[1 Stock Per Day PLUS] The truth about the AI market revealed by the Anthropic IPO prospectus. The accounting structure of the '$80.7 billion loss' and the 5 key stocks …
Hello, this is Hal.
Thank you for always reading my note.
On October 26, 2026, frontier AI development company Anthropic publicly filed its Form S-1, a registration statement for its initial public offering, with the U.S. Securities and Exchange Commission. An application has been made to list on the Nasdaq market under the ticker symbol ANT, and as the first pure-play frontier AI company to enter the stock market—ahead of OpenAI, which is projected to go public in 2027—this is a landmark event.
Following this S-1 filing, sensational headlines have been circulating in the media and on social media, such as ‘A massive $80.7 billion loss in the first half of 2026 alone’ and ‘A mega-IPO with an expected market capitalization of $1.5 trillion to $2 trillion.’ I believe many people are looking at these superficial numbers and growing increasingly wary, wondering, ‘Are AI companies just bubbles bleeding massive losses?’
However, when you carefully read through the financial statements and notes in the S-1 one by one, two significant facts emerge that are completely different from the impression reported in public news.
The first fact is that Anthropic’s core business has already achieved a dramatic operating profit, and its cash on hand has barely decreased. point.
And the second fact is that the company has committed to purchasing computing resources totaling $637 billion over the next 10 years—approximately 100 trillion yen—with about two-thirds of that being non-cancellable, fixed contracts. point.
In other words, what this S-1 proves is that AI development is not a bubble without substance, but has entered a phase of generating explosive revenue. At the same time, it highlights the industrial structure itself: Anthropic is taking on the risk of demand fluctuations, while $637 billion in guaranteed cash flow is being forced into the infrastructure and semiconductor companies behind it.
In this article, the first half will provide a detailed explanation of the accounting mechanics behind the ‘$80.7 billion loss’ based on primary S-1 information, the company’s core earning power, the economic law where model price cuts lead to rapid sales growth, and the validity and structural risks of the expected $1.5 trillion to $2 trillion market capitalization.
In the second half, the paid portion, I will extract a population of 12 small-to-mid-cap U.S.-listed stocks that will receive direct special demand from this $637 billion in committed spending and the spread of AI agents. After rigorous financial screening, I will thoroughly delve into the 5 ‘true beneficiary stocks’ that survived, covering their latest financial specs, target stock prices, and November earnings catalysts.
I hope you will read this carefully to the end as a basis for judgment to block out the noise of superficial public news and accurately read the next flow of funds from the primary financial data.
Chapter 1: The accounting reality of the ‘$80.7 billion loss in the first half’ and the dramatic operating profit of the core business
First, let’s confirm the truth behind the data most discussed in the market: ‘a net loss of $80.748 billion recorded in the first half of 2026.’ To conclude, this massive loss does not indicate cash outflow from operations at all.
Let’s trace the company’s core business growth speed and changes in revenue structure chronologically from the income statement included in the S-1.
Confirmed revenue trends
-
Full year 2023: $41 million
-
Full year 2024: $386 million, approximately 9.4 times the previous year
-
Full year 2025: $4.585 billion, approximately 11.9 times the previous year
-
First half 2025: $1.163 billion
-
First half 2026: $16.286 billion, approximately 14.0 times year-on-year
Revenue for the first half of 2026 reached $16.286 billion, approximately 14 times that of the same period in the previous year, meaning that in just six months, the company has earned 3.55 times its total revenue for the full year 2025. This is a pace more than four times faster than OpenAI’s 253% revenue growth rate for the full year 2025.
Furthermore, looking at the quarterly trends, the acceleration of growth and the improvement in profit margins become even clearer.
Quarterly trends in revenue, compute resource cost ratio, and operating profit/loss
-
Q1 2025: Revenue $376 million / Compute resource cost-to-revenue ratio 251.1% / GAAP operating loss $935 million
-
Q2 2025: Revenue $787 million / Compute resource cost-to-revenue ratio 176.6% / GAAP operating loss $1.2 billion
-
Q3 2025: Revenue $1.382 billion / Compute resource cost-to-revenue ratio 148.1% / GAAP operating loss $3.089 billion
-
Q4 2025: Revenue $2.04 billion / Compute resource cost-to-revenue ratio 144.4% / GAAP operating loss $2.836 billion
-
Q1 2026: Revenue $4.731 billion / Compute resource cost-to-revenue ratio 71.6% / GAAP operating loss $1.667 billion
-
Q2 2026: Revenue $11.555 billion / Compute resource cost-to-revenue ratio 48.1% / GAAP operating profit of $974 million
Note that the figures for Q3 2026 are left blank in the original S-1 and are scheduled to be finalized in future revisions, but according to reports, it was stated in an investor presentation that revenue reached $16.8 billion, an increase of approximately 45% from the previous quarter.
