Anthropic’s revenue figures are impressive, so why is the AI sector still declining?
Anthropic’s latest revenue figures, which have drawn global attention, have finally been released. Third-quarter revenue reached $16.8 billion, a nearly 46% increase quarter-over-quarter. The company reported its first adjusted profit, with free cash flow turning positive during the quarter.
In the same period of 2025, this figure stood at only $787 million. By the first quarter of 2026, revenue had risen to $4.73 billion, reaching $11.5 billion in the second quarter.
Over a span of just over six months, revenue surged from under $5 billion to $16.8 billion. In a single quarter, it exceeded 3.6 times the full-year total for 2025, while also crossing the break-even threshold. Such figures would be considered textbook examples of success in any technology company’s performance report.
The Nasdaq Composite Index fell 1.88%, closing at 26,692.87 points, marking its largest single-day decline since July. The S&P 500 Index dropped 1.11%, closing at 7,718.94 points. The Philadelphia Semiconductor Index plummeted by more than 4%, accumulating a decline of approximately 10% over six trading sessions, with the market capitalization of its constituents evaporating by over $1 trillion. Computing power assets suffered sharp declines: CoreWeave plunged as much as 16% intraday, Dell fell nearly 12%, Oracle dropped 9%, and NEBIUS declined 11%. The optical communication and memory chip sectors faced collective pressure.
The following day, the sell-off spread to Asia. South Korea’s KOSPI Index opened with a steep 5.59% drop, falling below the 5,700-point mark and triggering circuit breakers. SK Hynix declined by more than 7%, Samsung Electronics fell by over 6%, and Japan’s Nikkei 225 Index dropped 3%.
Market sentiment was extremely poor.
Anthropic has demonstrated that real demand exists, that customers are willing to pay for its products, and that its business model is crossing the break-even point. However, at the same juncture, the macroeconomic environment and financial pipelines supporting this massive technological narrative are simultaneously exposing vulnerabilities across multiple dimensions.
Resurgent inflation and a 5.2% risk-free rate
The immediate selling pressure on global tech stocks that day stemmed primarily from re-emerging inflation data.
The September Consumer Price Index (CPI) released by the U.S. Bureau of Labor Statistics showed signs of a reversal. The September CPI rose 3.7% year-over-year, higher than August's 3.4% and the market expectation of 3.6%, while increasing 0.6% month-over-month to hit a five-month high. Core CPI rose 0.3% month-over-month and 2.6% year-over-year. The primary driver was energy, with gasoline prices surging 9.1% month-over-month and 34.2% year-over-year, and heating oil rising 74.2% year-over-year. Brent crude oil surpassed $106 per barrel.
Inflation has not only reaccelerated but is also spreading to a broader range of goods and services.
Market pricing rapidly shifted gears. The CME FedWatch Tool indicated that the probability of a rate hike at the Federal Reserve’s late-October meeting jumped from approximately 22% to 44%. The yield on the 10-year U.S. Treasury note rose to 5.188%, while the 30-year yield touched 5.522%. Money markets began to fully price in two rate hikes by the end of January next year, and the U.S. Dollar Index refreshed its highest level since May 2025.
In a research report, Goldman Sachs provided a set of estimates indicating that the construction boom in AI data centers and underlying infrastructure has directly contributed approximately 0.5 percentage points to the Federal Reserve’s preferred inflation metric.
The massive expansion of computing power has increased electricity load, driving up prices for transformers, copper cables, high-bandwidth memory chips, and data center land leases, thereby entering the consumer price basket. Rising prices have dampened expectations for monetary easing, pushed up nominal interest rates, and consequently raised the discount rate for future cash flows.
Many long-term capital commitments signed in 2024 under low-interest-rate assumptions must now seek funding in an environment with the highest cost of capital since 2007. Revenue addresses the numerator, while interest rates determine the denominator. When the denominator rises faster than the numerator expands, asset-level valuations face broad downward revisions.
Financing channels are becoming congested at the entity level.
Beyond the rise in macro interest rates, the financing pipeline for the AI infrastructure supply chain itself has shown a series of stranding signals.
Over the past week, three data center listing plans in the Asia-Pacific region were consecutively hindered.
Firmus Grid, an Australian data center operator backed by NVIDIA, withdrew its IPO on October 9. The company had originally planned to raise up to approximately $5.5 billion, which would have been one of the largest listings in Australia in thirty years. The company’s current revenue is approximately $51 million, yet the proposed issuance valuation was as high as $30.6 billion, implying a price-to-sales ratio of nearly 600x. Although the company showcased about $72.8 billion in future intended contracts, its actual built capacity stands at only 5%, with remaining projects requiring hundreds of billions of dollars in construction funding. Even after the offer price was lowered by 25% during the book-building period, institutional demand remained sluggish, leading to the final withdrawal.
