On May 18, Matrix Ultra Intelligence officially launched its third-generation flagship humanoid robot, MATRIX-3, in Shanghai.
Zhang Haixing, founder and CEO of Matrix Ultra Intelligence, told our reporter in an exclusive interview that MATRIX-3 will begin mass production and delivery in Q3 this year, with a full-year delivery target of around 1,000 units. Some early deliveries could happen as soon as the end of June.
According to the release, MATRIX-3 boasts 33 full-body degrees of freedom, with its dexterous hand offering 27 degrees of freedom—slightly exceeding Tesla’s Optimus at 22. The joint scheme uses a series-parallel linear actuator architecture, with three linear actuators in the thigh and two in the calf per leg. The body is covered in 3D knitted fabric with embedded distributed tactile sensing networks, paired with the self-developed embodied large model Matrix Wave.
The disclosed information also revealed that Matrix Ultra Intelligence is already capable of delivering 5,000 comprehensive test units within the year, and has begun deep stress testing in real industrial scenarios. With ongoing upgrades to its production line, the company expects to reach a 100,000-unit production capacity by the first half of 2027, leveraging scale to accelerate the adoption of universal labor.
The pricing announced this time is also above the norm. The fully flagship MATRIX-3 humanoid robot starts at 580,000 yuan, while the MATRIX-3 PRO—equipped with the world’s first 27-DOF dexterous hand, MATRIX HAND—starts at 680,000 yuan, both including a one-year basic service package.
Behind all this is a market that’s rapidly splitting. Right now, China’s humanoid robot industry is at that sweet spot between “mass production year” and “delivery year.”
According to IDC, global humanoid robot shipments in 2025 will hit about 18,000 units—a 508% jump year-over-year. IDC further predicts that by 2026, China’s market will be close to $1.3 billion, more than doubling from the previous year.
Goldman Sachs analysts noted in a report that 2026 could be the key year for scaling and resetting expectations. Major players are setting delivery targets for 2026–2027 that are several times higher than 2025—jumping from hundreds or thousands of units to tens of thousands.
In this race, Matrix Ultra Intelligence is taking a different path. Founded in Shanghai in April 2024, the company is one of the few in China to truly master the linear actuator technology route.
But Zhang Haixing has a label he can’t easily shake: he was personally invited by Elon Musk to lead the founding of Tesla’s China Design Research Center, and he was deeply involved in the early R&D and design of the Optimus humanoid robot.
This background brings attention to Matrix Ultra Intelligence, but also skepticism. Some say it’s just “following Tesla,” others think “a designer’s background” means a lack of technical depth. Zhang’s response? Pretty blunt: “We made old Ma’s bragging a reality.”
Before the launch, our reporter had an in-depth conversation with Zhang. In this talk, he addressed all the doubts about the tech route, the high price, and the “designer building robots” tag. He also walked us through his views on data collection, the GPT Moment for embodied intelligence, and the big industry shakeout.
Below is the full transcript of the interview, lightly edited for clarity:
Linear Actuators: The Harder Road
Journalist: MATRIX-3 uses linear actuators—not many companies in China go that route; most choose rotary ones. Why this path?
Zhang Haixing: The core difference is what each is good for. Rotary joints have mature algorithms and are easy to debug. A team can grab an open-source design and put together a demo—running, flipping, backflips—no big deal. But they’re more for show. When it comes to actually carrying heavy loads and working for long periods, the motors and reducers wear out fast under impact, and the precision for fine manipulation isn’t enough.
Linear actuators are the opposite—hard to develop, but real productivity tools. Tesla uses linear actuators, and we do too. The reason is simple: they can actually do real work.
Journalist: So why are different companies going different ways?
Zhang Haixing: Two main factors push most away. First, the supply chain is weak. High-performance lead screws, planetary rollers—these core technologies were controlled by European and Japanese companies. Some Chinese suppliers are making them, but it’s still early stages—cost isn’t much better, and performance might be 20–30% less.
For example, a European lead screw sells for 15,000 yuan; a Chinese one goes for 12,000. With similar cost but worse performance, you can’t get an edge. Right now, only two or three companies in China can make these core parts, and costs are stuck where the supply chain is.
Second, the algorithm barrier is huge. Rotary joint motion control algorithms came from quadruped robot dogs—lots of open-source projects from top universities to reference. You can build and tune hardware easily. But linear actuator motion control? Almost no open-source resources.
Our team actually ran the whole-body motion control algorithm for this series-parallel architecture from scratch, and we’ve patented it. The algorithm works with any body type—if we switch to pure rotary or hybrid joints in the future, it migrates seamlessly.
Journalist: That brings up a real question. Linear actuators handle loads well, but they cost a lot. Roughly what’s the BOM for a single MATRIX-3? Is there business sense in sticking with a high-cost route?
