Anthropic's rating pressure and the cash flow strain of Big Tech's AI capital expenditure

On October 30, foreign media reported that global investment giant Blackstone changed its expectations for Anthropic's future credit rating. Previously, Blackstone believed that this AI company, which is about to go public, was expected to obtain an investment-grade rating by the end of this year; now it expects that even after completing an IPO, Anthropic may remain in the BB category, known as "junk grade."

On the same day, S&P changed its credit rating outlook for Meta from "stable" to "negative," warning that its expanding AI investments could cause free operating cash flow to remain negative for consecutive years.

Oracle had already been downgraded in July of this year, leaving it just one step away from junk grade.

Financial data are also beginning to show pressure.

In the first three quarters of this year, the combined operating cash flow of Microsoft, Alphabet, Amazon, Meta, and Oracle hit a record high, but capital expenditures nearly doubled, reaching $556.9 billion and exceeding operating cash flow for the first time. The five companies' combined free cash flow turned from positive to negative, and share buybacks plunged 83%.

Rating agencies are beginning to guard against the risks, while Wall Street is still continuing to lend. Increasingly complex financing arrangements are binding more companies' interests together with debt. But whether the massive investment in AI can ultimately translate into sufficient cash returns remains an unavoidable question for the entire industry.

A.

In June of this year, the U.S. credit rating agency KBRA assigned Anthropic an issuer credit rating of BB+, just one notch below the lowest investment grade, BBB-, still within speculative grade—commonly known as "junk grade."

If one looks only at the operating data Anthropic subsequently disclosed, it is hard to associate it with "junk grade."

According to the IPO prospectus made public on October 26, in the first half of this year Anthropic achieved revenue of $16.286 billion, a year-on-year increase of 1,300%, while its operating loss narrowed to $693 million.

On the customer side, as of the end of September, more than 11,900 customers had contributed at least $100,000 in revenue to Anthropic over the past 12 months, compared with only about 950 a year earlier; among them, more than 1,900 contributed at least $1 million.

In terms of cash reserves, Anthropic's performance is not poor either: as of the end of June, Anthropic held $112.4 billion in cash and cash equivalents, as well as about $19.3 billion in short-term investments, totaling more than $130 billion. However, this huge cash reserve was accumulated mainly through external financing—in the first half of this year alone, the net cash obtained by Anthropic through financing activities reached $110.8 billion, mainly from preferred stock issuance.

However, such impressive growth still failed to meet the expectations of some equity investors.

According to foreign media reports, because Anthropic's quarter-on-quarter revenue growth slowed from about 145% to 45% in the third quarter, some shareholders and potential IPO investors have begun discussing lowering its expected listing valuation from about $2 trillion to $1.5 trillion.

At the same time, Anthropic's computing power spending is expanding rapidly.

In the first half of this year alone, the company's computing power and infrastructure operating expenses reached $8.945 billion, equivalent to 55% of revenue in the same period, an increase of about 283% year on year.

The prospectus discloses that Anthropic has committed to purchasing at least $114.5 billion, $110 billion, and $44.9 billion in computing power from Google, Amazon, and Microsoft, respectively, over the coming years, with total minimum purchase commitments reaching $269.4 billion.

The performance periods of the relevant contracts extend to between 2033 and 2036, and most include minimum payment obligations. Even if actual computing power usage is lower than the agreed level, Anthropic still must pay the corresponding fees.

In addition, Anthropic has signed a computing power procurement agreement with SpaceXAI totaling up to $120.5 billion, with a term extending to 2029, which can be terminated with 90 days' notice after the initial non-cancellable period ends. The equipment finance lease agreement signed with Broadcom involves about $161.2 billion in lease payments, excluding interest, and is in principle non-cancellable during the lease term.

These contracts span many years and have different payment terms, but all mean that Anthropic will need to bear enormous computing power and infrastructure expenditures in the future.

Whether the corresponding revenue can continue to grow remains difficult to determine.

