Samsung and SK Hynix's Investment in Anthropic: Different Strategic Agendas

Samsung Electronics and SK Hynix, alongside Micron, joined Anthropic’s Series H ($65 billion) as strategic infrastructure partners. However, their motives differed: among the three memory makers, only Samsung possesses a foundry, opening the door to potential logic chip production. The investment amounts remain undisclosed.

Date: 2026-06-10 · Last checked: 2026-09-27 · Figures based on reports from May–September 2026

While reading the investment announcement released by Anthropic on May 28, one line made me pause.

It stated that the technologies of the newly joined memory partners play a critical role in the global supply of memory, storage, and logic chips.

There were three partners: Samsung Electronics, SK Hynix, and U.S.-based Micron. All three are memory companies.

However, only one of them operates a foundry capable of manufacturing logic chips. This means that while the three companies appeared on the same list on the same day, they were actually carrying different strategic agendas.

At first, I read it simply as news that two Korean semiconductor giants had become shareholders in an AI company.

Looking again, however, the question split in two: Why would memory companies insist on buying equity in a client? And what does that single word—applicable only to Samsung Electronics—really point to?

How large was the scale of Anthropic’s Series H?

Anthropic raised $65 billion in its Series H round, valuing the company at $965 billion post-money.

According to Korea Herald reports, this is more than double the $380 billion valuation in February, making it the most valuable AI company, surpassing OpenAI's $852 billion valuation at the end of March.

The round was co-led by Capital Group, Coatue, and Singapore’s GIC.

Samsung Electronics, SK Hynix, and Micron were listed separately as ‘strategic infrastructure partners.’ The amount invested by each of the three companies has not been disclosed. While there has been talk of ’trillion-won investments’ in Korea, no confirmed figures exist.

Anthropic’s valuation has never risen this sharply. I have summarized the trends over the past year and a half in the table below.

Anthropic Valuation Trends (by round, post-money)

Date	Round	Valuation
March 2025	Series E	$61.5 billion
September 2025	Series F	$183 billion
February 2026	Series G	$380 billion
May 2026	Series H	$965 billion

Revenue has jumped in tandem. Anthropic announced in early May that its annualized revenue had exceeded $47 billion and projected its first operating profit in the second quarter. Given that its annualized revenue was approximately $9 billion at the end of 2025, it has grown more than fivefold in just five months.

Why did the three memory companies want to become Anthropic shareholders?

While profit is one motivation, the core reason is volume and information. The way AI models operate reveals why.

Large Language Models (LLMs) shuttle billions of parameters from memory to computing units every time they generate a single token.

No matter how fast the computing unit is, if the memory cannot supply data on time, the processor remains idle.

That is why the performance of modern AI accelerators is determined by how many layers of HBM (High Bandwidth Memory) are stacked next to the chip.

HBM is a product made by stacking thin DRAM dies vertically and connecting them with Through-Silicon Vias (TSVs). It is difficult to manufacture and, consequently, expensive.

Anthropic’s business model overlaps with this.

Claude is now used less as a chatbot for simple conversations and more as an ‘agent’ that can write code, search documents, and work autonomously for hours. Agents run inferences longer and more frequently. The longer the inference, the faster memory hits its limit.

The scale of computing resources Anthropic is purchasing is proof of this. They signed a contract with Amazon worth up to $100 billion for approximately 5GW of capacity, and entered a new multi-gigawatt TPU contract with Google and Broadcom starting in 2027. Broadcom’s filing cites approximately 3.5GW.

HBM is required for every one of those accelerators. For memory companies, Anthropic is already a top-tier indirect customer.

It is generally estimated that SK Hynix holds more than half of the HBM market, with Micron in pursuit. When a supplier becomes a shareholder of its client, the level of dialogue changes. They can learn what kind of memory the next-generation model will require and in what quantities, before specifications are finalized. The market interpreted the participation of Micron and SK Hynix as a defensive move—an effort to gain early insight into future memory requirements and stay ahead in next-generation standards.

Honestly, I found this surprising. News about Korean semiconductors has always been read through the lens of ‘who supplies more HBM to NVIDIA’.