Zhang Haixing: The standard MATRIX-3 starts at 580,000 yuan. High costs mainly come from the supply chain—same reason as before; domestic industrialization is too early, so cost is comparable to Tesla’s.
The key to lowering cost is scale—not by cutting performance or reducing degrees of freedom to fake it. When the whole industry hits 100,000 or even a million units, the cost of these core parts gets flattened a lot.
Our internal estimate is that once the domestic supply chain catches up, even with no other design changes, the terminal price can drop to 100–200 thousand yuan. And for the entire humanoid robot industry, the ultimate target is 30–50 thousand yuan—like buying a home appliance.
Journalist: The humanoid robot scene in China is buzzing—companies are running marathons, sports events, even showing up on the Spring Festival Gala. You guys didn’t take that route. Was it a deliberate choice, or just not your thing?
Zhang Haixing: There’s a clear path split among domestic companies. And that split comes down to the underlying hardware architecture.
Most products that focus on showy performances use rotary joints as the main. The advantage of rotary joints is mature algorithms and low barriers. Teams can grab open-source info and put together hardware that walks, runs, jumps—it’s great for demos, for showing off flips or high-dynamic moves. That’s why everyone goes that way, competing on flashy stunts harder and harder.
But in our view, that architecture isn’t a real productivity tool. Linear actuators are hard to develop, but they shine in high-load carrying and repeatable precision at the end effector.
Let me give you a concrete idea: MATRIX-3 can easily push or pull 200 kg loads. More critical is end-point precision. Rotary joints tend to break under high load, and the longer the joint chain, the more the end-point error gets amplified. Linear actuators don’t jitter, so you can do fine operations—using tools, pushing buttons, assembly work.
Journalist: The outside perception is that you really don’t do much show performance. But that also brings another question: Is Matrix Ultra Intelligence just telling a story, just in a different way?
Zhang Haixing: We’re not promising a price or volume that will never happen. We can say Q3 mass production because we’re already producing—first customer deliveries are expected as early as the end of June.
Journalist: Optimus has postponed its mass production plans several times. How can you guarantee delivery in Q3 this year—not just an empty promise?
Zhang Haixing: We can deliver because of China’s incredibly strong manufacturing supply chain. That’s why our iteration speed is crazy fast.
We’re in China, close to suppliers. If we find a problem during testing, we might drive to the factory that afternoon and spend two or three days with the supplier, fixing it on the spot. We’ve already achieved full vertical integration from component design to final assembly. Tesla can’t do that.
Based on the current state, Optimus can’t mass produce this year. Their completion level isn’t as high as ours. We really made old Ma’s bragging come true. Globally, we’re the top players in linear actuator technology.
It’s not us telling a story—we’re using the efficiency advantage of China’s supply chain to reach the industry’s expected milestones ahead of schedule. We’re dead sure about delivering around 1,000 units this year.
Journalist: Matrix Ultra Intelligence uses the slogan “Building China’s Optimus Humanoid Robot.” Add your background as a former Tesla China executive, and people see you as a pure Tesla follower. How do you feel about that? What’s your real tech moat?
Zhang Haixing: We don’t mind that label, honestly.
In a new industry’s early days, following the leader is just business common sense. Musk helped the whole industry educate the global market and explore scenarios. Once he proves a path works, it’s like he’s laying tracks for us—we just hook up the train cars. China can follow much faster.
But in terms of specific engineering, we’re not exactly the same. Take the hip joint: Tesla’s Optimus has two protruding round rotary motors, which forces its legs to spread far apart, leaving a big gap when walking. In MATRIX-3, we just removed those two motors and replaced them with a lateral linear lead screw. The legs can close together like a normal person’s—a more advanced architecture with better load capacity.
Even more core is the algorithm—the whole-body motion control for linear actuators under series-parallel architecture. Almost no open-source in China. We’re one of the few teams that actually made it work. Still, it’s hard to truly shake off the Tesla label.
Journalist: How big is your team now? What’s the composition?
Zhang Haixing: Over 100 people—not too many yet. Roughly 60% are in algorithms, the rest are teams for mechanical, electrical, and processing/production.
Journalist: What’s the target scenario for MATRIX-3? How do you define it internally?
Zhang Haixing: Internally, we position the third-gen’s core scenario as service. Shopping malls, restaurants, hotels, office front desks, government service halls, exhibition guides—any B2B scenario where you need to interact with people.
Journalist: What about factory scenarios? Will you consider going into factories? Many companies see that as the biggest commercial path.
Zhang Haixing: Industrial manufacturing is a huge market, but in my view, for dark factories that don’t need human-robot collaboration, traditional robotic arms, AGVs, or wheeled-legged robots are already efficient enough. Industrial scenarios don’t need interaction—humanoid is overkill. The real value of humanoid is in complex environments with changing tasks and where you have to deal with people.
