Some deals change direction, not just numbers.
In the spring of 2026, Samsung Electronics and SK Hynix poured trillions of won into an AI company in Silicon Valley.
The media called it an ‘investment.’ That’s not wrong. But it doesn’t capture the true weight of this event.
The air in the room where this deal was struck must have been different.
For decades, these companies, who have manufactured the world’s best memory semiconductors, have always existed as components slotted into someone else’s accelerators.
This was the moment they signed a contract to step into the very center of the design process for the first time.
I. The Price Tag of Superintelligence: A $965 Billion Question
Let’s face the numbers first.
Anthropic raised $65 billion in its Series H funding round in 2026. Its post-money valuation is $965 billion.
This means its valuation has skyrocketed by more than 2.5 times in just a few months, from $380 billion in February 2026.
To put these numbers into perspective, let’s change the scale.
$965 billion. That’s more than half of South Korea’s annual GDP.
It’s nearly double Samsung Electronics’ market capitalization.
Even some of the world’s oldest investment banks combined don’t reach this figure.
Yet, Anthropic doesn’t make semiconductors. It has no factories. It doesn’t produce a single physical product.
What this company creates is a language model that writes, codes, and mimics thought—or, in some ways, performs thought itself.
So, what is $965 billion the price of?
It’s the price of control over a system that can replace or augment human cognitive abilities.
More coldly, it’s also the price of the premium that can be charged by those who supply the power, cooling, and silicon needed to run that system.
Samsung Electronics and SK Hynix’s participation in this round is precisely an attempt to move from the position of ‘supplier’ to that of ‘designer.’
II. Why Dario Amodei Left OpenAI
Every company’s DNA is forged from what its founders feared most.
In 2021, Dario Amodei was a key research vice president at OpenAI. His sister, Daniela, was head of safety.
They saw the world the same way, but they also began to believe, around the same time, that the path OpenAI was taking would not make that world safe.
After Microsoft’s $10 billion investment was finalized, OpenAI accelerated its commercialization. The speed of commercialization outpaced the speed of safety verification.
Faster product launches began to take precedence over
substantive research into the potential for advanced language models to evolve in biased or hostile directions, beyond human control.
In protest, the Amodei siblings, along with several other key researchers from OpenAI, left.
The company they founded is called Anthropic.
And the first thing they created at Anthropic was not a more powerful language model.
It was a set of principles for controlling language models.
This framework, called Constitutional AI, fundamentally redesigned how AI systems are trained.
The previous method involved human evaluators reviewing AI responses and providing correction instructions.
-> This structure inherently infused human bias and led to exponentially increasing costs as scale grew.
Instead, Anthropic created a ‘constitution.’
It’s a codified set of 60-75 clear ethical principles derived from the Universal Declaration of Human Rights, major global corporate safety guidelines, and copyright regulations.
And it made the AI itself critique and revise its own responses based on this constitution.
This was not just a technical differentiation.
It was a structural solution to the problem Dario Amodei feared most at OpenAI—the phenomenon of AI subtly evading human oversight.
[IMAGE: A split-image composition: on the left, a chaotic tangle of neural network visualization with red warning indicators; on the right, an ordered, luminous lattice structure with clean blue pathways representing constitutional constraints. Style: digital art, scientific visualization aesthetic, sharp contrast between entropy and order, suggesting the core tension of AI alignment research.]
Institutions that are among the most conservative and security-conscious globally flocked to the company born from that fear.
Major global financial institutions, government agencies, and enterprise clients. As of May 2026, Anthropic’s annualized revenue exceeded $47 billion.
Quarterly revenue alone reached $10.9 billion. For the first time since its inception, profitability was becoming visible.
Obsession with safety became profitable. And that profit became the force attracting Korean semiconductor companies.
III. How Claude Evolved: From Sonnet to Mythos
The name of Anthropic’s language model is Claude.
When the Claude 3 generation emerged in the spring of 2024, the industry classified it as a competitor to GPT-4.
