DeepSeek and Unitree’s $14M Deal Signals AI’s Shift to Infrastructure-Level Waste

Woofun AI reports that the convergence of artificial intelligence and robotics has reached a critical inflection point, defined not by capability but by the strategic devaluation of computational and physical resources. The core event anchoring this shift is the strategic placement investment announced on August 6, wherein DeepSeek committed 140.8 million yuan to acquire new shares in Unitree Robotics ahead of its IPO on the STAR Market.

This transaction is not merely a financial injection but a structural alignment between two entities led by Liang Wenfeng and He Yi, respectively, who have independently pursued a singular engineering philosophy: the radical reduction of marginal costs. The collaboration signals the emergence of what Chinese internet users and foreign media alike term the 'Death Zone,' a threshold where technology becomes so inexpensive that its waste is no longer economically significant.

By integrating DeepSeek’s model training services with Unitree’s embodied intelligence applications, the partnership seeks to dismantle the traditional premium pricing of intelligence, transforming it from a scarce resource into a disposable utility. This strategic placement marks the first time these two Hangzhou-based companies have synchronized their cost curves, suggesting that the future of AI lies not in maximizing performance per dollar, but in minimizing the cost of failure.

The concept of the 'Death Zone' was crystallized following the release of DeepSeek V4 Flash on July 31, a model that disrupted existing market benchmarks through its unprecedented price-to-performance ratio. Charts that plot model intelligence levels against usage costs, V4 Flash achieved an intelligence score of 50, placing it firmly among the world’s top-tier models.

However, its average cost per test was merely 3 cents, positioning it at the extreme top-left corner of the coordinate system—a region previously occupied only by models with significantly lower capabilities. This pricing structure created a competitive vacuum where models offering superior intelligence were prohibitively expensive, while cheaper alternatives lacked the necessary sophistication. Bloomberg subsequently highlighted this dynamic in its coverage of the pricing pressure Chinese models exert on American AI companies, explicitly using the term 'Death Zone' to describe the threshold below which competitors cannot sustainably operate.

Since the release of the V2 version in 2024, DeepSeek has consistently pursued a strategy of reducing the cost of each computational 'thought,' decoupling price increases from complexity gains. This approach contrasts sharply with the broader industry trend of emphasizing larger parameters and higher benchmark scores, instead prioritizing the achievement of equivalent intelligence at a fraction of the cost.

Parallel to DeepSeek’s computational cost reduction, He Yi has spent over a decade in Hangzhou applying similar engineering principles to robotics, a discipline historically constrained by high hardware expenses. During his graduate studies, He Yi developed XDog, a project initiated with limited funding that forced a reliance on low-cost components rather than industry-standard solutions. Unable to afford expensive hydraulic systems, he utilized low-cost motors; unable to purchase ready-made drive boards, he designed and programmed his own control systems.

This early experience instilled a foundational mindset: prioritize cheaper alternatives before considering established industry practices. He Yi recalled that in 2010, the total mechanical parts cost for his bipedal robot was only 200 yuan, a figure that underscores the extreme frugality required at the inception of Unitree’s technology stack. Years later, when questioned about the limits of cost-cutting, He Yi asserted, 'Don’t compare cost-cutting with us; we can cut costs even further.'

This statement reflects a deep-seated commitment to engineering efficiency over market positioning. For a long time, despite both being based in Hangzhou, DeepSeek and Unitree operated on separate trajectories—one lowering the cost of thinking, the other lowering the cost of action. The August 6 investment announcement marked the convergence of these paths, aligning their respective efforts to make expensive technologies affordable through structural innovation rather than temporary subsidies.

The strategic investment details reveal a clear division of labor and mutual dependency between the two entities. DeepSeek’s 140.8 million yuan investment in Unitree’s new shares is accompanied by specific cooperation terms that prioritize cross-integration of their technologies. Unitree has agreed to give priority to DeepSeek for model training services and technical solutions, effectively making DeepSeek the primary intellectual engine for its robots.

Conversely, DeepSeek will prioritize Unitree for the development of robots and embodied intelligence applications, providing a physical platform for its models to interact with the real world. This arrangement is not driven by the sheer size of the investment, which is modest in the context of today’s AI industry, but by the alignment of their long-term cost curves. The partnership raises a fundamental question about the nature of AI: if both the cost of machine thinking and acting continue to drop, what does intelligence become?

The answer suggests a shift from a premium service to a commoditized utility. By combining DeepSeek’s computational efficiency with Unitree’s hardware frugality, the collaboration aims to lower the barrier to entry for embodied intelligence, allowing for more widespread adoption and experimentation. This strategic alignment indicates that the value of AI is no longer measured solely by its accuracy or speed, but by its accessibility and the ability to scale without proportional cost increases.

DeepSeek’s structural efficiency is rooted in its technical architecture, which prioritizes parameter activation over total parameter count. The DeepSeek-V2 model, released in 2024, featured a total of 236 billion parameters, but only 21 billion were activated per token processed. This sparse activation mechanism resulted in a 42.5% drop in training costs compared to the previous generation, a 93.3% decrease in KV Cache usage, and a 5.76 times increase in maximum generation throughput.

These metrics demonstrate that cost reduction is not achieved through brute force scaling but through architectural optimization. With the subsequent release of V3, this approach was further refined. The model required 2.788 million H800 GPU hours for training, a significant computational investment that was offset by optimizations such as hybrid experts, low-precision training, and improved communication protocols.

