DeepSeek's funding round, price hikes, and Unitree robotics investment

On August 6, 2026, DeepSeek told developers a 'significant' increase to its API pricing is coming, without specifying the amount or the date ^[1]. That vague warning immediately produced three incompatible explanations, and none of the three platforms tracking the story agree on which one is right.

The web coverage frames it as an infrastructure story: DeepSeek says the hike reflects rising compute costs and capacity bottlenecks as demand for its reasoning-heavy V4-Flash and V4-Pro models surges, with agent-style reasoning tasks consuming dozens to hundreds of times the compute of an ordinary chat request ^[2]. Reporting also connects the pricing move to DeepSeek's plan to fund its own data center buildout, including a large facility in Inner Mongolia, using part of its new capital raise ^[3].

X read it differently. Dax (@thdxr), a well-known infrastructure and developer-tools builder, argued that his team has been able to reproduce DeepSeek's current prices on rented GPUs - meaning DeepSeek almost certainly isn't 'losing money' at today's rates, and the hike looks more like traffic shaping by a service that's simply overloaded with demand.

Reddit leaned toward a third read entirely: that DeepSeek's rock-bottom pricing was never sustainable to begin with, and this is the moment the subsidy runs out. Discussion centered on whether the hike reflects a genuine compute shortage, deliberate demand management, or margins that were unsustainable from the start, with heavy comparison to how competitors like OpenAI, Meta, and China's own GLM, Mimo, and MiniMax have repriced their own models. Notably, this isn't DeepSeek's first pricing move this cycle - it introduced a peak-and-off-peak scheme in July 2026 that already doubled costs during busy weekday hours, ahead of the broader increase now being flagged ^[4]. None of the three theories is mutually exclusive, but the fact that infrastructure economics, deliberate rationing, and subsidy exhaustion are all live theories about the same line item says something about how opaque frontier AI pricing has become, even at a company that built its brand on radical transparency around cost.

DeepSeek didn't just enter the AI market in January 2025 - it detonated the pricing model everyone assumed was fixed. The R1 reasoning model matched or beat OpenAI's o1 on key benchmarks at roughly 27 times lower cost, a launch that triggered a global AI price war and wiped roughly $600 billion off Nvidia's market cap in a single day ^[5]. That was DeepSeek's entire identity: frontier-level reasoning at a price point no incumbent could match.

Eighteen months later, the company that started the price war has raised prices twice in a single month. It quietly introduced a peak-and-off-peak pricing structure in July 2026 that effectively doubled API costs during busy weekday hours ^[4], and now - as of early August - it has warned of a second, larger increase on top of that ^[1]. Current list prices for DeepSeek-V4-Flash sit at $0.14 per million input tokens and $0.28 per million output tokens, with the more capable V4-Pro at $0.435 and $0.87 respectively ^[6]- already several times R1's original $0.07-per-million-token launch price, before the newly announced hike is even applied.

The reversal lines up with a broader shift in how DeepSeek is funding itself. Its first-ever external funding round, roughly $7-7.4 billion at a 350-450 billion yuan post-money valuation, closed only in June 2026 ^[7]- meaning DeepSeek operated almost entirely off founder Liang Wenfeng's own hedge fund, High-Flyer, through its first year of global relevance. What hasn't changed as outside capital comes in is who's in charge: Liang controls roughly 84.29% of DeepSeek's equity through direct and indirect shareholding and holds nearly 100% of exercisable voting rights ^[14], through a fund-pool structure in which outside investors - apart from China's state-backed National AI Fund - inject capital into a limited partnership Liang personally manages, rather than investing directly in DeepSeek, under a five-year lock-up ^[15]. That structure was set up for the first round and is being carried forward unchanged into the current one, restarted August 5, 2026, with Liang himself contributing roughly 20 billion yuan - about 40% of the new round's target. Whatever the immediate trigger for the price hike, the larger pattern is unmistakable - a company built on being the cheap alternative is now raising money and prices at the same time, while its founder tightens rather than loosens his grip, the clearest sign yet that efficiency alone stopped being enough to compete at the frontier ^[8].

DeepSeek's language models are very good at code, math, and multi-step reasoning - and they have no way to see the physical world. That gap is the strategic logic behind DeepSeek taking a 140.8 million yuan (~$20.8 million) stake - 933,399 shares, or 2.31% of the strategic placement, locked up for 36 months - in Unitree Robotics' Shanghai STAR Market IPO ^[9]. The investment came bundled with an operating agreement: Unitree will give DeepSeek preference for model-training services and technical solutions, while DeepSeek favors Unitree when it buys robots or explores embodied-AI applications ^[10].

That's not a passive financial bet. Building AI that can act autonomously in physical space requires training data that doesn't exist in text form - a robot's sensorimotor experience, not internet text. Partnering with a humanoid-robot maker gives DeepSeek a direct pipeline into that kind of data, plus Unitree's mechanical-engineering and motion-control expertise, in exchange for access to DeepSeek's reasoning models ^[9]. It also lands DeepSeek inside a company that just became China's first mainland-listed humanoid robot manufacturer, after Unitree priced its Shanghai shares at 150.80 yuan to raise about 6.1 billion yuan at a roughly 61 billion yuan valuation ^[11].

The relationship predates the IPO by more than a year: the founders of both Hangzhou-based companies appeared together at a high-profile meeting with President Xi Jinping in February 2025 ^[12], a detail that reads differently once you see the two companies show up as co-investors eighteen months later. Reddit's coverage of DeepSeek this week barely touched the robotics angle at all - the closest any post came was a single passing reference to 'AI robots' buried inside a post mostly about the price hike and data center buildout - which suggests the Unitree relationship hasn't yet registered with DeepSeek's developer community the way the pricing news has.

Line up the numbers and DeepSeek's past nineteen months look like two distinct financial eras. The company launched R1 in January 2025 pricing tokens as low as $0.07 per million - about 27 times cheaper than the equivalent OpenAI model at the time ^[5]. It stayed almost entirely self-funded through Liang Wenfeng's hedge fund, High-Flyer, until its first external round closed in June 2026: roughly $7-7.4 billion at a 350-450 billion yuan post-money valuation ^[7]. Now, barely two months later, DeepSeek has resumed talks for a second round targeting close to $8 billion (up to 50 billion yuan) at a pre-money valuation near 500 billion yuan (~$74 billion), with Monolith Management - an early backer of rival Moonshot AI that has raised about $289 million for Chinese AI bets - reportedly in talks to join ^[13].

Run the Unitree side of the ledger alongside it and the scale gap is the story. DeepSeek's 140.8 million yuan (~$20.8 million) stake bought 2.31% of Unitree's IPO placement ^[9]- a rounding error next to the $8 billion DeepSeek itself is raising, but enough to secure a seat inside China's first mainland-listed humanoid robot company, now valued at roughly 61 billion yuan after its Shanghai debut ^[11]. Three numbers, read together, tell the real story: a first-round valuation in the 350-450 billion yuan post-money range that multiple outlets converge closer to the $50-60 billion mark ^[16], a second-round target set two months later at roughly 500 billion yuan (~$74 billion) pre-money - figures that aren't directly comparable, since one is measured after the round closed and the other is where the new round's talks are starting, not a simple 'grew from X to Y' jump - a robotics stake worth a fraction of a percent of that raise, and a pricing line that's about to move for the second time in a month. That's a company betting big on compute and infrastructure while making a comparatively small, targeted bet on physical-world data - and charging its existing customers more to help pay for both.