Semiconductor stocks plummet, yet Anthropic wants to create a 2nm chip

Written by Boyang, Tencent Technology

Anthropic is developing its own chips and, like Musk, has chosen Samsung over TSMC.

The Information cites three informed sources stating that Anthropic has initiated early work on self-developed AI chips and is discussing with Samsung to manufacture the chips using the latter's advanced 2nm process and packaging technology.

The aforementioned sources said that Anthropic is also determining the chip's functions, performance metrics, and its adaptation schemes in servers or clusters. Currently, the company has started collaborating with several design service providers, but it is far from detailed design, testing, and mass production. The sources also reminded that Anthropic may still abandon the project midway.

Public information shows that Anthropic's chip team has already begun to form. For example, the company hired Clive Chan, an early member of OpenAI's custom chip team, on June 9.

Clive Chan, an early member of OpenAI's custom chip team, has joined Anthropic.

Anthropic has been very cautious in publicly stating its self-developed chip plans, emphasizing that future computing power expansion will still primarily rely on AWS's Trainium, Google's TPU, and NVIDIA's GPU, while not commenting on its own chip roadmap.

In fact, the idea of Anthropic developing its own chips had been rumored for months. Reuters reported in April that the computing demands of Claude were growing too fast, and available computing power could not keep up, prompting Anthropic to consider making its own chips. At that time, it was still just a preliminary idea. Now, it seems that things have progressed a step further.

It is worth noting that the news of Anthropic's chip development comes at a time when global chip stocks have experienced a sharp decline over two trading days. The Philadelphia Semiconductor Index has seen a cumulative drop of nearly 12% over two days, the largest since June 5. Memory chip stocks have been particularly hard hit, with Samsung Electronics dropping 9.1% overnight and SanDisk down 14.1%.

01 A Foreshadowing from Financing and a Breakthrough in Samsung's Foundry Business

It is not surprising that Samsung has become a partner of Anthropic.

In May of this year, Anthropic completed a $65 billion Series H financing round, with Samsung Electronics participating as a strategic investor, along with memory chip giants SK Hynix and Micron Technology.

At that time, the supply of memory chips was tight, forcing consumer electronics manufacturers like Apple to raise prices. One direct benefit of this investment was that it tied Anthropic to the chip suppliers essential for its growth.

However, Samsung's calculations go beyond financial returns.

The company is a leader in the memory chip field but has lagged behind TSMC in the foundry sector. TSMC's advanced production lines remain the industry standard for cutting-edge AI chips, but the current demand for AI chips has strained TSMC's capacity, providing Samsung with an opportunity to attract customers using its 2nm technology.

According to The Information, Google is already considering having Samsung produce some of its future TPUs. Securing Anthropic as a client would be a significant victory for Samsung's foundry business.

Two informed sources revealed that Anthropic is evaluating Samsung's 2nm manufacturing process and advanced packaging facilities. "2nm" is one of the most advanced chip manufacturing technologies in the industry, allowing for denser and more energy-efficient processors. Advanced packaging places the main processor closer to high-speed memory, enabling faster data movement within the chip.

Samsung is a manufacturing partner for NVIDIA's HBM memory and LPU chips, and in turn, Samsung relies on NVIDIA's software to produce chips. The two companies are jointly building an AI chip factory in South Korea. If Anthropic joins the client list, Samsung gains another chip in its hand to catch up with TSMC in advanced processes.

Samsung's willingness to pursue Anthropic is backed by South Korea's national commitment to the semiconductor industry.

The South Korean government recently announced a ten-year investment plan led by Samsung Group (the parent company of Samsung Electronics) and SK Group (the parent company of SK Hynix), with a total investment of $518 billion to build four memory chip factories in South Korea.

On July 2 alone, Samsung announced an investment plan for the Chungcheong region, exceeding 140 trillion won, approximately $90 billion. Among these, Samsung Electronics will build a next-generation HBM production base in Chungcheong and Wonju, adding five production lines. Since the AI boom, HBM has been in short supply, and after the expansion, Samsung could increase revenue by billions of dollars in the coming years. Samsung Electro-Mechanics is also expanding its AI server packaging substrate capacity in Sejong City while investing in advanced packaging technology research and development.

Samsung's participation in Anthropic's financing round is a bet on becoming a long-term hardware partner for the latter.

02 Inference Chips: A Breakthrough Point in a Money-Burning Loop

Anthropic's first 2nm chip is expected to focus on inference, similar to OpenAI.

Custom inference chips can allow Anthropic's models to run faster, consume less power, and save on cloud billing, while also enabling hardware optimization specifically for Claude.

Previously, OpenAI chose Broadcom to design a chip named "Jalapeño," aimed at more efficiently running large language models. OpenAI claims that this chip has higher energy efficiency and better performance per watt than competing products.

OpenAI collaborates with Broadcom to design the inference chip named "Jalapeño."

