The Great Earnings Divide: How Moore Threads and Unitree Told Opposite Profit Stories on the Same Day

On the evening of October 28, 2026, two of China's most closely watched AI hardware companies filed quarterly reports within hours of each other on the Shanghai Stock Exchange's STAR Market. The timing was coincidental. The contrast was not.

Moore Threads, the domestic GPU maker routinely described as "China's Nvidia," reported the first quarterly profit in its six-year history. Revenue surged 1,773.9% year-on-year to 1.55 billion yuan. Net profit attributable to shareholders reached 49.26 million yuan — arriving a full year ahead of the company's own 2027 profitability target.

Unitree Robotics, the humanoid robot manufacturer that listed in August as the A-share market's first embodied AI stock, delivered its first complete quarterly report as a public company. Revenue grew 69.93% to 665 million yuan. But net profit fell 49.55% from the previous quarter, and non-GAAP profit for the first nine months declined 18.78% year-on-year as the company poured money into research at a pace that would alarm most investors — if the stock had not risen more than 50% in October alone.

One question unites these two companies: can China's AI hardware champions convert state-fueled urgency and capital market euphoria into durable, self-sustaining businesses? Their answers, delivered on the same day, point in sharply different directions. Together, they sketch the contours of China's AI hardware moment more honestly than either could alone.

Moore Threads: From -70% Gross Margin to Black Ink

The numbers Moore Threads reported on October 28 would have seemed fantastical four years ago. In 2022, the company's gross margin was negative 70% — every GPU it sold generated less than a third of what it cost to produce. Revenue that year was 46 million yuan, against more than a billion yuan in annual R&D spending.

The Q3 2026 report tells a different story entirely.

The nine-month cumulative picture reinforces the inflection. Revenue for January through September reached 3.29 billion yuan, up 319.1% year-on-year. Trailing twelve-month revenue hit 4.01 billion yuan. Gross margin, which stood at 65.5% in 2025, has settled into the mid-fifties — still among the highest in the domestic GPU industry, and a figure that signals genuine pricing power rather than subsidy-driven discounting.

The quality of this quarter's profit matters as much as its existence. In Q1 2026, Moore Threads recorded its first attributable profit of 29.36 million yuan, but non-GAAP profit was still negative 54 million yuan — the gap filled by government subsidies and fair-value adjustments. Q3 is the first quarter where the core business itself is contributing to the black, with non-GAAP profit of 20.07 million yuan. The company has, in its own words and its auditors' ledger, crossed from "story" to "answer."

What Is Fueling the GPU Engine

The proximate driver of Moore Threads' surge is straightforward: scale delivery of its MTT S5000 full-function GPU, which serves both training and inference workloads, into China's insatiable AI computing market. Cloud products account for 97.5% of revenue, and the customer list spans internet companies, cloud providers, model-as-a-service operators, telecommunications carriers, and government computing centers.

The single largest validation came in September, when JD Cloud announced plans to purchase a 100,000-GPU cluster from Moore Threads, with deployment scheduled through 2027. For a domestic GPU industry that has spent years fighting the perception of being a "Plan B" for when NVIDIA supply is constrained, the JD deal carried symbolic weight approaching its commercial weight. A single order of that magnitude does not just fill a revenue line — it signals that a hyperscaler has concluded the product is viable at production scale.

Equally important is the software ecosystem play. Moore Threads has committed to Day-0 support for major open-source models — DeepSeek, GLM, MiniMax, Kimi, and Qwen — and earned official adaptation from core inference frameworks including vLLM, SGLang, Slime, RLinf, LMCache, and llm-d. In GPU markets, hardware without ecosystem support is silicon sculpture. Day-0 model compatibility means a customer deploying DeepSeek or Qwen on Moore Threads hardware does not wait weeks for porting — the stack is ready when the model ships.

The company's inventory position offers a forward-looking corroboration. At the end of H1 2026, inventory stood at 3.55 billion yuan, up 1.7 times from the start of the year — raw materials and work-in-progress stockpiled against contracted deliveries. For a chip company, inventory buildup of that scale is either a warning or a promise. Given the delivery pipeline and the JD commitment, the market is reading it as the latter.

Unitree: Growing Fast, Profiting Less, Betting Bigger

If Moore Threads' report answered a question — can domestic GPUs make money? — Unitree's report asked one: what is the humanoid robot business actually worth right now?

The headline numbers appear strong. Revenue grew 69.93% year-on-year to 665 million yuan. For the first nine months, revenue reached 1.82 billion yuan, up 55.71%. The company holds 7.35 billion yuan in cash following its August IPO. Total assets expanded 216.5% from year-end. Inventory grew 177.2% to 1.02 billion yuan, which the company attributes to advance stocking for anticipated demand.