The most notable point here is the ratio of compute resource and infrastructure costs to revenue. Anthropic’s compute resource costs include not only the costs of providing inference services to customers but also the training costs for developing and learning next-generation frontier models.
In Q1 2025, the company spent $2.51 in compute resource costs to generate $1 in revenue, but due to explosive revenue growth, strong operating leverage kicked in. By Q2 2026, compute resource costs per dollar of revenue dropped to $0.48, a ratio of 48.1%, which is less than one-fifth of the previous level. The gross profit margin, calculated based on inference provided to customers, surged from 40% for the full year 2025 to 85% in the first half of 2026, and even after deducting all training costs for next-generation models, the gross profit margin for Q2 2026 was a positive 51.9%.
As a result, in Q2 2026, the company achieved a clear turnaround to profitability with a GAAP-based operating profit of $974 million, an operating margin of 8.4%. The non-GAAP adjusted operating margin, excluding stock-based compensation expenses, reached 19.0% for the second quarter alone.
So, despite the core business generating such high profitability, why did the final profit/loss for the first half of 2026 result in a loss of as much as $80.748 billion? Breaking down the details from the S-1 income statement, it looks like this:
Breakdown of the $80.748 billion net loss for the first half of 2026
-
Core business GAAP operating loss: -$693 million
-
Interest income from cash management: +$1.111 billion
-
Fair value changes in convertible bonds, derivative liabilities, and future stock acquisition rights: -$72.388 billion
-
Other non-operating expenses such as valuation losses on note purchase agreements: -$8.757 billion
The total core business GAAP operating loss for the first half, including the Q1 loss, was only $693 million, which was fully absorbed by the $1.111 billion in interest income generated from cash on hand, effectively creating a cash surplus of $418 million. Additionally, the non-GAAP adjusted operating profit for the first half was a surplus of $1.618 billion, and the adjusted net profit was a surplus of $2.671 billion.
The $72.388 billion, which accounts for approximately 90% of the net loss, and the $8.757 billion in other non-operating expenses are accounting valuation losses related to convertible bonds and future stock acquisition rights issued by Anthropic in past fundraising rounds, which involve zero cash outflow.
Under U.S. accounting standards, the rule is that as a company’s stock valuation rises, the book value of the liabilities for these financial instruments—which will be converted into stock in the future—increases, and that increase must be recorded as a valuation loss on the income statement.
The truth is that because the company’s valuation soared from $965 billion at the time of the May 2026 funding round to the $1.5 trillion to $2 trillion range expected for the IPO, a total of $81.145 billion in non-cash valuation losses was generated on the books.
As proof, looking at the balance sheet as of June 30, 2026, out of the $141 billion in cumulative funds raised by Anthropic since its founding, a full 93.4%, or $131.651 billion, remains preserved as cash, cash equivalents, and short-term investments without being spent.
There are zero regular interest-bearing borrowings, and a $15 billion credit facility remains entirely unused.
Chapter 2: The law of price cuts exploding sales and the asymmetry of the $637 billion contract
Another important fact revealed by the S-1 is empirical data on the Jevons paradox, where lowering the unit price of AI models leads to an explosive increase in revenue.
Some in the stock market have argued the pessimistic view that “if AI model performance converges and a price war on token unit costs occurs, the revenue of AI development companies will shrink.” However, as the Jevons paradox in economics demonstrates, when technological progress increases resource utilization efficiency and lowers unit prices, uses that were previously not cost-effective are suddenly unlocked, and total consumption explodes beyond the extent of the price drop, causing total revenue to surge.
Every time Anthropic introduces a new generation model, it has intentionally lowered the “effective cost per task” for customers by 25% to 45% by combining token unit price reductions with efficiency features like 1 million token long context windows and prompt caching.
Track record of cost reduction and accelerated revenue growth through model deployment
Claude Opus 4.5, launched on November 24, 2025: Immediately after launching with a lower unit price and cost per task than the previous generation, the quarter-over-quarter revenue growth rate accelerated sharply from 48% in Q4 2025 to 132% in Q1 2026, and recorded a 144.2% increase quarter-over-quarter in the following second quarter.
Claude Fable 5.1, launched on September 1, 2026: Reduced general task costs by approximately 25% and reduced costs for complex coding and advanced autonomous agent tasks by up to 45%.
