Anthropic’s plan to break free from Nvidia: Samsung chips, SK Hynix memory, and 12+ GW of custom compute

Samsung & SK Hynix chip supply agreements

On July 25, 2026, Anthropic CEO Dario Amodei confirmed that the company has signed supply agreements with both Samsung Electronics and SK Hynix . The deals include potential long-term memory chip contracts and investment arrangements, though full details have not yet been formally released by the companies . The announcements came during a San Francisco AI summit where South Korean semiconductor makers and U.S. AI firms announced a combined $950 billion in chip supply and infrastructure partnerships over five years .

Samsung — Anthropic is in early-stage discussions to use Samsung’s 2-nanometer foundry and advanced packaging technology to manufacture a custom AI chip . Samsung has also made a strategic investment in Anthropic as part of deepening the relationship .

SK Hynix — SK Group Chairman Chey Tae-won disclosed at a San Francisco AI event that Anthropic formally submitted a request for chip raw material supplies to produce its own semiconductors . The requests involve memory chip supply for Anthropic's custom chip ambitions, though analysts note that the current disclosures describe a useful infrastructure relationship rather than a confirmed HBM purchase or dedicated allocation .

Custom semiconductor ambitions

Anthropic’s push into custom silicon began in April 2026, when Reuters first reported that the company was investigating building its own AI chips in response to the acute shortage of Nvidia GPUs and soaring compute costs .

• The Samsung talks center on designing a custom AI training/inference chip that would reduce Anthropic’s dependence on Nvidia and lower costs .
• SK Hynix would supply the memory components needed to complement Anthropic’s own chip design .
• The plans remain early-stage; Anthropic could still decide to only buy chips rather than design them, and has not yet publicly committed to a specific design or assembled a dedicated engineering team .

Broader compute strategy: 12+ GW across multiple partners

Anthropic’s path to compute self-sufficiency rests on four major pillars:

1. Google & Broadcom TPU agreement (3.5 GW)

In April 2026, Anthropic signed a landmark deal with Google Cloud and Broadcom to secure 3.5 gigawatts of next-generation TPU compute capacity, slated to come online starting in 2027 . This deal, worth several hundred billion dollars, gives Anthropic priority access to Google’s custom Tensor Processing Units designed by Broadcom, with availability contingent on Anthropic's commercial growth targets . Broadcom separately extended its TPU development agreement with Google through 2031 . The vast majority of the new compute will be sited in the United States, hosted in Anthropic’s own facilities .

2. AMD partnership (2 GW)

In July 2026, AMD committed to investing up to $5 billion in Anthropic. In return, Anthropic signed multi-year procurement agreements to purchase tens of billions of dollars in AMD’s latest Instinct MI450 server GPUs, deployed in Helios rack-scale systems, securing 2 gigawatts of computing capacity. The first gigawatt is scheduled to come online in early 2027 .

3. Amazon AWS relationship

Anthropic has historically relied on chips rented from Amazon (AWS Trainium and Inferentia), Google (TPUs), and Nvidia (GPUs) . While the Samsung/SK Hynix agreements mark a move toward self-built silicon, Anthropic continues to use AWS infrastructure for training and serving Claude models.

4. $65 billion funding & strategic investments

In May 2026, Anthropic raised $65 billion in Series H funding, at a post-money valuation of $965 billion, with participation from Samsung Electronics and SK Hynix alongside other strategic investors . The world’s three largest memory chipmakers — Samsung, SK Hynix, and Micron — appeared together on the same AI company’s funding list for the first time . This capital fuels both the custom chip effort and the massive compute prepayments, with Samsung’s investment alone reported to be in the multi-trillion-won range .

5. Data center infrastructure

Anthropic is building its own data center capacity to house the 3.5 GW of TPU compute and, eventually, its own custom chips. The Google/Broadcom deal specifies that Anthropic will host the TPU capacity in its own facilities, giving the company direct control over its compute infrastructure rather than relying solely on third-party cloud . This is an expansion of Anthropic's November 2025 commitment to invest $50 billion in U.S. domestic infrastructure .

The three-horizon strategy

Anthropic’s compute strategy spans three horizons:

1. Near-term (2026–2027): Reliance on contracted TPU/Broadcom capacity (3.5 GW from 2027) and AMD GPUs (2 GW from early 2027), plus continued use of AWS and Google cloud.
2. Medium-term (2027+): Custom chip design manufactured by Samsung’s 2nm foundry with memory from SK Hynix. This is still early-stage and not guaranteed.
3. Long-term: Buildout of its own data centers to house all of the above, giving Anthropic direct control over its infrastructure.

Together, these agreements represent a total compute pipeline of over 12 GW across Google, AMD, and custom silicon pathways, making Anthropic one of the largest consumers of AI compute in the world . The South Korean chip deals are a critical enabler: Samsung provides the fabrication and SK Hynix provides the memory, together allowing Anthropic to eventually field its own silicon and break what has been near-total dependence on rented infrastructure .