The delayed listing of AirTrunk, a portfolio company of Blackstone, serves as an even more significant signal.
AirTrunk is a hyperscale computing infrastructure platform jointly acquired by Blackstone and the Canada Pension Plan Investment Board in 2024 for AUD 24 billion. Its assets span Australia, Japan, Singapore, Malaysia, Hong Kong, China, and India. The company originally planned to raise approximately USD 1.5 billion through a REIT structure in Singapore, positioning itself as a defensive AI infrastructure asset offering stable leases and dividend yields.
The listing plan was ultimately postponed until 2027. With the 10-year U.S. Treasury yield hovering around 5.2%, even a proposed increase in the distribution yield to 6%—providing only about 78 basis points of credit and liquidity premium—failed to attract institutional investors seeking certainty. In the same week, Vitro, a data center operator under a Philippine telecommunications carrier, also deferred its listing plans to 2027, while SB Energy, SoftBank’s energy subsidiary, suspended its IPO roadshow.
More indicative of capital flow trends than the cooling equity financing market is the cautious stance adopted by the credit market.
On October 12, AI cloud infrastructure platform Volta Infrastructure launched a six-year syndicated loan facility totaling USD 5.012 billion. Of this amount, a USD 1.3 billion term loan was underwritten by JPMorgan and Goldman Sachs, whereas the USD 3.712 billion specialized loan intended for the procurement of approximately 36,000 NVIDIA GPUs was not underwritten by these two major banks and was instead distributed directly to the public market. This unrated hardware-backed loan was priced at SOFR plus 625 to 650 basis points, bringing the overall yield to maturity close to 11.5%.
While the two top-tier investment banks were willing to act as arrangers to earn fees, they refused to retain exposure to the rapidly depreciating hardware assets on their balance sheets.
CoreWeave’s CDS Volatility and the Revaluation of Computing Clouds
If traditional banks are tightening underwriting risk controls, the secondary and derivatives markets are beginning to flash red lights for highly leveraged computing lease providers.
On October 12, market data showed that the quote for CoreWeave’s five-year credit default swaps (CDS) surged to nearly 881 basis points, up from around 547 basis points in late June.
A CDS is essentially an insurance contract purchased by investors against debt default. When the quote for a CDS jumps by hundreds of basis points in a short period, it means bondholders must pay significantly higher insurance premiums to protect against or hedge default risk.
Trace pricing data shows that CoreWeave’s unsecured high-yield bonds maturing in 2032 have fallen below 80% of par value, with secondary market yields to maturity surging to nearly 15%. The yield on its 2031-maturity bonds, carrying a 9.75% coupon, has similarly risen to 14.9%.
When a company’s borrowing costs are pushed to the 15% level, capital markets are no longer assigning a premium for growth assets but rather pricing them as distressed assets. Based on current CDS levels, the market-implied five-year probability of default exceeds 50%.
Elon Musk described these figures as “scary” on social media. Analysts have labeled CoreWeave the “subprime tranche” of the AI boom, warning that persistently high refinancing costs will place significant pressure on the valuations of asset-heavy lessors. Although CoreWeave’s management insists that its order backlog is sufficient to support debt servicing, the stock market also experienced severe volatility, with shares plunging by as much as 16% during intraday trading.
The potency of this signal lies in its ability to transcend individual companies and impact the financing conditions of the entire NeoCloud sector.
As the most aggressively expanding group in this round of computing infrastructure build-out, NeoCloud relies heavily on balance sheet arbitrage. By pledging NVIDIA GPUs to Wall Street investment banks and private credit funds to secure billions of dollars in loans, purchasing chips, and then leasing them to frontier large-model startups, these firms cyclically amplify their debt leverage.
With the leader’s CDS spreading to 881 basis points and bond yields touching 15%, the temperature of the entire funding pipeline has shifted dramatically. Creditors are beginning to reassess credit risk across the entire NeoCloud sector, not only significantly raising risk premiums for new loans but also tightening loan-to-value ratios and collateral terms. For computing cloud providers that need to continuously roll over debt and fund capital expenditures for next-generation chips, the sharp rise in funding costs has directly compressed their existing arbitrage margins.
On one side, frontier laboratories present grand blueprints to the outside world; on the other, intermediaries bearing the burden of purchasing cards and leveraging assets are charged distress-level default premiums by capital markets. These two contrasting images displayed at the same trading desk have directly altered buyers’ risk appetite.
Public statements by the heads of Broadcom and Intel
It was at the DBS Asian Future Frontiers Conference held in Singapore on October 12 that the tightness in capital markets was explicitly highlighted before global investors.
