the gigawatt-scale rollout, whether performance-per-dollar gains materialize as described, how fast inference demand grows across Meta’s platforms, and whether the company extends custom silicon beyond the current MTIA roadmap. The potential upside is straightforward, since even modest efficiency improvements could translate into meaningful savings at Meta’s scale of operation. The risk lies in execution. Designing competitive AI chips requires substantial capital, multi-year development cycles, and tight coordination with external partners like Broadcom and TSMC, any of which could slip. For now, Meta’s message to the market is that custom AI hardware is becoming a core pillar of its long-term cost strategy, not a side experiment. *Disclaimer: This content was partially produced with the help of AI tools.*