These technical choices reflect a deliberate strategy to maximize efficiency within the constraints of available hardware. By focusing on the cost of each token rather than the total size of the model, DeepSeek has created a pricing structure that is sustainable in the long term. This technical foundation enables the company to offer competitive pricing without relying on unsustainable subsidies, thereby creating a stable market position that competitors find difficult to replicate.

Woofun AI data shows that the pricing principles adopted by DeepSeek and Unitree reflect a shared skepticism toward short-term promotional strategies. In May of this year, DeepSeek removed the 75% promotional discount from its V4-Pro model and made it a permanent price, signaling a shift from temporary incentives to structural cost advantages. Liang Wenfeng has explained that their pricing principle is based on actual costs, avoiding long-term losses and refraining from seeking exorbitant profits.

This approach ensures that prices are sustainable and reflective of the true value of the service. Similarly, He Yi holds a firm stance on profitability in robotics. When interviewed by WanDian in 2025, he was asked whether it was still profitable to sell the H1 robot at 90,000 USD, given that the more affordable G1 model started at just 99,000 RMB. He responded, 'Commercial activities must generate reasonable profits. Cost has always been our KPI for everything—we’re here to make money.'

This statement underscores the importance of cost management as a core business metric. He rarely mentions the common notion that production scale naturally lowers prices, instead emphasizing that the lower limit of robot prices is determined by specific engineering choices, such as motor types, reducer designs, and component integration. This industrial approach to cost reduction involves cutting every penny in half, from saving centimeters on circuit boards to reducing the number of components per joint.

Unitree’s sales data illustrates the impact of this cost-reduction strategy on market adoption. In 2023, the company launched its first full-sized humanoid robot, the H1, selling only 5 units at an average price of 593,400 RMB. The following year, the smaller G1 model entered the market, with sales rising to 410 units and the average price dropping to 260,700 RMB. By the first nine months of 2025, Unitree had sold 3,551 humanoid robots, with the average price falling further to 167,600 RMB. This trajectory demonstrates a clear correlation between price reduction and volume increase. As prices dropped, new groups of buyers emerged, including universities, development teams, and enterprises that previously could not afford such technology. Devices costing hundreds of thousands of yuan required extensive budget approvals and justification, while models costing over 100,000 RMB became accessible to a broader range of institutions.

This shift from specialized budgets to regular purchases indicates that Unitree’s robots are transitioning from experimental prototypes to practical tools. The reduction in price is not merely a marketing tactic but a result of continuous engineering improvements, such as optimizing motor types and reducer designs. These incremental changes, when combined, directly affect the final price, enabling the company to capture a larger share of the market.

The embodied intelligence training cycle represents the most expensive aspect of robotics development, involving both real-world interaction and computational processing. Unitree is currently working on world models and VLA (Vision-Language-Action) systems, which require robots to interact with the real world to generate data. Half of this cycle takes place in the physical realm, where robots reach out, walk, and pick up objects, with humans or other systems providing guidance.

This process creates sequences of successes and failures, which are then recorded as trajectory data. The other half of the cycle occurs in the computing domain, where data is cleaned, labeled, and fed into models to train new strategies. After inference and validation, the updated models are sent back to the robots for further testing. This iterative process is expensive on both ends: expensive hardware limits the number of robots that can be used for training, while expensive computing resources constrain the amount of data that can be processed.

Consequently, robot experiments have traditionally been precious, with researchers carefully planning data collection and validating trajectory data to maximize the value of each trial. The high cost of resources tends to focus research on problems with the highest chance of success, ignoring the tedious, everyday issues that robots will face in real-world applications.

The application of Jevons’ Paradox provides a historical context for understanding the impact of cheap technology on consumption patterns. In 1865, British economist William Stanley Jevons observed that as steam engine efficiency improved and less coal was needed to complete a task, Britain’s coal consumption did not decrease but increased. This phenomenon, known as Jevons’ Paradox, occurred because cheaper steam power enabled new applications that previously were too costly, leading to a net increase in resource use.

Similarly, the internet and electricity have become so inexpensive that their costs are no longer factored into daily decisions. In 1911, Alfred North Whitehead wrote in A Mathematical Introduction that 'The progress of civilization lies in continuously expanding those important operations that we can perform without thinking.' This principle applies to AI as well: as intelligence becomes cheaper, it will be used for less important tasks, leading to a broader integration into daily life.

Currently, AI users still calculate token costs and choose suppliers based on minor price differences, indicating that the technology has not yet reached the stage of ubiquitous, unthinking use. The collaboration between DeepSeek and Unitree aims to accelerate this transition by reducing the cost of both thinking and acting, thereby enabling a wider range of applications and a higher tolerance for error.

The conclusion of this analysis is that the true value of the 140.8 million yuan investment lies in its potential to redefine the cost of mistakes in AI and robotics. Today, every piece of real-world data must be carefully collected, and every round of training must be meticulously selected, because the cost of failure is high.

However, as token costs and humanoid robot costs continue to drop, intelligence will transform from an expensive capability into infrastructure.

This shift will allow machines to make more mistakes, accumulate more failures, and learn from a broader range of experiences. The future of AI will not be marked by a single breakthrough but by the gradual realization that robots can spend afternoons practicing box-picking in warehouse corners, failing hundreds of times without drawing special attention. When machines can think, act, and err without causing economic concern, intelligence will have truly become infrastructure. This transition requires not just technological advancement but a cultural shift in how we value computational and physical resources. The partnership between DeepSeek and Unitree is a critical step in this direction, aligning the cost curves of thinking and acting to create a future where AI is cheap enough to be wasted.