TSMC is buying stakes in Anthropic and OpenAI worth up to $1 billion

Taiwan Semiconductor Manufacturing Company (TSMC) has decided to invest US$500 million in equity stakes of both OpenAI and Anthropic, two leading artificial intelligence (AI) companies. This move will grant TSMC, the world's largest contract chipmaker, direct financial ties to the AI labs that are driving demand for its business. Board members authorized these investments during a meeting on Tuesday.

This strategic shift for TSMC represents a significant departure from its earlier stance, which dismissed the idea of investing in end-market consumers like Nvidia. In July, chairman and CEO C.C. Wei had stated that the foundry would not pursue investing in end-market consumers, asserting that existing customer arrangements were functioning well.

TSMC's pivot towards AI development comes at a time when the foundry is expanding its focus beyond its immediate fabless customers, such as Nvidia and Advanced Micro Devices (AMD), to gauge true end-user demand amidst a significant surge in capital spending across the industry. Earlier this year, TSMC dispatched teams to the US to assess data center construction and grid capacity, integrating real-world power constraints and hyperscaler balance sheets into its factory expansion plans.

Industry analysts have noted that TSMC's five-year forecasting framework now accounts for AI model demand and emphasizes the financial resilience of established cloud providers to mitigate potential overcapacity risks during cyclical corrections. TSMC's equity investments in AI companies may also help the company manage supply bottlenecks by offering a closer look at the compute needs of its customers.

Anthropic stated in an updated IPO filing that TSMC's capacity constraints had occasionally delayed chip deliveries. The company's prospectus revealed that TSMC subsidiary TSMC Partners would acquire non-voting Class N shares in a US$26 billion private placement alongside investors like Google and Nvidia. Similarly, OpenAI has been collaborating with Broadcom to deploy 10 gigawatts of in-house custom accelerators, further blurring the lines between chip designers and AI model makers.