Anthropic Warns AI Chip Shortages Could Constrain Growth as Infrastructure Deals Reach $637 Billion

Three memory suppliers sit at the heart of Anthropic's expansion

Anthropic has identified Samsung Electronics, SK hynix and Micron as critical supply risks in its filing for a US initial public offering, warning that shortages of advanced memory could restrict its ability to develop and operate Claude. The warning comes alongside seven disclosed computing infrastructure agreements whose stated amounts add up to about $637 billion.

The company is pursuing a stock market listing with a reported valuation target of up to $2 trillion. Its registration statement, submitted to the US Securities and Exchange Commission, describes a business whose growth depends on securing vast quantities of chips from a supply chain concentrated among a few manufacturers.

Samsung, SK hynix and Micron dominate production of high bandwidth memory, or HBM. This memory sits close to powerful AI processors and moves large amounts of information quickly. An accelerator can perform calculations at great speed, but it still needs memory to deliver the data those calculations require.

Anthropic warned that manufacturers could give other large customers priority or commit output through exclusive agreements. Even as suppliers expand factories, demand could grow faster than production. For Claude users and businesses building services around it, insufficient computing capacity could affect how quickly new models and products become available.

Revenue is rising faster than computing costs

Anthropic disclosed computing and infrastructure costs of $8.945 billion for the first half of its reporting year, compared with $2.334 billion in the corresponding period a year earlier. That is an increase of about 283%, or roughly 3.8 times the previous amount.

Revenue grew faster, reaching $16.286 billion from $1.163 billion. Those figures represent approximately fourteenfold growth. Computing and infrastructure costs were equivalent to about 55% of revenue in the latest period, compared with about 201% a year earlier.

The comparison shows that revenue has expanded faster than this particular expense category. Anthropic nevertheless needs to finance a much larger computing operation in absolute dollars: the increase in these costs alone was $6.611 billion.

Training new AI models and answering customer requests both consume computing resources. Anthropic must divide its capacity between developing future products and operating the services that generate revenue today. Its filing makes access to hardware part of the commercial challenge, alongside the quality of its models.

What the $637 billion in agreements covers

Anthropic's seven major disclosed infrastructure agreements span equipment leases, cloud services and access to computing capacity. Their stated amounts total about $637 billion, with different payment structures and conditions attached to individual arrangements.

The largest disclosed agreement is an approximately $161.2 billion lease for AI accelerator equipment with Broadcom. Google cloud agreements account for at least $114.5 billion, Amazon agreements for at least $110 billion, and Microsoft for at least $44.9 billion. Agreements involving SpaceXAI, Nscale and Lambda also form part of the total.

These arrangements combine minimum spending commitments, equipment lease payments and amounts that depend on the volume actually supplied. The filing left its aggregate computing spending commitment blank, while setting out individual agreements. The practical financial burden will therefore depend on each contract's terms, duration and delivery conditions.

Anthropic has traditionally used cloud services operated by major technology companies. It is now also leasing accelerators separately for deployment in dedicated data centers. That gives the company more ways to obtain computing power, while adding equipment and infrastructure obligations to its business.

More chip choices, shared manufacturing constraints

Anthropic uses Nvidia and AMD graphics processing units, Amazon's Trainium processors and Google's tensor processing units, or TPUs. All are specialized processors that can perform the large numbers of calculations needed for AI, although their designs and software differ.

Anthropic says it has developed technology to route workloads among different chip families, choosing the most efficient option for a particular job. The company described the benefit of this approach in its registration statement:

Our multi-chip compute approach provides us with substantially greater chip access than relying on one platform and enables us to opportunistically capture the price-performance benefits of each chip family and generation,

Access to several chip designers gives Anthropic more purchasing options. The manufacturing chain remains concentrated further upstream, however. Different accelerator designs can still depend on the same memory producers and the same advanced chip factories.

Earlier details of Nvidia's custom memory manufacturing plans illustrate that connection. Nvidia asked Samsung, SK hynix and Micron to use TSMC's 3 nanometer manufacturing process for base dies in its planned NVHBM memory. A base die is the chip underneath the memory stack that handles connections and logic functions. The proposal would connect several competing memory suppliers to the same manufacturer for an important component.

Taiwan remains a major point of exposure

Anthropic identified Taiwan Semiconductor Manufacturing Company, or TSMC, as another major supply risk because so much advanced chip manufacturing is concentrated there. TSMC manufactures accelerators designed by companies such as Nvidia, making its production capacity important even to AI companies that buy computing services rather than chips directly.

The filing said TSMC had previously been unable to satisfy advanced chip demand fully, affecting Anthropic's ability to obtain the chips it needed on time. It also listed earthquakes, typhoons, electricity and water constraints, and tensions between Taiwan and China as threats to supply.

A military conflict or blockade could disrupt deliveries for an extended period, according to Anthropic. Such an interruption would reach beyond a single supplier because processors, memory and packaging must arrive together before computing systems can operate.

