Anthropic IPO: What Public Company SEC Filings Reveal

Anthropic's own prospectus is still confidential, but its investors, customers and partners have been disclosing it in their SEC filings all year. If you're on an SEC reporting or technical accounting team, those filings are a free case study in how to account for a hot private stake, and how to write about depending on one AI vendor.

Key takeaways

• Anthropic has filed nothing public with the SEC yet. It confidentially submitted a draft S-1 on June 1, 2026. Everything on EDGAR comes from other companies.
• Amazon runs two accounting models for one investee: convertible notes at fair value through other comprehensive income, and nonvoting preferred stock at cost with adjustments for observable price changes. By June 30, 2026 the two carried about $190 billion combined.
• Salesforce and Zoom use the measurement alternative too. Anthropic's funding rounds are the "observable price changes" that move their earnings.
• Only 5 of 11 AI-using companies we reviewed name Anthropic in Item 1A. The rest name it in the business section and keep risk factors generic, until a government action forced several to name it in mid-2026.

What do SEC filings reveal about Anthropic before its IPO?

Anthropic's IPO filing isn't public, so the real disclosure trail runs through other registrants. Anthropic says it has confidentially submitted a draft registration statement on Form S-1. Draft statements stay private until the issuer files publicly, so EDGAR shows no Anthropic S-1 yet. For background on the confidential route, see our explainer on the Anthropic S-1 and what EDGAR shows now.

What EDGAR does hold is a steady stream of 10-Ks and 10-Qs from companies that invest in Anthropic, build on its Claude models, or sign deals with it. Read together, they show four kinds of disclosure:

How does Amazon account for its Anthropic investment?

Amazon splits one investment across two measurement models. Its Anthropic convertible notes are "classified as available-for-sale and reported at fair value with unrealized gains and losses included in 'Accumulated other comprehensive income (loss)'" as Level 3 assets, per its Q1 2026 Form 10-Q. When notes convert to nonvoting preferred stock, that piece moves to cost with adjustments for observable price changes.

The filings don't cite codification numbers. Reading the language, the notes follow the available-for-sale debt model in ASC 320 and the preferred stock follows the ASC 321 measurement alternative. That mapping is ours, so treat it as inference.

What happens when the notes convert?

Conversion is where the gains hit the income statement. Amazon's FY2025 Form 10-K explains that on conversion the notes are reclassified "to equity investments accounted for at cost," with the unrealized gain moved out of AOCI "to 'Other income (expense), net.'"

1. Before conversion: fair-value gains sit in AOCI and bypass earnings.
2. At conversion: the accumulated gain is recycled into other income. In 2025 that reclassification was $5.6 billion.
3. After conversion: the preferred stock only moves on an observable transaction or an impairment. Those upward adjustments were $7.7 billion in 2025.

Together they drove most of Amazon's $15.2 billion of 2025 other income. The 10-K says the net gain "is primarily from an upward adjustment for observable changes in price relating to our nonvoting preferred stock in Anthropic."

How big is the stake now?

Source: Amazon Q2 2026 Form 10-Q.

In Q2 2026 alone Amazon recorded "upward adjustments of approximately $50.5 billion" to the preferred stock, and $62.8 billion for the first half, all "related to Anthropic's fundings." Its six-month tax provision "included $15.9 billion of net discrete tax expense primarily attributable to the upward adjustments to our investments in Anthropic."

Key takeaway: a private stake under the measurement alternative is quiet until the investee raises money, then one financing round can swing a quarter's earnings and tax line.

What else does Amazon disclose?

• Variable interest entity: Amazon holds a variable interest in Anthropic but doesn't consolidate it "because we are not the primary beneficiary." Maximum exposure "is generally limited to the current carrying values of these investments and any future funding commitments."
• Funding commitments: a financing facility "not to exceed $20.0 billion," drawn as new convertible notes or, after a listing, common stock. Buying $5.0 billion of Series H preferred cut the available amount to $15.0 billion.
• IPO mechanics: on an IPO, outstanding notes "would be converted to nonvoting common stock, subject to our ownership cap," and Amazon expects "a customary lock-up period."
• Commercial ties: AWS and Anthropic expanded their existing commitment "by more than $100.0 billion over 10.0 years," per the Q2 2026 10-Q.

How do Salesforce, Zoom and Nvidia disclose their Anthropic stakes?

