DeepSeek turns an AI price war into an IPO margin test

DeepSeek’s decision to make a 75 per cent discount on V4-Pro permanent gives investors a more concrete way to price the next wave of AI flotations. The issue is no longer only whether OpenAI is preparing to file confidentially for an IPO or whether Anthropic can reach $10.9 billion in second-quarter revenue at a $900 billion valuation. Public investors also need to decide whether peak multiples still hold once a credible rival has shown it is willing to reset prices lower and keep them there.

For equity analysts, that distinction matters first. A temporary promotional cut can be dismissed as customer acquisition. A durable floor cannot. What looked like a hard-charging campaign tactic now reads more like a standing challenge to the revenue assumptions behind the broader AI listing queue, which Semafor said this week could include OpenAI and SpaceX in the same blockbuster cycle.

The pushback is easy to sketch. Cheap tokens are not the same as transferable enterprise demand. Regulated buyers may still hesitate before putting sensitive workloads on the lowest-cost option, especially when data governance, uptime and procurement risk matter as much as sticker price. Yet DeepSeek does not need to win every contract outright. It only needs to establish a visible outside option that forces premium labs to defend their pricing more often and with less room for error.

That is the margin test.

The price floor has moved

The important fact here is arithmetic, not rhetoric. DeepSeek’s own pricing page lists V4-Pro at $0.435 per 1 million input tokens and $0.87 per 1 million output tokens. In the company’s V4 preview release, long-context performance is part of the offer rather than an elite add-on. The lower price, in other words, is not attached to a stripped-down product line. It sits on a model DeepSeek is trying to sell as capable enough for a large share of routine enterprise work.

From there the story shifts from product excitement to capital allocation. A finance chief who believes routine summarisation, classification, coding assistance or customer-support drafting can be handled by a much cheaper model will treat the premium model as an exception tool rather than the default pipe. That is the builder case in this market: value moves up the stack, toward orchestration, routing, verification and fallback systems that decide when a more expensive model is worth using.

Ali Ghodsi, Databricks’ chief executive, made the point bluntly in CNBC’s analysis of the pricing fight:

“You can curb costs really well this way.”

• Ali Ghodsi, Databricks chief executive

The line lands because it captures the procurement logic better than a benchmark table does. Once buyers learn they can route the common path to a cheaper model and reserve frontier systems for difficult or high-risk prompts, the old hope of charging a premium on most usage starts to look brittle. DeepSeek does not need to match every frontier claim to change that behaviour. Being good enough, often enough, on workloads where token bills recur every day may be enough.

That makes this different from a standard tech price war. AI revenue narratives have rested on exceptionalism. Investors have tolerated huge private valuations in part because frontier-model groups were treated less like software vendors in a crowded category and more like scarce infrastructure owners with unusual pricing power. A permanent discount from a rival weakens that scarcity story. The market may clear faster than expected toward a split economy: low-cost base intelligence for broad use, premium reasoning only where the extra performance can be proved and charged for.

IPO maths gets harder

The shift lands awkwardly for the companies already closest to public scrutiny. OpenAI’s expected confidential filing is meant to turn private-market mystique into a public-market proposition. Sarah Friar, the company’s chief financial officer, framed the preparation this week as:

“good hygiene” for a company of OpenAI’s size to “look and feel and act” like a public company

• Sarah Friar, OpenAI chief financial officer

That governance point is sensible. The market question is tougher. Public investors are not being asked merely to buy revenue growth. They are being asked to underwrite the durability of that growth at prices high enough to support extraordinary valuations. CNBC put OpenAI’s private valuation above $850 billion. Separate CNBC reporting said Anthropic is targeting $10.9 billion in second-quarter revenue and what could be its first profitable quarter at a $900 billion valuation. Those are not normal software numbers. They imply a market still willing to believe demand growth and pricing discipline can coexist.

That case now needs more evidence. If model pricing starts to behave more like cloud pricing, with constant downward pressure and customers trained to arbitrage providers, the valuation argument shifts from scale alone to revenue quality. Which sales are sticky. Which workloads really require the frontier tier. Which margins survive once procurement teams stop assuming the best model belongs on every task. DeepSeek’s move does not answer those questions on its own, but it pushes them into the roadshow.

Timing makes the development more than an isolated jab. Investors are already being conditioned to absorb a cluster of huge AI-adjacent listings. The New York Times described SpaceX, OpenAI and Anthropic as racing to go public, while CNBC argued the eventual trading values could challenge the largest incumbents in public markets. In that setting, a permanent cut from DeepSeek does not just pressure one company’s sales plan. It raises the cost of believing the whole cohort deserves to float at the top end of the valuation range at the same time.

The defence may move up the stack

The skeptical case still deserves its due. For many western enterprises, particularly in regulated sectors, the cheapest model on paper may never become the default in practice. Procurement, compliance and legal teams can slow migration, and some customers will pay a premium for governance, auditability, domestic hosting or a provider with a longer record of uptime. That is why the story should not be reduced to a cartoon in which one Chinese price cut wipes out every premium lab’s future margin.

Even so, the more durable defence may no longer sit inside the model alone. Distribution, workflow embedding, enterprise controls and the ability to prove that a higher-cost system earns its keep on the narrow slice of tasks where errors are expensive may matter more. That is a smaller and more exacting investment case than many private rounds appeared to assume. It is also a more public-markets-friendly one because it asks harder questions about product mix, customer concentration and gross-margin quality instead of relying on narrative heat.

Anthropic’s own warning, cited by CNBC’s earlier analysis, was that rivals could end up “winning in global adoption on cost”. One question follows: which workloads move first? Probably the repetitive ones where finance teams can see the savings immediately, and where a fallback to a premium model can be automated rather than argued over. Once that architecture exists inside an enterprise, premium providers stop being universal utilities and start competing for the exception path.

None of that is fatal to the IPO story for OpenAI or Anthropic. High-growth companies can still list well if investors believe they own the high-value edge cases, the customer relationships and the software layer around the model. But DeepSeek’s announcement strips away one comforting assumption. Margin pressure is no longer hypothetical, and it is no longer something bankers can treat as a distant second-order risk. A permanent discount has turned the AI valuation debate away from abstract benchmark supremacy and back toward a much older public-market question: what happens to the multiple when the price umbrella comes down?