FOR MUCH OF 2026, the dominant fear hanging over Western artificial-intelligence labs was simple: that DeepSeek, a Chinese upstart, would commoditise the frontier. Its pricing was so low it seemed to defy the laws of economic gravity. Its V4 Flash model cost $0.14 per million input tokens and $0.28 per million output tokens -- a fraction of what Anthropic and OpenAI charged for their top-tier models. The fear was that the market-clearing price for frontier AI was heading toward zero, and that the West's richly valued labs would be dragged there with it.
On August 6th, DeepSeek announced a "significant" and "substantial" price increase for its API services. It gave no new rate card, no effective date and no specific multiplier. But the signal was clear enough. Even the lowest-cost provider in the market -- the one that had set the floor beneath which ByteDance, Tencent and every other Chinese rival were forced to price -- could no longer sustain its own economics. The market-clearing price for frontier AI is being discovered from below. And it is higher than the near-zero narrative assumed.
That matters enormously for two companies that are not yet public. Anthropic, which confidentially filed its S-1 on June 1st, is racing toward an October or November listing at a post-money valuation of $965bn. OpenAI filed its own S-1 on May 22nd but is leaning toward delaying its debut to 2027 to preserve a $1trn target, unsettled by tech-stock volatility and SpaceX's rocky arrival. Both have been priced -- in private markets, in analyst models and in the imagination of institutional investors -- with a competitive-risk discount. The fear was that Chinese labs would flood the market with near-free frontier tokens, eviscerating the pricing power that makes nine- and ten-figure revenue run-rates possible. DeepSeek's price hike does not eliminate that risk. But it narrows it.
To understand why, consider the pricing landscape as it stood before the hike. DeepSeek's V4 Flash, at $0.14 per million input tokens and $0.28 per million output tokens, was roughly 71 times cheaper on input and 179 times cheaper on output than Anthropic's Fable 5, which costs $10 and $50 respectively. OpenAI's GPT-5.6-sol, at $5 and $30, sat somewhere in between. On a standardised monthly workload of 1,000 requests per day, OpenAI's flagship cost 27 times more than DeepSeek's flagship. Even DeepSeek's more capable V4 Pro, at $0.435 and $0.87, was an order of magnitude below anything the Western labs offered at comparable capability.
DeepSeek's pricing is 36x–179x cheaper than Western labs, and the announced price hike will narrow this gap from below.
| model | Input price ($) | Output price ($) |
|---|---|---|
| V4 Flash | 0.14 | 0.28 |
| V4 Pro | 0.435 | 0.87 |
| Claude Haiku 4.5 | 1 | 5 |
| Claude Sonnet 5 | 2 | 10 |
| Claude Opus 5 | 5 | 25 |
| Claude Fable 5 | 10 | 50 |
| GPT-5.6-sol | 5 | 30 |
This pricing gap was not an academic curiosity. It translated directly into market share. Chinese open-weight models grew from under 2% of global token consumption in late 2024 to more than 50% by June 2026, according to OpenRouter data. Uber exhausted its 2026 AI budget in four months. Cursor, a developer-tools company, shifted to usage-based pricing. The startup Lindy moved production workloads from Claude to DeepSeek. The cost curve, as one chief executive put it, had "crashed to the ground."
The "DeepSeek danger zone" was the market's shorthand for this dynamic: the fear that Chinese models, by pricing at or below cost, would make the frontier a commodity before Western labs could recoup their enormous capital outlays. It was a credible fear. The same Milk Road analysis that tracked the 50% token-share milestone argued that value was shifting away from the models layer toward scarce infrastructure -- power, memory, packaging, optical interconnect. If the models themselves were interchangeable, the billions that Anthropic, OpenAI and their investors had poured into frontier capability looked increasingly like a sunk cost with no pricing-power payoff.
