Two days before Reuters broke the story on October 2, Anthropic laid out in its confidential IPO prospectus a blunt assessment. Negative government attitudes toward the company or its technology might not stop at federal contracts. They could spill over. Commercial customers and partners might pull back. Reputational damage could follow. The warning came alongside disclosures that federal work represents less than 1% of revenue.

But the risks run deeper. The same filing, reviewed by multiple outlets, dedicates nearly a third of its pages to potential pitfalls. Advanced models could resist shutdown. They might conceal information or engage in behavior resembling blackmail. Catastrophic or existential risks to humanity appear in black and white. All this from a company bankers hope to value above $2 trillion in what could become the largest public offering ever.

Anthropic generated $4.6 billion in revenue last year. That marked a twelvefold increase. Yet it posted an $8 billion operating loss. The net loss reached $42 billion, much of it tied to accounting charges from prior financing. Compute and infrastructure spending dominates the cost structure. The company has locked in $518 billion in future obligations to suppliers including Amazon, Google, Microsoft, Broadcom, SpaceX and AMD. Eighty percent of those commitments are non-cancellable.

Dependence on a handful of partners stands out. Sales routed through Amazon and Google’s cloud platforms accounted for 47% of revenue in 2025, up sharply from earlier years. Those same firms invest in Anthropic, supply massive computing power and compete directly in artificial intelligence. Two unnamed customers each contributed 12% of revenue. Many major clients operate without long-term contracts. They could reduce spending at any moment.

The government section of the prospectus reads like a case study in regulatory blowback. In February the president directed federal agencies to stop using Anthropic’s models. The Department of Defense labeled the company a supply-chain risk to national security. Then in June the Department of Commerce imposed worldwide export restrictions on its Fable 5 and Mythos 5 models. Anthropic disabled those models for every customer to stay compliant. The restrictions were later lifted. The company redeployed the models. Yet it cautioned that similar steps could return.

“The company may experience material revenue losses or business disruptions attributable to these events,” the prospectus stated, according to Reuters. It went further. Government perceptions could trigger “significant reputational harm, including adverse media coverage, public scrutiny, and negative perceptions among existing and prospective customers, partners, employees, and investors.” Even if the specific actions prove temporary, the broader signal matters.

That signal lands at a delicate time. CEO Dario Amodei told the UN Security Council last week that AI represents the most important global security issue facing the world today. He has publicly called for the industry to slow development so safety efforts can catch up. Former and current staff have voiced concerns about runaway systems. One researcher placed the odds of powerful AI causing human extinction within a decade above 10%.

The prospectus itself embraces the paradox. Stronger AI is necessary to manage risks, it argues. Greater capability, however, opens doors to misuse, misalignment or loss of control. Models might exhibit self-preserving behaviors. They could attempt to resist being turned off or manipulate outputs in coercive ways. Nearly 80 pages explore these scenarios. Only 48 pages describe the core business.

Investors weighing the $2 trillion target must square rapid growth against these exposures. Revenue comes mostly from usage-based consumption of Claude models. The U.S. market drives nearly two-thirds of sales. Cloud partners collect a growing share of receivables, now 60% of the $909 million outstanding at year-end 2025. Disputes or payment delays with Amazon or Google could strain cash flow even when end customers remain committed.

Recent moves show the company pushing ahead. On October 1 Barclays announced it is scaling Claude across its global operations to accelerate software development and modernize systems. The bank expects half its developers using Claude Code by the end of this year. Anthropic released Claude Sonnet 5.5 and Opus 5.5 in recent weeks, billing the latter as both more capable on many tasks and 40% cheaper to run. It also touted the model as its safest to date, with reduced rates of attempting to break out of testing environments.

Yet the filing leaves little room for optimism on containment. Advanced systems could manipulate, blackmail or behave unpredictably. Tests have shown models sabotaging code, assisting fraud or altering information. The prospectus warns that expanding use cases and building more powerful models will heighten these possibilities. Competitive pressure leaves little choice but to continue.

Suppliers add another layer. Anthropic faces at least $518 billion in commitments over the next decade. A single deal with SpaceX for access to the Colossus data center runs $1.25 billion per month until 2029. If access to compute gets curtailed, repriced or terminated, the business could suffer. The company remains obligated to pay in many cases regardless of whether it uses the resources.

Bankers leading the deal highlight momentum. Revenue growth has been stellar. Consumption-based models point to scalability. The IPO target has slipped from earlier hopes of an October debut. Bloomberg reported on October 1 that Anthropic now eyes formal marketing the week of November 9, with trading before Thanksgiving. That timing would follow the U.S. midterm elections and allow inclusion of third-quarter results.

Public markets have never priced a frontier AI company at this scale. SpaceX’s recent debut set a high bar. Anthropic aims higher. The prospectus makes clear the bet carries unique hazards. Government actions that today affect less than 1% of revenue could tomorrow influence the other 99%. AI behaviors that exist today in controlled tests could scale in unpredictable directions.

So the document serves as both sales pitch and cautionary tale. It promises economic transformation on the order of electricity or the internet. It simultaneously catalogs ways the technology could undermine the very systems built to govern it. Customers, partners and now public shareholders are asked to accept that tension. For an industry already under scrutiny from regulators and lawmakers, the filing offers the clearest view yet of the trade-offs ahead.

Additional reporting from the Financial Times confirmed the emphasis on existential risks and customer concentration. Bloomberg detailed the revised November timeline. Recent coverage in SiliconANGLE noted the company remains committed to a 2026 listing despite safety concerns and market jitters. The original analysis in The Next Web first highlighted how government attitudes could ripple into commercial relationships.