Journalist: But here’s a contradiction: Linear actuators’ biggest advantage is high load and end-point precision. Yet the service scenarios you mentioned—front desk, barista—don’t really need those advantages. Isn’t that a mismatch of resources?
Zhang Haixing: It does seem contradictory, but it’s about redundancy—and redundancy isn’t a mistake. Right now, hardware is ahead of the models—that’s an industry-wide reality. Model capability doubles roughly every 3 to 6 months; in three to five years, fine end-effector manipulation will mature, and then the advantage of linear actuators will fully pay off.
You’re selling hardware now to prepare for future model capabilities—not designing hardware for what today’s model can do. The other way, waiting for the model to catch up and then finding your hardware is insufficient—that’s the real cost.
Journalist: What’s the current order backlog for MATRIX-3? Including overall delivery status for all generations?
Zhang Haixing: Third-gen unit orders have exceeded 100 units. The full-year target is 1,000 deliveries, with the fastest expected by the end of June.
Since the first-gen was launched in 2024, we’ve shipped several hundred units. Combined with the third-gen, we expect cumulative shipments of all models to be under 2,000 by the end of this year.
Journalist: Other companies are already charging toward tens of thousands, while your target for this year is 1,000. No worries?
Zhang Haixing: You can’t just look at volume alone. There are several dimensions. Some companies ship 10,000 units, but that might be a collection of several robot types—each averaging just a few thousand. In our view, that’s low-quality growth.
Also look at unit price. At 5–600,000 yuan per unit, we’re in a completely different revenue quality compared to someone averaging 100,000 per unit. We think the real thing to watch is who can first reach 100,000 units in a single category.
One more thing: the simpler the SKU, the better. Just piling up various SKUs to inflate the number doesn’t matter.
Journalist: You mentioned MATRIX-3’s unit price is 50–60,000 yuan. Who’s the target customer?
Zhang Haixing: Unit price is between 500,000 and 600,000 RMB. Main customers are government bodies and partners along the supply chain.
Journalist: Why would customers pay 50–60,000 yuan? There are robots on the market for 20–30,000 or even less.
Zhang Haixing: Let’s separate things. Products under 10,000 yuan are not true intelligent productivity; they’re more like smart toys with cameras, or educational tools—not the same category. It’s like comparing a four-wheeled car to a bicycle. Some industrial half-robots are better for pre-programmed actions, but their generalization and sensors are tailored for specific scenarios, not built for AI-driven tasks.
Different price ranges mean different capabilities. It’s like buying a car: luxury and budget cars have a huge price gap, but both are cars. Same with phones—different price points give users different experiences.
Our robot can actually work, and it can dynamically adapt to different tasks under AI. Why do some well-known big clients choose us after seeing so many robots? Because they want the best quality. They won’t cheap out and hurt their brand’s image just to save a few hundred thousand.
Journalist: Does the high price mean future sales focus overseas? Are overseas markets more willing and able to pay?
Zhang Haixing: In the next 3 to 5 years, we aim for a 50/50 split between domestic and overseas markets, with overseas slightly more.
Journalist: MATRIX-3 comes with 3D knitted biomimetic skin. There are two ways robots “wear clothes”: one is integrated fabric at manufacturing, the other is adding outfits later—like what some rental companies do. What’s the logic behind yours?
Zhang Haixing: The fabric skin trend—Figure and 1X are already doing it in North America. In China, Fourier and others are too.
We think skin makes human-robot interaction better. Robots with aluminum or plastic shells feel cold, heavy, and intimidating—they keep people at a distance. We want you to actually want to interact with a robot you buy—more warmth, more homely. It’s a design choice, and a consideration for user acceptance.
Countdown to Home Entry
Journalist: In the industry, there’s a debate: which is more important—the “brain” (end-to-end large model) or the “cerebellum” (motion control)? When executing complex physical tasks, which one decides?
Zhang Haixing: Both are essential, but if I have to prioritize—at the current stage of physical execution, the cerebellum and underlying hardware are the foundation.
No matter how smart the brain model is, no matter how well it breaks down a task, if during execution there’s jitter, or it doesn’t grip tightly, or it presses a button wrong—the causal chain breaks and the entire task fails. It’s a strict system engineering problem.
For example, in testing, a robot might suddenly fall. You check the algorithm code line by line—nothing wrong. Then you open the hardware and find a cold solder joint on a circuit board from a third-party supplier. Because robot control is a high-frequency response at hundreds or even thousands of hertz, a tiny hardware defect can cause loss of balance.
So it’s system engineering. Whoever can refine that reliable hardware, middleware, and software cerebellum system first is qualified to talk about grafting a super brain. If the chassis is unstable, even the most advanced end-to-end model is useless. Hardware reliability is the prerequisite for all model generalization.