TSMC has been expanding outside Taiwan. Its overseas factory expansion included production starting at its first Phoenix and Kumamoto facilities in late 2024, alongside construction in Dresden and a larger US investment program. Anthropic's warning concerns continued access to the advanced production capacity its computing plans require.

Samsung's Broadcom agreement connects memory to future demand

Samsung and Broadcom signed a memorandum of understanding in July to pursue more than $200 billion in cooperation involving memory and foundry operations through 2030. A foundry manufactures chips designed by other companies. The planned cooperation includes HBM for Broadcom's next generation of AI accelerators.

Broadcom's large equipment lease agreement with Anthropic creates a commercial connection between the AI company's expansion and Samsung's potential memory business. Neither the amount of HBM destined for Anthropic nor the value of individual delivery contracts has been confirmed.

The Samsung agreement is a memorandum setting out intended cooperation. Its scale points to the companies' ambitions, while specific orders and deliveries will determine how much business reaches the factories.

The disclosed agreements extend across several years:

• July: Samsung and Broadcom signed their memorandum to pursue memory and foundry cooperation.
• Through 2030: The Samsung and Broadcom cooperation plan targets more than $200 billion in business.
• Through 2033: Anthropic disclosed commitments of $114.5 billion tied to Google's TPUs.
• Through 2036: Anthropic disclosed commitments of $110 billion tied to Amazon's Trainium processors.

Anthropic also said it could buy more than $20 billion in AMD chips. Its arrangements with SpaceXAI, Nscale and Lambda provide additional routes to Nvidia computing capacity, broadening the set of companies involved in its infrastructure plans.

Forecasting demand creates a second financial risk

Anthropic said it has underestimated its computing requirements before, affecting product development and deployment plans. Forecasts are difficult because the company must predict both how much computing future models will require and how heavily customers will use them.

A shortage can force a company to ration capacity between research and customer services. Buying too much can leave expensive equipment or contracted capacity underused. Agreements lasting into the next decade require decisions well before future demand is known.

Anthropic described the competitive danger in its filing:

If we fail to secure sufficient compute, or if we fail to effectively direct the limited compute we have across our workloads to develop and deploy models with capabilities at least comparable to those of our competitors, we risk falling behind the technological frontier,

The technological frontier means the best capabilities currently available in the industry. For Anthropic, staying close to that level requires both enough computing capacity and decisions about which projects receive it. The filing links those operating choices directly to its ability to compete.

Trade policy and competitors can affect access

Anthropic warned that much of its hardware is manufactured outside the United States, exposing it to tariffs, import restrictions and trade disputes. These measures can add costs or change where equipment can be delivered and used.

US controls on exports of advanced chips and manufacturing equipment to China and other destinations create another source of uncertainty. Anthropic said tighter restrictions or policy changes could affect the availability and price of computing resources.

Competitors' investment decisions also matter. Rival AI companies can design chips, operate data centers or negotiate exclusive hardware agreements, securing capacity on terms that may be harder for others to match. This process is often called vertical integration: a business takes responsibility for more stages of producing and delivering its service.

Anthropic's relationships already connect investment with hardware supply. Nvidia, AMD, Amazon and Google hold stakes or options to acquire stakes in the company. Those ties place its computing purchases within a wider network of financial and strategic interests.

A hardware warning within a wider risk filing

Anthropic's registration statement also addressed the misuse of AI for cyberattacks, biological weapons and government surveillance. It warned about losing control of advanced models and about systems that could pose catastrophic risks.

The company separately identified the Trump administration's designation of Anthropic as a supply chain risk as an example of government action affecting its business. That designation concerns Anthropic itself; the chip procurement discussion focuses on its access to the equipment needed to train and run models.

No listing date was identified in the information disclosed about the IPO plans. The reported valuation target of up to $2 trillion places the company's rapid revenue growth alongside substantial infrastructure obligations and dependence on a small number of manufacturers.

Anthropic's central operating challenge is concrete: it must secure enough processors, memory and data center capacity, at manageable cost and on workable schedules, to keep developing Claude while serving a growing customer base. Its purchasing strategy spreads orders across several platforms, while its risk disclosures show how closely those platforms remain connected through manufacturing.

Key Points

• Anthropic named Samsung, SK hynix and Micron as key memory suppliers whose capacity and allocation decisions could constrain its growth.
• Seven major infrastructure agreements have stated amounts totaling about $637 billion, combining leases, spending commitments and amounts dependent on delivery volumes.
• First half computing and infrastructure costs rose 283% to $8.945 billion, while revenue reached $16.286 billion.
• The company uses chips from Nvidia, AMD, Amazon and Google, but remains exposed to shared memory and manufacturing constraints.
• TSMC's production concentration in Taiwan creates exposure to natural disasters, infrastructure problems and geopolitical disruption.
• Samsung and Broadcom plan more than $200 billion in cooperation through 2030; specific HBM deliveries for Anthropic have not been confirmed.
• Anthropic's disclosed computing commitments extend through 2033 for Google TPUs and 2036 for Amazon Trainium.