In September of this year, a banking syndicate led by Morgan Stanley increased Anthropic's revolving credit facility from $3.25 billion to $15 billion and extended the term to September 2031. This means Anthropic can borrow from the banks within the facility, repay, and then borrow again.

But the banks also imposed corresponding conditions: under the agreement, Anthropic needs to pledge the assets of the company and some important subsidiaries as collateral, and some subsidiaries must also provide guarantees for the borrowing. If the company cannot repay its debt as agreed, the banks can dispose of the collateral assets according to the contract or seek recourse from the guarantors. In addition, the agreement also restricts Anthropic from incurring new debt, selling assets, and distributing funds to shareholders.

Only when Anthropic obtains investment-grade debt ratings from at least two of the three major rating agencies—Moody's, S&P, and Fitch—with a "stable" rating outlook and meets other contractual conditions can it apply to release these collateral and guarantees. Even if Anthropic successfully goes public, these financing constraints will not be automatically lifted.

Blackstone's latest judgment further reflects market doubts about Anthropic obtaining an investment-grade rating in the short term—previously, Blackstone expected Anthropic to be likely to obtain an investment-grade rating by the end of this year, but now it believes that even after completing an IPO, the company may still remain in the BB category.

Equity investors look at growth, while debt investors look at credit. Anthropic, which is pushing ahead with its listing, is caught in the middle, under considerable pressure from both sides.

B.

Beyond Anthropic, technology giants with mature business support are also beginning to face credit pressure brought by AI capital expenditures.

According to foreign media reports, in the first three quarters of this year, the operating cash flow of five technology giants—Microsoft, Google parent Alphabet, Amazon, Meta, and Oracle—totaled $542.9 billion, up 29% year on year and a record high. But the capital expenditures of these five companies nearly doubled over the same period, reaching an astonishing $556.9 billion.

This is the first time since relevant statistics began in 2011 that the combined capital expenditures of the five companies have exceeded their operating cash flow.

After deducting capital expenditures, the combined free cash flow of the five companies turned from positive $143.5 billion in the same period last year to negative $14 billion.

Rating agencies are concerned that such massive capital expenditures will continue for many years and that the companies' future financial pressure may further intensify.

On October 30, S&P changed its credit rating outlook for Meta from "stable" to "negative," while maintaining its AA- credit rating unchanged.

According to its third-quarter financial report, Meta's operating cash flow for the quarter reached $32.79 billion, but capital expenditures were as high as $34.91 billion, causing free cash flow to turn negative at $2.12 billion.

Meta expects full-year capital expenditures this year to reach $135 billion to $145 billion, while S&P predicts that Meta's capital expenditures may exceed $225 billion in 2027, with a free operating cash flow gap of $71 billion.

However, Meta does not agree with S&P's judgment. The company said AI has already improved the performance of its advertising business and will create new revenue opportunities in the future.

Oracle, meanwhile, has already been actually downgraded.

On July 9 this year, S&P downgraded Oracle's credit rating from BBB to BBB-, just one step away from "junk" status. The reason also points to the massive capital investment, debt, and long-term lease burdens brought by AI infrastructure construction.

On October 28, Oracle announced its financing and cash flow plan for the next four years: capital expenditures are expected to peak in fiscal years 2027 to 2028, and free cash flow will not turn positive until the second half of fiscal year 2028. To support this transition, the company plans to raise no more than $20 billion in additional financing through capital markets by fiscal year 2030, including $10 billion each in equity and debt financing.

In addition to affecting credit ratings, continuously expanding capital expenditures are also beginning to squeeze the space for technology giants to return cash to shareholders.

Reuters statistics show that in the first three quarters of this year, the total stock buybacks of the five companies plunged from $81.8 billion in the same period last year to $13.6 billion, a drop of 83%. Among them, Alphabet and Meta did not conduct any stock buybacks in the first three quarters of this year.

Including cash dividends, the five technology giants returned only 9.3% of operating cash flow to shareholders, down from 27.4% in the same period last year and the lowest level in at least 15 years.

Technology giants are still making money, but the money they make is increasingly insufficient to cover their spending.

C.

Although credit rating agencies have issued warnings one after another, financing activity on Wall Street has not stopped because of this.