This time, however, memory companies bypassed the GPU company and partnered directly with the model company. A component manufacturer has moved one step closer to the end-user.

The one word that applies only to Samsung: ‘Logic Chip’

Up to this point, all three companies had the same calculation. But Samsung Electronics has one extra line.

In its announcement, Anthropic wrote that its partners play a “critical role in the supply of memory, storage, and logic chips.”

Logic chips are made in foundries, or contract manufacturing plants. Among the three partners, Samsung Electronics is the only one that operates a foundry.

Consequently, in Korea, this was interpreted as a signal that Samsung could produce custom AI chips for Claude, in addition to memory.

Anthropic did not single out Samsung. The announcement grouped all three companies together; the interpretation was made by the industry.

People are hopeful about this interpretation because of the current state of Samsung Foundry.

Although it held the No. 2 spot in the global foundry market in 2025 with a 7.2% share, the gap with No. 1 TSMC is 62.7 percentage points.

The division has been in the red for years. Recently, however, its order list has changed. It secured Tesla’s next-generation autonomous driving chips (AI5/AI6) and is producing Groq’s 3rd-generation LPU (inference-dedicated chip). All three are AI computing chips—not general-purpose semiconductors, but chips designed by specific companies for their own models. If Anthropic joins this list, Samsung Foundry, previously called a ‘TSMC alternative,’ could aim for the position of a dedicated ‘AI company chip factory.’ This is why Samsung’s investment carries a different weight despite being on the same list as SK Hynix and Micron.

How does Samsung Foundry compete with TSMC?

In terms of process technology alone, it is difficult to beat TSMC.

What Samsung can offer is a different kind of value: the ability to complete logic chip production, HBM production, and advanced packaging that combines both as a single package within one company.

TSMC is the world’s best foundry, but it does not make memory.

To complete an accelerator, one must receive a logic chip from TSMC and HBM from SK Hynix or Samsung, then combine them at the packaging stage.

When multiple companies are involved, schedules become fragmented. In an era of chronic AI chip shortages, these coordination costs are significant. If Samsung can deliver a finished package at once, the client can reduce negotiation channels and predict delivery times more easily. For companies where opening a data center on time is more urgent than a few percentage points in chip speed, this is a quite persuasive proposal.

However, this is Samsung’s logic, not a confirmed advantage.

Tom’s Hardware viewed TSMC as the ultimate winner in the custom AI chip competition. They manufacture most of the cutting-edge wafers and control CoWoS packaging, which attaches the HBM. For Samsung’s turnkey strategy to work, yield must come first.

What SK Telecom’s Anthropic investment showed us first

There is a precedent for this investment method: SK Telecom, from the same SK Group.

In August 2023, SK Telecom invested approximately $100 million (about 130 billion KRW) in Anthropic and formed a partnership to co-develop a telco-specific language model. At the time, Anthropic had just released Claude 2. Three years later, the results were reflected in the numbers. As of the end of June, the [book value of the Anthropic stake](https://www.newspim.com/news/view/20260824001138) in SK Telecom's semi-annual report was 3.505 trillion KRW, up from 1.3762 trillion KRW at the end of last year—an increase of over 2 trillion KRW in half a year. This includes additional purchases of approximately 139.9 billion KRW made in the first half of this year. [Hana Securities estimated the stake's valuation in June at $2.9 billion](https://www.sisajournal.com/news/articleView.html?idxno=384587), or about 4.35 trillion KRW at the time. It is roughly a 30-fold return on the initial investment.

However, what is more noteworthy than the return is the structure. From the start, SK Telecom bundled money and cooperation. What they aimed for was not just recovering the investment, but securing a foothold in the growth of Anthropic. Samsung Electronics and SK Hynix entering as ‘strategic infrastructure partners’ follows the same logic. Memory suppliers have stepped inside the very place where memory demand is being created.

How Anthropic turned ‘safety’ into money

Then why specifically Anthropic? Technology alone doesn’t explain it. It is difficult to say that Claude outperforms competing models in every area. The difference lies in the trust that corporate clients feel.

Anthropic’s starting point was safety.