But the profit picture complicates the narrative. Q3 net profit of 113 million yuan fell 13.37% year-on-year and 49.55% from Q2's 224 million. Nine-month non-GAAP profit of 349 million yuan declined 18.78% year-on-year. Operating cash flow dropped 54.69% to 194 million yuan as procurement, payroll, and R&D expenses scaled with revenue.

None of this is accidental. Unitree is deliberately converting margin into capability. R&D spending grew 249.58% year-on-year to 19.1% of revenue — a ratio that dwarfs most hardware companies and approaches the spending intensity of software firms. The company is hiring AI researchers at a pace that is reshaping its internal culture, which was historically dominated by mechanical and controls engineers.

Wang Xingxing's All-In Bet on the Robot Brain

The strategic logic behind the spending surge is embodied — literally — in founder and CEO Wang Xingxing's recent public statements. "My biggest mistake over the past decade and a half was not believing in AI enough," he said at the Shanghai Industry Fair in October. "Almost everyone in the world doesn't take AI seriously enough or believe in it enough, including myself."

That belief is now directing Unitree's capital allocation. In Q3, the company released UnifoLM-WLA-1.0, a foundation model for humanoid robots built on a World-Language-Action architecture, with its embodied reasoning base constructed on Alibaba's Qwen3-VL-4B. It follows the January open-sourcing of UnifoLM-VLA. The company has begun releasing model weights and data, positioning itself not just as a robot manufacturer but as a platform for embodied intelligence.

The hardware pipeline is equally aggressive. G1+, an upgraded version of what Wang calls "the world's best-selling humanoid robot" with cumulative production of 10,000 to 20,000 units, launched in September. The H2-X, designed for hazardous environments including emergency rescue, debuted in October. Japan Airlines continues its pilot program using G1 robots at Haneda Airport. And at the company's partner ecosystem conference, Unitree launched the "YuChuang Plan," committing to incubate more than 200 ecosystem companies over five years across industrial manufacturing, education, entertainment, retail, and emergency response.

Wang has been characteristically direct about the endgame. He defines the industry's "ChatGPT moment" as the point when a robot can complete 80% of major tasks in 80% of unfamiliar environments using only voice or text instructions — and he believes it could arrive as early as 2027. "Physical AI is showing signs of being cracked," he said. "Fast: one to two years. Slow: two to three years."

*Unitree's bet: hardware margins today are less important than owning the embodied AI platform layer tomorrow.*

Whether that timeline is prescient or promotional is the central uncertainty of Unitree's investment thesis. The company is spending as if the moment is certain, while its profit statement acknowledges it has not yet arrived.

Head-to-Head: Two Companies, Two Phases, One Question

Placed side by side, the earnings reports reveal a structural divergence that transcends any single quarter.

The valuation column deserves scrutiny. Moore Threads traded at roughly 180 times sales in August — a multiple that made even NVIDIA's 2024 peak of around 35 look conservative. That has since compressed to approximately 67 times trailing revenue as earnings materialized. Unitree trades at roughly 117 times annualized revenue, a multiple that prices in years of flawless execution on a technology — general-purpose embodied AI — that does not yet exist at commercial scale.

Both companies are, in essence, pre-paid futures contracts on China's AI hardware ambitions. The difference is that Moore Threads has begun delivering the underlying asset, while Unitree is still building the delivery mechanism.

The Memory Wall Both Companies Are Running Into

Behind both earnings stories looms a third data point that received less attention but may prove more consequential. Chinese cloud providers — including Alibaba, Tencent, and ByteDance — are reportedly negotiating ten-year DRAM supply contracts with Samsung and SK Hynix, extending to 2035. That is two to three times the duration of typical long-term supply agreements. Samsung has reportedly signed five major datacenter clients; SK Hynix, roughly ten. An estimated 70% to 80% of next year's memory production is already committed under long-term contracts.

The signal is unambiguous: China's largest technology companies expect the AI infrastructure buildout to continue for a decade, and they are locking up the one input — memory — that cannot be quickly replicated domestically. SK Hynix's nine-month profit surge of 527% confirms who is capturing the most reliable profits in the AI hardware stack: not the GPU designers or the robot makers, but the memory fabs that supply them both.

This backdrop adds nuance to the Moore Threads story. Domestic GPU demand is being driven not just by export controls on NVIDIA hardware but by the same capex wave that is pushing Alibaba, Tencent, and ByteDance to hoard memory. The compute buildout is real, it is massive, and it is not abating — even as Wall Street grows nervous about AI infrastructure financing, with semiconductor stocks selling off on credit-risk concerns.

