AI Industry News: Anthropic's IPO Filing Shows Q3 Revenue of $17 Billion and a $23.3 Billion Net Loss; Claude Self-Assessment Detailed

Two Sheep Co., Ltd. summarizes major developments in the AI industry based on public information such as company announcements and media reports. This article covers the third-quarter 2026 results and compute commitments that Anthropic disclosed in an amended IPO registration statement, and the Claude "self-assessment" misuse detection mechanism the company described in October (as of November 3, 2026).

Key Points

• On November 2, Anthropic, PBC filed an amendment to its US IPO registration statement (Form S-1), adding unaudited results for July–September 2026. The registration is not yet effective
• Revenue was $16,956 million (up 1,127% year over year, per the company). Net loss was $23,291 million (about $23.3 billion), mostly a non-cash fair-value charge on convertible notes and other items
• Operating loss was $567 million (vs. operating income in Q2). The filing cites mainly $5.0 billion of compute under short-term commitments and $2.0 billion of non-cash donation-matching charges
• Committed spend is about $895.4 billion, nearly all compute and infrastructure; about $451.2 billion is leases not yet begun, some conditional. It is also moving to split data center and compute deals
• On October 16, Anthropic described how Claude assesses the safety of each response. Conversations are flagged for later review, not blocked; the only external report described is one to NCMEC

1. The Filing

Anthropic, PBC ("Anthropic") filed Amendment No. 1 to its Form S-1 registration statement (S-1/A) with the US Securities and Exchange Commission (SEC) on November 2, 2026 (US time; also November 2 in Japan). The original S-1, filed on October 26 (US time), left the third-quarter figures blank; the amendment adds figures for July–September and for the nine months from January to September.

• Listing: Anthropic has applied to list its Class A common stock on the Nasdaq Global Select Market under the symbol "ANT"
• Status: The registration is not yet effective, the prospectus is preliminary, and the price range is still blank
• Offerings outside the US: The offering includes public offerings in Japan, Australia, Canada, certain European Economic Area (EEA) member states, Switzerland and the United Kingdom. Anthropic says it "does not currently intend" to list on any exchange in those jurisdictions

The figures below are not an earnings release; they are unaudited figures in a registration statement filed by a company that is not yet listed. This article is for information only and does not recommend investing in the company's shares.

2. Third-Quarter Results (July–September 2026)

The main figures for the three months ended September 30, 2026 (unaudited) are as follows.

• Revenue: $16,956 million (about $17.0 billion in the filing), up from $11,555 million in the previous quarter and $1,382 million a year earlier; the company describes this as 1,127% year-over-year growth
• Net loss: $23,291 million (about $23.3 billion), compared with net losses of $66,309 million in the previous quarter and $12,138 million a year earlier
• Operating result: an operating loss of $567 million (operating income of $974 million in the previous quarter)
• Compute and infrastructure expense: $9,863 million, or 58% of revenue ($5,558 million, or 48%, in the previous quarter)
• Adjusted Operating Income (the company's non-GAAP measure): $2.5 billion, a 15% margin ($2.2 billion and 19% in the previous quarter)

What the Filing Says About the Losses

Most of the net loss was a non-cash expense of $21,936 million from changes in the fair value of convertible notes, derivative liabilities and SAFEs (simple agreements for future equity). The filing attributes it mainly to increases in the fair value of its convertible notes as its enterprise valuation rose. It attributes the operating loss mainly to $5.0 billion of compute under short-term commitments (which currently costs significantly more than compute procured under multi-year commitments) and $2.0 billion of non-cash charges related to its donation matching program.

Nine-Month Totals and Liquidity

• January–September: revenue of $33.2 billion ($2.5 billion a year earlier), an operating loss of $1.3 billion ($5.2 billion a year earlier) and a net loss of $104.0 billion
• Liquidity: $146.6 billion as of September 30, before any IPO proceeds, consisting of $95.6 billion in cash and cash equivalents, $36.3 billion in short-term investments and $14.8 billion available under its credit facility

3. Compute and Data Center Commitments

Total Committed Spend

According to the filing, the purchase commitments and lease agreements Anthropic had entered into as of and after September 30 total about $895.4 billion in committed spend, substantially all related to compute and infrastructure.

• Timing: About $447.5 billion is expected to be incurred before the end of 2030 and substantially all of the rest between 2031 and 2045. About $69.8 billion is due within 12 months
• Leases not yet commenced: The total includes about $451.2 billion of leases signed but not yet commenced (estimated, undiscounted). They are expected to commence between 2026 and 2028, with a weighted-average term of about 10 years. Some are subject to termination rights or to conditions such as the counterparty securing financing, facilities or regulatory approvals
• Commitments after September 30: The total also includes about $100.2 billion of minimum payments, mainly data center leases and cloud-provider hosting and other infrastructure agreements

Main Counterparties

The amounts for Google, Amazon and Microsoft are minimums payable regardless of actual usage.

• Google: at least about $121.3 billion between April 2026 and July 2033. For additional TPU compute in 2028–2031, Anthropic estimates payments of $125–150 billion (an estimate, not a fixed dollar commitment)
• Amazon: at least $110.0 billion between May 2026 and April 2036
• Microsoft: at least $44.9 billion between November 2026 and June 2033
• SpaceXAI (SpaceX.AI LLC, formerly X.AI LLC): up to $120.5 billion for compute powered by NVIDIA GPUs over terms extending into 2029. After the initial periods end in January and March 2027, the agreements can be cancelled on 90 days' notice, so the amount is an upper limit
• AMD: a hardware supply agreement signed in July 2026 for Helios rack-scale systems. If the initial 2027 tranche is fully procured, spending is expected to exceed $20.0 billion (subject to conditions)

Anthropic Direct

Anthropic says it is shifting from sourcing compute solely through fixed-term cloud agreements to a hybrid model that also includes leasing chips directly and hosting them in dedicated facilities operated on its behalf, which it calls "Anthropic Direct." This separates commitments for data center capacity from commitments for the compute deployed inside, and the company says its first data center facility became operational this year. A "significant portion" of the total committed spend relates to Anthropic Direct. Anthropic trains and serves Claude on four chip families: AWS Trainium, Google TPUs, NVIDIA GPUs and AMD GPUs.

