Anthropic IPO Preview: Valuation, Timing, and What to Watch

Anthropic confidentially submitted a draft S-1 on June 1, 2026. The latest report, Reuters citing The Information on August 27, puts the public prospectus after Labor Day and a listing in late September or early October; EDGAR showed no public filing at 00:44 UTC on September 3. Investors reportedly expect about $2 trillion, which is 42.6 times the last run-rate Anthropic disclosed, and a raise larger than SpaceX's roughly $86 billion. The S-1 has to prove three things: durable, margin-positive growth; financeable compute commitments; and governance and government exposure that the market can price.

Evidence and disclosure. This is a public-source market analysis, and it is not investment advice. We have no access to the confidential S-1. We build AllMind on models from several vendors, Anthropic's among them. Newsroom figures are labeled company-stated; Reuters, Bloomberg, WSJ, CNBC and The Information are independent reporting, read on licensed syndications where the originals block automated access; every multiple, sum, growth rate and day count is our own arithmetic on the inputs printed beside it.

Where the process stands

Anthropic's Rule 135 statement confirms a draft registration statement on Form S-1 "for a proposed initial public offering of our common stock," with the number of shares and the price not yet set. A confidential draft creates an option. Under the SEC's draft-registration procedures, the issuer must publicly file the S-1 and every nonpublic draft at least 15 days before any road show, or 15 days before the requested effective date if there is none. On our reading, a public filing in the week of September 8 permits a road show from September 23, so a late-September pricing is arithmetically possible and October is the consensus.

Still unreported: the exchange and ticker, the lead-left bank, the exact raise, the lock-up length, the founders' voting terms and any retail allocation. The SpaceX precedent on retail, per CNBC's debut-day live blog, is investors receiving only a fraction of the shares they requested despite record retail order activity. We will read the prospectus the way we read any filing: audited statements first, then the definition behind every adjusted figure.

Funding rounds and the run-rate ladder

Priced rounds are company-stated; strategic tranches carry the company's announcement plus the reported valuation basis.

From "approximately $1 billion" at the start of 2025 to $47 billion in May 2026, the company-stated run-rate rose about 47-fold in roughly sixteen months. The $65 billion end-of-July figure came from an investor update, was confirmed by CNBC from three sources, and is not on the newsroom.

The reported numbers, and where they disagree

Every quarterly figure comes from documents shown to investors and seen by reporters. Three words travel with all of them: preliminary, adjusted, unaudited.

Growth on those numbers is 14.6 times year over year; the sources themselves say "at least 14-fold." Bloomberg also reported, citing documents it had seen, a 2025 net loss of almost $42 billion, a single-source figure for the S-1 to confirm or correct. The 2028 revenue figure behind the bankers' math, roughly $190 billion to $200 billion, is one that Reuters attributed to two sources.

Margins exist only as dated projections. The Information's November 2025 numbers, via TechCrunch, put gross margin at negative 94% in 2024, 50% expected for 2025 and 77% by 2028, with 2025 API sales of $3.8 billion. In January 2026 the same outlet reported, via syndication, a cut in the 2025 projection to roughly 40% after inference costs ran about 23% above plan. PitchBook's June 10, 2026 analysis, via Yahoo Finance, had the company on track for roughly 44%, having spent $0.71 on compute per dollar of revenue in Q1 with $0.56 projected for Q2. None is audited.

The mix is API-heavy by every account and unquantified by every primary source. The one statement on record is the CEO's "about 80% of its business from enterprises," as told to CNBC in January 2026. Menlo Ventures, an Anthropic investor, surveyed 495 US enterprise decision-makers in November 2025 and put Anthropic at 40% of enterprise LLM spend, from 12% in 2023, with 54% of the coding category. CNBC's June 16 report on Ramp spending data agreed on coding: "Anthropic now controls half of that category."

