Anthropic Warns of HBM Shortage Risk in $637 Billion IPO Filing

Anthropic has put a possible high-bandwidth memory shortage at the top of its risk list, naming Samsung Electronics, SK hynix and Micron Technology as the three suppliers that may not keep up with demand. The warning sits in the IPO registration statement the company filed with the U.S. Securities and Exchange Commission on the 26th, according to industry sources cited on the 27th. The same filing lays bare the scale of Anthropic's compute commitments: a simple sum of seven major AI infrastructure contracts exceeds $600 billion.

The numbers behind that sum matter for anyone tracking memory demand. The seven disclosed contracts add up to roughly $637 billion, or about 845 trillion won. The single largest is an AI accelerator equipment lease with Broadcom worth about $161.2 billion — roughly 214 trillion won. Cloud agreements follow close behind: at least $114.5 billion with Google, at least $110 billion with Amazon and at least $44.9 billion with Microsoft. Deals with SpaceXAI, Nscale and Lambda round out the list.

Don't read that total as a fixed investment figure. The contracts mix minimum spending obligations, equipment lease payments and amounts that vary with actual volumes supplied, so the real outlay will move with usage.

What does the supply warning actually say?

Anthropic's filing states that most high-performance memory used in AI computing infrastructure — HBM above all — comes from SK hynix, Samsung and Micron. The company said there's no assurance the three can deliver the volumes it needs on time, even as they expand capacity. Shortages could get worse if large customers or competitors lock up a particular supplier's output in advance. The concentration of leading-edge chip manufacturing at Taiwan's TSMC appears in the filing as a separate risk factor.

Why the demand picture is broadening

Anthropic built and ran its models on cloud services from the big tech platforms until recently. It has since diversified: it now leases AI accelerators separately and deploys them in dedicated data centers. The chip mix spans Amazon's Trainium, Google's tensor processing units and GPUs from Nvidia and AMD. As infrastructure spending spreads across multiple chip types rather than clustering on Nvidia alone, HBM demand rises with each of them.

The cost side is climbing just as fast. Anthropic's computing and infrastructure costs hit $8.945 billion — about 11.9 trillion won — in the first half of this year, up 283% year over year. Revenue still outpaced it, jumping from $1.163 billion to $16.286 billion, roughly 21.6 trillion won.

What's in it for Korean memory makers?

For Samsung and SK hynix, the filing points to a demand tailwind they can't fully book yet. Samsung signed a memorandum of understanding with Broadcom — Anthropic's largest single counterparty — in July, targeting cooperation worth more than $200 billion, or about 265 trillion won, across memory and foundry through 2030. HBM supply for Broadcom's next-generation AI accelerators sits among the planned cooperation areas.

It's still an MOU. Neither the HBM volume bound for Anthropic-linked systems nor the size of any individual delivery contract has been confirmed. An industry official put the logic plainly: AI companies spending enormous sums to secure computing power means memory demand is growing by that much.

For practitioners, the takeaway is straightforward. The largest AI developer financing event now on the calendar carries an explicit, SEC-filed admission that HBM supply is the choke point — and roughly $637 billion in contracts leaning on it.