Source: Shenlan Finance

The trajectory of artificial general intelligence (AGI) in the physical world is being reshaped by a “Tesla-connected” engineering team.

On May 18th, Matrix Superintelligence (Matrix), an embodied AI company, unveiled its full-sized humanoid robot, MATRIX-3, at its “2026 AI DAY” event in Zhangjiang, Shanghai. Alongside the global launch, Matrix announced pricing, production capacity plans, and a roadmap for commercialization. Led by Zhang Haixing, former founding leader of Tesla’s China Design Research Center, the team is striving to transition the “Chinese version of Optimus Prime” from laboratory prototypes to industrial-grade products.

However, what truly stands out—even more than the product launch itself—are the key signals Zhang Haixing shared during the media Q&A session: thousands of units to be shipped this year, a tenfold increase in production capacity next year, overseas orders already secured, and future subscription revenue projected to exceed 50%. These figures and insights paint a vivid picture of the commercial landscape for an embodied AI company.

1. Priced Starting at 580,000 Yuan, with 1,000 Units This Year and 10,000 Next Year

Standing at 170cm and weighing 65kg, MATRIX-3 is designed by Matrix Superintelligence as the “optimal solution for adapting to the human physical world.” Unlike many “laboratory prototypes” prevalent in the industry, MATRIX-3 features a hybrid architecture of rotational and linear joints, boasting over 70 degrees of freedom (DoF) across its body (including both hands)—27 DoF per hand and 33 DoF for the rest of the body.

Technologically, MATRIX-3 is powered by Matrix’s self-developed WAVE physical foundation model, achieving “zero-shot generalization” and “closed-loop error learning” through real-world data integration. Matrix claims its data collection and generation technology can accelerate data accumulation tenfold while reducing costs by 90%.

Pricing:

• MATRIX-3 All-Purpose Flagship Model: Starting at 580,000 Yuan
• MATRIX-3 PRO (with 27-DoF dexterous hand MATRIX HAND): Starting at 680,000 Yuan

Both models include a one-year basic service package.

Zhang Haixing confirmed on-site that Matrix Superintelligence’s Phase 1 production capacity plan targets 10,000 units annually, with full-link autonomous production achieved through its MFH Superintelligence Factory in Zhangjiang. When asked about this year’s shipment volume, he provided a more specific figure: “Our target this year is roughly a thousand units. We’re confident we can sell that volume—the third generation will be the mainstay.”

More critically, he hinted at next year’s guidance: “Our production volume next year won’t be in the low thousands. It will step up to a new level, adding another zero to the quantity.” In other words, Matrix Superintelligence has set its 2027 production capacity target at 10,000 units—a figure supported by its Phase 1 factory planning.

In terms of engineering cost reduction, the team has dramatically reduced the total number of screws in the robot from over a thousand in the initial prototype, a move seen as a “critical leap for Chinese humanoid robots transitioning from laboratory prototypes to industrial-grade products.”

2. Focusing on Real Productivity Scenarios, Not ‘Data Collection’ Orders

When questioned about the industry’s widespread reliance on “data collection orders,” Zhang Haixing took a clear stance:

“Data collection factories are just one channel, but not the core or primary one. This approach has become overused—every company seems to be competing in data collection. While it can generate some orders, it’s not sustainable because it can’t translate into a continuous market demand for productivity. Data collection factories will eventually run out of scenarios.”

Matrix Superintelligence believes that relying on humans to collect one hour of data is fundamentally flawed—human effort should be reserved for high-precision calibration, while the majority of data should be generated by world models.

In terms of commercial outreach, Zhang Haixing revealed that the team has repurposed a complete autonomous vehicle sales system: “Our sales team is already experienced in selling high-end vehicles priced above 500,000 Yuan. These clients share strong commonalities with our previous target audience.” Currently, MATRIX has been deployed in scenarios such as Shanghai South Railway Station, the Global Developer Pioneer Conference, and Haidemann’s Lingang production line.