The most typical example is Blackstone: Blackstone lowered its expectation for Anthropic's future credit rating, believing that even if the company successfully goes public, it may still remain at junk level. But at the same time, Blackstone has just joined Bank of America, Citigroup, and Morgan Stanley to announce an AI computing power financing deal worth as much as $63 billion.

The deal uses a special financing structure: financial institutions provide funds to a newly established special purpose entity, which purchases Broadcom's AI chips and networking equipment and then leases them long-term to Anthropic.

Among them, Bank of America, Citigroup, and Morgan Stanley arranged about $43.5 billion in senior financing, which enjoys priority in the debt repayment order; Blackstone participated in about $9 billion in subordinated financing, which ranks later in repayment, bears higher risk, and is expected to yield more than 10%.

Wall Street has not ignored Anthropic's credit risk, but has priced the risk through different return and repayment arrangements.

There is also a special role in this deal—Broadcom, which provides equipment residual value support for the senior financing.

If the value of these chips and equipment declines in the future, Broadcom will need to bear the corresponding risk within the agreed scope, providing additional protection for creditors.

Through such arrangements, Anthropic can lock in future computing power in advance, financial institutions can expect interest income, and Broadcom has also obtained large-scale equipment orders.

But credit risk has also been allocated out. Anthropic needs to continue paying rent, financial institutions rely on future lease income to recover their investments, and Broadcom needs to bear the agreed equipment residual value risk.

Once Anthropic's future revenue growth falls short of expectations, these interrelated contracts and financing arrangements may all come under pressure.

Moreover, this transaction structure is not an isolated case.

Previously, Apollo and Blackstone had also participated in another $35 billion Anthropic computing power expansion financing, with Broadcom providing up to $29 billion in support for the deal.

In August this year, Nvidia disclosed that it was providing up to $105 billion in conditional guarantees for OpenAI's lease of a data center in Ohio, USA, built by SoftBank's SB Energy. This guarantee will take effect as the data center is phased into use, mainly covering lease and power payment obligations within the agreed scope, and may only be triggered in the event of specific defaults and other circumstances.

It is worth noting that both chip suppliers themselves have relatively high credit ratings. On October 27, Fitch affirmed Broadcom's long-term credit rating at A-, with the outlook maintained at "Positive"; Nvidia was upgraded by S&P from AA- to AA in June this year, and after disclosing the above guarantee in August, its AA rating was still maintained, with a "Stable" outlook.

Rating agencies have already noticed this change. On October 27, Moody's released a report pointing out that AI infrastructure construction is driving the rapid expansion of the private credit market, and the large volume of loans maturing between 2027 and 2029 may leave some financially weaker companies facing refinancing difficulties.

On the same day, the Bank for International Settlements (BIS) released a research report titled "Building AI on Credit," analyzing this financing model.

The research points out that, supported by large technology companies' lease commitments, asset guarantees, and equipment residual value, some lower-rated AI companies can also obtain investment-grade financing through special purpose entities, attracting funds from insurance companies and other institutional investors.

However, the BIS research found that the short-term impact of these financing developments on the U.S. Treasury and investment-grade corporate bond markets remains limited. But the research also warned that if AI-related debt continues to be issued at the current pace, the impact on long-term interest rates and financial stability may gradually expand.

As financing relationships become increasingly complex, companies across the AI industry chain have gradually formed a community of shared interests. All parties have an incentive to sustain this investment boom and, if necessary, may even continue to provide financing and extend debt maturities to prevent any link from suddenly collapsing. This mutual dependence can provide companies with more financing support, but it also makes risk increasingly difficult to isolate.

However, debt ratings can be raised through guarantees, but AI companies' own debt-servicing capacity will not automatically improve as a result. Wall Street has the ability to let "junk-rated" companies' projects borrow investment-grade money, but it cannot guarantee that AI will ultimately be able to earn that money back.

This article is from the WeChat public account "Zimu Bang" (ID: wujicaijing), author: Yuan Xinyue, published by 36Kr with authorization.