Anthropic Prospectus Breakdown: The AI Investment Logic Has Been Rewritten

A Prospectus You Must Read

When Anthropic filed that 335-page, 260,000-word registration statement with the SEC, it was no longer just another AI company going public. This document offers a rare, complete picture of how a top-tier large-model company makes money, spends money, signs long-term contracts, and views risk.

For investors holding QQQ, the S&P 500, or assets tied to Nvidia, Broadcom, Amazon, Google, Microsoft, or even data centers, power equipment, nuclear power, and gas turbines, Anthropic's financial data directly bears on whether the massive investments across the AI infrastructure chain can find real demand to absorb them.

Over the past three years, the market has mainly seen the earnings reports of the shovel sellers—GPUs, custom chips, cloud services, power equipment. Now, for the first time, we can fully see how much gold the shovel buyers have dug up.

Revenue Growing Faster Than Compute Costs: Unit Economics Are Starting to Hold

In Q2 2026, Anthropic's quarterly revenue reached $11.555 billion, up 1,368% year-over-year, while compute and infrastructure expenses grew 300% year-over-year. Revenue growing far faster than compute costs directly challenges the AI bears' favorite narrative: the stronger the model, the more GPUs it burns, and all revenue ultimately gets eaten by compute costs.

More critically, this quarter Anthropic disclosed for the first time quarterly operating income of about $974 million, an operating margin of roughly 8%. This means that, at least in the most important public sample available, an AI company has proven it can convert compute into increasingly higher economic output.

For the AI infrastructure sector, this is the first layer of significance: if the demand side cannot be profitable, upstream orders will eventually stop; but if the demand side begins to show operating leverage, the upstream capex cycle has a more solid foundation.

Unit Economics is the core analytical framework for assessing whether a single customer or product is profitable, typically calculating the difference between the revenue from a single transaction or user and the direct costs of acquiring and serving that user. In the AI large-model industry, the key variables in unit economics are the gap between inference cost (the compute, electricity, and bandwidth consumed per model call) and pricing. Only when revenue growth consistently outpaces compute cost growth does it show that compute investment is being effectively converted into gross profit, rather than falling into a "the more it grows, the more it loses" trap. Anthropic's swing to positive quarterly operating income is the most direct evidence that this model holds.

Enterprise Customers Using Deeply: Compute Demand Expanding Nonlinearly

As of the end of September 2026, Anthropic disclosed data on more than 11,900 enterprise customers. Even more striking is the spending growth of the same cohort of customers a year later—simply put, a customer that spent $100 last year might spend $900 this year.

Behind this number is a fact more important than net revenue retention itself: enterprises don't adopt AI by buying a software seat and stopping there; they try a little first, and once they find it works, they keep adding volume. First let a few dozen programmers use Claude Code, then expand to hundreds; then bring in customer service, research, legal, and data analysis; then upgrade AI Agents from answering a single question to working continuously for tens of minutes, hours, or even days in the future.

As a result, the compute demand generated by an enterprise customer is not linear growth—it is more like an ever-expanding workload. This is exactly why Anthropic's customer data matters to Nvidia, Broadcom, HBM, networking, data centers, and power companies—what you see is not 11,900 customers, but more than ten thousand entry points that are continuously amplifying compute consumption.

Net Revenue Retention (NRR) measures the ratio of revenue contributed by the same cohort of customers one year later relative to one year earlier, and is a core health metric for SaaS and subscription businesses. NRR above 100% indicates existing customers are expanding usage rather than churning; exactly 100% means standing still. Traditional enterprise software NRR of 110% to 130% is already considered excellent, while Anthropic's disclosed spending growth among the same cohort can reach several times, meaning NRR is far above the industry's normal range. This "customers don't churn and keep adding volume" characteristic is the fundamental difference between AI-native applications and traditional software seat-based business models, and the direct source of nonlinear compute demand expansion.

Long-Term Contracts and Revenue Mismatch: Moat or Lead Weight?

This is where the document truly gets dramatic. Anthropic discloses that, for future model training and inference, it must lock in compute 12 to 18 months or even longer in advance, rather than waiting until next month when GPUs are short to go buy them.

According to statistics from multiple contracts disclosed by the company, Anthropic is already involved in more than $600 billion in long-term compute and infrastructure arrangements. One five-year equipment leasing arrangement alone carries total rent of about $125.2 billion, excluding interest. The company also disclosed that it has deposited restricted cash with Broadcom-related arrangements to lock in compute costs.