The three-tiered system of Opus, Sonnet, and Haiku was reasonable but not revolutionary.
However, Claude 3.5 Sonnet, released in June 2024, was a different story.
It lowered prices and increased performance. That alone was enough to shake the market.
And in October of the same year, the world witnessed for the first time an AI directly controlling a computer screen.
Claude moved the mouse, clicked buttons, and typed on the keyboard.
This technology, named ‘Computer Use,’ was the moment the boundary between AI and software was erased for the first time.
In February 2025, ‘Hybrid Reasoning’ was introduced.
It was a system that adjusted the depth of thought in real-time based on the difficulty of the question.
Simultaneously released, Claude Code was an agent that performed complex programming tasks from the terminal using only natural language commands.
By 2026, Claude ascended to another dimension.
Opus 4.6 opened a massive 1 million token context window. The ‘Agent Team’ function, where multiple AI agents collaborate as a team, was permanently integrated.
-> This was a transition from a simple chatbot to a system that autonomously performs complex enterprise tasks.
And in April 2026, Anthropic created a model they decided not to disclose to the world.
IV. Claude Mythos: The Unreleasable Weapon
Project Glasswing.
Few people know this name. There’s a reason Anthropic decided to keep it strictly confidential.
Claude Mythos Preview is a special model designed for the cybersecurity field.
It’s not just a security diagnostic tool. It’s a neural network system that autonomously analyzes entire software systems and probes for potential vulnerabilities.
The results Mythos showed in its preliminary safety tests were too sensitive to be disclosed.
It analyzed the code of major operating systems and web browsers deployed worldwide and found thousands of previously undiscovered zero-day vulnerabilities in just a few weeks.
More shocking is the fact that it unearthed a vulnerability hidden for 27 years in the OpenBSD kernel, which had been fending off hackers globally for decades.
Its success rate in generating complex exploits was over 83%.
To understand what this number means, we need to pause for a moment.
It’s not just about finding vulnerabilities. It means that when attempting to connect multiple small vulnerabilities to build a complete hacking pathway, it succeeds more than 8 out of 10 times.
This model automates tasks that would take state-level hacker groups months to accomplish.
Anthropic chose containment. They permanently froze the public release of Mythos to the general public and developer community.
Instead, they launched Project Glasswing, a private alliance composed exclusively of national core infrastructure operators and top-tier defense companies that have passed rigorous government regulatory approvals.
This decision itself is a message.
Anthropic has consistently stated that its goal is not to create more powerful AI.
Its goal is to create AI that can be safely controlled.
In a way that proves that claim, they have put their most powerful tool into a drawer.
Companies that understand this decision have invested in Anthropic.
V. The Memory Bottleneck: The Unspoken True Limit of AI
Most discussions about the AI industry revolve around algorithms, parameters, and data.
But one of the most crucial stories is rarely discussed.
Electricity and memory.
When large language models perform inference, they constantly move billions of parameters between memory and computational units.
The speed of this movement determines the practical limit of the entire system.
No matter how fast the GPUs or TPUs are, if the memory cannot supply data in time, the computational power cannot be utilized more than half its capacity.
This is the structural position of High Bandwidth Memory (HBM) in AI infrastructure.
HBM is structured by stacking thin DRAM chips vertically and then penetrating them with ultra-dense interconnects (TSVs, Through-Silicon Via) to form a single block. Its data transfer speed is dozens of times faster than regular memory. However, its manufacturing difficulty is extremely high.
It requires aligning hundreds of micro-bumps with precision of tens of micrometers and drilling thousands of micro-holes vertically through the entire silicon chip.
SK Hynix holds a monopolistic position in this market.
Over 50% market share in the HBM market. A major supplier of HBM for Nvidia’s H100 and H200. This position stems not only from technological superiority but more fundamentally from time. It’s the result of running far ahead before competitors could catch up.
Anthropic’s pursuit of 5GW to 10GW of data center infrastructure. Thousands of AI accelerator cards will be plugged into this. HBM chips will be mounted on each card. With SK Hynix already securing most of the supply for this volume, Anthropic’s equity investment is a financial lock on that relationship.
