Meta Scraps Advanced AI Chip, Loses Key Talent
Meta has reportedly scrapped an advanced in-house AI training chip after hitting design roadblocks, highlighting the immense difficulty of competing in custom silicon. The setback is compounded by talent migration, with two founding members of AI startup Thinking Machines Lab departing for Meta, suggesting a strategy shift towards acquiring talent rather than just building hardware.
This isn't Meta's first attempt at a custom AI accelerator to hit a wall. The now-scrapped advanced training chip, codenamed "Olympus," follows a previously canceled second-generation chip known as "Iris." These efforts are part of the broader "Meta Training and Inference Accelerator" (MTIA) program, which has struggled to produce a viable in-house alternative for high-performance AI training. The Olympus chip was deemed too risky by executives who feared its complex design and potential software instability could put them further behind competitors like Google and OpenAI. The design, which was to be led by the team from Rivos, a startup Meta acquired, aimed for compatibility with Nvidia's dominant CUDA software but proved too difficult to manufacture at scale. In response to the internal setbacks, Meta is aggressively turning to external suppliers. The company has signed a multi-billion dollar, multi-year deal to rent Google's Tensor Processing Units (TPUs) and has also announced major partnerships with both Nvidia and AMD to power its next-generation AI infrastructure. This pivot represents a massive capital expenditure, with Meta committing up to $135 billion for its 2026 AI infrastructure build-out. The talent acquisition from Thinking Machines Lab involves Christian Gibson, a former OpenAI engineer who worked on the original ChatGPT and specializes in AI training supercomputers, and Noah Shpak, an AI engineer from Character.AI and X. Their hiring signals a focus on acquiring proven expertise in building and scaling large-scale AI systems, rather than just chip design. This is part of a larger pattern of Meta poaching top talent from the high-profile startup. Another co-founder, Andrew Tulloch, joined Meta in 2025, bringing the total number of founding members who have departed Thinking Machines Lab for Meta to four. The startup, led by former OpenAI CTO Mira Murati, raised $2 billion at a $12 billion valuation last year but has consistently faced poaching from larger tech firms.