The other investors lean on the measurement alternative, and their disclosures get specific once the stake turns material. Salesforce and Zoom now name Anthropic, quantify the carrying value, and tie gains to specific financing rounds.

A few details worth copying, or avoiding:

• Salesforce flags concentration in its policies note. Anthropic went from about 22% of its strategic portfolio in January to about 45% by July 31, 2026.
• Zoom names the trigger. Its gain was "based on valuation adjustments using information including Anthropic's financing round announced on May 28, 2026." It also explains that the gain drove a higher tax provision.
• Microsoft stays out of its periodic reports. It named a "$3.2 billion gain from our investment in Anthropic" only in its July 29, 2026 earnings release, filed on Form 8-K.

Key takeaway: if a single private holding becomes a large share of a portfolio or moves earnings, investors name it, size it, and tie the remeasurement to the transaction that caused it.

How do companies disclose dependence on Anthropic's models?

Most keep the vendor's name out of their risk factors. Of the 11 companies we reviewed that build on Anthropic's models, only 5 name Anthropic in Item 1A: Figma, Tenable, Roper, Expensify and Veeva. The other 6 (Zoom, Box, Asana, Kaltura, 8x8 and Braze) name it in Item 1 Business as a product partner, then refer only to "third-party AI models" in risk factors.

The risk factors that do name providers follow a recognizable template. Roper's FY2025 Form 10-K has the fullest version:

"These providers may change their terms of service, increase pricing, discontinue services, experience outages, decline to provide certain indemnities, or make changes to their AI models that adversely affect our products or operations."

In its Q2 2026 Form 10-Q, Roper updated that risk to add "and/or change pricing models." It's a small edit, and exactly the kind Part II, Item 1A of Form 10-Q exists for.

What made companies start naming Anthropic?

A government action. In mid-2026 10-Qs, several companies disclosed that the federal government had designated Anthropic a supply chain risk, and spelled out their exposure:

• Figma (Q2 2026 10-Q) disclosed that federal agencies were directed to stop using Claude, "the large language model on which we have built the AI features embedded in our governmental offerings."
• Tenable (Q2 2026 10-Q) warned the designation "could force us to decouple or replace integrated technologies on short notice."
• Veeva (Q2 FY2027 10-Q) added that access to frontier models "may also be subject to sudden suspension or restriction by providers or by governments."

That's the lesson for any team drafting AI risk factors: a generic vendor risk reads fine until one vendor becomes the story. Then the SEC's expectation that risk factors be specific, and that 10-Q updates reflect material changes, pulls the name onto the page. Our guide to risk factor vs MD&A disclosure covers where each piece belongs.

None of the 11 put a dollar figure on vendor dependence. Figma came closest, with "significant purchase commitments with a limited number of foundational AI model providers" and "a material portion of our AI features" relying on them.

How do content owners account for AI licensing and the Anthropic settlement?

The same kind of AI deal lands in revenue three different ways, and a settlement with Anthropic sits off the books entirely. Reddit, Wiley and News Corp all license content to AI companies. None names an AI counterparty for its licensing revenue, but their ASC 606 choices diverge.

The timing choice shows up fast. Wiley's point-in-time model made its AI license revenue fall to $13.7 million in the quarter ended July 31, 2026, from $28.9 million a year earlier, per its Q1 FY2027 10-Q. Reddit's ratable model smooths the same kind of deal into $92.1 million of remaining performance obligations. For more on the judgment calls, see our piece on ASC 606 risks in AI revenue recognition.

The Anthropic authors' settlement stays unrecognized. Wiley and News Corp's HarperCollins are both class members. News Corp says that because "the timing and amount of any potential proceeds receivable under the settlement remain subject to the claims process, the Company has not yet recognized any gain." That's textbook gain-contingency treatment: disclose it, don't book it until it's realized.

How does SpaceX's S-1 disclose its Anthropic compute deal?

SpaceX discloses the deal as a third-party customer contract, with the dollar terms in the prospectus summary and a thinner version in the financial statements. Its S-1/A filed June 1, 2026 describes "Cloud Services Agreements with Anthropic PBC" for compute capacity that "includes approximately 325,000 NVIDIA GPUs."

"Pursuant to these agreements, the customer has agreed to pay us $1.25 billion per month through May 2029, with capacity ramping in May and June 2026 at a reduced fee. After the initial three-month period, the agreements may be terminated by either party upon 90 days' notice."

