DeepSeek, the Hangzhou lab that built its brand on undercutting every frontier AI company in the world, spent this week quietly taking that brand apart. On Wednesday it warned developers that its API prices are about to rise "significantly." On Thursday it resumed an $8 billion funding round and confirmed it had bought into the Shanghai IPO of humanoid-robot maker Unitree. Three moves, one signal: the era of dirt-cheap DeepSeek is ending.
The Price Hike: No Rates, No Date, Big Signal
The company published a notice saying it plans to raise prices across its API services "in the near future," warning that the increase could be substantial. It has not released a new price schedule, named the affected models, or set an effective date. The change is planned, not yet in force.
The timing is hard to read as coincidence. The notice lands roughly a week after V4-Flash-0731, the 284-billion-parameter lightweight model that drew global attention precisely because it paired strong performance with unusually low inference prices.
Current published rates give a sense of the base:
- V4-Flash: $0.14 per million cache-miss input tokens, $0.28 per million output tokens.
- V4-Flash cached input: $0.0028 per million tokens.
- V4-Pro: $0.435 per million cache-miss input tokens, $0.87 per million output tokens.
- Artificial Analysis pegged V4-Flash at about $0.03 per task on average, versus $3.15 for Anthropic's Claude Fable 5.
Epoch AI, cited by the South China Morning Post, put DeepSeek's pricing an order of magnitude below Zhipu AI's GLM-5.2 first-party API. That gap is what made DeepSeek the default low-cost option for developers who wanted capable open-weight models without US-tier bills.
DeepSeek has not explained the hike. The likely context: surging demand for V4-Flash-0731, plus a July pricing change that introduced peak and off-peak rates. Whatever the exact trigger, the direction of travel is now unmistakable.
The Money Moves: An $8B Round and a Robot Stake
Bloomberg reported on Thursday that DeepSeek has resumed its second funding round, seeking close to $8 billion at a valuation near 500 billion yuan, roughly $74 billion. Investment firm Monolith Management is in talks to participate. The round was first reported in July, then suspended weeks later; now it is back on the table, still ahead of a possible mainland IPO.
The second move is a bet on hardware. According to a stock-exchange filing disclosed on August 6, DeepSeek invested 140.8 million yuan, about $20.8 million, in Unitree Robotics' initial public offering on the Shanghai STAR Market, where the robot maker trades under ticker 688836. DeepSeek receives 933,399 shares, a 2.31% stake in the strategic placement, locked up for 36 months.
The filing also reveals a procurement pact. Unitree will give DeepSeek preference when buying model-training services and technical solutions; DeepSeek will favor Unitree when buying robots or exploring embodied-AI applications. Both companies are based in Hangzhou.
The logic is straightforward. Unitree's IPO priced at 150.80 yuan per share and raised about 6.1 billion yuan, near 61 billion yuan in valuation, making it the first mainland-listed humanoid-robot maker in China. Its machines walk, run, and perform choreographed routines, but reliably translating instructions into physical actions in unfamiliar rooms is still the industry's hard problem, and it demands physical-world training data that is far scarcer than text.
That is where DeepSeek's models come in, and where the lab's gap shows. DeepSeek's flagship strength is language: coding, math, reasoning, at aggressive prices. Reuters noted that its commercial line has no integrated vision-language system, even as it publishes separate research like VL2. A steady supply of robots generating embodied data is what Unitree offers in return for priority access to DeepSeek's training services. Unitree's offering priced rich, at a price-to-earnings ratio of 219 against an industry average of about 39, and subscriptions open August 10.
What It Means for the AI Price War
First, the hike does not automatically end DeepSeek's cost advantage. A large percentage increase from such a low base could still leave it cheaper than most rivals. Its mixture-of-experts architecture and training methods keep inference costs structurally low, as University of Hong Kong professor Zhenhui Jack Jiang told the South China Morning Post.
Second, the direction matters more than the size. DeepSeek's aggressive pricing pushed competitors including ByteDance and Tencent to cut their own rates. A substantial increase gives those companies room to raise prices again, or to hold theirs and pick off developers reconsidering DeepSeek.
The broader read: the lab that ignited the race to zero is now raising prices, raising capital, and planting a flag in robotics, all while an IPO looms. DeepSeek is signalling it wants to be a durable commercial giant rather than a permanent discount brand. Whether its rivals let it have it both ways is the next chapter of the story.
Comments