Dario Amodei was OpenAI’s VP of Research, and his sister Daniela Amodei was VP of Safety and Policy. In late 2020, they left OpenAI with fellow researchers and founded Anthropic in 2021. Their flagship technique is ‘Constitutional AI.’ Instead of humans grading every response, they provide a list of codified principles and have the AI critique and revise its own answers based on those principles. This method was published in a 2022 paper.

This attitude was most dramatically demonstrated in Project Glasswing, announced on April 7, 2026.

Anthropic stated that its new model, Claude Mythos Preview, found thousands of zero-day vulnerabilities across all major operating systems and web browsers.

The oldest was a bug that had been hidden for 27 years in OpenBSD, which is known for its security. What is interesting is that Mythos is not a security-dedicated model, but a general-purpose one. As its coding and reasoning capabilities improved, its hacking ability followed.

Anthropic did not release this model to the public. Instead, they opened it for defensive purposes only to 12 founding partners and about 40 infrastructure/open-source organizations, including AWS, Apple, Broadcom, Cisco, CrowdStrike, Google, JPMorgan Chase, the Linux Foundation, Microsoft, NVIDIA, and Palo Alto Networks.

They created the most powerful tool, then put it in a drawer.

When financial firms or government agencies adopt AI, the first thing they ask is not performance, but controllability. The guarantee that the system won’t move in an unexpected direction, the assurance that their data won’t leak.

If Constitutional AI is the technical basis for that assurance, the decision to seal Mythos is proof shown through action. That trust leads to corporate revenue, which leads to gigawatt-scale computing resource purchases, which in turn leads to memory and chip orders. What Samsung and SK Hynix bought was one link in this chain.

Anthropic’s own chips and Samsung Electronics: What has changed since the publication (as of September 2026)

When I first wrote this in June, ’logic chip’ was in the realm of interpretation. In the three months since, some things have become facts.

In early August, Anthropic officially confirmed it is putting together an in-house chip design team.

The goal is to design models and hardware together to run Claude faster and more efficiently.

This plan was first reported by Reuters in April, and in June, The Information reported that Anthropic had discussed manufacturing partnerships with Samsung Electronics.

This means a single word in an investment announcement is evolving into concrete discussions.

In my original piece, I read this trend as an ‘attempt to break free from NVIDIA dependency,’ but this needs to be corrected.

Anthropic is already mixing GPUs from NVIDIA and AMD, Google TPUs, and Amazon Trainium, and they have clarified that their in-house chip is an extension of a multi-chip strategy, not a replacement for existing suppliers. It is not an escape, but a cost battle. For a company processing billions of tokens a day, even a small reduction in cost per query adds up to a large sum.

The numbers for SK Telecom have also grown. After Anthropic submitted confidential listing documents to the U.S. SEC in June, the stock market began treating SK Telecom as a de facto proxy investment for Anthropic.

What remains undecided in the Samsung Electronics/SK Hynix investment in Anthropic

So, if you ask if the Korean semiconductor industry has occupied a seat on the inside of AI design, the most accurate answer is that they have ‘opened the door.’

Three things have been confirmed: the three memory companies became Anthropic shareholders, Anthropic decided to make its own chips, and Samsung Electronics is mentioned as a potential manufacturing partner. More remains unconfirmed: when the chips will come out, which process will be used, and whether the actual contract will go to Samsung. It usually takes several years from the establishment of a design team to the production of mass-market chips. If Samsung Foundry’s yields fall short of expectations, the volume can go to TSMC at any time.

There has been an old hierarchy in the semiconductor industry.

Designers decided what to make, memory companies produced to set specifications, and prices were determined by economic cycles.

This investment does not overturn that hierarchy. It simply bought a seat where they can have a say before specifications are finalized. While all three companies appeared on the same list on the same day, Samsung has a slightly broader scope for what it can say at that table because it owns a foundry. Whether that difference translates into actual purchase orders will be revealed in the filings of the coming quarters.

Source: Self-produced image

Notice: This article is based on public information at the time of writing and is not a recommendation for specific investments or financial products. Figures and policies may change. (Last checked: 2026-09-27)

※ This is not an article recommending the purchase or sale of specific stocks or assets.