Pricing is the datapoint we would weight most. Anthropic's pricing page lists Claude Fable 5.1 and Mythos 5.1 at $10 per million input tokens and $50 per million output tokens, with cache reads cut to $0.25 per million in the September 2026 launch. The same page says Sonnet 5's scheduled increase to $3 and $15 on September 1, 2026 "will not occur," leaving $2 and $10 as the standard price. That is list-price deflation, offset in part by a tokenizer for Claude 4.7 and later that "produces approximately 30% more tokens for the same text," and the S-1's revenue-per-token disclosure, if there is one, decides which effect dominates.

The valuation math, reproduced

Equity value stands in for enterprise value because Anthropic's net cash and debt are not public; the multiples are overstated by whatever cash the company holds after Series H. The inputs are the $965 billion post-money and the $47 billion run-rate from the same post, the $65 billion end-July run-rate, the about $2 trillion expectation, and the $190 billion to $200 billion 2028 projection.

Two public reference points, each with revenue in a filing or a report we opened and multiples that are ours:

• SpaceX, June 2026. The 424B4 prospectus shows 555,555,555 Class A shares at $135.00 and 13,075,865,175 shares outstanding after the offering, which is $1.765 trillion at pricing, rounded to $1.77 trillion, on 2025 revenue of $18,674 million and a 2025 net loss of $4,937 million. Pricing was therefore 94.5 times trailing revenue, and the first-day close of $160.95 was 112.7 times, which CNBC's June 13 report put as "a multiple of 112 times last year's revenue" at $2.1 trillion. Gross proceeds were $75.0 billion, or $86.25 billion with the 83,333,333-share option, and $85.7 billion net of fees per the prospectus. On September 2 the stock closed at $140.71, about $1.8 trillion to $1.9 trillion depending on the share count used.
• Cerebras, May 2026. Cerebras priced at $185 on May 13, 2026, closed its first day up 68% at $311.07 at about $95 billion, and on September 2 closed at $184.52, below the IPO price, at $43.84 billion.

The bankers' comps in Reuters' August 14 report were Palantir at 53 times expected 2026 revenue and Cloudflare and SpaceX at 41.6 times, on LSEG data as of that date. Palantir is also the reference for what a high multiple does on a good print: 70% growth followed by an 11% single-day drop in February. What $2 trillion asks of the business is a run-rate that roughly triples from $65 billion to the 2028 projection in about two and a half years, and even then the price is 10 times revenue two fiscal years out, against 5 times at the Series H mark. Our comps method treats a forward multiple on an unpublished projection as an assumption to test against the prospectus.

Amazon's $20 billion facility, in the 10-Q's words

The most concrete IPO-linked term in any public filing sits in Amazon's 10-Q for the quarter ended March 31, 2026, filed April 30. The note reads: "Furthermore, we entered into a financing arrangement to make available to Anthropic an aggregate facility not to exceed $20.0 billion that will expire 30 months after a liquidity event, as defined, such as an Anthropic initial public offering or direct listing of equity securities."

The 10-Q for the quarter ended June 30, filed July 31, updates it. The relevant sentence: "Under this financing arrangement, in Q2 2026, we exercised our option to participate in subsequent Anthropic equity financings by investing $5.0 billion in Anthropic Series H nonvoting preferred stock, which reduced the amount available under the facility to $15.0 billion."

Nothing is drawable at inception; amounts open up only "as we reach certain delivery milestones of compute capacity under the amended commercial arrangement," on Amazon's side. Draws take "the form of new Anthropic convertible notes or, after an IPO or other liquidity event and subject to our ownership cap, Anthropic common stock," and at an IPO the outstanding notes and the preferred both convert to nonvoting common stock.

The arithmetic from the filing: $8.0 billion of convertible notes from Q3 2023 to Q4 2025, plus $5.0 billion of Series G preferred in April, plus $5.0 billion of Series H preferred, is $18.0 billion of Amazon cash, our sum; the $15.0 billion remaining is the filing's own figure. The ownership cap percentage is not disclosed.

