The Hidden Risks Behind Anthropic's Nearly $100 Billion Gamble
Anthropic's CEO Dario Amodei once wrote that to create a sufficiently powerful AI is like "gathering a group of geniuses in a data center." However, the company's latest disclosed IPO prospectus reveals another side of reality: to achieve this vision, Anthropic is relying on a fragile and high-risk financing mechanism.
Anthropic's CEO Dario Amodei. Image source: Oliver Contreras---AFP via Getty Images
This week, Anthropic updated its prospectus, adding third-quarter financial data. Affected by the continuous rise in computing power costs, the company's profit margins are narrowing; at the same time, the company disclosed additional commitments for data center construction expenditures: planning to invest up to $895.4 billion by 2045.
Some suppliers providing chips and cloud services to Anthropic are also financing parties for related procurement projects. For example, Broadcom finalized a $63 billion equipment financing agreement: if Anthropic defaults, Broadcom will bear the corresponding losses; after all equipment is delivered and put into use, Broadcom's maximum potential payout could reach about $43 billion.
Relying on such backstop clauses, Anthropic can obtain computing resources without paying the full price of the equipment upfront. Google and AMD have also linked part of their funding to Anthropic's leasing or deployment of their hardware. However, once AI demand declines, hardware manufacturers, while facing pressure on their own revenue and market valuations, will also need to fulfill their guarantee obligations, paying funds to lenders and partners.
Who bears the risk?
In the $63 billion transaction announced on October 26, Bank of America, Citigroup, Morgan Stanley, and Blackstone will jointly fund the establishment of a new company responsible for procuring Broadcom chips and network equipment, which will then be leased to Anthropic, with equipment expected to be delivered in 2027. Anthropic will pay rent in installments, while investors will rely on rental income to recover their principal.
If Anthropic is unable to pay rent on time and defaults, Broadcom will make up the difference as agreed: 72.5% of the remaining lease debt minus the proceeds from selling the equipment. For example, if the remaining debt is $100 and the equipment sale only recovers $60, Broadcom would need to compensate $12.5.
D.A. Davidson's head of technology research, Gil Luria, told Fortune magazine that Broadcom had to reach such agreements, or else orders would be taken by competitors like Nvidia, which also provides financing guarantees for top chip procurement customers. Luria stated, "To avoid falling behind in the competition for customers supported by Nvidia, large semiconductor manufacturers like Broadcom and AMD also have to follow suit and provide support to buyers." He upgraded Broadcom's rating to "Buy" on October 27.
After a debt default occurs, whether investors can recover their funds largely depends on whether these chips can be subleased to other customers. Andrew Judich of credit rating agency KBRA, after comparing Nvidia GPUs with Broadcom's Google-specific AI chip TPU, told Fortune magazine, "Custom TPUs are designed for a more specialized ecosystem, and the potential user base is more limited, so the risk of subleasing is higher compared to Nvidia GPUs."
Judich pointed out that Broadcom's residual value backstop agreement can help investors recover funds when equipment sale or sublease income falls short of expectations, "but essentially this shifts part of the risk to the guarantor." Therefore, when evaluating this financing, it is crucial to examine the credit status of the guarantor company, the legal effectiveness of the agreement terms, and the coverage of compensation.
Broadcom's CEO Hock Tan. Broadcom has agreed to provide guarantees for Anthropic in a financing agreement aimed at supporting the AI company's procurement of Broadcom-produced AI chips TPU. Image source: David Paul Morris---Bloomberg via Getty Images
Stein Van Nieuwerburgh, a professor of real estate and finance at Columbia Business School, believes that hardware manufacturers with diversified main businesses, abundant cash flow, and not solely relying on AI business have the capacity to absorb losses. However, if AI market demand slows down, it will not only trigger guarantee payouts but also impact the hardware sales, cloud business revenue, and investment valuations of the aforementioned partners. "Independent contracts will transmit the same potential demand shock to the balance sheets of multiple companies. A single loan may seem safer, but the overall system's risk concentration is rising and harder to assess."
Anthropic warns in its prospectus that the sustainability of its computing power expansion may heavily depend on whether a few cloud service providers, partners, and investors are willing and able to provide support. Among them, the two core partners, Amazon and Google, have already exceeded their operating cash flow in capital expenditures in the AI field. Over the past 12 months, Amazon's free cash flow was negative $42.4 billion, mainly due to increased investments in AI equipment; Google's parent company Alphabet reported a negative free cash flow of $9.1 billion in the third quarter. Although both companies still possess strong cash-generating capabilities and smooth financing channels, continuously providing blood transfusions to Anthropic will compete for funds with their own AI infrastructure projects.
Dario Amodei once described a scenario where an AI company's growth rate could be extremely fast, but the scale of computing power procurement still exceeds its capacity. In a podcast interview in February, he explained the financial risks with an extreme hypothesis: a company procures computing power based on the expectation of a tenfold increase in revenue each year; but if actual revenue only grows fivefold, it may be unable to settle the payment. Even if the future market space is huge, the growth rate may still lag behind the procurement arrangements made based on previous growth expectations. "If the forecast deviates by two or three years, it could be enough to drag the company down," he stated.
Complex financing conditions
Under an agreement reached in March 2026, Anthropic can issue up to $42 billion in convertible notes to Broadcom to pay for eligible quarterly rent, thereby preserving cash. Interest will be added to the debt principal, and noteholders can convert the notes into Anthropic equity at a discount.
