News · 2026-07-23

Anthropic plans up to two gigawatts of AMD chips, with AMD committing up to $5 billion back

AMD announced that Anthropic plans to deploy up to 2 gigawatts of MI450-series capacity in AMD Helios systems, with the first gigawatt beginning in the first half of 2027. Alongside it, AMD says it has committed to make a strategic equity investment of up to $5 billion in Anthropic in the future. Both numbers are ceilings on forward plans, and the distinction matters more than the size.

Key facts

• The capacity: up to 2 GW of MI450-series deployment, first gigawatt beginning H1 2027.
• The money: an equity investment of up to $5 billion, described as a future commitment with no stake, schedule, or terms disclosed.
• The configuration: MI455X GPUs, "Venice" CPUs, Pensando networking, and the ROCm software stack; Anthropic already uses MI355X.
• Primary source: AMD's announcement.

Two gigawatts is a genuinely enormous figure -- comparable to the continuous electrical draw of a mid-sized city, and among the largest single AI capacity plans any company has put on paper. But the sentence structure in the release is doing careful work, and it should be preserved rather than rounded off. "Up to 2 gigawatts" is a cap on a plan. "Beginning in the first half of 2027" means nothing is being installed now. And "committed to make a strategic equity investment of up to $5 billion in Anthropic in the future" is not the same sentence as "AMD invested $5 billion in Anthropic," which is how it has been widely paraphrased.

What is concrete is the configuration. Anthropic would run MI455X GPUs paired with AMD's "Venice" CPUs and Pensando networking, on the ROCm software stack, inside Helios rack-scale systems. Anthropic already uses the current-generation MI355X, so this is an expansion of a working relationship rather than a first date.

The software detail is the one that should interest anyone tracking whether AMD can actually take share. For years the practical obstacle to AMD in AI has not been silicon, it has been CUDA -- the mature Nvidia software ecosystem that every training and inference framework was written against. ROCm is AMD's answer, and a named frontier lab committing to multi-year, gigawatt-scale deployment on it is a stronger signal about ROCm's maturity than any benchmark AMD could publish. Chips are a purchasing decision. A software stack is a hiring and rewriting decision, and labs do not make those lightly.

Anthropic's chief compute officer Tom Brown framed the move as diversification, saying that having multiple hardware options lets the company map each workload to the hardware suited for it. Notably, the release contains no exclusivity claim and no suggestion of replacing Nvidia. Frontier labs run heterogeneous fleets, and reading a supplier addition as a supplier swap is a persistent error in AI infrastructure coverage.

The circularity question is worth raising, because it recurs across this sector. A chip vendor committing equity to a customer who commits to buying its chips creates a loop where a portion of the vendor's revenue is funded by the vendor's own balance sheet. That pattern is now common across AI infrastructure deals, and sits alongside the off-balance-sheet financing structures funding the wider buildout, and it is why the "up to" and "in the future" qualifiers on the $5 billion matter: an uncapped, completed investment would make the loop tighter than a capped future one. AMD discloses no schedule, no stake, and no chip-pricing terms, so the actual tightness of the loop is not knowable from the release.

The same day, AMD announced a separate partnership with Cerebras to split inference across two chip architectures, giving the company two distinct plays: bulk training and inference capacity for a frontier lab, and a specialized low-latency inference product for agentic workloads.

The distinction between capacity for training and for inference is not broken out in the release either, which matters because the two have very different utilization profiles. For context on the market these announcements sit in, Intel reported second-quarter revenue of $16.1 billion, up 25% year over year, with CEO Lip-Bu Tan calling it the company's strongest revenue growth in more than 15 years and attributing unprecedented compute demand to AI (release). That causal framing is management's characterization rather than an independent market finding, but the revenue figure is audited and dated -- which is more than can be said for most numbers in AI infrastructure this week.

The honest caveat: every number in the AMD announcement is forward-looking. Capacity beginning in 2027, an investment "in the future," a plan capped at "up to." Deployment plans of this size are routinely revised, delayed, or quietly reduced, and the first checkable milestone -- did the first gigawatt actually begin in the first half of 2027 -- is nearly a year away.

Primary source, verified: read the paper →

Ground Truth. (2026, July 23). Anthropic plans up to two gigawatts of AMD chips, with AMD committing up to $5 billion back. Ground Truth. https://groundtruth.day/news/anthropic-plans-2gw-of-amd-chips.html

@misc{groundtruth:anthropic-plans-2gw-of-amd-chips,
  title  = {Anthropic plans up to two gigawatts of AMD chips, with AMD committing up to $5 billion back},
  author = {{Ground Truth}},
  year   = {2026},
  month  = {jul},
  url    = {https://groundtruth.day/news/anthropic-plans-2gw-of-amd-chips.html}
}

Topics: hardware · amd · anthropic · datacenter · compute