The reason DeepSeek is now raising prices is instructive. The company is funding a 1-gigawatt AI data centre in Inner Mongolia, the kind of facility that Nvidia's chief executive, Jensen Huang, has estimated costs around $50bn at global prices, though Chinese construction costs are typically lower. Below-cost pricing is a strategy that works until capital requirements overwhelm it. A 1GW data centre is not a marginal expense; it is a bet-the-company commitment that demands revenue to match. DeepSeek had set the market floor. Now its own costs are breaking through that floor.
To be sure, the hike is still a sketch, not a finished rate card. DeepSeek has not disclosed the magnitude -- "significant" and "substantial" are adjectives, not multipliers. The 2x peak-hour surcharge introduced in July applies only to specific daytime windows in Beijing time and does not constitute the general price increase. No Western lab has yet responded with a pricing move of its own, and it is possible that ByteDance or Tencent, seeing an opportunity, will hold the line and absorb the customers DeepSeek's increase might shed. The degree to which the gap with Anthropic and OpenAI narrows remains unknown until actual rates are published.
Yet the direction of travel matters more than the magnitude. The market's worst-case assumption was that the price of frontier tokens would keep falling, driven by Chinese labs with state tolerance for indefinite losses and a strategic interest in commoditising the West's most valuable technology export. That assumption is now in question. If the price floor is rising rather than falling, the gap between DeepSeek and the Western labs narrows even if Anthropic and OpenAI do nothing. The "death zone" narrative -- in which Chinese pricing makes Western models economically obsolete -- becomes harder to sustain.
For Anthropic, the timing could hardly be better. The company's annualised revenue run-rate has surged from $9bn in December 2025 to $47bn by May 2026 -- a fivefold increase in under six months. It has overtaken OpenAI in enterprise market share, capturing 34.4% in April compared with OpenAI's 32.3%. Its confidential S-1, filed on June 1st, sets the stage for a public listing as early as October, with a post-money valuation of $965bn from its May Series H round and a target market capitalisation above $1trn.
Anthropic's annualised revenue run-rate quintupled from $9B to $47B in under six months, underscoring the explosive growth underpinning its $965B pre-IPO valuation.
| period | Annualised revenue run-rate ($B) |
|---|---|
| Dec 2025 | 9 |
| Feb 2026 | 14 |
| Mar 2026 | 19 |
| Apr 2026 | 30 |
| May 2026 | 47 |
That revenue trajectory is extraordinary. It also explains why the competitive-risk discount matters so much. At a $965bn valuation, the market is not just pricing Anthropic's current run-rate. It is pricing the assumption that Anthropic can sustain premium pricing on its models -- that Claude Code, its enterprise relationships and its safety-branded differentiation can command margins that justify the capital. A world in which DeepSeek keeps cutting prices indefinitely is a world in which that assumption breaks. A world in which DeepSeek is forced to raise prices is a world in which the assumption looks more defensible.
OpenAI's situation is both more precarious and, in one respect, easier to assess. The company lost roughly $1.22 for every dollar of revenue in the first quarter of 2026. HSBC estimates it may need $207bn in new financing by 2030. Its decision to lean toward a 2027 IPO rather than accept a lower valuation this year reflects a calculation that the market is not yet willing to meet its $1trn target. If DeepSeek's price hike helps narrow the competitive-risk discount, it may also narrow the gap between what OpenAI's backers think the company is worth and what public-market investors are prepared to pay. That is not a guarantee of a successful listing. It is a reduction in one of the forces pushing the other way.
The broader lesson is that the economics of frontier AI are being discovered in real time, and the discovery is not unfolding as the pessimists predicted. The near-zero pricing that frightened markets in early 2026 was never a stable equilibrium. It was a land-grab financed by venture capital, state subsidy and, in DeepSeek's case, a willingness to price below cost that the cost of a 1GW data centre has now made untenable. The market-clearing price is higher than the market feared. That is not a reason to buy Anthropic or OpenAI at any price. But it is a reason to narrow the discount that the fear of Chinese commoditisation had embedded in their valuations. The price of cheap AI, it turns out, is not so cheap after all.
Commentaires
Pas encore de commentaires