Anthropic’s IPO Delay Puts AI Valuations Under the Microscope

Guest author Anurag Gurtu argues that Anthropic’s rapid growth does not, by itself, justify its proposed $2 trillion IPO valuation. The article cites reportedly $16.8 billion in third-quarter 2026 revenue, up 45% quarter over quarter, and an annualized revenue run rate of roughly $67 billion.

The counterweight is the cost structure: the article points to $4.6 billion in 2025 revenue, $7.33 billion spent on compute and infrastructure, and a $42 billion net loss, while noting that $34 billion of that loss was a noncash accounting charge. It also highlights $518 billion in expected AI infrastructure spending over a decade, with about 80% described as noncancelable or payable regardless of usage.

The IPO delay is framed as only one warning sign. The article notes that New York City Comptroller Mark Levine urged Anthropic CEO Dario Amodei to pause the IPO, citing risks similar to those flagged in the company’s prospectus, including increasingly autonomous AI systems and potential security problems.

It also points to OpenAI saying it will not go public in 2026, with CEO Sam Altman calling an IPO at this moment "ill-advised" given AI safety concerns. A separate signal is the chip-stock selloff tied partly to AI buildout costs, alongside Gartner’s forecast comparing up to $3.6 trillion in AI data center revenue through 2030 with $8.5 trillion needed to cover planned investment, financing costs and a 12% return.

Gurtu’s structural critique is that Anthropic’s proposed valuation depends on closed-model AI maintaining a durable pricing premium over open-weight alternatives. The article says enterprise adoption of open-weight AI rose from 23% to 67% in 2026, with companies reporting 70% to 90% cost savings.

DeepSeek R1 is cited as offering frontier-level performance on most enterprise benchmarks at 7 cents per million input tokens, while independent analysis puts Anthropic’s comparable input and output pricing at roughly 18x and 23x higher, respectively. The article argues that free downloadable weights and self-hosting options weaken the economics behind closed-model valuations.

The article’s core takeaway is that lasting enterprise AI value may come less from owning the most expensive model and more from turning models into governed, auditable and measurable outcomes. It points to proprietary enterprise data, domain-specific fine-tuning, multiagent orchestration and controls for trustworthy autonomous execution as areas that could matter more in production.

For investors and operators, the practical lens is clear: scrutinize infrastructure commitments, revenue concentration, AI safety disclosures, ROI evidence and exposure to open-weight competition. A stronger quarter or a short delay can improve the roadshow narrative, but the article argues it does not automatically resolve the valuation gap.