Claude Opus 5.5, launched on September 22, 2026: Compared to the previous generation Opus 5, it reduced input/output unit prices by 20% and cache reading unit prices by 60%, while maintaining the same level of performance as Fable 5.1 and reducing cost per task by approximately 40%.
With the dramatic reduction in cost per task, corporate AI usage has completely shifted from “one-off chats where humans input questions” to “AI agents that autonomously write code and execute tasks 24 hours a day.”
Claude Code, an autonomous development product that appears 50 times in the S-1 text, has become a standard tool in corporate engineering organizations. In Q3 2026, the top 5% of users utilizing Claude Code spent about 20 times more than those who did not. Furthermore, within Anthropic itself, Claude wrote over 90% of the contributions to the codebase in September 2026, and revenue per employee over the last 12 months jumped approximately 4.7 times from $1.4 million at the end of 2024 to over $6.6 million by the end of September 2026.
Due to this expansion in usage, the revenue retention rate for corporate customers as of September 30, 2026, recorded a staggering figure of approximately 900%. This means that, on average, customers who spent $1 or more a year ago are paying about 9 times that amount one year later.
The number of large customers spending over $100,000 annually has surged approximately 12.5 times from over 950 companies a year ago to over 11,900 companies, with over 1,900 companies spending over $1 million annually, and approximately 250 ultra-large customers spending over $10 million annually.
Furthermore, dependence on specific large customers is rapidly decreasing; the revenue share accounted for by customers other than the top 5 expanded from approximately 64% in Q1 2026 to approximately 79% in Q3, building a robust customer base that involves the majority of Fortune 500 companies.
Valuation of $1.5 trillion to $2 trillion and Anthropic’s own greatest risk
So, is it rational to directly purchase Anthropic stock, which is expected to go public as early as mid-November, at an assumed market capitalization of $1.5 trillion to over $2 trillion?
When calculating the valuation, the Price-to-Sales Ratio (PSR) against the annualized run rate of $46.22 billion in Q2 2026 revenue would be approximately 32.5x at a market capitalization of $1.5 trillion, and approximately 43.3x at $2 trillion.
Even based on the annualized run rate of $67.2 billion derived from the reported Q3 figure of $16.8 billion, a high multiple of approximately 22.3x to 29.8x for the PSR would be applied.
While this might seem justifiable at first glance given the high growth rate, a close examination of the S-1 notes reveals that there are significant structural risks in Anthropic’s own stock that cannot be overlooked.
This is the extreme asymmetry of ‘holding long-term computational resource purchase obligations totaling $637 billion over the next 10 years, despite 80% of revenue being usage-based billing without lock-in’.
Of the $16.286 billion in revenue for the first half of 2026, fixed subscription contracts accounted for only about 20%, or approximately $3.2 billion, while the remaining 80%, or approximately $13.1 billion, was usage-based billing charged according to the volume of tokens consumed by customers.
Since usage-based billing lacks contractual ties to lock in customers, there is a risk that revenue could flow out immediately if competitors release better models at lower prices or if government regulations are imposed.
In fact, in June 2026, Anthropic experienced a situation where it had to temporarily suspend the provision of both models to all customers after the U.S. Department of Commerce issued export restrictions for the top-tier models Fable 5 and Mythos 5 to the entire world.
While there is no long-term guarantee on the revenue side in this way, procurement contracts for computational resources on the cost side must be arranged 12 to 18 months in advance, and most of these are non-cancellable fixed payment contracts.
The total commitment for computational resources with the seven major companies through 2036 disclosed in the S-1 reached approximately $637 billion, and about two-thirds of that, between $417 billion and $431 billion, are non-cancellable contracts that must be paid even if the computational resources are not used.
The breakdown of the counterparties and terms for the $637 billion total contract amount revealed from the S-1 is as follows.
Broadcom: Approximately $161.2 billion (25.3% of the total)
The contract period is 5 years, with some operations starting in 2026 and additional portions starting sequentially from 2027. This is a fully non-cancellable contract and is definitive data showing that Anthropic is procuring custom AI semiconductors and high-speed network equipment designed by Broadcom directly through leasing, not just via the cloud.
This is a contract through 2029, scheduled to be fully operational in December 2026. After the initial non-cancellable period ending in January and March 2027, it can be cancelled with 90 days’ prior notice, but after full operation this December, payments of approximately $3.1 billion per month, or approximately $9.3 billion per quarter, will occur.
