July 3 – NUPIAO has learned that Samsung Electronics is rapidly becoming the core manufacturing hub for global tech giants’ in-house AI chips (ASICs), with its long-term accumulated orders potentially approaching 50 trillion won.
Sources indicate Meta is pushing forward a collaboration with Samsung Foundry to design and produce its next-generation ASIC, a deal valued at over 10 trillion won. Meta has already locked in Samsung as a partner for its self-developed AI accelerator, MTIA, and plans to mass-produce hundreds of thousands of units using Samsung’s cutting-edge 2-nanometer process. The two sides have yet to finalize the exact details of their cooperation.
On July 3, Samsung Electronics (005930.KS) shares jumped nearly 8% on the back of the AI foundry order buzz, and the market has raised its full-year revenue forecast for the company’s wafer fabrication division.
Meta has been pouring money into AI infrastructure, lifting its annual capital budget this year to between $115 billion and $135 billion. Over 60% of that is funneled into AI servers and homegrown computing chips. To meet the immense demand for inference and lightweight training, Meta needs to secure stable 2nm wafer capacity from multiple suppliers.
Designed specifically for social recommendation and multimodal large model inference, the MTIA series is Meta’s core hardware for building an independent computing base and reducing its reliance on off-the-shelf GPUs.
The previous two generations of MTIA chips were mainly manufactured using TSMC’s 3nm process. However, TSMC’s 2nm and 3nm high-end capacity is now booked out through 2027, pushing global tech leaders to proactively diversify their foundry networks.
Rigid demand for high-end AI chips is intensifying the scramble for 2nm and 3nm advanced process capacity. Designing custom ASICs has become a standard play for cloud providers and AI unicorns. Industry statistics show that in 2026, the combined AI-related capital expenditure of Microsoft, Google, Amazon, and Meta will exceed $700 billion, with nearly 70% flowing into AI servers and tailor-made computing chips.
Each of these companies is simultaneously advancing in-house chip development and third-party procurement to hedge against the supply constraints of a single vendor and long-term computing cost pressure.
Another US AI powerhouse, Anthropic, is also evaluating using Samsung’s 2nm process to develop chips.
On July 2, NUPIAO learned that Anthropic has held talks with Samsung Electronics about a potential manufacturing partnership. That said, Anthropic has publicly stated it will not abandon its multi-hardware architecture that includes Google TPUs, Amazon Trainium, and Nvidia GPUs.
As a core rival to OpenAI, Anthropic is expanding rapidly behind its Claude series of large language models. The company has just closed a $65 billion Series H funding round, pushing its valuation to $965 billion. With computing costs soaring, relying solely on Nvidia GPUs and Amazon Trainium chips makes it hard to control the expense of model iteration.
Anthropic has officially kicked off a custom AI chip development project, aiming to slash long-term computing procurement costs through its own ASIC. Its hardware team has been beefed up and recently poached a core engineer from OpenAI’s custom chip effort to lead the architecture design.
However, Anthropic is still mapping out the processor’s functional positioning, performance level, and how it will be deployed inside servers or clusters. The company has contacted several chip design houses but hasn’t entered the detailed design phase yet. At this stage, it is closely comparing process technologies from multiple foundries, with Samsung’s 2nm node and its accompanying packaging lines among the top options being evaluated.
As the most advanced process currently ready for mass production, 2nm technology can significantly increase transistor density and reduce operating power. Paired with advanced packaging, it can slash data transfer latency between compute units and high-bandwidth memory, easing the notorious congestion in AI clusters.
Back in May, Anthropic completed its Series H round, and Samsung Electronics, SK Hynix, and Micron participated as “strategic infrastructure partners.” Among the three memory giants, Samsung is the only one with logic chip wafer foundry capabilities, which already had the market anticipating an order windfall.
If the deal materializes, Samsung will add another heavyweight customer to a roster that already includes Tesla, Nvidia, and Apple.
This also marks a strategic payoff for Samsung’s foundry business as it continues to chase the industry leader. TSMC has long held over 70% of the global advanced foundry market, but Samsung is fighting back with its unique GAA technology and new 2nm fab capacity. By the first quarter of 2026, Samsung’s 2nm yield reportedly exceeded 60%. While still a hair short of the 70% threshold for profitable mass production, it’s already enough to take on initial volume orders from big customers.
Meanwhile, Samsung’s foundry unit has recently tweaked its supply strategy to prioritize existing client orders while selectively taking on new ones. Its 4nm process capacity is virtually sold out, and even next year’s capacity is fully booked. Some of its 8nm lines are also running at near-full capacity.
The South Korean government recently unveiled a semiconductor investment plan as part of its three “mega national projects.” The plan encourages Samsung Electronics, SK Hynix, and other companies to make massive investments in the semiconductor industry, physical AI, and AI data centers.
On June 29, South Korea’s Minister of Trade, Industry and Energy, Kim Jeong-gwan, announced that the country will invest 800 trillion won (about 3.52 trillion yuan) to build four memory chip wafer fabs in the southwestern Gwangju and Jeolla regions, creating the country’s second-largest semiconductor cluster after the capital area. Samsung and SK Hynix will jointly contribute 81 trillion won to set up an advanced packaging base in Chungcheong Province, specifically designed to meet the expansion needs of AI high-bandwidth memory (HBM) production.
On the same day, Samsung officially announced its investment plan, totaling 2,655 trillion won (about 11.68 trillion yuan), with 2,030 trillion won earmarked for its semiconductor cluster in Yongin and Pyeongtaek, South Korea.
