BusinessMarkets, money & strategy
Anthropic files for Nasdaq IPO, warning in its prospectus that its AI could pose catastrophic risk to humanity
Anthropic's S-1 for a Nasdaq listing as "ANT" shows fast revenue growth, a large, mostly accounting-driven loss and roughly $417bn in compute commitments. It came the same month as a leaked federal AI bill drafted by Anthropic and a cofounder's claim that the singularity has begun.
From the MIT Technology Review · AI article of October 26, 2026 and cited sources: Anthropic, PBC filed its S-1 on 26 October 2026, with Goldman Sachs and Morgan Stanley as underwriters. [1] [2] [5] [6]
SOURCES[1] MIT Technology Review: Will the real Anthropic please… technologyreview.com[2] Yahoo Finance: IPO live updates finance.yahoo.com[5] Anthropic: S-1 announcement anthropic.com
AT Pulse research deskPublished · Updated · 5 min read · AI-assisted analysisAI-assisted, not human-edited. Methodology Corrections Who owns this site
- What happened
- Anthropic, PBC filed its S-1 on 26 October 2026, with Goldman Sachs and Morgan Stanley as underwriters. [1] [2] [5] [6] [9]
- Original source
- MIT Technology Review · AI article, independent perspective
- Brief by
- AT Pulse research desk AI-assisted; publisher and contact on the desk page
- When
- October 26, 2026 article published
- Covered
- Oct 28 edition
- Evidence
- 12 sources across 11 sites
- 1MIT Technology Review: Will the real Anthropic please…technologyreview.com
- 2Yahoo Finance: IPO live updatesfinance.yahoo.com
- 3The Next Web: Anthropic to rewrite leaked AI billthenextweb.com
- 4Import AI 477 (Jack Clark)jack-clark.net
- 5Anthropic: S-1 announcementanthropic.com
- 6SEC EDGAR: Anthropic, PBC Form S-1sec.gov
Reach: not measured yet
Anthropic, PBC filed its S-1 on 26 October 2026, with Goldman Sachs and Morgan Stanley as underwriters. Share count and price range are not yet set. Revenue for the first half of 2026 was $16.3bn, compared with $4.6bn for all of 2025. The net loss of $80.7bn came mostly from accounting charges, and Q2 brought the first operating profit, $974m. Compute obligations total about $417bn, including up to $120.5bn with SpaceX through 2029. The risk factors say the technology could pose catastrophic or existential risk. They also say debates over model sentience could lead Anthropic to adopt costly safeguards that go beyond legal requirements. [1] [2] [5] [6] [9]
A draft federal AI bill written by Anthropic leaked on 16 October. It would let the government block a model only for imminent risks above $1bn in damage or 50 casualties, and it would preempt some state laws, including California's SB 53. Critics say auditors would face no licensing standards and that labs could pay their own auditors. David Sacks argues the opposite: lower thresholds would reach too many companies. Anthropic said its preemption wording was imprecise and promised further proposals on standards for independent evaluators. [3] [7] [12]
Leadership messaging is running ahead of the research. At the IMF on 17 October, Jack Clark said the singularity has begun, pointing to benchmark scores he credited to Claude Opus 5.5 and a rival model. These are company leadership claims, and one benchmark uses unpublished problems that outsiders cannot reproduce. In remarks at the Vatican, Chris Olah described evidence of introspection in Claude models. Anthropic's own paper calls that introspection highly unreliable and limited to artificial setups. Dario Amodei defends the 'race to the top' strategy but concedes some decisions could have been handled better. [4] [7] [8] [10] [11]
What is documented: as a public company, Anthropic will owe quarterly disclosures, and the filing shows very large compute commitments.
Read the full assessment
What follows: buyers and regulators should watch whether stated safeguards, possible usage limits and the bill's evaluator rules survive shareholder pressure.
Sources are gathered; writing the dossier now.
Executive brief
Anthropic filed its IPO prospectus on 26 October 2026. Its risk factors say the company's own technology could pose "catastrophic or existential risk to humanity" (Yahoo Finance). They also say debates over model sentience could push it into costly self-imposed limits (MIT Technology Review). The filing came the same month as a leaked Anthropic-drafted federal AI bill and a cofounder's claim that "the singularity has begun." MIT Technology Review asks which Anthropic investors and regulators are dealing with: a safety organisation, a profit-driven frontier lab, or both.
What changed and event timeline
Draft bill leaks
Politico published Anthropic's draft federal AI safety bill. It lets the government block a model only for imminent risks above $1bn in damage or 50 casualties, and it preempts some state laws ().
Clark tells the IMF the singularity has begun
In the Per Jacobsson Lecture in Bangkok, Jack Clark argued that AI is now speeding up science faster than humans can follow ().
Anthropic promises revisions
Anthropic's Anthony Cimino said the preemption language was "not precise enough." He promised further proposals on requirements for independent evaluators ().
- Late Oct 2026
Amodei answers his critic
Dario Amodei replied to Encode's Nathan Calvin, defending the "race to the top" strategy. Zvi Mowshowitz dates the reply 27 October, but MIT Technology Review was already quoting it in an updated 26 October piece ().
S-1 filed
Anthropic, PBC filed to list Class A shares on Nasdaq as "ANT," with Goldman Sachs and Morgan Stanley as underwriters. The share count and price range were not yet set (;).