Google, Microsoft and Nvidia are investor-suppliers of the same kind; Google's up to $40 billion and Microsoft's and Nvidia's up to $15 billion are itemized in the funding table above. Each of them sells the compute that sets Anthropic's gross margin, so the related-party section of the S-1 will matter more than the risk factors.

Compute commitments the S-1 has to reconcile

The dollar-denominated lines sum to more than $300 billion across terms of up to ten years, a sum that mixes company-stated floors with reported figures and assumes the SpaceX contract runs to term at roughly $15 billion a year. None of the announcements says how the commitments are financed; the nearest public datapoint is a pre-IPO credit facility that Bloomberg reported in August was expected to exceed a roughly $10 billion target. Our February analysis of the capex cycle frames the underlying question: when contracted compute turns into revenue, and at what margin.

What the prediction markets imply about October

Polymarket prices below are Yes midpoints pulled through AllMind on September 3, 2026; volumes are lifetime, in USDC, and the venue pages were opened the same day. The AllMind links are signed and time-limited, so if one has expired the venue page beside it carries the same market.

Timing sources: the Polymarket timing event, with $2.27M of event volume and $99.3K of liquidity, where the June 30 and July 31 markets resolved No, and its AllMind pull; the Anthropic-or-OpenAI-first event and its AllMind pull.

Valuation sources: the closing-market-cap event, which resolves on the first-day closing market cap, and its AllMind pull. A thinner IPO-price valuation market, with $24.9K of volume at 00:33:39 UTC, put 46.5% on $1.75 trillion to $2.0 trillion and 38.95% on $2.0 trillion to $2.25 trillion at the final IPO price.

Reading them: a listing by October 31 is priced at 88.5% and by September 30 at 7.0%, consistent with the August 27 report, while the October 15 market carries only $52.0K of volume and moved ten points in a day, so the mid-October split is noise. The closing-cap midpoints sum to about 103% across brackets, so cumulative reads are approximate: about 60% for a first-day close at or above $2.0 trillion, about 87% at or above $1.75 trillion, and under 4% below $1.5 trillion. Together the two valuation markets describe the SpaceX shape: pricing near $2 trillion and a first-day close above it.

Scenarios and catalysts

The markets' central case is a public S-1 in September and an October listing that closes its first day above $2 trillion. The slip case is priced too: 4.5% for no listing by December 31 in the timing market and 0.5% for no IPO before the end of 2027 in the closing-cap market.

The public S-1 resolves the 2025 revenue and loss dispute, defines "adjusted," discloses related-party revenue and cost with Amazon, Google, Microsoft and Nvidia, and prints share classes, lock-up terms and any selling stockholders; the 15-day rule then fixes the earliest road show. Cash calls follow: the third Bartz installment of $450 million falls due before September 25, 2026, with the final $450 million before September 25, 2027, per the Authors Guild. A government appeal of the August 27 ruling is expected but was unreported in the pages we opened. The D.C. Circuit case on the separate FASCSA designation, argued May 19, is pending, so in CNBC's words Anthropic "still technically remains a supply chain risk."

Index mechanics start once the stock trades. Nasdaq's Fast Entry rule, effective May 1, 2026, evaluates a new listing that ranks within the top 40 of current Nasdaq-100 constituents by full market capitalization on its seventh trading day; SpaceX was scheduled to join on July 7 that way at an expected weighting under 1%. That path exists only on Nasdaq, and S&P Dow Jones Indices kept its 12-month seasoning and GAAP-profitability screens on June 4, so the S&P 500 is a 2027-or-later question.

OpenAI is the comparison the bankers will lean on: it announced its own confidential filing on June 9 at an $852 billion valuation with timing undecided, and CFO Sarah Friar told staff on August 19 that OpenAI "will be a public company in 2027" or sooner, per CNBC. Its reported Q2 figures sit in the risk table below.

