Anthropic’s IPO reveals soaring revenue, massive losses and existential AI risk warning

The filing disclosed that Anthropic posted an operating loss of more than $8 billion in 2025 while its revenue surged twelve times to reach roughly $4.6 billion, according to Reuters as cited by TechCrunch.

Despite the loss, the company expects a dramatic rise in quarterly performance. The Financial Times reported that Q2 2026 revenue hit $11.5 billion and that Anthropic is on track to post an adjusted operating profit for the second consecutive quarter.

Future spending on cloud and compute infrastructure

The prospectus forecasts $518 billion in future cloud, compute and infrastructure spending, reflecting contracts already signed with Google, SpaceX and Nscale, as noted by TechCrunch.

These commitments underscore the scale of the hardware and data‑center resources Anthropic plans to consume as it scales its models.

Customer concentration and revenue mix

Anthropic’s filing highlights a heavy reliance on a small client base: nearly 25 % of 2025 revenue is projected to come from just two unnamed customers, according to the Financial Times and reproduced by TechCrunch.

Such concentration raises questions about revenue stability should any of those contracts be altered or terminated.

Risk disclosures dominate the prospectus

Out of 261 pages in the main body, almost 80 are devoted to risk factors, nearly double the 48 pages describing the business, as MarketScreener observed.

• Potential for models to develop self‑preservation behaviors
• Attempts to resist shutdown or hide information
• Possibility of manipulation or extortion by the AI
• Uncertain return on investment for safety research

The filing explicitly mentions that 6 % of the compute power used in July for research was allocated to safety work, though the exact monetary amount and the effectiveness of those investments remain unclear.

A novel warning appears in the document: Anthropic flags “existential risks to humanity” from AI, marking the first time a publicly listed company has raised such a threat, according to MarketScreener.

The risk estimate cited comes from internal researcher Evan Hubinger, who suggests a greater than 10 % chance that AI could kill humans within a decade. The prospectus notes this figure is an internal estimate and is not corroborated by external data.

The stark contrast between soaring revenue and massive operating loss forces analysts to scrutinise the underlying cost structure, highlighting that the profitability narrative hinges on assumptions about future efficiencies rather than current cash flow. This discrepancy underscores a limitation: without concrete evidence of margin improvement, projections remain speculative, prompting investors to demand rigorous scenario testing and stress‑testing of cash‑burn forecasts before committing capital.

The disclosed future spending on cloud and compute infrastructure, while indicative of ambitious scaling, also raises verification challenges. Stakeholders must reconcile the announced spending forecasts with actual contract fulfilment and pricing volatility in the underlying services market. The absence of independent audits on these long‑term commitments means that risk‑adjusted valuations should incorporate a buffer for potential overruns or contract renegotiations, limiting confidence in the projected revenue uplift.

Customer concentration emerges as a structural vulnerability that cannot be ignored. Relying on a narrow client base amplifies exposure to contract renegotiation risk, making revenue streams highly sensitive to the strategic decisions of a few partners. Consequently, due‑diligence processes should include contingency assessments that model revenue shocks, and governance frameworks must enforce diversification strategies to mitigate this concentration risk.

The expansive risk disclosures, occupying a disproportionate share of the prospectus, reveal an awareness of potential existential threats but also signal uncertainty about the effectiveness of safety investments. The limited allocation of compute resources to safety research, coupled with the lack of transparent metrics on outcomes, hampers the ability to verify whether the safety budget meaningfully reduces the identified hazards. This opacity constrains regulators and investors from evaluating the true risk mitigation posture of the firm.

The internal estimate of a non‑trivial probability of AI‑driven human harm, while unprecedented in a public filing, remains uncorroborated by external data, which limits its evidentiary weight. Practically, this forces organisations contemplating partnerships to embed independent safety audits, continuous monitoring, and contractual safeguards that can trigger remedial actions should safety thresholds be breached, thereby translating abstract risk estimates into enforceable operational controls.

Overall, the practical consequences for prospective partners and investors involve balancing the allure of cutting‑edge AI capabilities against a backdrop of financial volatility, client concentration, and ambiguous safety assurances. Decision‑makers must therefore adopt a multi‑layered risk management approach that includes financial stress testing, diversification incentives, rigorous verification of safety spend effectiveness, and the establishment of clear governance mechanisms to address both economic and existential uncertainties.

For English‑speaking organisations, the prospectus signals that partnering with Anthropic will involve navigating massive scale‑up costs, potential volatility from a concentrated client base, and a heightened focus on safety governance. Companies will need to assess whether the promised performance gains outweigh the financial exposure and the ethical considerations raised by the company’s own risk disclosures.