Capabilities and access
There is no model launch in this story. Clark credited Claude Opus 5.5 with 67.7% on Humanity's Last Exam. He credited a rival system, GPT-6-Astra, with 98% on TheoPhysEval, a set of unpublished theoretical-physics problems (Import AI 477). The IPO is the access event. Retail investors can buy shares once a price is set, and governance stays inside a public benefit corporation (Anthropic).
Technical analysis for researchers and developers
- Benchmark claims: Clark's figures are headline scores in a speech.
- Introspection research: Jack Lindsey's paper on emergent introspective awareness tested Claude models by injecting concepts directly into their internal activations.
- Bill thresholds: The draft's harm triggers ($1bn, 50 deaths or serious injuries) and its three risk categories (bioweapons, model escape, cyberattack) would decide which evaluations carry legal weight (The Next Web).
Read the full section
- Benchmark claims: Clark's figures are headline scores in a speech. TheoPhysEval uses unpublished problems, so outsiders cannot reproduce it, and none of the reviewed sources describe how the scores were measured (Import AI 477).
- Introspection research: Jack Lindsey's paper on emergent introspective awareness tested Claude models by injecting concepts directly into their internal activations. It found the models' introspection "highly unreliable," prompt-dependent and limited to artificial setups. Opus 4 and 4.1 scored best (Transformer Circuits).
- Bill thresholds: The draft's harm triggers ($1bn, 50 deaths or serious injuries) and its three risk categories (bioweapons, model escape, cyberattack) would decide which evaluations carry legal weight (The Next Web).
Claims and evidence
- Financials: First-half 2026 revenue was $16.3bn, against $4.6bn for all of 2025.
- Compute: The S-1 discloses about $417bn in compute obligations (Yahoo Finance). That includes up to $120.5bn with SpaceX through 2029 (Yahoo live).
- Valuation: MIT Technology Review says the IPO values Anthropic above $1.5tn.
Read the full section
- Financials: First-half 2026 revenue was $16.3bn, against $4.6bn for all of 2025. The net loss was $80.7bn, mostly accounting charges, and Q2 brought the first operating profit, $974m (Yahoo Finance).
- Compute: The S-1 discloses about $417bn in compute obligations (Yahoo Finance). That includes up to $120.5bn with SpaceX through 2029 (Yahoo live).
- Valuation: MIT Technology Review says the IPO values Anthropic above $1.5tn. Other coverage describes $1.5–2tn as Wall Street's expectation, since no price range has been filed (Yahoo live).
- The singularity claim: This is Clark's own interpretation, based largely on benchmark gains. No independent evaluation in the reviewed sources supports it.
Context and prior work
- Chris Olah told a Vatican panel that Anthropic finds "evidence of introspection" and emotion-like internal states.
- Calvin's critique listed nine charges, including large compute deals, Anthropic's retreat from unilateral safety commitments in version 3 of its Responsible Scaling Policy, and the draft bill (Zvi Mowshowitz).
- Anthropic played a role in California's SB 53, the state law its draft bill would preempt (Zvi Mowshowitz).
Read the full section
- Chris Olah told a Vatican panel that Anthropic finds "evidence of introspection" and emotion-like internal states. Reports date that panel to 25 May 2026 (The Decoder), while MIT Technology Review says it happened "last year."
- Calvin's critique listed nine charges, including large compute deals, Anthropic's retreat from unilateral safety commitments in version 3 of its Responsible Scaling Policy, and the draft bill (Zvi Mowshowitz).
- Anthropic played a role in California's SB 53, the state law its draft bill would preempt (Zvi Mowshowitz).
Limitations, safety and contested findings
- Introspection framing: A critic says Olah's Vatican remarks dropped the paper's caveats about unreliability and confabulation (Granier).
- The bill, from both sides: Critics say evaluators would face no licensing standards and blocking a model would be too hard.
- Amodei's defence: He cites costly choices, such as delaying the Mythos model and restricting sales in China.
Read the full section
- Introspection framing: A critic says Olah's Vatican remarks dropped the paper's caveats about unreliability and confabulation (Granier). The paper itself makes no claim about subjective experience (Transformer Circuits).
- The bill, from both sides: Critics say evaluators would face no licensing standards and blocking a model would be too hard. David Sacks argues the opposite: lower thresholds would "sweep in most of Silicon Valley" (The Next Web). Tech Times says labs could pay their own auditors (Tech Times).
- Amodei's defence: He cites costly choices, such as delaying the Mythos model and restricting sales in China. He concedes some items "could have been handled better" (Zvi Mowshowitz).
Business and practitioner implications
- Disclosures: Becoming a public company brings quarterly reporting. Buyers of Claude gain clearer insight into Anthropic's finances and roughly $417bn in compute obligations.
- Usage limits: The S-1 says Anthropic could adopt safeguards "beyond existing legal requirements," with sentience debates as one trigger.
- Compliance: If the bill passes as drafted, state rules such as SB 53 could be preempted, and compliance would move to self-selected auditors.
Read the full section
- Disclosures: Becoming a public company brings quarterly reporting. Buyers of Claude gain clearer insight into Anthropic's finances and roughly $417bn in compute obligations.
- Usage limits: The S-1 says Anthropic could adopt safeguards "beyond existing legal requirements," with sentience debates as one trigger. That signals possible future limits on how models can be used.
- Compliance: If the bill passes as drafted, state rules such as SB 53 could be preempted, and compliance would move to self-selected auditors.
- Evaluation: Treat Clark's benchmark gains as claims from the company's own leadership, not as procurement-grade evaluations.