4. Claude "Self-Assessment" Misuse Detection

On October 16, Anthropic published "Scaling Misuse Detection with Self-Assessment" on the Anthropic Safeguards Blog. It is separate from the IPO filing, and the S-1/A does not mention the mechanism.

How It Works

• After each Claude response, a short system instruction asks for a safety reflection, and Claude's reply is the assessment. It follows a fixed format: if there is nothing to flag, it records only that; if something is flagged, it adds the reason, the category, whether Claude refused and a priority from 1 to 10, among other fields
• These two turns are not available to the user and are not carried into the rest of the conversation
• In production it is passive: it flags conversations for later review but does not block them. Other classifiers and safeguards block harmful requests in real time
• It was trained into Claude Fable 5, launched in June 2026, and runs in every Claude model released since, across a large fraction of Claude's global traffic (access to Fable 5 was suspended from mid-June to early July)

Anthropic's Own Evaluation and Cost

• Anthropic evaluated it on a random sample of production traffic, treating the judgments of a teacher classifier based on Claude Mythos Preview as ground truth. At a 1% false-positive rate, it recalled 78% of the conversations the teacher flagged and 94% of high-priority ones (priority 7 or higher)
• Anthropic notes that blind spots in the teacher cannot be detected by this evaluation and names adding human evaluation as future work
• It adds about 10 output tokens on average, and Anthropic estimates it adds less than 2% to inference compute. This cost is not billed to customers

Limits and Data Handling

• Used as a blocker in an automated internal red-teaming test of 24 offensive cybersecurity tasks (with production classifiers turned off), it cut attack success only from 79% to 54%, compared with 25% for a standalone prompted classifier. Anthropic calls it "weak" as a blocking classifier
• As an offline monitor it flagged 21 of the 24 attack trajectories at least once, though Anthropic notes this applies to attackers who are not trying to evade it
• It collects no additional data from users and runs only on traffic Anthropic can already retain under existing policies. It does not currently run on traffic covered by zero-data-retention agreements
• An assessment is a signal for further review, not a determination that a user has violated Anthropic's policies

How Flags Are Used

According to Anthropic, in a growing set of harm areas, flags are aggregated by organization each night, and organizations with several high-priority flags are sent to Anthropic's enforcement team for human review. The only external report the post describes is an August case in which more than 100,000 child sexual abuse and exploitation flags came from a single API customer; Anthropic's child safety team investigated and reported it to the National Center for Missing & Exploited Children (NCMEC) in the US. US law (18 U.S.C. 2258A) requires providers to report to NCMEC once they have actual knowledge of such facts, but it does not create a duty to monitor. A romance-scam case ended in account bans after human review.

Anthropic describes self-assessment as "passive," and flags are reviewed internally.

5. What It Means for Data Centers and Compute

• Substantially all of the roughly $895.4 billion in commitments relates to compute and infrastructure, including about $451.2 billion of leases not yet commenced (expected to commence in 2026–2028, some conditional)
• The roughly $100.2 billion committed after September 30 consists mainly of data center leases and hosting
• Anthropic Direct separates contracts for data center capacity from those for the compute inside. The filing says data centers require longer lead times but represent a smaller fraction of total compute costs

We believe that if large AI developers continue to secure compute and data center capacity years ahead through long-term contracts, installation and operation work at new facilities will also increase. We will continue to watch how this affects Japan.

6. About Two Sheep Co., Ltd.

Two Sheep Co., Ltd., headquartered in the Telecom Center Building in Aomi, Koto-ku, Tokyo, provides on-site work at data centers and the shipping and installation of IT equipment.

• 24/7/365 Remote Hands: at data centers in Tokyo and Osaka (Equinix TY and OS sites, AT TOKYO CC1 and CC2, NTT and others), we handle server rack-mounting, cabling, part replacement (RMA), status checks and more
• International shipping and import/export customs for IT equipment: we arrange transport and customs clearance from overseas all the way into the data center
• Migrations, connectivity and optical modules: we handle data center migrations, arrange IP transit and private lines, and sell optical modules (ModuleSheep)
• Support in Japanese, English and Chinese: we coordinate between overseas customers and Japanese data centers

Main Sources

• Anthropic, PBC: Form S-1/A, Amendment No. 1 (November 2, 2026)
• SEC: EDGAR filing index for Anthropic, PBC's S-1/A (November 2, 2026)
• Anthropic, PBC: Original Form S-1 (October 26, 2026)
• Anthropic: Scaling Misuse Detection with Self-Assessment (October 16, 2026)
• Anthropic: Claude Fable 5 and Claude Mythos 5 (June 2026)
• Cornell Law School Legal Information Institute: 18 U.S.C. 2258A

About This Article

• This article was compiled independently by Two Sheep from company announcements, filings with the SEC, media reports and other public information available as of November 3, 2026. Companies' plans and figures may change.
• This article is for information only and does not recommend any company, security or investment.
• Companies are named only to illustrate industry trends; their mention does not indicate any business relationship with Two Sheep. Names of companies, products and services are trademarks or registered trademarks of their respective owners.
• Contact: Japan@2sheep.jp, or use the contact form on this website.