From TSMC to Samsung, Why Is Meta’s AI Chip Foundry Route Shifting?

TradingKey - On July 3, Eastern Time, according to industry sources, Meta ( META) is in talks with Samsung for a custom AI chip foundry order valued at over 10 trillion Korean won, planning to mass-produce hundreds of thousands of sets of its self-developed AI accelerator, MTIA, using a 2-nanometer process, which covers two generations of products: MTIA 450 and MTIA 500.

The first two generations of MTIA chips were both manufactured by TSMC ( TSM ). The reason for shifting orders to Samsung this time is not just a shortage of production capacity.

Why is Meta considering outsourcing its chip manufacturing to Samsung?

The main reason for considering a shift in orders to Samsung is that TSMC's production capacity is extremely tight and fully booked through 2027, and Meta simply cannot afford to wait. In the initial stage of mass production, monthly capacity is about 35,000 wafers, which is expected to rise to 140,000 wafers by the end of the year. Apple ( AAPL ), Nvidia ( NVDA ), AMD ( AMD ), Qualcomm ( QCOM) are all competing for capacity, and the waitlist has already extended beyond 2027. In 2026, the combined AI-related capital expenditures of tech giants Microsoft ( MSFT ), Google ( GOOGL ), Amazon ( AMZN ), and Meta are projected to exceed $700 billion, with nearly 70% flowing into AI servers and customized computing chips.

Meta's chip iteration pace is also accelerating, compressing the industry's usual one-to-two-year generation cycle down to six months. The roadmap released in March this year shows that MTIA 300 is already in production, MTIA 400 is being deployed in data centers, MTIA 450 is planned for large-scale launch in early 2027, and MTIA 500 will follow closely behind.

The performance improvements are equally aggressive. Generation by generation: the FP8 computing power of MTIA 400 is five times that of MTIA 300 (a 400% increase), and the bandwidth is increased by 51%; MTIA 450's bandwidth doubles to 18.4TB/s; the MX4 computing power of MTIA 500 reaches 30 PFlops. In less than two years, from MTIA 300 to MTIA 500, the bandwidth increased from 6.1TB/s to 27.6TB/s (an increase of about 4.5 times).

With a pace of one generation every six months, a single foundry simply cannot handle it. Meta's capital expenditure for this year has been revised upward to $125 billion to $145 billion, with over 60% allocated to AI servers and self-developed chips. If any issues arise with the foundry, tens of billions of dollars would be tied up. Finding Samsung as a second supplier is an inevitable choice.

There are also industry rumors suggesting that Meta may have established a deeper cooperative framework with Samsung's System LSI division, where Samsung might participate in joint development starting from the design stage. However, this claim has not yet been confirmed by either party.

Samsung's 2nm Yield Rate Soars from 20% to 60% in Six Months

While Samsung Foundry has historically lagged behind TSMC, the gap is narrowing rapidly at the 2-nanometer node.

In the second half of 2025, Samsung's 2nm yield was only about 20%, but it has surpassed 60% by the first quarter of 2026, climbing 40 percentage points in half a year. However, this yield rate has still not reached Qualcomm's 70% threshold for stable mass production, and Qualcomm has already decided to have TSMC exclusively manufacture its next-generation flagship chip. For Meta, a 60% yield is acceptable for initial mass production, but a clear gap remains compared to the requirements of top-tier customers.

Samsung's Taylor plant in the U.S. is expected to start 2nm mass production in the second half of 2026, with a monthly capacity target of 50,000 wafers, making it more convenient for local customer delivery.

In terms of pricing, Samsung's 2nm foundry quotes are reportedly about 30% cheaper than TSMC's. For Meta, the core logic behind self-developed chips is cost reduction, as a successful custom chip program can save billions of dollars in procurement costs annually, and Samsung's price advantage aligns perfectly with this demand.

On the capacity side, Samsung's foundry division is essentially sold out of 4nm capacity, and some 8nm lines are also close to full capacity. A South Korean industry official revealed that starting with Tesla's ( TSLA) AI chip orders last year, Samsung's wafer foundry has seen orders for AI server semiconductors enter a phase of rapid growth, with the current order backlog reaching approximately 50 trillion won, and an operating profit is expected as early as the third quarter of this year.