Google has TPUs, Amazon has Trainium, Microsoft has Maia, and Meta is also developing its own. Anthropic is nearly the last major AI lab still fully relying on others' chips. The temptation for autonomy in this field is too great.

In February of this year, Anthropic CEO Dario Amodei discussed the logic of computing power decisions in an interview with podcast host Dwarkesh Patel.

Amodei said, "If I assume revenue continues to grow tenfold each year, by the end of 2026 it would be $100 billion, and by the end of 2027 it would be $1 trillion. Based on that, I could order $1 trillion worth of computing power. But what if I'm wrong? What if the growth multiplier is 5 times instead of 10 times? If revenue in 2027 is only $800 billion instead of $1 trillion, then I would go bankrupt. No hedge could save me."

Therefore, Amodei must find a balance for Anthropic: enough computing power must be purchased to ensure it can handle high growth, but it must not over-expand to the point where a slowdown in growth would crush it under early investments.

Amodei revealed that the global computing power built this year is about 10 to 15 gigawatts, roughly tripling each year. By 2028 or 2029, the industry will be investing tens of trillions of dollars annually in computing power. He believes this figure reflects the trend the industry is moving towards.

Developing self-owned chips is like adding chips to this gamble.

On one hand, self-developed chips can save a significant amount of costs in a cluster of several gigawatts by improving inference efficiency. On the other hand, when everyone is scrambling for processors, data centers, and electricity, having one's own chips provides more leverage in negotiations.

03 Supply Chain "Matryoshka": Who is Dependent on Whom?

Anthropic wants chip independence, but this web is becoming increasingly intricate.

Programmer Benny Lam pointed out a phenomenon on social media: every American AI lab is looking for Asian foundries to collaborate on custom chips, resulting in not breaking free from dependency but rather tightening the supply chain.

Anthropic needs Samsung's wafer fabs to reduce its dependence on NVIDIA, while NVIDIA needs Samsung's wafer fabs to meet its expanding capacity demands. Samsung, in turn, needs to retain both major clients to keep its foundry business competitive against TSMC.

The three parties are tied together, and none can do without the others.

Lam commented that the trend of custom chips should have allowed AI labs to gain more independence, but every new collaboration between American labs and Asian foundries is making this network denser and harder to dismantle.

Tech blogger Vaibhav Sisinty also noted Samsung's unique role in this game.

He wrote that Samsung's investment in Anthropic's $65 billion financing round is not just for financial returns. Samsung is a chip manufacturer, memory supplier, and advanced packaging factory; its investment is a bet on becoming Anthropic's hardware partner. With millions of queries daily, Anthropic must pay for computing resources each time it calls, resulting in astronomical bills. Whoever can control their own inference chips controls the speed and cost of scaling.

04 NVIDIA's Fortress Remains Strong

Designing a competitive AI chip is extremely complex, involving architecture, software, manufacturing, testing, and large-scale deployment; each link cannot be rushed. Anthropic cannot shake off its existing hardware partners in the short term.

From Anthropic's current computing power layout, its strategy is already quite clear. The company has consistently adhered to a multi-cloud rental approach and has not tied itself exclusively to NVIDIA's hardware like OpenAI or xAI. It uses chips from Amazon, Google, and NVIDIA simultaneously and is negotiating chip collaborations with Microsoft and the UK startup Fractile.

Anthropic's ties with Amazon and Google are particularly deep.

In April 2026, Anthropic expanded its cooperation with AWS, committing to invest over $100 billion in the next decade, securing 5 gigawatts of new computing power capacity, covering custom chips like Trainium2, Trainium3, and Trainium4. Amazon has cumulatively invested tens of billions of dollars in Anthropic.

At the same time, Anthropic signed long-term agreements with Google and Broadcom to obtain approximately 3.5 gigawatts of next-generation TPU computing power starting in 2027. Previously, in 2025, the two parties had already signed agreements worth hundreds of billions of dollars, involving millions of TPUs and over 1 gigawatt of capacity.

Claude is also one of the few cutting-edge models that can be used on AWS, Google Cloud, and Azure simultaneously.

And it doesn't stop there. In May 2026, Anthropic rented the entire capacity of the SpaceX Colossus 1 computing cluster from xAI.

This cluster, located in the Memphis data center, houses over 220,000 NVIDIA GPUs, with a power exceeding 300 megawatts, and a monthly rent of about $1.25 billion. The total contract value is worth hundreds of billions of dollars, extending to around 2029. The two parties are even discussing the idea of deploying multi-gigawatt computing power in orbit.

A broad layout and parallel multi-supplier strategy is Anthropic's way of diversifying risk.

But currently, the dealer at the table is still NVIDIA. According to The Information, despite the ongoing increase in financing and design activities in the inference chip market, NVIDIA's market share has actually risen to 74% in recent years. CEO Jensen Huang has consistently asserted that NVIDIA's chips are more efficient than any alternatives for inference tasks.