Under normal circumstances, this batch of convertible notes has an annual interest rate of 9%, with interest added to the principal, and the conversion price set at 90% of the market price. If the agreed revenue targets are not met, the interest rate will directly rise to 13%, and the conversion price will be lowered to 85% of the market price, allowing the same notes to convert into more shares, further diluting the equity of existing shareholders. If performance thresholds are not met over the long term, Anthropic will completely lose its eligibility for subsequent financing. Anthropic stated that the revenue thresholds set by the agreement are far below recent actual growth rates: the company's revenue grew by 45% quarter-over-quarter and over 1000% year-over-year in the third quarter.
Google and AMD have also linked funding support to hardware delivery: when Anthropic signs the TPU leasing agreement, part of Google's funding can be activated; AMD's investment will be disbursed after GPU delivery and deployment milestones are met.
Van Nieuwerburgh told Fortune magazine, "Equity investments linked to hardware make Amazon and AMD bear risks, but do not make them full guarantors." If Anthropic is unable to repay, Broadcom and Google will bear losses within the limits set by the backstop agreement, while risks beyond that range will be borne by the lenders.
Anthropic stated that the company has continuously optimized the Claude large model over the years, enabling it to be compatible with chips from multiple manufacturers including Amazon, Google, Nvidia, and AMD, thus flexibly allocating computing power loads based on cost and performance. However, funds tied to a single supplier's equipment cannot be automatically transferred to other manufacturers. This may weaken Anthropic's flexibility in choosing chips for different workloads.
Can IPO fundraising provide greater flexibility?
As of September 30, Anthropic held $146.6 billion in liquid funds, covering cash, short-term investments, and available bank credit, sufficient to cover current payments and expansion expenses; however, if bank credit is utilized, the company's debt level will also rise.
Market analyst Tchinkar believes that ample liquidity is beneficial for bondholders, but whether Anthropic's debt can achieve an investment-grade rating "depends on how rating agencies assess its enormous future commitment expenditures, short operating history, and complex business relationships."
He stated, "The company has now disclosed a forward commitment amount of $895 billion, higher than our initial estimate of about $600 billion." The massive expenditure commitments and complex cooperation networks will limit the company's ability to obtain low-cost financing at investment-grade spreads (i.e., the lower premium that investment-grade borrowers typically pay over benchmark rates).
"The core issue is whether the compound growth rate of revenue can outpace the cost of computing power and cover future commitment expenditures." He continues to monitor two signals: whether Anthropic will further increase the forward commitment amount and how its liquidity situation evolves. "Utilizing revolving credit lines is also a potential warning signal," he added.
Anthropic has launched its own cloud computing infrastructure project, Anthropic Direct, separating data center site leasing from large-scale chip procurement, thereby reserving ample time to assess market demand before locking in equipment orders. The company can first secure dedicated infrastructure operated by third parties, and then finalize chip orders. The prospectus shows that the preparation cycle for data centers is longer than for internal hardware deployment. Once the self-built dedicated infrastructure is officially put into operation, the company can use low-cost long-term computing power to replace more expensive short-term leasing resources.
In the AI data center demonstration project modeled in Van Nieuwerburgh's paper, computing power equipment and supporting systems account for about two-thirds of the initial investment, with the remaining portion for the data center building and power supply.
Even if it first leases the data center and delays chip procurement, Anthropic will still need to pay for idle data center space. Van Nieuwerburgh pointed out that even if the power supply for the data center is secured, "if the final data center demand does not meet expectations, the data center will become stranded assets" (editor's note: fixed assets that lose economic value and cannot recover investment costs due to technological advancements, changes in market demand, or policy adjustments), meaning its business revenue cannot cover the investment costs, and it will be difficult to find a buyer.
Anthropic stated that to reduce reliance on partners for funding infrastructure and equipment, part of it depends on whether it can obtain financing from independent channels based on its own credit. Funds raised through the IPO can be directly used for computing power procurement without waiting for specific suppliers' hardware to reach delivery or deployment milestones. Additionally, the Anthropic Direct plan allows for deferred equipment procurement, enabling self-raised funds to reduce reliance on equipment suppliers' financing support when comparing hardware and financing options.
IPO fundraising will significantly enhance Anthropic's initiative in hardware selection and financing negotiations. But as Van Nieuwerburgh said, "While supplier financing can support infrastructure implementation, ultimately, repaying debt still relies on revenue generated by end customers." (Fortune Chinese Network)
· Anthropic's latest IPO prospectus shows that its cumulative data center expenditure commitments could reach $895.4 billion by 2045, heavily relying on a complex supplier financing backstop mechanism, reflecting the high financial leverage and risks behind the AI arms race, and the market should be wary of potential systemic crises under aggressive expansion.
To cope with competition, chip suppliers like Broadcom have transferred some market risks to their own balance sheets through equipment financing guarantees, revealing hidden concerns beneath the apparent prosperity of the AI industry chain: once demand falls short of expectations, the intertwined guarantee network can easily trigger a chain reaction.
Amazon and Google are facing pressure on free cash flow due to AI infrastructure, and the ongoing financial support for Anthropic will compete for capital with their own projects, highlighting the pain points faced by tech giants: as the contradiction between their own cash generation ability and AI capital expenditure intensifies, whether the compound revenue growth rate can outpace the enormous computing power costs will determine the sustainability of future commercialization.
Compiled by Wealth Chinese Network, with edits and adjustments to the original text.