[KongchangAI]
· 7 min read· 3,535 words

Deep Dive into Anthropic's S-1: The $850 Billion Compute Bet and AI's New Valuation Logic

Deep Dive into Anthropic's S-1: The $850 Billion Compute Bet and AI's New Valuation Logic

Anthropic's S-1 reveals a bet on short-term usage growth with rigid long-term compute contracts—efficiency determines moat or leverage.

Anthropic's S-1 discloses full financials for the first time: H1 2026 revenue of $16.29B, Q2 up 144% QoQ, operating profit positive, but the $80.7B net loss is mainly non-cash valuation losses. Gross margin jumped from 39.5% to 85.4%, though methodology matters. Total compute commitments are ~$850B via cloud purchases, equipment leases, and third-party GPUs, with actual cash spending still early. Revenue per $100M compute rose from $0.40 to $2.08. The core contradiction is the timing mismatch between rigid costs and elastic revenue, making post-IPO valuation highly elastic.

Anthropic's filing of its S-1 has dropped a bombshell on the tech world. This isn't just the first time a frontier AI lab has laid out its complete financials for public market pricing—it's also a public stress test of AI capital expenditure cycles, technical roadmaps, and governance models. The information density of this document far exceeds the sum of all rumors from the past two years.

This metric uses

Revenue Rocket: From $4.5B to $16.2B

Let's start with the most stunning growth curve. Full-year 2025 revenue was $4.585 billion, and in the first half of 2026 it skyrocketed to $16.286 billion. Q2 alone brought in $11.555 billion, a 144% quarter-over-quarter increase—a growth rate with virtually no precedent in software industry history.

More critically, the profitability inflection point. Q2 operating income was $974 million, an operating margin of about 8%; adding back stock-based compensation, adjusted operating income was $2.235 billion, a margin of about 19%. The market's earlier expectation of Q2 2026 operating profitability was delivered.

But then there's a number that will make you jump: a net loss of $80.748 billion for the first half. Don't panic—$72.388 billion of that is non-cash valuation change losses from the fair value changes of convertible note derivatives and SAFE instruments. The more the company's valuation rises, the higher the book liabilities of these instruments, and the larger the recorded losses. In other words, this isn't money lost from operations—it's an accounting loss caused by the company being too valuable. Actual operating losses for the same period were only $693 million—this distinction is crucial, as many people get misled by the headlines.

Fair Value Changes of Convertible Note Derivatives and SAFEs: Convertible note derivatives typically refer to embedded derivative clauses in convertible bonds, such as conversion rights and redemption rights, whose fair value increases as the company's valuation rises and must be recognized as liabilities and recorded in current profit and loss under accounting rules. SAFE (Simple Agreement for Future Equity) is a financing instrument commonly used by startups, where investors put in funds first and convert to equity under agreed conditions upon specific trigger events (such as an IPO or the next financing round). In accounting treatment, fair value changes of SAFEs and convertible note derivatives are often recorded as non-cash losses or gains. When a company's valuation rises rapidly, the book liabilities of these instruments increase significantly, generating large book losses, but this does not affect the company's actual operating cash flow. Understanding this helps distinguish accounting-level valuation fluctuations from genuine operating losses.

Gross Margin Leap and Hidden Concerns About Revenue Quality

The gross margin performance is equally astonishing. First-half operating costs were $2.380 billion against $16.286 billion in revenue, for a gross margin of about 85.4%. In 2025, that figure was 39.5%. Going from 39.5% to 85.4% in one year is an enormous leap.

But mind the methodology. The S-1 doesn't break out API business gross margin separately, and Anthropic classifies channel fees as sales expenses rather than costs, which significantly affects comparability with other companies' gross margins. So 85.4% is a calculated gross margin, not a fully-loaded gross margin. If you mechanically deduct all sales and marketing expenses, only about 62% of revenue remains to cover R&D and management.

On cash flow, first-half operating cash inflow was $3.929 billion, already positive. As of the end of June, cash plus short-term investments totaled about $131.7 billion. First-half net financing inflow was $10.774 billion, so cash reserves are substantial. But the company itself notes that directly building infrastructure will increase upfront capital expenditures and cash needs.