These two exaggerated numbers are not the point; what really matters is the timing mismatch: the revenue side is mainly usage-based—customers pay more if they use more today, and can use less if budgets change; but much of the cost side consists of long-term minimum purchase contracts and long-term leases, which must be paid whether used or not.

This is the core risk of the AI infrastructure cycle—revenue is flexible, costs are increasingly fixed. If future demand continues to explode, these long-term contracts are a moat; if demand falls short of expectations, they instantly turn from a moat into the heaviest lead weight.

Chip Landscape: GPUs No Longer Win Everything; Custom ASICs Gain External Validation

Anthropic explicitly treats compute as a unified, interchangeable pool, using AWS Trainium, Google TPU, Nvidia GPUs, and AMD GPUs, dynamically allocating based on price and performance.

The most direct positive for Nvidia is that this document proves frontier AI compute demand is real and solid—Anthropic had severely underestimated compute demand, and at one point capacity shortages affected product development and deployment. In other words, their biggest mistake was not buying too much, but buying too little.

But the prospectus also contains the most cautionary line for Nvidia long term: the total AI compute pie will keep growing, but who gets how much will be re-competed. For Nvidia, the real question is not whether its share will decline, but whether the total pie can grow faster than its share declines over the long term.

For Broadcom, Anthropic's multi-chip architecture is essentially a real-world, large-scale validation of the custom ASIC route. If specialized chips like Google TPU and AWS Trainium can handle large-scale training and inference of frontier models, then the custom accelerator and networking infrastructure ecosystem is no longer just a small experiment by cloud vendors to optimize internal costs.

For Amazon, the relationship with Anthropic is special—it is simultaneously an investor, a cloud provider, and part of the chip supply chain. If Anthropic puts more and more workloads on Trainium, AWS gains not just cloud revenue, but the opportunity to retain within its own system some of the economic value that would otherwise go to third-party GPUs. The same applies to Google—Anthropic's willingness to put TPU into a unified compute pool is a high-value validation vote for TPU's external usability.

AMD is more like an option on a second supply chain—Anthropic's inclusion of AMD GPUs in its multi-chip architecture shows that customers are unwilling to tie their fate to a single supplier.

ASIC (Application-Specific Integrated Circuit) is a chip customized for a specific task, as opposed to general-purpose parallel computing chips like Nvidia GPUs. The generality of GPUs lets them serve almost all AI models, but the cost is that some transistors and energy go to non-current tasks; ASICs deeply optimize circuit and memory structures for specific model architectures or training/inference workloads, theoretically achieving higher energy efficiency and lower unit cost at the same process node. Google TPU and AWS Trainium both fall into this category. Their risks are lower flexibility, high upfront tape-out costs, and long architecture iteration cycles—once model structures change dramatically, the advantages of specialized chips may shrink quickly. Therefore, Anthropic's inclusion of custom ASICs in a unified compute pool is seen as important validation of this route under real frontier workloads.

AI Bottlenecks Spreading into the Physical World: From Chips to the Power Grid

If you trace Anthropic's compute procurement all the way down, you'll find it doesn't stop at code—it stops in a very clunky physical world: chips go into servers, servers go into data centers, data centers need switches and high-speed interconnect, need cooling, UPS, transformers, transmission, and finally need a power plant somewhere to actually deliver the electricity.

Anthropic devotes nearly a third of the prospectus to discussing risks separately, even specifically addressing power, water, data center construction permits, and equipment delivery. The company also mentions that AI companies may in the future need to bear the costs of grid upgrades and new power plants.

This means AI's bottleneck is gradually spreading from "do we have GPUs" to "do we have enough of the entire industrial system." A software company building large models ultimately affects networking equipment vendors, cooling and power supply providers, electrical equipment companies, gas turbine manufacturers, nuclear power assets, and even electricity prices and grid investment in some parts of the United States.

Data center power constraints are an often-underestimated part of the current AI infrastructure discussion. A large AI data center can consume hundreds of megawatts, comparable to a small or medium-sized city, with highly concentrated and volatile loads, imposing requirements on grid interconnection capacity, transformers, switchgear, and backup power far beyond traditional data centers. In the United States, building new transmission lines and grid interconnection approvals often take years, and lead times for key equipment such as gas turbines and transformers have also lengthened. Therefore, the pace of compute expansion depends not only on chip supply but is also constrained by power interconnection, water cooling, and construction permits. This is why the prospectus devotes extensive space to power, water, and permitting risks—a software company's growth curve is ultimately constrained by the delivery speed of physical infrastructure.

