Infrastructure

Anthropic Asked a Memory Maker for Its Own Chips

SK Group's chairman said on 25 July that Anthropic approached SK Hynix for supplies to build its own semiconductors. That adds a claimant to the queue you are in.

InfrastructureBy Servola Tech Desk2026-07-264 min read

AI-assisted, edited by humans. | Editor-in-Chief: Leon Soliman | Editorial standards

Anthropic Asked a Memory Maker for Its Own Chips

Key takeaways

A model lab designing its own silicon does not reduce demand on the memory pool, it adds a new claimant to it, so the supplier concentration that already sets your server prices gets one more large buyer ahead of you rather than one fewer.
Treat high-bandwidth memory, not the accelerator, as the constrained part in any 2027 infrastructure plan, because that is where the queue forms and where a new entrant lands first.
Note that the memory maker being asked to supply is also an investor in the company asking, and ask your own vendors to disclose equity relationships in their supply chain before you sign a multi-year commitment.
When a vendor tells you capacity is tight, ask which component is tight and who else is in that queue, because a generic shortage story is not something you can plan against and a named bottleneck is.

A chairman says the quiet part at a conference

Chey Tae-won, chairman of SK Group, told an audience at an artificial intelligence event in San Francisco that Anthropic had approached SK Hynix for supplies to make its own semiconductors. He called the sight of AI developers harbouring chip ambitions remarkable. The remark was reported on 25 July, and it is the kind of disclosure that arrives sideways, from a supplier rather than from the company doing the buying.

Why it matters: the request centres on high-bandwidth memory, the stacked memory that sits beside an AI accelerator and feeds it. HBM is the part of the AI hardware stack with the least slack in it. SK Hynix is one of a very small number of firms that can make it at volume, and it is already allocating that output between the largest buyers in the world.

There is a second fact worth holding alongside the first. SK Hynix took part in Anthropic's Series H funding round in May 2026, alongside Samsung and Micron. The company being asked to supply is also a shareholder in the company asking. That is not improper, and it is common in semiconductors, but it is a relationship an owner should know exists before reading any statement from either side about supply availability.

Custom silicon adds a claimant, it does not remove one

The intuitive reading of a lab building its own chip is that it steps out of the queue for someone else's. That is close to the opposite of what happens.

A custom accelerator still needs memory stacked next to it, and it needs the same HBM from the same handful of suppliers. What changes is that the lab now buys memory directly instead of buying a finished accelerator that already had memory allocated to it. The claim on the memory pool does not disappear. It moves upstream and becomes visible, and it competes with every other direct buyer, which now includes the hyperscalers, the accelerator vendors, and anyone building sovereign capacity in Europe.

The bottom line: the constrained resource in AI infrastructure stopped being the accelerator some time ago. It is the memory that sits beside it, and every announcement about a new chip programme should be read as a new order on that specific supply, not as relief anywhere.

What this means for a European infrastructure plan

Most European firms are not buying HBM directly. They buy servers, or they rent capacity, and the shortage reaches them as a delivery date that slips and a quote that does not hold. That indirect exposure is precisely why the queue matters.

If you are refreshing on-premise compute in 2027, the question to put to your integrator is not whether GPUs are available. It is which memory configuration the quoted machine ships with, and what the lead time on that configuration is if you order in the fourth quarter. Ask for the second answer in writing. A vendor who will commit to an accelerator date but not to a memory-configuration date has told you where the risk sits.

Yes, but: if you rent rather than buy, this reaches you as price and as instance availability in your region rather than as a delivery date. The equivalent question for a cloud contract is whether your committed capacity is guaranteed in a named European region or merely in the provider's global fleet. Those two commitments look identical in a summary and behave very differently when supply tightens.

Three questions to put to your vendors this quarter

Which component is short, by name. A supplier who says capacity is tight has told you nothing actionable. A supplier who says the HBM configuration on this SKU has a twenty-week lead time has given you a date to plan around and a thing to design out of the specification if you need to.

Who else is in this queue. You will not get a customer list, and you should not expect one. You can reasonably ask whether your order sits in the same allocation pool as the vendor's largest accounts, and whether allocation is by contract date or by volume. The answer tells you what your order is actually worth in a squeeze.

What equity relationships exist upstream. Ask your hardware vendors to disclose whether any supplier in the chain is also an investor in, or an investee of, another party to your contract. The Anthropic and SK Hynix relationship is disclosed and visible. The ones worth worrying about are the ones that only surface when allocation gets decided and you cannot work out why your order moved down the list.

Frequently asked questions

Does Anthropic building its own chip make AI compute cheaper?

Not for a third-party buyer, and not soon. A custom accelerator programme reduces one company's dependence on one vendor. It does not add memory capacity to the market, and in the near term it adds a large direct buyer to the same constrained memory supply that everyone else draws on.

Why does high-bandwidth memory matter more than the accelerator?

Because it is the part with the least spare capacity. Accelerator supply has widened as more designs reach production, while HBM is made at volume by a very small number of firms and cannot be substituted from ordinary memory. When AI hardware slips, the memory configuration is usually the reason.

Should a supplier being an investor in its customer worry me?

It is not improper and it is normal in semiconductors, where capital and capacity often move together. It is simply information you should hold. Equity relationships shape how allocation gets decided when supply is scarce, and an owner signing a multi-year commitment is entitled to know which relationships exist above them in the chain.

Every lab that announces its own silicon is announcing a new order on the same memory. The bottleneck is not moving away from you. It is moving upstream, where it becomes harder to see and easier to lose your place in.

AnthropicSK HynixSK GroupChey Tae-wonHBMhigh-bandwidth memorycustom silicon

SK Hynix Raised $26.5B in a Record Nasdaq Debut

SK Hynix priced its Nasdaq ADR debut at USD 149 on July 10, 2026, raising about USD 26.5 billion, the largest US listing ever by a foreign company, and the stock jumped 17 percent on day one. What it means for the memory market and compute buyers.

4 min read

Infrastructure

Anthropic's Fix for Claude Outages Arrives in 2027

AMD will invest up to $5bn in Anthropic and supply 2 gigawatts of MI450 chips. The relief for Claude's rate limits is real, but the first gigawatt is dated to 2027, and your chip supplier now owns a stake in its own customer.

4 min read

Law & Regulation

The EA Buyout Now Hinges on Brussels

The largest take-private in history has cleared EA shareholders and US antitrust. The last real gate is the EU - and not classic merger review, but the Foreign Subsidies Regulation, facing its first blockbuster test.

3 min read

← All articles