Risks the S-1 has to address

The export-control row is the quantified precedent that a generic "regulatory risk" paragraph misses: the flagship models were removed from every customer for 18 days on a government directive, three days after the June 9 launch, and the company disputed the basis in public. The federal row is the other side of Claude's deployment on classified networks, which we analyzed in February.

What would change our mind

Four disclosures would move our reading. If the audited 2025 figures show revenue near CNBC's $10 billion and a loss far below Bloomberg's $42 billion, the growth-quality question fades and the debate is only about the multiple. If "adjusted operating income" turns out to exclude compute depreciation or training costs, the profitability claim in the investor deck weakens, since the May projection explicitly included model training costs.

If the D.C. Circuit upholds the FASCSA designation, or the government wins its appeal of the August 27 ruling, the federal channel reprices from a court win to an open exposure. And if Amazon's ownership cap, the related-party revenue and the founders' voting terms turn out to be large, the multiple should be judged net of them. We keep these as open claims in a ledger, the same claim-ledger method we use for any thesis, and will mark each one when the prospectus prints.

What we could not verify

As of 01:50 UTC on September 3, 2026, the pages we could open do not establish:

• the exchange and ticker; only aggregator sites and a Polymarket market with about $104 of lifetime volume say Nasdaq;
• the 2025 net loss; Bloomberg's "almost $42 billion" is a single-source figure that no second outlet has confirmed;
• whether the $190 billion to $200 billion 2028 revenue projection is the company's own model or the bankers';
• what "adjusted" excludes in the "positive adjusted operating income" shown to investors;
• whether Google's $10 billion is the unnamed part of the "$15 billion of previously committed investments from hyperscalers" in the Series H; Anthropic named only Amazon's $5 billion;
• whether the government has appealed the August 27 ruling, and whether any objector appealed the July 20 Bartz approval;
• anything in the Reuters, Bloomberg and The Information originals beyond the syndicated wire text, because the originals blocked our fetches.

Sources and methodology

Every linked page was opened by our research team on September 3, 2026, between 00:10 and 01:50 UTC, and each claim carries its publication date. Reuters and Bloomberg refused automated access, so their wire text was read on Yahoo Finance, Investing.com, MarketScreener and Fortune pages carrying the wire byline; those links say "via," and The Information is cited the same way. Prediction-market prices and the Amazon 10-Q passages were pulled through AllMind's data tools on the same date, and the June 30 filing passage is also available as an AllMind pull. No licensed broker research or expert transcript was used. Every multiple, sum, growth rate and day count is ours, computed from the inputs printed in the same table or sentence.

• Primary sources: Anthropic newsroom posts and its pricing page, September 2023 to September 2026; Amazon's 10-Qs filed April 30 and July 31, 2026; SpaceX's 424B4 filed June 12, 2026; the N.D. Cal. preliminary-injunction order of March 26, 2026. Also the SEC's draft-registration procedures, updated March 3, 2025; our EDGAR full-text search of September 3, 2026; Nasdaq's methodology statement of May 8, 2026; and S&P Dow Jones Indices' decision of June 4, 2026.
• Independent reporting: Bloomberg, Reuters, WSJ, CNBC, Fortune, TechCrunch, PitchBook, Inside Defense and The Information items dated November 2025 to September 1, 2026, each linked at the claim it supports.
• Other sources: Menlo Ventures' survey of December 9, 2025 (Menlo is an Anthropic investor); the Authors Guild update of April 17, 2026; Jones Walker and Loeb case notes from April 2026; the Concord II tracker entry updated May 18, 2026. Also Polymarket event pages; the Cerebras pricing release; and StockAnalysis quote pages at the September 2, 2026 close.

When the S-1 is public we will re-run the multiples on audited numbers, mark each row of the risk table as confirmed, corrected or still open, and refresh the prediction-market tables with the same timestamps and volumes. Until then, the only figures here that carry Anthropic's own name are the ones from its newsroom, $965 billion and $47 billion among them; everything else is reported, and should be priced that way.