It is reported that Meta plans to throw a $6.5 billion AI chip order to Samsung, which may change the exclusive foundry pattern of Taiwan Semiconductor

According to the Zhitong Finance APP, after a month of rumors about "cooling cooperation," Meta (META.US) and Samsung Electronics have experienced a dramatic epic reversal in their collaboration in the field of custom AI chips. According to the Seoul Economic Daily, Meta is in deep negotiations with Samsung Foundry, the wafer foundry division of Samsung Electronics, planning to sign a massive contract worth up to 10 trillion Korean won (approximately $6.54 billion) for the large-scale production of Meta's next-generation custom AI chips.

Saying goodbye to TSMC's exclusive foundry, Meta teams up with Samsung System LSI

Insiders revealed that the core of these negotiations is Meta's third-generation self-developed AI training and inference accelerator (MTIA). Previously, Meta's first and second-generation MTIA chips were exclusively manufactured by TSMC (TSM.US). However, due to the ongoing extreme shortage of advanced wafer foundry capacity globally, coupled with Meta's internally set aggressive R&D cycle of "iterating every six months," the enormous workload has overwhelmed Meta's internal design team.

To address this, Meta has established a joint development framework with Samsung Electronics' System LSI department, responsible for logic chip design. The two parties will jointly design and have Samsung's foundry manufacture this next-generation AI accelerator based on ASIC (Application-Specific Integrated Circuit) architecture.

After the "pause storm," the giants' multi-supplier strategy begins to emerge

It is noteworthy that this cooperation has experienced significant ups and downs over the past month. On June 4 of this year, local media in South Korea reported that Meta requested to "temporarily halt" the custom chip development project with Samsung's System LSI department, almost simultaneously with news of a cooling project between OpenAI and Samsung for developing an inference NPU (Neural Processing Unit). The withdrawal of two major clients raised doubts about Samsung's advanced process foundry strategy.

However, the latest developments released today indicate that the early project adjustments did not kill this deal but instead evolved into the current negotiation for a large-scale production contract worth $6.5 billion.

Industry observation: Silicon Valley is experiencing a wave of "de-NVIDIA" and "de-TSMC"

Analysts point out that if this $6.5 billion contract is finalized, it will become one of the most significant milestones for Samsung's wafer foundry business in recent years.

Currently, AI giants in Silicon Valley, such as Meta, OpenAI, and Anthropic, are making every effort to implement a "multi-vendor diversification" strategy. On one hand, each company is working to reduce its absolute dependence on NVIDIA's general-purpose GPUs through self-developed ASIC chips; on the other hand, the giants are also urgently seeking a "Plan B" to divert large orders from TSMC.

Recently, the tech media The Information also reported that AI unicorn Anthropic is in talks with Samsung, planning to use Samsung's 2nm (SF2P) process to produce its self-developed chips From OpenAI partnering with Broadcom to develop chips, to Anthropic and Meta secretly meeting with Samsung, the global AI computing power supply chain in the second half of 2026 is evolving from "domination by Taiwan Semiconductor" to a more complex game among giants