Customer Diversification: An Overlooked Key Improvement

There has been a persistent view that Anthropic's revenue relies on a few top customers. The newly disclosed data provides a strong rebuttal: the top five customers' revenue share dropped from about 36% in Q1 to about 21% in Q3, and each of the top five customers hit a quarterly revenue high. Concentration declined not because big customers fell off, but because other customers grew faster—this is a completely different quality of information.

Customer count expansion is also dramatic. Customers contributing more than $100,000 in revenue over the past 12 months exceeded 11,900 by the end of September, versus only about 950 a year earlier. There were more than 1,900 customers above $1 million and about 250 above $10 million. Net revenue retention is about 900%—the same cohort of existing customers is generating 9x the quarterly revenue of a year ago. This figure is unheard of in the SaaS industry.

But note the methodology: this metric uses quarter-over-quarter revenue for the same customer cohort and excludes fixed subscription revenue from Pro and Max, counting only usage. So it reflects existing customers using more, not existing customers paying more to renew. This metric is naturally volatile because about 80% of the revenue structure is usage-based pricing, with only 20% from subscriptions.

This leads to revenue quality concerns. The company's annualized revenue metric multiplies the most recent 28 days of usage revenue by 13, plus monthly subscription revenue multiplied by 12. Multiplying 28 days by 13 essentially annualizes the most recent month's usage directly, making it extremely sensitive to short-term fluctuations. The S-1 also discloses that customer spending during the two weeks spanning the new year in early 2026 fell about 20% versus the prior two weeks—two weeks of usage fluctuation can noticeably change the annualized revenue figure. This is one of the core issues in the later pricing debate: revenue is consumption-based, and contract visibility is very low.

There's also a structural feature: cloud channel share. Revenue generated through cloud marketplaces like AWS, Google Cloud, and Azure rose from 11% in 2023 to 47% in 2025 and 48% in the first half of 2026—nearly half. This portion is recognized on a gross basis—the cloud platform collects first, deducts fees, then settles with Anthropic. So it affects both gross margin methodology and cash flow timing. Model companies and cloud providers are now mutually dependent.

Costs are certain

Net Revenue Retention (NRR) and Consumption-Based Revenue: Net Revenue Retention (NRR) is a core SaaS industry metric that measures the change in revenue from the same customer cohort after one year, typically including the combined effects of upgrades, cross-sells, downgrades, and churn. An NRR above 100% means revenue from existing customers is growing. The ~900% NRR in this article is extremely high, but note its methodology: it counts only usage-based revenue, excludes fixed subscription revenue, and uses quarterly year-over-year comparison. Consumption-based revenue refers to customers paying based on actual usage rather than a fixed subscription fee; its advantage is that it grows with customer usage, and its disadvantage is high volatility and low contract visibility. Annualized revenue (ARR) is estimated by multiplying the most recent 28 days of usage revenue by 13; this short-term annualization method is extremely sensitive to usage fluctuations—two weeks of spending changes can significantly affect the ARR figure. Therefore, behind high NRR and high growth also lie revenue quality risks.

The $850 Billion Compute Bet: Breaking Down Every Dollar

Now for the most important part: total compute spending commitments of approximately $850 billion. This number needs to be broken down.

As of the end of June, fixed or minimum commitments under non-cancelable purchase agreements were about $265.6 billion, plus lease obligations not yet commenced of about $321.4 billion, totaling $587 billion. Then after the end of June, long-term agreements were added or amended involving future minimum payments of about $245 billion, plus short-term non-cancelable payments of about $21 billion. So the total reaches the $850 billion magnitude.

Broken down by supplier, the three major clouds' minimum purchase commitments total $169.4 billion: Microsoft $44.9 billion, Google $114.5 billion, Amazon $110 billion. Contract terms extend to 2033, 2033, and 2036 respectively, all including minimum payment obligations. There's also incremental TPU compute procurement from Google and other suppliers for 2028 to 2031, with expected spending of $125 to $150 billion.