The nature of negotiations changes the moment a supplier becomes a shareholder of its customer.
VI. Samsung Electronics’ Different Calculation: The Single Word ‘Logic Chip’
While SK Hynix focused on securing memory supply volume, Samsung Electronics aimed for something else.
A single word in the announcement sent a shiver through industry insiders. Among the list of suppliers that Anthropic mentioned in explaining its partnership with Samsung Electronics was ‘Logic Chip.’
Micron is a memory-specialized company. SK Hynix is also. Neither company possesses foundry facilities. Samsung Electronics is the only one on this strategic partner list with massive fabs capable of directly mass-producing logic semiconductors using ultra-fine advanced processes.
This was a declaration.
It signified that Anthropic has designated Samsung Foundry as its key manufacturing partner when designing and mass-producing its own AI-specific accelerator chips (ASICs).
Samsung Foundry’s current position is stark. As of 2025, its market share is 7.2%. This is incomparable to TSMC’s 69.9%.
Years of advanced process yield issues, major customer departures, and continuous deficits in the foundry business were the objective reality.
However, the tide has recently turned.
They secured orders for Tesla’s next-generation autonomous driving AI processors, AI5 and AI6. They were selected as the production base for Grok’s inference processor, acquired by Nvidia.
And now, the possibility of adding Anthropic to their order list has opened up.
These three cases have something in common.
They are all chips for AI computation. This is a signal that Samsung Foundry is concretizing its positioning as an ‘AI-specific silicon manufacturer.’
VII. One-Stop Turnkey: What Only Samsung Can Offer
Samsung Electronics’ logic against TSMC is simple: We can do everything in-house.
Logic chip manufacturing (foundry), HBM memory production, and advanced packaging to integrate these two into a single package.
Samsung can handle these three within a single fab complex. This is the ‘One-Stop Turnkey’ strategy.
TSMC is the world’s best foundry. However, TSMC does not make memory. To complete an AI accelerator of Anthropic’s scale, TSMC would need to procure HBM from Samsung or SK Hynix. The complexity of this logistics and coordination, in the current situation of chronic AI semiconductor shortages, leads to severe delivery delays.
If Samsung can independently deliver finished accelerator packages,
Anthropic is freed from the complexity of managing a dual supply chain, dealing with TSMC and SK Hynix separately.
This is why it’s not purely a debate about technological superiority.
Even if process yields are lower than TSMC’s, supply chain simplification and delivery predictability offer different kinds of value to customers.
And for companies that need to build large-scale data centers, the cost of delivery uncertainty can be far greater than the difference in unit performance.
There isn’t only one way to beat TSMC. It’s not always about having superior processes.
It’s also possible to create a market by offering value that TSMC does not. This is the logic Samsung is betting on.
VIII. SK Telecom’s Three-Year Gamble: How $100 Million Became 4 Trillion Won
To understand this massive investment, we need to go back three years.
In 2023, SK Telecom invested $100 million in Anthropic.
At the time, Anthropic’s valuation was $5 billion. Claude 1.0 hadn’t even been released yet, and the investment fervor in generative AI hadn’t reached its current level.
In the first half of 2026, the book value of SK Telecom’s Anthropic shares exceeds 4 trillion won. This is a return close to 30 times the initial investment.
It was the biggest jackpot in the history of the Korean IT industry for a single unlisted venture investment.
However, what’s more noteworthy in this story is not the size of the profit, but its cause.
SK Telecom’s choice of Anthropic in 2023 was not just a financial bet.
SK Telecom simultaneously signed a technical partnership with Anthropic to jointly develop a telecommunications provider-tailored language model.
They organized the ‘Global Telco AI Alliance (GTAA)’ centered around Anthropic, connecting global telecom companies like SoftBank, Singtel, and Deutsche Telekom.
By investing money, they also secured a position within the value chain.
Financial investment was bundled with strategic partnership.
