Two gigawatts of MI450 capacity and up to $5 billion in equity, with no warrants attached. The structure of this deal says more about AMD’s position than the headline number does.
AMD has signed its third gigawatt-scale AI customer in nine months. What changed this time is who paid whom. Under the partnership announced on Wednesday, Anthropic will deploy up to 2 gigawatts of AMD Instinct MI450 Series GPUs inside AMD Helios rack-scale systems, with the first gigawatt beginning deployment in the first half of 2027. AMD has committed to a strategic equity investment of up to $5 billion in Anthropic, released as deployment milestones are met.
The Wall Street Journal, which reported the agreement before it was confirmed, put the value of the hardware itself in the tens of billions of dollars and said Anthropic will buy some of the chips outright for its own data centres while leasing the rest of the capacity.
Set the equity terms against the two deals that came before. When AMD landed OpenAI last October and Meta in February, it handed each company warrants for up to 160 million shares, exercisable at a cent apiece if purchase and performance targets were hit. Between them, those warrants account for roughly a fifth of AMD. Anthropic received none. AMD is putting cash into its customer rather than giving away equity to win the order.
Jefferies analyst Blayne Curtis, who holds a $615 price target on the stock, said the terms mattered more than the win itself. Investors seemed to take the same view. AMD climbed around 10% on Wednesday, part of a roughly $85 billion gain in market value over two sessions that carried its market capitalisation past $900 billion.
Third in a queue that AMD still has to deliver on
The commitment is real. Whether it arrives on schedule is a separate question, and Anthropic is not first in line. OpenAI’s initial gigawatt of AMD capacity is due in the second half of this year. Oracle is scheduled to begin deploying 50,000 MI450 GPUs this quarter. Microsoft committed on Monday to ramping Helios on Azure. Meta’s 6-gigawatt agreement starts with its own gigawatt this year.
Counting Anthropic, AMD’s announced AI compute commitments now sit at roughly 14 gigawatts, all of it depending on a rack-scale platform that has not yet shipped in volume. The scale of what AMD is trying to close is easier to see in the market share numbers. Futurum Group estimates cited by CNBC put Nvidia above 95% of the data centre GPU market, leaving AMD with something near 4.5%. AMD’s data centre segment turned over $5.78 billion in the first quarter, up 57% year on year, which is strong growth against a base that is still small relative to the ambition. Second-quarter results land on August 4.
The gigawatts get built in Asia
None of this capacity is manufactured in Santa Clara, and that is where the timeline gets interesting. The MI450 accelerator die is expected to use TSMC’s 2nm-class N2P process, with the EPYC “Venice” CPUs alongside it among the first enterprise processors on TSMC’s 2nm node, according to reporting from Commercial Times and TechPowerUp.
Packaging moves to CoWoS-L, TSMC’s advanced process for bonding memory and logic onto a single module. Samsung has reportedly secured a lead supply role for HBM4, the stacked memory that sits alongside the accelerator and largely determines how fast data reaches it. Chosun Biz puts the MI450 at 432GB per GPU. SK hynix retains the larger share of Nvidia’s Rubin volumes.
Assembly runs through the same Taiwanese base that builds most of the world’s AI infrastructure. AMD announced more than $10 billion in Taiwan ecosystem investments in May, naming Sanmina, Wiwynn, Wistron and Inventec as the ODMs building Helios systems, with substrate and PCB work going to Unimicron, Nan Ya PCB, Kinsus and AIC.
That supply base is where the risk concentrates. HBM production at all three major suppliers has been effectively sold out through 2027, and advanced packaging capacity is allocated years in advance. A 2-gigawatt commitment beginning in 2027 is a claim on capacity that is already spoken for several times over.
What Anthropic is actually buying
The compute is the visible half of the arrangement. The software is the part that matters longer term.
Anthropic already runs AMD’s current-generation MI355X GPUs, a detail the two companies confirmed publicly for the first time this week. The new agreement adds a multi-year engineering collaboration in which Claude will be used to tune workloads for Instinct chips and speed up development of ROCm, the AMD software layer that lets AI researchers actually program the hardware. AMD will also adopt Claude across its own engineering teams.
ROCm is the reason AMD keeps losing frontier customers to Nvidia. Nvidia’s equivalent, CUDA, has had two decades to accumulate tooling, documentation and developer habits, and most AI researchers have never written code for anything else. Faster silicon on its own does not close that gap. Getting a frontier lab to point its own models at the problem is a cheaper route than hiring for it.
Anthropic now draws on roughly six distinct silicon sources for Claude. The company has committed more than $100 billion over ten years to AWS for up to 5 gigawatts of Trainium capacity, signed a multi-gigawatt agreement with Google and Broadcom in April, contracted Nvidia compute through Microsoft Azure, and agreed in May to pay SpaceX $1.25 billion a month through 2029 for the full capacity of its Colossus 1 data centre in Memphis.
“Running across a diversified range of hardware lets us map the right workloads to the right hardware,” said Tom Brown, Anthropic’s co-founder and chief compute officer. That is a polite way of saying no single supplier gets to set the price.
In short, AMD has spent nine months buying its way into that conversation. This is the first time it did so without handing over a piece of itself.