On third-party GPUs, SpaceXAI is the largest, with a contract spending cap of $120.5 billion and a monthly fee of about $3.1 billion once all capacity is delivered, expected to complete delivery in December 2026, with termination possible on 90 days' notice after the irrevocable period ends in early 2027. What about xAI and Lambda? xAI up to $44.6 billion, Lambda about $41.3 billion, both billed on actual delivered capacity.

On hardware equipment, Broadcom's lease scale is large: five-year PPU equipment finance lease rents total about $161.2 billion, with the capacity cap even higher in 2026 excluding interest, capacity expected to be delivered from 2027, typically paying rent once equipment is ready for service. AMD involves Helios rack-scale systems and Instinct MI450 GPUs, with first capacity planned for delivery in 2027; if fully executed, expected spending over the agreement period exceeds $20 billion, and AMD has also committed to subscribing for up to $5 billion of Anthropic shares when conditions are met—the shovel seller taking a stake in the miner.

But note a contrast: as of the end of June, the company's future fixed or minimum guaranteed purchase commitments were about $265.6 billion, while cash spending on fixed asset purchases in the first half of 2026 was only $480 million. In other words, actual hardware cash investment is still at a very early stage; the bills really start coming due over the next few years.

Expected spending over the agreement period exceeds $20 billion

Supporting these are financing facilities with $77 billion undrawn as of the end of June: Broadcom $42.1 billion, Google $20 billion, Amazon $15 billion. But these facilities come with conditions—drawdowns require meeting revenue, compute, delivery, or other conditions. Broadcom's $42 billion is mainly for paying equipment rent, typically at a 9% interest rate; if some revenue conditions are not met, the rate can rise to 13%, and the conversion discount also increases. This is a very critical leverage design: miss revenue targets, and financing costs rise. This is the two-sided nature of vendor financing—it supports growth during expansion but amplifies pressure during a slowdown.

Vendor Financing and Minimum Purchase Commitments: Vendor financing refers to suppliers (such as cloud providers and hardware manufacturers) providing credit, leasing, or deferred payment arrangements to customers to help them purchase their products. In the AI compute sector, cloud providers and chip companies lock in future capacity and support customer expansion through long-term contracts, equipment finance leases, and minimum purchase commitments. A minimum purchase commitment is a customer's commitment to purchase a minimum amount of services or products over a future period, typically non-cancelable and legally binding. For suppliers, such arrangements reduce demand uncertainty; for customers, they lock in costs in advance but also constitute rigid liabilities. If customer revenue falls short of expectations, these commitments can become a heavy burden. In this article, financing facilities from suppliers like Broadcom also come with conditions, such as interest rate increases if revenue targets are missed, reflecting the two-sided nature of vendor financing: providing leverage during expansion and amplifying pressure during a slowdown.

Efficiency Hedge: Is Compute a Moat or Leverage?

With so much cost discussed, we need to see whether there's an efficiency hedge. There is, and it's obvious.

Revenue per $100 million of compute and infrastructure expense rose from $0.40 in Q1 2025 to $1.4 in Q1 2026 and $2.08 in Q2—a 5.2x increase. Q3 was reportedly about 1.7, mainly affected by ARM's growth slowdown, with some recovery expected in Q4. This metric is actually the most important one in the entire S-1, because it determines whether compute is a moat or leverage.

On internal compute allocation, training and inference are roughly 7:3. What about profit sharing with cloud providers? Rough estimates suggest that after deducting cloud computing costs, Anthropic and cloud providers split gross profit roughly 60:40.

The economic value of model iteration is also key. Over 80% of Q3 revenue came from model versions released within the prior six months. Opus 5.5 reduced customer per-task costs by about 40% versus Opus 5 while maintaining its own margins. So it's not simply cutting prices to gain volume—it's cost reduction from efficiency gains. In September, Fable 5.1 typical task costs fell about 25%, and Opus 5.5 fell about 40%; the company says this comes from a combination of price cuts and efficiency gains while maintaining margins. Concerns about price wars can be eased, but price wars in low-end models can't be ruled out.