This is the same logic behind Samsung Electronics and SK Hynix participating in this round under the same title of ‘Strategic Infrastructure Partner.’
The goal is not to recoup the investment.
The goal is to position themselves within the structure of Anthropic’s growth as it expands.
IX. The 5GW War: A Battle for Energy and Silicon Supply Chains
A significant portion of the $65 billion Anthropic raised will go directly into data center infrastructure.
An exclusive data center capacity contract with Amazon Web Services (AWS) for up to 5GW.
Another 5GW dedicated distributed computing network being pursued with Google and Broadcom.
Combined, these two pillars require up to 10GW of power demand solely for Anthropic’s AI operations.
10GW. That’s equivalent to the output of 10 nuclear power plants.
To realize this scale, Anthropic knows it must reduce its reliance on Nvidia GPUs.
Nvidia exercises absolute supply control in the global AI accelerator market and dictates prices.
When this much infrastructure is concentrated in one company, its operating costs become directly dependent on the GPU manufacturer’s pricing.
This is the background for Anthropic’s internal review of developing its own custom AI semiconductor (ASIC).
Just as Google created TPUs and Amazon created Trainium, designing a chip optimized for Anthropic’s own model characteristics could be cheaper and faster in the long run.
The entity designing that chip is Anthropic. If Samsung Foundry becomes the entity manufacturing that chip, the logic of this investment is complete.
X. The Ripple Effect: Signals to KOSDAQ Equipment Companies
When large-scale semiconductor facility investments are decided, the impact isn’t limited to semiconductor companies.
When Samsung Electronics or SK Hynix expands new lines or upgrades existing ones,
The equipment installed in them is mostly supplied by small and medium-sized equipment companies with specialized technologies.
These companies’ very survival depends on the investment decisions of the prime contractors.
HPSP possesses the unique technology for ‘High-Pressure Hydrogen Annealing (HPA)’ equipment, which heats and processes materials under high pressure exceeding 20 atmospheres using 100% hydrogen gas. This is a core process equipment that heals defects at transistor interfaces in ultra-fine processes below 3nm. The exclusivity of this technology enables an average operating profit margin of 53.7% per quarter.
PSK Holdings is at the top of the market for reflow and cleaning equipment in advanced packaging processes. These are machines that precisely melt fine solder bumps and remove organic contaminants in the packaging process, which physically combines logic chips and HBM for AI accelerators. Anthropic’s data center expansion directly increases demand for packaging processes.
Hanmi Semiconductor is a company at the root of SK Hynix’s HBM manufacturing. It leads the global market for ‘Dual TC Bonder,’ a key HBM equipment that stacks thinly sliced DRAM chips vertically. As HBM production volume increases, so does the demand for this equipment.
The stock prices of these companies react proactively to large investment decisions by Samsung Electronics or SK Hynix.
Before contract announcements are made, the market already reads this direction.
XI. Three Futures: Paths by Scenario
Investment is always a bet on uncertainty.
Let’s simulate the impact of Anthropic’s large-scale investment on the Korean semiconductor ecosystem through three paths.
Scenario A: Best Case — “Korea Will Outsource AI Brains”
Samsung Foundry’s GAA 2nm process yield fully stabilizes, and an exclusive contract for manufacturing Anthropic’s next-generation Claude-specific accelerator logic chip is signed. SK Hynix’s HBM4 samples pass Anthropic’s data center suitability tests independently, securing a long-term exclusive supply contract.
In this case, Samsung Electronics resolves its persistent foundry deficit and begins a rally towards a stock price of over 100,000 won. SK Hynix’s annual operating profit reaches its peak. Key equipment companies like HPSP and PSK Holdings are revalued with a corporate value multiple of over 35x.
Scenario B: Neutral Case — “Territorial Expansion Amidst Stable Dualization”
Anthropic pursues a dualization strategy, splitting next-generation accelerator orders between Samsung and TSMC to diversify risk. SK Hynix maintains its dominance in the HBM market with over 60% share.