Impact on Index Investors: You Don't Own Anthropic Stock, But Your Account Is Already Affected by It

Anthropic itself is part of the chain of funders, suppliers, and beneficiaries in this AI capex cycle. Microsoft, Amazon, and Alphabet are spending; Nvidia and Broadcom are selling compute; TSMC is manufacturing; and a large group of networking, storage, data center, and power companies are carrying the baton downstream.

The market's answer to the question of whether AI capex ultimately pays off will affect both sides: on one side are hyperscale cloud vendors—if returns are insufficient, massive capex will compress free cash flow and valuations; on the other side are the shovel sellers—once cloud vendors hit the brakes, upstream orders will immediately be repriced.

Thus, a problem that originally seemed to be Anthropic's own will transmit along the capex chain to the Nasdaq, and then to the S&P 500. This is why saying you don't own Anthropic stock doesn't mean Anthropic has nothing to do with your account—if you hold QQQ, in a sense you already hold a whole basket of derivative exposure to "whether Anthropic can turn compute into revenue."

The Debate Escalates: From "Is AI Real?" to "Who Gets the Profits?"

This document does not make the AI bubble debate simpler; instead, it makes the debate more sophisticated.

The bulls can say: demand is real, revenue has surged from tens of millions to billions of dollars, enterprise customers are increasing, the same cohort of customers may spend many times more a year later, and compute costs as a share of revenue are still declining.

The bears can also say: true, but precisely because everyone sees this curve, the entire industry is signing unprecedented long-term compute contracts—if real demand over the next decade is even slightly below today's most optimistic forecasts, fixed costs will crush returns.

Neither view is nonsense. What's truly interesting is that the AI bull market has moved from stage one to stage two: the core question in stage one was "Is AI real?", and the core question in stage two has become "Of course AI is real, but who will ultimately capture the profits?"

Four Questions to Watch Next

First, can Anthropic's revenue continue to outpace compute costs? If so, the unit economics of the AI application layer will become increasingly credible.

Second, will enterprise customers continue to significantly increase usage after reaching a high base? A 900% net revenue retention rate cannot last forever, but if it declines to a still very high level, it shows that Agents are genuinely expanding compute demand.

Third, who will the multi-chip architecture allocate new value to? How much can Nvidia still take, and how much can Google TPU, AWS Trainium, Broadcom custom ASICs, and AMD each take?

Fourth, and most important: can the compute actually used by the industry over the next three to five years catch up with the long-term contracts already signed and the data centers under construction? If the answer is yes, today's seemingly outrageous capex may look conservative in a few years; if the answer is no, the AI revolution can still continue, but the AI infrastructure sector will experience a very traditional cyclical reckoning.

Technically Correct Does Not Equal Investment Correct

Two years ago, the biggest risk was that AI might be just a flashy theoretical project that enterprises simply wouldn't pay for. Now that risk can be said to be almost nonexistent for mainstream AI large models. But a more classic and more dangerous capital market question has emerged: a technological revolution can be 100% real, and yet the people who invest in it can still lose a lot of money.

Railroads were real, the internet was real, fiber optics were real, and smartphones were real. Many historical bubbles were never because the underlying technology was fake, but because capital ran too fast on a real trend, saw too much, and built out the next decade's demand three years ahead of time.

So when studying AI infrastructure now, you can no longer just ask how much demand is growing; you also need to ask one more question: is supply expanding faster, or is demand expanding faster? If AI revenue doubles every year while compute, power, and data center supply grow only 30%, the infrastructure chain will remain in shortage; but if all hyperscale cloud vendors, model companies, and sovereign funds simultaneously see today's growth curve and build five- and ten-year capacity according to the most optimistic scenario, then even if the AI revolution succeeds, some segments could absolutely see price wars, declining utilization, and collapsing returns.

This is the distance between technically correct and investment correct. Anthropic's prospectus does not end the AI bubble debate; on the contrary, it pushes the debate to a stage more worth discussing for investors: who will ultimately capture the profits created by this revolution, and how great a price have we already paid in advance to get them.