Meta Plans Cloud Business, AI Compute Reselling Expected to Generate Tens of Billions in Annual Revenue

The shift of orders to Samsung is driven by a deeper motive: Meta's desire to sell its computing power.

In early July 2026, it was reported that Meta was preparing to build a cloud infrastructure business named "Meta Compute," planning to sell AI computing power and model access to external customers. The business model has two paths: first, hosting its self-developed large models like Muse Spark for developers to access, benchmarking against AWS Bedrock; second, directly leasing "bare computing power," mimicking CoreWeave.

Zuckerberg mentioned during an earnings call this May that almost every week, external companies reach out to Meta wishing to obtain API services or asking if they can buy computing power, even expressing willingness to pay prices higher than Meta's procurement costs.

Meta has a unique advantage in running a cloud business: its data centers and GPUs were originally built to support Facebook, Instagram, WhatsApp, and Llama model training, rather than being sunk costs invested specifically for a cloud business. The marginal cost of selling computing power externally is extremely low, giving it much greater pricing flexibility than AWS. While others must calculate depreciation and amortization for cloud services, Meta only needs to calculate the incremental electricity costs.

Following the announcement, Meta's stock price surged 8.8% in a single day, adding approximately $127 billion to its market capitalization. Bank of America analysts Justin Post and Nitin Bansal calculated that if Meta can monetize its surplus computing power at an annual lease rate of $10 billion to $15 billion per gigawatt, it could generate an additional $30 billion to $45 billion in potential annual revenue—enough for the market to re-evaluate the return logic of its $145 billion capital expenditure.

Wells Fargo used data center power consumption as a proxy indicator for computing power scale to perform calculations, offering an even larger space for imagination: Meta's total computing power scale will increase from 7.5 gigawatts in 2025 to 21.2 gigawatts in 2028, with dedicated AI computing power increasing from 1.5 gigawatts to 13.2 gigawatts. Calculated at an annualized revenue of $20 billion per gigawatt of resold computing power, the annualized revenue of this business is expected to reach $264 billion by 2028.

The market remains divided on this. Bank of America pointed out that Meta's self-developed chip progress lags behind Amazon, Microsoft, and Google, and Meta itself still needs to purchase about 1.6 GW of computing power externally from third parties like Crusoe. Whether a company that still needs to outsource computing power can simultaneously resell it externally is currently the biggest point of controversy.

Samsung's 2nm orders projected to surge 130% year-on-year, with turn to profitability expected in Q3

For Samsung, if this order from Meta worth over 10 trillion won materializes, it will directly fill the initial idle capacity of its 2nm production line and provide an important endorsement for securing major clients in the future. Anthropic is also reportedly in talks with Samsung regarding a foundry partnership for its self-developed chips.

Over the past few years, Samsung's foundry business has been continuously suppressed by TSMC. In the first quarter of this year, TSMC's market share exceeded 70%, while Samsung's remained in the single digits, and its foundry division has suffered consecutive losses since 2023.

However, the situation is turning around. Samsung's 4nm capacity is basically sold out, and its 8nm is also close to full capacity. As Tesla, Anthropic, and Meta place orders successively, its medium- to long-term order backlog is expected to approach 50 trillion won.

In May this year, Anthropic completed a $6.5 billion Series H funding round, with Samsung Electronics participating as a "strategic infrastructure partner." Samsung is the only one among the three major memory manufacturers with logic chip foundry capabilities, and this investment also laid the groundwork for subsequent orders.

Samsung internally forecasts that the number of 2nm-related orders in 2026 will surge by more than 130% compared to 2025. TSMC's 2nm capacity for 2026 has basically been fully booked by Apple, Nvidia, and AMD, and Samsung happens to have surplus 2nm capacity at this time. If the orders from Meta and Anthropic can land as scheduled in the second half of 2026, Samsung's foundry division is expected to turn profitable in the third quarter of 2026.