In his speech, Broadcom CEO Hock Tan offered a frequently cited assessment. He argued that the biggest constraint on the future of the AI industry may no longer be production capacity, but capital.
Hock Tan reviewed the industry supply chain over the past two years, noting that engineering challenges such as high-bandwidth memory shortages, insufficient wafer fabrication capacity, and slow grid connections are gradually being overcome. Wafer shortages are expected to ease over the next two to three years, and newly built power generation facilities are progressively connecting to the grid. However, investor patience and the cost of capital are becoming hard constraints. He stated bluntly that many investors entered the market with extreme expectations, creating an environment akin to a casino. Furthermore, companies still need to provide more compelling evidence regarding the conversion efficiency of AI applications in their actual business operations.
Intel CEO Lip-Bu Tan expressed similar views at the same event. He pointed out that the most pressing challenges currently are rising capital costs and investors' stringent scrutiny of return on investment (ROI).
Both semiconductor leaders acknowledge the genuine demand for computing power, but they jointly highlight a prerequisite: the sustainability of growth in the coming years ultimately depends on whether downstream customers can prove that AI delivers tangible commercial returns. If the industrial penetration of frontier models falls short of expectations, the procurement pace for upstream computing infrastructure may be adjusted.
DBS Bank CEO Piyush Gupta provided a financial perspective in an interview with Bloomberg Television. He stated that risks associated with AI-related investments are being repriced, with lenders demanding higher return compensation and stricter transaction structures. DBS Bank has led approximately $20 billion in syndicated loans related to computing power this year, but typically retains only a minimal exposure share in these transactions.
The tap on capital has not been completely turned off; rather, the flow of funds and willingness to bear risk have shifted.
Lowered revenue expectations for large language models
Meanwhile, some easily overlooked adjustments are beginning to emerge within the narrative surrounding frontier large language models.
First is the transparency of revenue recognition. Data updated by OpenAI for institutional investors shows that its annualized revenue reached approximately $50 billion by the end of September, lower than the widely circulated market estimate of $70 billion. The downward revision stems from the new methodology excluding revenue shares from cloud partners.
This move has triggered cross-comparisons in the market. If OpenAI adopts a net revenue basis excluding channel partner shares, Anthropic’s gross revenue—including flows through Amazon and Google Cloud channels—loses a comparable benchmark. As the direction of expectation corrections for core industry metrics shifts from upward revisions to downward adjustments to remove inflation, valuation systems face realignment.
Regarding the pace of technological development, frontier laboratories themselves are signaling a slowdown. In early September, OpenAI CEO Sam Altman explicitly stated that the company would not pursue an initial public offering (IPO) in 2026, citing the need for more time to address model safety issues. Around the same time, Anthropic CEO Dario Amodei published an article calling for the entire industry to proactively slow the pace of capability advancement to allow a window for safety alignment research.
The restrained stance adopted by the two companies is directly linked to abnormal incidents exposed in recent security tests. As industry leaders begin to call for a slowdown, the capital market’s linear projection of endless acceleration in computing power demand has lost a layer of emotional support.
The Balance Sheet Behind the Impressive Figures
Let us return to Anthropic’s performance report, which shows $16.8 billion.
Based on the disclosed annualized run rate of $65 billion as of late July, the third-quarter revenue of $16.8 billion met the baseline threshold. The adjusted gross margin remained at approximately 85%, the adjusted operating profit margin exceeded 11%, and both operating cash flow and free cash flow turned positive.
However, market scrutiny has focused on the growth slope.
From the first to the second quarter of 2026, the company’s revenue grew by approximately 143% quarter-on-quarter. From the second to the third quarter, the sequential growth rate slowed to 46%. It took only two quarters to transition from doubling acceleration to a moderated growth pace.
The annualized run rate for the third quarter was approximately $67 billion, representing only a slight increase from the $65 billion disclosed in late July. This suggests that monthly net additions may have already shown a flattening trend between August and September. The core assumption underpinning the ultra-high valuation multiples granted by the market is often exponential expansion with an extremely steep slope. Once this slope changes, the benchmarks of valuation models face recalculation.
Based on the current annualized revenue of $67 billion, the IPO target valuation of $2 trillion corresponds to a price-to-sales ratio of approximately 30x. Five months ago, Anthropic completed a financing round with a post-money valuation of $965 billion, just surpassing OpenAI in terms of valuation. Planning to enter the public market with a doubled valuation of $2 trillion just months later represents a leap that has caused hesitation among potential cornerstone investors.
Underwriters are using models based on forward-looking revenue for 2028 to anchor the valuation. Anthropic has presented institutional investors with a 2028 revenue target ranging from $190 billion to $200 billion. Based on a forward revenue estimate of $200 billion, a market capitalization of $2 trillion implies a forward price-to-sales ratio of approximately 10x.
