Inside Samsung Electronics’ foundry division, the phone-chip orders that once set the calendar have given way to a different kind of customer. Meta Platforms is in talks with Samsung to place a custom AI chip order worth about 6.5 trillion won, roughly $6.5 billion, according to people familiar with the matter, the latest sign of how the artificial intelligence boom is remaking the contract manufacturing business.
The order would cover production of the next generation of Meta’s in-house accelerator, the Meta Training and Inference Accelerator, or MTIA, which the company designed to handle its AI workloads without relying entirely on Nvidia’s GPUs. Meta’s first two MTIA generations were manufactured by TSMC. Industry sources said the third generation is planned for Samsung’s most advanced process, with a production run that could reach several hundred thousand units.
The talks come as Samsung’s foundry business finds itself in an unfamiliar position: sold out. The company’s 4-nanometer capacity has been booked ahead of schedule, and its overall backlog of AI chip orders has grown past 50 trillion won, according to people familiar with the division’s operations. Samsung has introduced a quota system that prioritizes core clients, including Tesla and Anthropic, while new customers face waiting lists.
The shift is strategic as much as commercial. Samsung has spent years trying to close the manufacturing gap with TSMC, which dominates the market for advanced chips. Winning Meta’s business would validate Samsung’s newest process technology for large-scale AI production and give the company a second anchor customer in the custom silicon market, alongside Tesla’s automotive chips and the inference accelerators it builds for the startup Groq.
Meta’s interest in a second supplier is easy to read. The company is spending tens of billions of dollars a year on AI infrastructure and has said it wants to cut its dependence on Nvidia. MTIA is the vehicle for that ambition, and a single-source relationship with TSMC would simply replace one dependency with another. Samsung offers an alternative line, competitive pricing and the political comfort of a supply chain that does not run entirely through one foundry in Taiwan.
Anthropic is pursuing a similar path. The AI company, which has been in talks with Samsung over custom chips of its own, invested alongside Samsung in a relationship that industry executives describe as part of a broader realignment: hyperscalers and AI labs no longer accept that advanced chips come from one place. The result, analysts said, is that Samsung’s foundry business is being pulled from the smartphone supply chain into the center of the AI build-out.
The numbers help explain the enthusiasm in Seoul. A backlog above 50 trillion won, roughly $35 billion, would give Samsung’s foundry division years of visibility and support the turnaround that investors have been waiting for since the business lost ground to TSMC in advanced nodes. Samsung’s share price has responded to the AI order flow, and executives have begun talking publicly about foundry as a growth engine rather than a problem child.
For Meta, the deal is part of a wider push that includes a planned cloud service selling AI compute to outside customers and a target of building data centers with 5 gigawatts of capacity by 2030. Custom chips are central to that plan: Meta needs accelerators tuned to its own workloads, and it needs them in volume. Samsung, with its combination of memory, manufacturing and advanced packaging, is positioned to supply more of the stack than TSMC alone.
The talks are not finished, and people familiar with the discussions cautioned that pricing and capacity details remain unresolved. Samsung has been selective about which orders it accepts, and Meta has been known to negotiate hard. A deal of this size would also need to clear the usual regulatory and supply-chain reviews, though neither company has signaled obstacles.
Samsung’s strategy for winning this business rests on more than price. The company is one of the few manufacturers that combines foundry, memory and advanced packaging under one roof, and its high-bandwidth memory, or HBM, is already sold into the AI market. For a customer like Meta, buying accelerators and memory from the same supplier shortens the supply chain and simplifies the design process. Samsung executives have described this integration as the company’s main advantage against a pure-play foundry like TSMC.
The process road map supports the pitch. Samsung has said it is concentrating its advanced-node program on stabilizing its newest gate-all-around technology, deferring some planned expansions to get yields right before pushing volume. Industry analysts read that as a deliberate choice: rather than chase TSMC across every node, Samsung wants to win on a few process generations where its technology is competitive and its capacity is real.
Industry analysts said the deeper story is the diversification of the AI chip supply chain. A decade ago, foundry customers chose a partner and stayed; today, the largest buyers are deliberately splitting orders across suppliers to protect themselves against disruption and price pressure. Samsung’s 4-nanometer lines being full, and its backlog growing past 50 trillion won, are evidence that the strategy is working.
For Samsung, the Meta order would mark the completion of a long transition, from a component supplier for phones to a manufacturing platform for the AI economy. For Meta, it would be another step away from dependence on a single chip maker. Both companies are betting that the era of the one-stop foundry is over.