*The memory grab behind the earnings: China's cloud giants are locking up DRAM supply through 2035, signaling a decade-long compute buildout.*

What the Divide Reveals

The temptation is to declare a winner. Resist it.

Moore Threads' profit is real and hard-earned, but the company still faces the deepest competitive moat in technology: NVIDIA's CUDA ecosystem, which has absorbed two decades of developer tooling, library optimization, and institutional knowledge. Day-0 support for vLLM and SGLang is necessary but not sufficient. The company's path from ¥4 billion in revenue to the scale where it can fund next-generation architecture development from operating cash flow remains long and uncertain.

Unitree's profit compression is rational but not costless. The company is spending ahead of demonstrated demand for general-purpose humanoid robots, in a market where its own CEO admits the technology is one to three years from a breakthrough that may or may not arrive on schedule. Its 60.4% gross margin gives it room to maneuver, but margins compress when volume scales — as the humanoid robot unit price decline from 593,400 yuan in 2023 to 166,400 yuan in 2025 already illustrates.

What the same-day earnings divide actually reveals is that China's AI hardware stack is maturing unevenly, layer by layer. Compute silicon is first to reach financial self-sufficiency, because the demand is immediate, the substitution market is enforced by export policy, and the product cycle is measurable in quarters. Embodied AI is years behind, because the demand is speculative, the substitution market does not yet exist, and the product cycle — as Wang Xingxing himself frames it — depends on a model capability breakthrough that no amount of capital expenditure can guarantee.

The market is pricing both companies as if the discipline phase will never arrive. At some point, it will. The question these twin earnings reports pose — and cannot yet answer — is whether Moore Threads' newly discovered profitability will compound long enough to fund genuine architectural independence, and whether Unitree's newly discovered spending will produce a robot brain before its newly acquired cash runs out.

Both companies are betting they have more time than the skeptics think. On October 28, only one of them had the numbers to back it up.

Social Voices

Xueqiu (investor forum):

摩尔线程这个盈利质量还是要拆开看，扣非才2000万，大头还是靠规模冲上来的。毛利率从65%降到54%说明S5000放量是以价换量。国产替代的故事是真的，但估值已经把这个故事讲到2028年了。

>

*"You have to break down the quality of Moore Threads' profit. Non-GAAP is only 20 million — the bulk came from scale. Gross margin dropping from 65% to 54% shows the S5000 ramp is trading price for volume. The domestic substitution story is real, but the valuation has already priced that story through 2028."*

Zhihu (tech discussion):

宇树Q3的研发费用翻了两倍多，这不是在优化现有产品，这是在赌一个还没被验证的技术路线。要么三年后回头看是天才，要么是教科书级的过度扩张。没有中间地带。

>

*"Unitree's Q3 R&D more than tripled. This isn't optimizing existing products — it's betting on an unvalidated technical path. Either it's genius in hindsight three years from now, or it's a textbook case of overexpansion. There is no middle ground."*

Weibo (social media):

王兴兴说物理AI一到两年内会被击穿，我信他一半。G1确实卖得好，但"会跳舞的机器人"和"会打螺丝的机器人"之间隔着一个模型革命。宇创计划孵化200家企业倒是务实，先把生态搞起来再说。

>

*"Wang Xingxing says physical AI will be cracked in one to two years. I believe half of it. The G1 sells well, but between a 'robot that dances' and a 'robot that tightens screws' there is a model revolution. The YuChuang Plan incubating 200 companies is more pragmatic — build the ecosystem first."*

X (global investor):

Everyone is focused on the GPU and robot earnings but the real story is the DRAM contracts. Alibaba, Tencent and ByteDance signing 10-year deals through 2035 tells you China Inc is hoarding compute inputs before export controls get worse. The memory suppliers are the picks and shovels of this gold rush.

Hacker News (technical community):

Day-0 support for DeepSeek, Qwen, SGLang and vLLM is genuinely impressive for a GPU vendor that was founded in 2020. But CUDA's moat isn't model compatibility — it's the 20 years of accumulated tooling, profiling, and developer muscle memory. Moore Threads doesn't need to beat NVIDIA on specs. It needs to beat the ecosystem lock-in. That's a much harder problem.

WeChat (retail investor group):

两家公司市值加起来5500亿，超过了很多国家整个芯片行业的市值。我们在给2026年的收入付2035年的价格。亢奋是真的，风险也是真的。

>

*"The two companies' combined market cap exceeds the entire chip industry of many countries. We're paying 2035 prices for 2026 revenue. The euphoria is real. So is the risk."*