Whether the public markets are willing to place full trust in forward-looking assumptions for two years out constitutes the biggest point of divergence. New Constructs, an independent research firm, has publicly raised doubts, arguing that to sustain a $2 trillion valuation, its steady-state net profit would need to approach NVIDIA’s levels. The firm believes that the excessively high listing valuation reflects liquidity pricing driven by primary market capital seeking exit opportunities, rather than a true reflection of corporate value.
More stringent financial constraints are hidden within undisclosed IPO filings. Disclosures by Reuters reveal that Anthropic has committed to investing at least $518 billion in AI infrastructure over the next decade, partnering with six computing power and cloud platform providers. Approximately 80% of these contracts contain irrevocable clauses or take-or-pay provisions requiring full payment regardless of actual usage.
A breakdown of the contracts shows that the company has committed approximately $161.2 billion for equipment leases from Broadcom. The agreement with Google is valued at $111.1 billion, with a performance period spanning from April 2026 to July 2033, explicitly stipulating that any shortfall in actual computing power consumption must be made up. The partnership with Amazon amounts to approximately $110 billion, with an execution period until April 2036, also including strict take-or-pay clauses. The computing power agreement signed with Microsoft is valued at approximately $31.4 billion.
Compared to the Stargate infrastructure cost-sharing mechanism jointly built by OpenAI with partners such as SoftBank and Oracle, Anthropic bears an extremely high level of individual performance liability. Using revenue streams that are still in the incubation phase to shoulder rigid payment obligations exceeding $500 billion essentially transfers the risk of future computing power consumption entirely onto its own balance sheet. If the expansion of procurement budgets for downstream enterprise software falls short of expectations, these irrevocable minimum payment commitments will directly translate into sustained rigid cash outflows.
The quality of revenue is also under scrutiny. The company's top two customers contributed approximately 24% of its revenue share, largely through flexible usage contracts lacking long-term lock-in mechanisms. In 2025, about 47% of the company's revenue came from the cloud channels of Amazon and Google, with platform revenue-sharing ratios hovering around 20%. This means that gross revenue does not fully convert into retained cash on the balance sheet. Furthermore, these two major cloud channel partners are simultaneously investors, suppliers, and direct competitors with their own self-developed models.
On the product pricing front, pressure from the global open-source ecosystem and closed-source competitors is gradually becoming apparent. Data from Artificial Analysis shows that the unit price for calls to Anthropic's flagship model is more than double that of comparable products from OpenAI. Chinese open-source and open-weight models, leveraging extreme cost control and usability, are prompting overseas enterprises to restructure their AI budgets. Research by Ramp, a software spending analytics firm, indicates that corporate clients are curbing unrestrained model usage and shifting toward more cost-effective alternatives. As hybrid token prices across the industry continue to decline, maintaining an 85% gross margin will heavily depend on the rate of improvement in inference efficiency outpacing the rate of decline in unit prices.
In the enterprise deployment sector, data from Counterpoint Research shows that Anthropic's revenue share in the large language model market reached 31.4% in the first quarter of 2026, ranking first in the industry, with leading institutions such as Blackrock and Barclays having integrated the Claude system. However, in the consumer traffic segment, statistics from Sensor Tower show that ChatGPT holds a 46.4% global share, Gemini has risen to 27.7%, while Claude accounts for only 10.3%. Competition in underlying large models is highly dynamic; the window of technological leadership may not necessarily cover the depreciation cycle of multi-decade computing power contracts.
The over-the-counter (OTC) market has provided immediate feedback. The price of ANTH, the OTC trading ticker for Anthropic established on Hyperliquid, has retreated from above the $2 trillion mark to the range of $1.89 trillion to $1.94 trillion.
The $16.8 billion disclosed by Anthropic and the 1.88% decline in Nasdaq serve as mirror images of each other.
On one hand, there is the astonishing commercial expansion of frontier large model enterprises, with revenue jumping from $4.7 billion to $16.8 billion within half a year, achieving underlying profitability and crossing the cash flow break-even point. This proves that the willingness of enterprises to pay for advanced intelligent tools is genuine, and the direction of technological advancement remains unchanged.
On the other hand, there is a rational restructuring in the capital markets, characterized by $518 billion in irrevocable purchase guarantees, credit default swap spreads soaring to nearly 881 basis points, computing power intermediary borrowing costs approaching 15%, successive delays in the listing of data center assets, and benchmark Treasury yields pressing toward 5.2%.
Revenue figures confirm that demand is indeed surging. However, capital markets are voting with their feet, reflecting the true cost required to support this future.