Anthropic warns of memory supply risks in $637 billion infrastructure plan

The supply chain bottleneck

Anthropic, the developer of the Claude AI model, recently filed a registration statement with the U.S. Securities and Exchange Commission. In this filing, the company explicitly identified a potential shortage of high-bandwidth memory (HBM) as a significant risk factor. HBM is a type of high-performance memory used specifically in AI computing infrastructure to handle large amounts of data quickly.

The company named three specific manufacturers as the primary sources for this critical component: Samsung Electronics, SK hynix, and Micron Technology. Anthropic warned that these suppliers might not be able to meet its surging demand on time, even as the manufacturers work to expand their production capacity. This highlights a growing vulnerability in the AI supply chain where a few companies control the most essential components.

According to reports from Seoul Economic Daily, Anthropic expressed concern that the shortage could be exacerbated by market competition. The company noted that large customers or other competitors could preemptively lock up a particular supplier's output. This creates a situation where the sheer scale of AI development could lead to a 'winner-takes-all' scenario for the limited memory hardware available on the market.

A massive infrastructure spend

The scale of Anthropic's ambitions is reflected in its disclosed financial commitments. A simple sum of the company's seven major AI computing infrastructure contracts exceeds $637 billion, or roughly 845 trillion won. These contracts represent a massive investment in the physical backbone required to run and train modern generative AI models.

The largest single commitment in this group is a lease agreement for AI accelerator equipment with the U.S. chipmaker Broadcom. This deal is worth about $161.2 billion, or roughly 214 trillion won. Additionally, Anthropic signed cloud service agreements with major tech giants, including at least $114.5 billion with Google and at least $110 billion with Amazon.

The company also maintains a contract with Microsoft worth at least $44.9 billion. Other agreements included in the filing are with SpaceXAI, Nscale, and Lambda. While these figures are not fixed investment totals because they include minimum spending obligations and variable volume payments, they illustrate the enormous capital required to secure high-performance computing power.

Diversifying the hardware mix

Anthropic is shifting its strategy regarding how it secures and utilizes computing power. While the company previously relied primarily on cloud services from major technology firms, it has recently begun diversifying its infrastructure. This includes leasing AI accelerators separately and deploying them within dedicated data centers to gain more control over its resources.

The company's hardware strategy involves a wide variety of chips rather than relying on a single provider. It currently uses a range of processors, including Amazon's Trainium, Google's tensor processing units, and graphics processing units from both Nvidia and AMD. This diversification is intended to mitigate risks associated with any single chip manufacturer's production capacity or pricing.

As AI infrastructure spending spreads across these multiple types of chips, the demand for high-performance memory like HBM is expected to rise proportionally. Because HBM is a core component in many of these different types of AI accelerators, the memory shortage remains a central concern regardless of which chip brand Anthropic chooses to prioritize. The growth in demand is directly tied to the increasing complexity and scale of the models being developed.

Rising costs and market impact

The financial reality of building this infrastructure is reflected in Anthropic's rapidly climbing costs. In the first half of this year, the company's computing and infrastructure costs reached $8.945 billion, which is a 283% increase compared to the same period last year. This sharp rise underscores the difficulty and expense of securing the hardware necessary to stay competitive in the AI space.

Despite these high costs, the company's revenue is growing even faster. During the same six-month period, revenue rose to $16.286 billion, up from $1.163 billion a year earlier. This suggests that while the cost of entry and operation is skyrocketing, the commercial demand for the services Anthropic provides is accelerating at an even higher rate.

The massive spending by AI companies means that memory demand is growing in lockstep with these investments. This trend provides a potential boost for Korean memory makers like Samsung and SK hynix. For example, Samsung recently signed a memorandum of understanding with Broadcom to pursue cooperation worth more than $200 billion in memory and foundry operations through 2030, specifically looking at HBM supply for next-generation accelerators.