Why Meta Walked Away From Samsung for Custom AI Chips

In a significant strategic shift, Meta has suspended its custom chip development project with Samsung Electronics. The move, reported in early June 2026, was made at Meta's own request . While the exact reason remains officially undisclosed, the decision is not an isolated incident. It follows Meta's earlier cancellation of its most advanced in-house AI training chip, known as "Iris," due to design difficulties, forcing a pivot to a simpler design . Industry reports point to "strategic differences" between the partners as a key factor, suggesting a fundamental misalignment in direction .

This pause is not a retreat from custom silicon. Rather, it's a powerful consolidation of Meta's chip strategy around two proven partners: Broadcom for design and TSMC for manufacturing. This decision is made sharper by Samsung's well-documented struggles to compete with TSMC in the crucial advanced foundry market needed for cutting-edge AI chips.

A Strategic Pause, Not a Retreat

The suspension of the Samsung project fits a larger pattern of Meta decisively focusing its resources. Earlier in 2026, the company scrapped its "Olympus" AI accelerator project due to high execution risks in developing custom silicon . These setbacks haven't slowed Meta's ambition; they've redirected it. By pausing work with Samsung's System LSI division, Meta is choosing to avoid the execution risks associated with a partner that has faced persistent technology challenges, and is instead directing its multi-billion-dollar investment toward a tighter, more reliable ecosystem.

The Broadcom-TSMC Axis: Meta's Double Down

Meta's future in custom AI chips is now firmly anchored to two key partners. The relationship with Broadcom, already a primary ASIC design partner, was massively expanded in April 2026. The new multi-year, multi-generational deal extends through at least 2029 and has an initial commitment exceeding $2.3 billion, with Broadcom supplying up to 1 GW of chip capacity . Broadcom's CEO, Hock Tan, publicly confirmed the strength of the partnership on a Q1 2026 earnings call, stating, "Meta's custom accelerator MTIA road map is alive and well. We're shipping now" .

This design work is being realized through an aggressive manufacturing roadmap with TSMC. In March 2026, Meta announced it would deploy four new generations of its MTIA (Meta Training and Inference Accelerator) chips—the MTIA 300, 400, 450, and 500—by the end of 2027 . This blistering six-month cadence is up to four times faster than typical industry chip cycles . The MTIA 300 is already in mass production, and the MTIA 400 is completing testing ahead of data center deployment . Crucially, these chips are being manufactured by TSMC, not Samsung . This confirms that Meta's earlier 2024 exploratory talks with Samsung Foundry were reversed as the MTIA roadmap, built on a foundation of execution certainty, accelerated.

Samsung's Foundry Gap: Why Meta Chose Certainty

Samsung's inability to secure this business stems from a well-documented technology and execution gap with TSMC, which has made it a riskier partner for critical AI silicon.

The Yield Problem and Technology Gap

The core of Samsung's struggle is poor production yield on its advanced 3-nanometer (nm) and sub-3nm nodes, which has made it difficult to attract and retain major customers . Samsung was the first to start 3nm GAA (Gate-All-Around) production, but lower-than-expected yields caused clients like Google and Qualcomm to gravitate toward TSMC instead . Even as Samsung targets a 70% yield on its next-generation 2nm process, TSMC continues to widen its technological and market lead .

A Market Dominated by TSMC

TSMC's dominance is near-absolute. The company controls over 70% of the global foundry market, while Samsung's share has fallen to around 7-8% . TSMC's advanced processes are "sold out" through 2028, with Apple and Nvidia reserving nearly all capacity for 3nm and upcoming 2nm nodes . This has created a seller's market where TSMC can generate enormous AI-chip-driven profits, far surpassing Samsung's foundry results, while reinvesting over five times more capital into its foundry operations than Samsung .

Fighting for Scraps

The capacity crunch at TSMC would normally be a gift to Samsung, but the yield issues have prevented it from fully capitalizing. Samsung has been forced into aggressive tactics, such as bundling its high-bandwidth memory (HBM) products with foundry services at preferential terms to try and lure clients like MediaTek away from TSMC . While Samsung has scored some wins—such as producing Tesla's next-generation AI6 chip, a potential 2nm deal with AMD, and Nvidia's "Groq 3" LPU order—these are not enough to bridge the credibility gap with the largest AI hyperscalers like Meta that require massive, reliable output on the most advanced nodes .

The Big Picture: A Chain of Interlocking Decisions

Meta's decision to pause its Samsung project is a direct consequence of these interlocking forces:

• Execution Certainty: Meta needs a massive, rapid, and reliable supply of advanced AI silicon to support billions of daily AI interactions across its platforms. Samsung's yield problems introduced unacceptable risk.
• Design- Manufacturing Synergy: By concentrating on Broadcom for design and TSMC for fabrication, Meta leverages a deeply integrated partnership responsible for co-developing the industry's first 2nm AI accelerator as part of its MTIA 500 chip .
• Avoiding Distraction: Running a parallel, less-certain development path with Samsung's System LSI division would divide focus and resources.

Meta’s move is a clear-eyed, risk-management decision by a company spending tens of billions on AI infrastructure. It is not a commentary on the end of Meta’s chip ambitions but a proof point of how seriously it takes them. The company is betting its AI future on the partners that can deliver with the speed and certainty its roadmap demands, leaving Samsung on the sidelines of one of the industry's most consequential custom chip programs.