Artificial intelligence (AI) semiconductor demand is reshaping the global foundry landscape at an unprecedented pace. Samsung Electronics’ foundry division has shifted from scrambling for orders to being able to pick and choose. Its most advanced 4-nanometer process capacity is not only essentially sold out for this year, but its schedule for next year is already fully booked, with some 8nm lines also nearing full utilization. Behind this supply-demand reversal is explosive growth in AI server chip orders, pushing Samsung’s mid-to-long-term foundry order backlog to roughly 50 trillion won (approximately $35 billion).

Faced with this flood of orders, Samsung has activated a quota mechanism, concentrating limited advanced capacity on core clients such as Tesla, Meta, and Anthropic, while new customers face wait times. This strategic adjustment signals Samsung’s attempt to re-establish itself as a critical player in the advanced process market dominated by TSMC (2330.TW).

Full Capacity Triggers Quota System as AI Chips Drive Demand

According to market sources, Samsung Electronics’ foundry division has formally implemented a quota mechanism for certain process nodes. This means that due to surging AI semiconductor demand and a spike in orders from global big tech firms, Samsung is compelled to allocate capacity preferentially to existing core clients and high-priority processes, while selectively accepting orders from new customers.

Industry analysts note that the structural shift in demand is particularly pronounced. The demand center for advanced manufacturing processes has rapidly pivoted from smartphone application processors, which previously dominated the market, to AI accelerator chips, application-specific integrated circuits (ASICs), and high-performance computing (HPC) chips. Samsung’s foundry is currently producing autonomous driving chips for Tesla, AI inference chips for startup Groq, and is also expanding its collaboration with giants like Nvidia and Google.

A South Korean industry official revealed: “Starting with Tesla’s AI chip orders last year, orders for AI server semiconductors at Samsung’s foundry have entered a phase of rapid growth. The current order backlog stands at approximately 50 trillion won, and the business is expected to achieve an operating profit in the fourth quarter of this year.” The concentrated influx of orders allows the fab to focus its production lines on a few large-scale projects, which is far more operationally efficient than mass-producing a wide variety of different product types.

Big Tech Shifts Orders, Meta and Anthropic Bet on 2nm

The persistent supply shortage of TSMC’s advanced processes has opened a strategic window for Samsung. Clients who previously shifted to TSMC due to yield concerns are now re-evaluating Samsung, or adding it as a second source to diversify supply chain risk and enhance their bargaining power. The latest examples come from Meta and Anthropic.

Reports indicate Meta is in talks with Samsung’s foundry division for a next-generation ASIC design and production partnership valued at over 10 trillion won (approximately $7 billion). The first two generations of its custom AI accelerator, “MTIA,” were both manufactured by TSMC, but starting with the third generation launched this year, Meta has locked in Samsung as a core manufacturing partner. The plan is to use Samsung’s most advanced 2-nanometer process for a production run of hundreds of thousands of wafers. Meta is exploring a cloud service business to rent AI computing power to external enterprises and has set a target of building data centers totaling 5 gigawatts of capacity by 2030, making its need for custom chips extremely urgent.

Meanwhile, U.S. AI powerhouse Anthropic is also evaluating the use of Samsung’s 2nm process to develop custom AI chips. The super-unicorn, valued at $96.5 billion, is attempting to reduce its reliance on Nvidia GPUs and Google TPUs through its own chip designs. In May, Samsung participated as a strategic investor in Anthropic’s $65 billion Series H funding round, a capital tie that adds significant leverage in securing foundry orders. Anthropic’s long-term total investment in AI data centers is estimated at roughly $50 billion, about half of which will flow to AI semiconductor procurement. With its integrated capabilities in memory, foundry, and advanced packaging, Samsung is viewed as the largest potential beneficiary of this massive order.

From “Using Chips” to “Making Chips,” the AI Race Enters the Full-Stack Era

Anthropic’s chip-making ambitions are not an isolated case but a microcosm of a paradigm shift in the AI industry. On June 24, its competitor OpenAI, in partnership with Broadcom, unveiled its first custom AI inference chip, “Jalapeño.” This ASIC, designed specifically for large language model inference, went from design to tape-out in just nine months and is expected to reduce inference costs by about 50%. Google has its TPU, Amazon has Trainium, Microsoft has Maia, and Meta is accelerating the iteration of its MTIA. AI model developers are extending competition from model capabilities down to the underlying hardware infrastructure, seeking to control computing costs and supply chain lifelines through custom silicon.

To this end, Anthropic has begun recruiting, hiring Clive Chan, an early member of OpenAI’s custom chip team. However, the company is cautious in its public statements, emphasizing that future computing expansion will still primarily rely on AWS’s Trainium, Google’s TPUs, and Nvidia’s GPUs. Its custom chip project remains in the early planning stages, with specific functions, performance metrics, and server adaptation plans yet to be determined, leaving open the possibility it could ultimately be shelved.

For Samsung, this AI “chip-making” race represents a historic opportunity to catch up with TSMC. TSMC’s 2nm production schedule is reportedly extended through 2028 to 2029, while Samsung’s fab in Taylor, Texas, is expected to begin mass production by 2027, offering an integrated package of HBM4 memory and advanced packaging. If Samsung can successfully secure 2nm orders from Anthropic and Meta, it will gain crucial endorsements from heavyweight clients, paving the way for capturing more AI chip orders in the future.

However, challenges remain severe. Samsung’s advanced process yields still need to pass the ultimate market test. In its 1.4nm process roadmap, Samsung has made a strategic trade-off, postponing mass production from 2027 to 2029 to prioritize stabilizing its 2nm technology. This pragmatic strategy of “securing a foothold before leaping forward” sets the stage for the next phase of its process competition with TSMC and Intel.