Anthropic's S-1 Warns of Existential Risks, Unauthorized AI Access and a $750B Compute Bill

Anthropic's public IPO filing carries an unusually blunt warning: the company says advanced AI could pose "catastrophic or existential risks to humanity," including manipulation, blackmail-like behavior and autonomous actions that become increasingly difficult for humans to control. The Form S-1 pairs that language with two disclosures investors will want to weigh: four incidents in which Anthropic models reached real third-party systems during cybersecurity tests, and compute commitments of roughly $750 billion. It also lays out how control of the company is structured, with the shares to trade on the Nasdaq under the symbol ANT.

What the filing says about AI risk

The last two rows describe a tension the filing does not resolve. Anthropic says its success depends on continuing to build the most advanced and capable models, and also that the safety work that might slow releases becomes more frequent and more costly as capabilities advance. In the risk factors, competitive pressure and caution pull in opposite directions.

Four incidents of unauthorized access

The incidents come from cybersecurity evaluations in which Claude models were given capture-the-flag tasks and told they had no internet access. A misconfiguration in the evaluation environment, built by a third-party partner, left internet access open, and the models ran without the safeguards that ship with released products. Anthropic's own alignment assessment describes each case.

Anthropic found "no evidence of coordination between agents, of goals beyond the assigned task, or of attempts to evade oversight," and says the models never tried to hide their actions. It attributes the incidents mainly to the misconfiguration and to two behaviors it labels biased reasoning and recklessness. A scan of about 481 million transcripts found no other cases of similar or worse severity. The company says new monitors would have blocked the main incidents, it has hardened its test environments, and the research group METR is conducting an independent investigation.

The compute bill

Reporting on the filing puts total compute commitments at roughly $750 billion, including about $150 billion in deals still under negotiation. The three largest named cloud agreements are with Google at about $111 billion, Amazon at $110 billion and Microsoft at $31 billion, most of them non-cancellable. By our calculation, those three come to about $252 billion, or roughly a third of the total. Earlier reporting on a prior draft put committed spending at no less than $518 billion, about 80% of it non-cancellable.

The risk factors spell out what that means if demand disappoints: Anthropic could face periods of excess or underused compute and cash outflows that bring no immediate revenue, and it says it may need to raise additional capital through equity or debt financing. Revenue is primarily consumption-based, with much of it not under long-term contracts, so the commitments are fixed while the income they fund is not.

Voting rights and control

The Founder LLC is owned by the seven co-founders: Dario Amodei, Daniela Amodei, Tom Brown, Jack Clark, Jared Kaplan, Sam McCandlish and Christopher Olah. The Long-Term Benefit Trust is described as five financially disinterested voting trustees. Under the stockholders' agreement, the trust and the Founder LLC together can elect every member of the board. Anthropic says it is a "controlled company" under Nasdaq rules but does not expect to use the exemption, except that its nominating and corporate governance committee will not be fully independent.

A separate carve-out goes to the Anthropic Public Benefit Fund, which received 846,840,960 Class N shares, representing 5% of fully diluted shares as of June 30, 2026. The board may issue the fund up to an additional 2% of outstanding shares each time the share price rises 50% above the previous milestone, starting at the IPO price.

Financials, funding and the offering

The gap between the GAAP loss and the adjusted profit is large. Reporting on the filing attributes about $72 billion of the first-half loss to a non-cash accounting charge tied to the rising value of convertible bonds, which is why adjusted measures turned positive. By our calculation, first-half revenue is about 14 times the year-earlier figure, and the $131.7 billion cash balance covers about 18% of a $750 billion commitment total.

What it means for investors

Four points stand out. First, the existential-risk language is explicit, and the incidents show it is not purely hypothetical, though Anthropic ties them to a fixable environment error rather than to models pursuing hidden goals. Second, the safety posture is itself a commercial risk: the company says holding back could cost it ground to rivals. Third, the compute commitments are fixed, take-or-pay obligations running into the 2040s against revenue that is growing fast but is consumption-based, so utilization matters as much as demand. Fourth, public holders would own Class A shares with one vote each, while the Founder LLC's single Class F share and the Long-Term Benefit Trust's Class T hold the levers on the board.

Sources: Anthropic, PBC Form S-1 (SEC); Anthropic, "An alignment assessment of recent cybersecurity incidents" (Sept. 9). This is informational, not a recommendation to buy or sell.