Samsung Electronics’ foundry break-even point converges, and its stock price attempts to break through the upper bound of the 75,000-85,000 won range.
Scenario C: Worst Case — “The Gauntlet of Process Yields”
Samsung Foundry’s recovery of advanced GAA process yields falls significantly short of expectations, delaying delivery schedules. Anthropic postpones the adoption of its custom semiconductors and relies entirely on Google TPUs and TSMC’s sole production line.
Samsung Electronics faces renewed pressure for foundry restructuring, struggling precariously around the 60,000 won support level. KOSDAQ semiconductor small and mid-cap stocks experience sharp corrections of 30-40%.
Which of these three paths will be realized is not yet decided. However, one thing is common across all three paths.
The fact that Korean semiconductor companies have entered the inner circle of AI infrastructure decision-making through this investment remains unchanged.
XII. Value Chain Dynamics: Why Component Manufacturers Have Always Been at the Bottom
A one-sentence summary of the Korean semiconductor industry’s history: It creates the world’s best components but has never sat in a position to determine the value of those components.
This was not a problem of technology. It was a problem of structure.
There is an implicit hierarchy in the semiconductor industry’s value chain.
Designers (fabless) are at the top. They decide what to make, set performance standards, and control prices.
Manufacturers (foundries) are in the middle. They implement the designs they are commissioned to produce.
Memory manufacturers are at the bottom. They produce standardized products, and prices are determined by supply and demand cycles.
For a long time, Samsung Electronics and SK Hynix occupied the positions of memory manufacturers and foundries in this hierarchy.
No matter how advanced their technology, someone above decided what to make.
The equity investment in Anthropic is the moment this structure begins to crack for the first time.
Becoming a shareholder is not just about putting in money.
It means sitting in a position where you receive quarterly performance reports, communicate directly with management, and negotiate technical requirements from the early stages of the next-generation product roadmap.
It’s a move from a position where designers unilaterally dictate memory and foundry specifications to one where manufacturers can influence the direction of design.
This shift may seem minor. However, in an industrial hierarchy with decades of history, such a shift is akin to a revolution.
XIII. Escaping Nvidia Dependency: The Geopolitics of Supply Chain Self-Reliance
Nvidia controls 80% of the global AI accelerator market.
When a company like Anthropic, which needs to operate large-scale infrastructure, considers what this number means, it signifies a lack of options.
Even if they don’t want to use H100 chips, there’s no realistic alternative. Even if the price of H100 skyrockets, they cannot go elsewhere.
OpenAI, Google, Meta, and Anthropic are all grappling with the same question: How to break free from this dependency. The answer lies in designing their own chips.
Google began designing TPUs (Tensor Processing Units) a decade ago.
Amazon created Trainium and Inferentia.
Meta developed MTIA.
Anthropic’s consideration of custom ASIC development is an extension of this trend.
The logic of custom chip design is simple.
I know my model’s characteristics best.
Only I know which operations occur most frequently in Claude’s inference process, and which memory access patterns create bottlenecks.
A chip fully customized to those characteristics can be far superior in terms of energy efficiency and processing speed compared to a general-purpose GPU.
The problem is where to manufacture the designed chip. While TSMC is a natural choice, Samsung Electronics is offering other values like one-stop turnkey and strategic alliances.
This competition is not just a war of technology.
It’s also a war of supply chain trust, geopolitical risk diversification, and strategic relationships.
For customers conscious of the geopolitical risks associated with Taiwan, Korea offers a meaningful alternative.
XIV. The ‘Safety’ Brand: Why Anthropic Became Money
There’s one last question we need to return to.
Why Anthropic? Why not OpenAI or Google?
Technical superiority alone doesn’t explain it.
It’s difficult to claim that Claude is superior to GPT-4o or Gemini in every aspect. The competition continues, with advancements in certain benchmarks and setbacks in others.
The difference lies elsewhere: Trust.
When large financial institutions, healthcare systems, and government agencies worldwide adopt generative AI, their biggest concern is not performance. It’s controllability.