Three drafting choices stand out:

• Two levels of detail. The prospectus summary and Business section give the $1.25 billion monthly fee. The subsequent-events note in the financial statements says only that the customer "has agreed to pay a monthly fee through May 2029."
• The termination right does real work. At $1.25 billion a month, the full term is roughly $45 billion nominal (our math, not the filing's). The 90-day exit after three months means far less of it is locked in, which is why the filing gives the terms and doesn't present a total backlog figure.
• No related-party or concentration label. Anthropic isn't in the related-person transactions section, and the customer-concentration note covers FY2025, before the deal existed. Our look at how SpaceX framed its IPO covers the wider filing.

What do Form D filings reveal about trading in Anthropic shares?

At least 79 investment vehicles named after Anthropic have filed Form D since 2023, which is the paper trail of its pre-IPO secondary market. They include Hiive, Sydecar-administered series of CGF2021 LLC, and funds such as MW LSVC Anthropic, LLC. Excluding an unrelated hedge fund that shares the name, they account for about 104 Form D and Form D/A filings.

• Timing: first filings rose from 7 in 2023 to 30 in 2024, then 15 in 2025 and 28 in 2026 through September 29.
• Exemptions: 69 vehicles claim Rule 506(b), 10 claim Rule 506(c), and one relies only on the Investment Company Act 3(c)(1) exclusion.
• Size: the latest Form Ds report about $198 million sold to 2,418 investors. That's a floor, since several report indefinite offerings or no sales yet.
• Examples: MW LSVC Anthropic sold $18.4 million to 56 investors at a $100,000 minimum. HII Anthropic-01 sold $16.7 million to 35 investors and paid $817,078 in sales commissions.

Key takeaway: a Form D never names the company a vehicle invests in. The Anthropic link here rests on vehicle names, so treat these counts as a signal of demand, not a verified shareholder register.

What should SEC reporting teams take from these filings?

1. Tie private-stake remeasurements to the event. Zoom and Amazon name the financing round behind each adjustment. It's the clearest way to explain a gain that no market price supports.
2. Watch the tax line. Unrealized gains under the measurement alternative still create deferred tax. Amazon booked $15.9 billion of discrete tax expense in six months.
3. Disclose concentration before you're asked. Salesforce's 45% concentration sits in its accounting policies note, not buried in MD&A.
4. Name a critical AI vendor when the dependence is real. A generic "third-party models" risk factor may not hold up once that provider makes news.
5. Use Form 10-Q Part II, Item 1A for updates. Roper and Figma updated their AI risk factors in the quarter the facts changed.
6. Pick a license model and say why. Point-in-time versus ratable recognition decides whether AI licensing revenue arrives in lumps or smoothly, as Wiley and Reddit show.
7. Keep deal terms consistent across the document. SpaceX gives the monthly fee in the prospectus but not the notes. If your filing splits detail that way, make sure the pieces reconcile.

FAQ

Has Anthropic filed a public S-1?

No. Anthropic announced it confidentially submitted a draft Form S-1 on June 1, 2026. A draft stays nonpublic until the company files publicly, so EDGAR has no Anthropic S-1 yet.

Which public companies disclose an investment in Anthropic?

Amazon, Salesforce and Zoom report carrying values and gains in their 10-Qs. Nvidia disclosed a commitment to invest up to $10 billion. Microsoft named a $3.2 billion gain only in an earnings release, and Alphabet doesn't name Anthropic in its 10-K or 10-Q filings.

What is the measurement alternative?

It's the option to carry an equity investment without a readily determinable fair value at cost, adjusted for impairment and for observable price changes in orderly transactions of the same or a similar security. Salesforce, Zoom and Amazon's preferred stock all follow this approach.

Why does Amazon record Anthropic gains in both AOCI and earnings?

Its convertible notes are available-for-sale, so fair-value gains go to AOCI. When notes convert to preferred stock, the accumulated gain moves into other income, and later price changes on the preferred stock go straight to earnings.

Should our risk factors name our AI model provider?

Name it when your products depend on it in a way a reader would find material. The filings above show companies moving from generic language to named disclosure once a specific provider's risk became concrete.

How can I find every SEC filing that mentions Anthropic?

Use EDGAR full-text search for "Anthropic" and filter by form type. Expect gaps: the full-text index misses some filings, so read the filers you care about directly.