Overseas business has also expanded. “We’ve attracted many overseas users through official channels on foreign media forums and X (formerly Twitter), naturally facilitating our global expansion.” The focus for the second half of the year remains on the Chinese and Asia-Pacific markets, with subsequent expansion to the Middle East and Europe. When asked about domestic peers exporting to North America via third-party platforms, Zhang Haixing stated, “That’s a detour. We’ll exercise restraint—if we can’t achieve stable domestic delivery, we won’t easily enter new markets.”

3. Subscription Revenue to Exceed 50% in the Future

This segment contained the most insightful revelations during the interview, highlighting the most significant business model difference between Matrix Superintelligence and its peers.

“Hardware is currently the majority revenue stream for most embodied AI companies, but in the long run, it won’t exceed 50%. I believe 50-60% should come from ARR (Annual Recurring Revenue), i.e., subscriptions,” Zhang Haixing explained. He argued that automated labor is essentially “compute-powered,” with tokens representing labor units—costs that will be continuously charged to users through operators or manufacturers.

Beyond inference and compute fees, he envisioned a “robot version of the App Store”: “Imagine charging 96 Yuan annually for a Michelin-star-level tomato scrambled egg recipe. Is that reasonable? Absolutely. Who develops these Skills? We’ll collaborate with China’s best egg-fried-rice chefs, capture their data for training, and distribute it through our store with a roughly 30-40% revenue share.”

If this model succeeds, Matrix Superintelligence would evolve beyond a hardware company into a platform company attempting to replicate the iOS ecosystem.

4. Assessment of Figure and Timeline for Household Market Entry

Regarding Figure 03’s recent continuous livestreaming on assembly lines, Zhang Haixing offered an industry peer’s perspective:

“That was a milestone. It proved that Helix 02’s embodied model can operate continuously for 24 hours. However, the scenario was relatively simple, and the model would require retraining for other environments—it’s not highly generalizable. Figure has only been in this field for slightly over six months, so this represents deep optimization for this specific scenario.”

As for when household robots will enter the consumer market, Zhang Haixing provided a specific timeframe: “Aggressively, three years; optimistically, a median of five to eight years.” He revealed that Matrix Superintelligence plans to launch a dedicated, highly affordable household robot within the next two to three years.

Regarding the industry’s ultimate landscape, he expressed no anxiety about the current “shipment volume competition”: “There’s zero pressure. This industry hasn’t even truly opened up yet. It’s more like the automotive industry over a century ago, when hundreds of companies emerged in the U.S. and Europe before consolidating into Ford’s standardized model—the embodied AI industry will undergo a similar process.”

5. Future Landscape: China Leads in Hardware, Lags North America by 6-12 Months in ‘Brain’ Development

Zhang Haixing offered a candid assessment of the global competitive landscape:

“In terms of pure embodied AI models, North American companies are ahead. China’s top teams might lag by a quarter, with an average gap of six months to a year—take Physical Intelligence and OpenAI as examples. However, in physical robot bodies and hardware manufacturing, China is undoubtedly the global leader. This isn’t just catching up; it’s pulling ahead.”

China’s current weakness, in his view, lies in the return of top embodied AI scientists: “Some stay abroad and don’t come back, starting businesses or working in the U.S. Some do return, but not many.”

Nevertheless, he cited the Huawei vs. Tesla autonomous driving competition as grounds for optimism: “The market perception isn’t that stark. Huawei has already achieved outstanding results, approaching Tesla’s end-to-end black-box system—so it’s not a case of ‘America succeeds, China loses.’ That’s not how it works.”

Shenlan Finance New Media Cluster originated from the Shenlan Finance Journalist Community 15 years ago and is now a renowned domestic financial new media platform. Its accounts focus on China’s most valuable companies, cutting-edge industry developments, and emerging regional economies, providing valuable content for investors, corporate executives, and the middle class. Welcome to follow us.