The Core Contradiction: Timing Mismatch Between Rigid Costs and Elastic Revenue

How should this company be valued? The core contradiction isn't whether growth is fast—the market already expects that. The contradiction lies in the timing mismatch between an exponential demand narrative and highly rigid physical costs.

It's neither like traditional software—asset-light, high-margin, strong subscription visibility—nor like traditional heavy-asset cyclical industries where profit is determined by capacity utilization. What is it? It's a structure that uses long-term rigid contracts to support short-term variable usage revenue. Contracts fix costs 12 to 18 months in advance, while revenue still depends on actual usage to cover them. Costs are certain, revenue is uncertain, and about 80% of revenue is consumption-based with extremely low contract visibility; many purchases are non-cancelable and even carry cross-default clauses.

So usage and ARR are the decisive variables. Post-IPO elasticity will be enormous: when usage exceeds the fixed-cost coverage line, profit and cash flow improve rapidly, and multiples tend to expand; once it falls below the coverage line, fixed bills immediately become pressure and multiples contract sharply. And the coverage line itself moves as new capacity comes online and efficiency changes. What the market is really watching is whether usage growth can consistently outrun the cost ladder.

What's the founders' logic? According to the Scaling Law, future value far exceeds current value, so you must bet on compute in advance, like large upfront R&D in the pharmaceutical industry. What are the hedges? Efficiency gains, offloading excess capacity when necessary, and maintaining cash and financing options. But this is a bet with strong path dependence: demand, ambiguous model efficiency, declining inference costs, and the financing environment must all roughly align. If any one deviates significantly, fixed costs turn from a moat into leverage.

So the upper and lower bounds of pricing will be very wide. How to frame the upper bound? Closer to a successful AI infrastructure growth stock—during a phase of high demand growth and efficiency not yet fully realized, the market is more likely to assign high revenue multiples first; as efficiency continues to materialize, it shifts to EV/EBITDA or cash flow multiples. If usage can consistently cover and exceed the cost increases from rigid contracts, the previously locked-in compute becomes a barrier. What about the lower bound? Closer to a highly leveraged, low-visibility capacity cyclical stock—those rigid compute contracts turn from advantage into heavy burden, and valuation would be significantly discounted as a high-leverage, high-uncertainty cyclical asset. This hybrid model lacks a perfect comparable, so the market will certainly reference both sides and rapidly shift weights based on actual performance in usage efficiency and governance constraints.

Governance Structure: The Special Constraints of a Public Benefit Corporation

There's another very special aspect: governance structure. The founder entity holds 50.1% of total voting power through special shares, a long-term benefit trust has special director appointment rights, and it is legally a Public Benefit Corporation. The board is obligated to consider broader social interests—primarily safety-related—while pursuing shareholder interests. The founders have explicitly stated they will continuously reinvent the business, consolidate existing models less, proactively forgo potentially lucrative paths like advertising, and delay model releases in safety evaluations. These choices are already written into legal documents and risk factors. Public market pressure will continuously test how firm this balance point really is.

Additionally, the company has allocated shares equivalent to 5% of fully diluted shares as of the end of June to a public benefit fund and established a framework for additional donations tied to future stock price milestones. Common stock investors should consider long-term dilution.

Public Benefit Corporation (PBC): A Public Benefit Corporation is a special U.S. corporate legal form that requires the company to pursue one or more specific public benefit objectives while pursuing shareholder financial returns, as explicitly stated in legal documents. The board has an obligation to balance shareholder interests with the public benefit mission when making decisions, rather than maximizing shareholder interests as the sole objective. PBCs must periodically publish public benefit reports assessing their social and environmental impact. This structure is common among mission-driven tech companies like Anthropic. Its advantage is locking in the mission through legal form and resisting short-term market pressure; its disadvantage is potential governance conflicts with purely financial investors and the difficulty of quantitatively evaluating public benefit objectives. For public market investors, PBC governance constraints may affect strategic flexibility and long-term value, and should be considered in valuation.

