Meta Picks Samsung Foundry's 2nm to Build MTIA in $6.5B Deal

Meta will fabricate its next-generation MTIA AI accelerator at Samsung Foundry on 2nm in an order worth over $6.5 billion, breaking a two-generation TSMC streak.

Meta Picks Samsung Foundry's 2nm to Build MTIA in $6.5B Deal

Meta has tapped Samsung Foundry's 2nm process to produce the third generation of its custom AI accelerator, MTIA, in a deal industry sources value at more than 10 trillion Korean won (about $6.5 billion) — the biggest single win yet for Samsung's contract chip business.

MTIA T-Series Moves Off TSMC

Meta will fabricate MTIA at Samsung Foundry starting from the third generation, breaking a two-generation streak at TSMC. Meta and Samsung's System LSI division have set up a joint architecture team from day one to hit an aggressive six-month iteration cadence for the new chip, according to Seoul Economic Daily reporting on July 4. Meta plans to mass-produce "hundreds of thousands" of MTIA T-series ASICs to feed ranking, recommendation and generative-AI workloads across its data centers.

Why Samsung, Why Now

Two forces pushed Meta to hedge its foundry supply. TSMC's leading-edge capacity is booked out through 2027, and Samsung's 2nm quotes reportedly come in about 30% cheaper than TSMC's for the same node. That trade — modestly lower yields today for guaranteed capacity and lower unit cost — mirrors decisions Anthropic is now weighing for its own ASIC program, which Samsung also plans to build on 2nm. The MTIA and Anthropic wins push Samsung Foundry's medium-term order backlog toward 50 trillion won and give the division a credible path back to profit by Q3 2026.

Meta and Samsung Foundry 2nm MTIA production briefing

The Bigger Silicon Realignment

Meta already leans on external partners for compute silicon — from Nvidia GPUs to Qualcomm's Dragonfly C1000 CPU — but MTIA is where Chief AI Officer Alexandr Wang's Superintelligence Labs is trying to cut its dependence on merchant silicon. Rising demand for accelerators that can train models like Meta's in-training Watermelon frontier model and serve Muse Spark inference at scale is turning custom ASICs into a strategic asset. Meanwhile, TSMC's dominance is starting to fracture: after signing Sony on next-gen AI image sensors, the Taiwanese foundry is racing to add capacity, but hyperscalers are diversifying.

Neither Company Confirms

Meta and Samsung have not officially confirmed the deal — a Samsung spokesperson said "nothing has been decided" — but the negotiation is far enough along that Samsung's System LSI team is co-designing the chip. If finalized, the MTIA T-series would be Samsung's first flagship 2nm workload from a US hyperscaler and validate the process ahead of Tesla's own 2nm Dojo work also headed to Samsung's Taylor, Texas fab.

Reporting based on coverage from Seoul Economic Daily, Bloomberg, Benzinga, Seeking Alpha and Tom's Hardware.

Category: Partnerships

Tags: Physical AI artificial intelligence Partnership Semiconductors AI Infrastructure

Related Articles