This is a 7-year and 3-month contract from April 2026 to July 2033, which is in principle non-cancellable. There is an obligation to pay the difference if usage falls below the minimum guaranteed amount, and in addition to this amount, there are additional TPU procurement contracts with undetermined amounts estimated at $125 billion to $150 billion.
This is a 10-year contract from May 2026 to April 2036, which is also in principle non-cancellable and accompanied by an obligation to pay a minimum guaranteed amount. It secures demand for the next 10 years for the Trainium ecosystem, which is AWS’s custom AI chip.
Contracts have been signed with Microsoft for at least $44.9 billion (from November 2026 to May 2033, in principle non-cancellable), with AI-specialized cloud provider Nscale for up to approximately $44.6 billion (until 2033, in principle non-cancellable), and with GPU cloud provider Lambda for approximately $41.3 billion (until 2034, in principle non-cancellable). Furthermore, a contract has been signed to purchase AMD’s next-generation Helios rack systems starting in 2027, and expenditures will exceed $20 billion if the full amount is purchased. Combining Google’s additional TPU quota and the AMD quota, the total infrastructure demand generated by Anthropic alone reaches approximately $782 billion to $807 billion, or about 120 trillion yen.
Please try to look at this structure calmly from an investor’s perspective.
Buying Anthropic parent shares at a PSR of over 30x means simultaneously taking on the risk of fluctuating usage-based revenue and the risk of $637 billion in fixed costs. However, from the perspective of the infrastructure and semiconductor supply chain receiving funds from Anthropic, it has been legally confirmed that a high-quality customer—which holds $131.7 billion in cash on hand and will raise up to $100 billion more in an IPO—will reliably transfer a total of over $637 billion in cash flow over the next 5 to 10 years under non-cancellable terms.
And this massive influx of capital into giants like Broadcom, AWS, and Google, combined with the explosion in AI agent demand due to a 40% reduction in inference costs, will dramatically boost the performance of “specific small-to-mid-cap stocks” that provide essential components and infrastructure behind them.
The content beyond this point is for paid subscribers only.
In the paid section, we will thoroughly dissect the 5 “true beneficiary small-to-mid-cap stocks” selected through the following steps, based on the capital recycling routes identified from Anthropic’s S-1.
-
Chapter 3: Financial screening of 12 beneficiary candidates and the reasons why 7 were dropped
-
We apply 5 rigorous quantitative filters to 12 small-to-mid-cap stocks extracted from three areas: custom semiconductors/high-speed communications, data center power/cooling/nuclear fuel, and AI data construction/specialized cloud. We will clearly state the financial flaws of why stocks bought solely for their thematic appeal were dropped.
-
-
Chapter 4: A thorough dissection of the financials and catalysts for the “5 top stocks” that will monopolize the $637 billion recycling
-
For the top 5 companies that survived the screening, we will disclose all confirmed financial specs as of October 2026, valuations such as projected P/E and PSR, Wall Street target prices and upside potential, and specific stock price catalysts directly linked to the early November earnings announcements and the Anthropic IPO.
-
Please use this as a practical guide to get ahead of the curve and capture the top companies that will receive these confirmed massive capital expenditures, rather than jumping into the IPO frenzy at high prices.
Would you like to join the “Hal’s Asset Design Department” membership?
This is a very cost-effective plan where you can read almost all of my paid articles, including this one, for a monthly fee of 500 yen.
You are more than welcome to join and cancel after just one month to try it out, so please feel free to join.
▼ We also welcome your questions
Chapter 3: Financial screening of 12 beneficiary candidates and the reasons why 7 were dropped
The areas directly impacted by the total of $637 billion—and potentially up to $807 billion—in infrastructure investment executed by Anthropic, as well as the explosive spread of AI agents due to lower inference costs, can be broadly divided into three categories.
The first is the area of “high-speed interconnect and optical communication semiconductors” that connect hundreds of thousands of custom semiconductors with low latency in Broadcom’s $161.2 billion equipment leases, Google’s TPUs, and AWS’s Trainium clusters.
The second is the area of “power distribution equipment, liquid cooling systems, power semiconductors, and next-generation nuclear fuel” that support SpaceX AI’s $3.1 billion monthly massive clusters and data center expansions.
The third is the area of “AI training data construction and safety evaluation” that supports the development speed of Anthropic launching 250 models and features in 9 months, as well as “agentic cloud” services, for which demand is surging due to lower inference costs.
We extracted 12 highly relevant US-listed small-to-mid-cap stocks from these three areas as our population.
The subjects of our investigation were