The Old ROI Question: The 2028 Answer Remains Unclear

Back to the old question: AI investment ROI is unclear through 2028—what carries the baton? Ultimately, it all comes down to ROI.

This S-1 makes the "I" clearer: AWS's minimum purchase commitment extends to 2036, other major lines also cover beyond 2028, some with non-cancelable minimum payment arrangements, and the upper bound disclosed on an 8-year horizon is over $800 billion of investment. But the 2028 return question is unresolved.

Rough estimates suggest Anthropic's average annual compute spending over the next six years will be about $109.3 billion, plus annualized personnel, marketing, and operating expenses of $30 to $50 billion. To break even, it needs at least $150 billion in revenue. What's the ARR now? About $70 billion in September, with October ARR estimated to accelerate considering the Opus 5.5 effect.

What positive incremental factors are on the "R" side? Growth is spreading from top customers outward. On cross-selling, the proportion of large customers using both the Claude platform and at least one of their own applications rose from about 30% in Q4 2025 to over 50% in Q3 2026. Cowork's adoption speed at the same cap stage exceeds Claude Code. From developers calling models to employees using agents directly, there's room for per-customer value expansion. Industry penetration also shows signs: customers contributing at least $100,000 annually already cover more than half of the Fortune 500, including the top ten banks and 9 of the top ten pharma companies. In the first half of 2026, U.S. revenue still accounted for about 72%, and internationalization is also a potential growth driver.

On cross-selling

Future pricing depends on three things: whether non-coding revenue can scale, revenue and cash flow guidance for 2027-2028, and margins after new compute capacity caps come online. There's also short-term pressure: the company notes it will recently add short-term compute at higher unit prices, which will compress operating margins, so margins may decline in the short term. The 85.4% calculated gross margin also needs methodology scrutiny—cloud channel revenue is recognized gross, channel fees are listed under sales and marketing, and after mechanically deducting all functional sales and marketing expenses, about 62% of revenue remains to cover R&D and management.

Industry Chain Impact: 2027 Is a Key Inflection Point

What impact does Anthropic's massive procurement have on the mid- and downstream? It locks in forward capacity through three legs: long-term cloud service procurement, equipment finance leases, and third-party GPU compute partnerships. 2027 may see concentrated deliveries. The three major clouds' minimum commitments are relatively rigid, and Broadcom equipment lease payment timing mainly depends on the equipment cap. So 2027 is a key inflection point: large amounts of capacity are delivered in concentrated fashion, and depreciation and rent truly begin entering the financial statements.

On the competitive landscape, Anthropic's product advantage versus OpenAI has narrowed since the second half of the year, while the TAM for coding scenarios may slow in 2027-2028. AI industry penetration is rising, and cloud providers themselves are highly competitive. At the personal agent level, Anthropic's products haven't been tested at scale, and its consumer ecosystem is relatively weak. Overseas feedback suggests pharmaceutical AI still lacks revolutionary breakthroughs. With Microsoft, AWS, and Meta launching their own models and overseas open-source models closing the gap, model supply for non-coding scenarios will rise again. At that point, cloud providers' gateway advantage can access more models, helping long-term revenue share rise, and their advantage in one-stop service for top B2B customers will become more prominent. So after the AI hardware cycle, cloud providers may see a revenue and profit window—this is one judgment of the report.

One more detail worth noting: about 5% to 7% of revenue in the first three quarters came from unauthorized use (distillation). This proportion is not small; it is both part of revenue and part of risk factors.

Conclusion: Dynamic Weighting Between Two Narratives

The significance of this S-1 is that it provides a clearer standard—all future rumors will be compared against these numbers. It does raise confidence in long-term compute investment, but the ROI question of AI growth remains unresolved.

In one sentence: this is a company betting on short-term usage growth with long-term rigid contracts. If the bet is right, compute is a moat; if wrong, compute is leverage. What the market needs to do now is find a dynamic weighting between these two narratives.

#Anthropic S-1#AI valuation#compute spending#gross margin#net revenue retention#Public Benefit Corporation#vendor financing
Share:

Related articles