It’s a guarantee that the system will not operate in unpredictable ways. It’s confidence that their data is secure.
Constitutional AI is the technical basis for that guarantee.
Anthropic’s decision to contain the world’s most powerful cybersecurity tool is behavioral evidence of that guarantee.
Samsung Electronics and SK Hynix becoming shareholders in this company is not just about securing semiconductor volume.
It’s about becoming a beneficiary of the ‘safety’ brand value.
The more the world’s companies and governments trust Anthropic, the higher the value of the semiconductors Anthropic needs will also rise.
A structure where suppliers are financially linked to their customers’ growth. That is the true meaning of this investment.
XV. Return: To the Original Question
We said some deals change direction, not just numbers.
The contract signed by Samsung Electronics and SK Hynix with Anthropic is a promise to supply semiconductors, and a promise to invest money. But that’s not all.
It is a declaration by companies that, for decades, were in the position of implementing someone else’s designs, to step for the first time from the outside of design to the inside of design.
Of course, there’s no guarantee this will succeed.
Samsung Foundry’s yield issues remain a reality to be solved.
TSMC’s wall is still high. Whether Anthropic will actually proceed with custom chip design is also not yet confirmed.
However, the direction has become clear.
At this turning point where global AI infrastructure is shifting from a war of software to a war of hardware, Korean semiconductors have secured a position as participants, not observers.
We don’t yet know how much value that position will create.
But it’s clear that they can decide far more than they could have if they weren’t in that position.
And perhaps this is the beginning of a bigger question.
When the maker of the silicon that powers systems replacing or augmenting human cognition stands in a position to influence the direction of those systems—where lies the boundary between the physical producer of what we call ‘intelligence’ and its intellectual designer?
Silicon doesn’t think. But the makers of silicon will increasingly be involved in deciding what thinking beings can do.
References
- Constitutional AI: Technical Mechanisms and Ethical Guidelines for AI Alignment — Anthropic Research (Dario Amodei)
- Revolution and Threats in Cybersecurity Paradigms: The Background of Project Glasswing’s Launch — ArmorCode Security Research Institute (Nikhil Gupta)
- Global Foundry Market Analysis and Outlook for Advanced GAA Process Adoption — Digitimes Research, Taiwan (Minji Kim)
- HBM and Advanced Packaging Technology Roadmaps and Silicon Valley Big Tech Supply Chain Changes — Korea Semiconductor Industry Association Policy Report (Sang-yi Byeon)
- Global Telco AI Alliance (GTAA)’s Strategy for Creating Custom Business Based on Generative AI — SK Telecom Economic Research Institute (Jae-heon Jeong)
- Anthropic Series H Funding Round: Strategic Analysis and Market Implications — Sequoia Capital Research Division
- Advanced Packaging Technology Landscape: CoWoS, HBM Integration, and the AI Chip Race — SEMI Industry Report 2026
- Constitutional AI: Harmlessness from AI Feedback — Anthropic Research (Bai et al., 2022)
- Samsung Foundry vs TSMC: Competitive Dynamics in the 3nm and 2nm Era — Morgan Stanley Semiconductor Research
- High-Bandwidth Memory Architecture and the AI Inference Bottleneck — IEEE Solid-State Circuits Journal
- Global AI Infrastructure Investment: Data Center Power Demand and Supply Chain Implications — BloombergNEF Energy Storage & AI Report 2026
- SK Telecom’s Anthropic Investment and the Telco AI Alliance Model — Goldman Sachs Asia Technology Research
- The Economics of Custom AI Silicon: Why Hyperscalers Build Their Own Chips — Bernstein Research Special Report
- Zero-Day Vulnerability Landscape and AI-Assisted Cybersecurity — CrowdStrike Global Threat Intelligence Report 2026
- Korean Semiconductor Equipment Ecosystem: HPA, Advanced Packaging, and the AI Demand Cycle — KB Securities Semiconductor Deep Dive Report