Journalist: Speaking of large models, the whole industry struggles to get high-quality real-world data. Whether pure simulation data or teleoperation data collection, both have bottlenecks. How do you assess current data collection paths? How does Matrix Ultra Intelligence tackle this?
Zhang Haixing: Data is indeed the biggest pain point for embodied intelligence right now. Recently, there was a consensus that we must rely on teleoperation to collect real machine data—some even called for accumulating 100 million hours of data. But in reality, that approach has terrible ROI. First, teleoperation equipment is expensive. Second, different robots have different joint signals, torques, and noise, making the collected data not cross-platform.
So the most universal, high-quality data is actually human first-person task data. We don’t need complex full-body teleoperation rigs—just give a human worker a camera device to record hand and visual actions, and then use a vision model to map human motions to robot joints. Of course, this data is also the most expensive.
Our current judgment is that as world model capabilities improve, we might not even need to collect 100 million hours of real data. Gather a core set of high-quality data, then use a powerful vision and world model to synthesize and amplify—60–70% of needs can be covered this way. That cuts investment a lot and avoids the sunk cost of changing tech routes and losing all data.
Our strategy is to be disciplined, not the first mover. American companies have deep pockets—they can run 100 experiments at once. Even if 99 fail, one optimal solution is enough. We don’t have that scale. So our approach: let them run. Wait 3–6 months for the best practices to emerge, then use the distilled results to take the correct path directly.
That time lag is visible inside the industry, but ordinary users don’t feel it. Look at Huawei’s autonomous driving—it came half a year to a year later than Tesla, but became the best solution in China without burning that much cash.
Journalist: The “ChatGPT moment” for embodied intelligence is a hot topic. What’s your benchmark, and how long do you think it’ll take?
Zhang Haixing: I think aggressively in about three years, conservatively maybe five years.
Two trends happening simultaneously: one, code automation tools are getting incredibly powerful—these super tools are like “alien technology” writing code for you, accelerating the evolution of foundation models and world models. Two, hardware is iterating and getting cheaper, approaching the threshold where B and C end users will pay.
When you buy a silicon-based companion that comes out of the box with 100 basic skills (like grabbing a drink from the fridge, folding clothes), and you can expand to 1,000 skills through something like an App Store, without needing a developer—that’s the real tipping point. That’s the GPT Moment.
Before that, specialized and vertical scenario skills will be proven out first—like logistics sorting, unmanned stores, unmanned coffee shops. Some companies are already doing that this year, working to some degree, but the skills don’t transfer across scenarios—each needs retraining.
Journalist: When do you think robots will actually enter households?
Zhang Haixing: I’d say many companies are already starting to get products into homes this year, though forms vary—and not necessarily for housework. Those little robots around 10,000 yuan—they’re entering homes, just not doing much work. More for data collection and companionship.
True home robots that can work and make people willing to pay—based on overseas pace, I think next year is possible.
Journalist: Are you referring to NEO? They said early this year for mass production but it seems delayed?
Zhang Haixing: Yeah, I even placed a pre-order. They took my deposit but haven’t updated delivery schedule—haven’t produced it.
Journalist: Does Matrix Ultra Intelligence have any plans for the consumer market?
Zhang Haixing: We plan to launch a smaller, consumer-oriented robot next year. Price around 10,000 yuan. It has emotional value—can chat with kids, talk with elderly, and be controlled remotely via an app. But it won’t do much physical work.
The point isn’t just sales—it’s about accumulating real home-scenario data, building consumer stickiness, and preparing for a full-sized home robot later.
Journalist: Won’t selling a small consumer robot dilute the premium image you’re building in the B2B market?
Zhang Haixing: Not at all. Our premium B2B line will continue to move forward at the same time—the two lines don’t conflict. You’ll see our high-end robots working in real scenarios. Because of that high-end line, like a sports car brand, selling a consumer robot actually makes us more credible to consumers.
The Real Fight After the Bubble
Journalist: Some domestic robot companies are heading to the capital markets. Unitree’s STAR IPO was accepted, and others are rumored. How do you see this wave of public listings? What does it mean for the industry?
Zhang Haixing: I think it’s good for the industry. Early listings create an anchoring point for pricing. Once you go public, people can dissect your revenue, hardware quality, model capabilities, customer base—the whole picture. That gives the rest of the industry a benchmark.
Right now, top players are in late-stage valuations between 10 billion and 40 billion yuan. Is it expensive or cheap? No clear anchor. With public company comps, that question gets answered.
But this industry is still very early. In my view, no company in China has truly broken out yet. The whole track is still in a messy stage.
Journalist: How do you assess the current overall financing frenzy and valuation levels? Do they match real progress, or is there a bubble?
Zhang Haixing: From a primary market perspective, valuation bubbles are real. Companies getting to 10–20 billion are fueled by capital—everyone invests, pushing it up. It’s quite detached from commercial fundamentals. Hardware quality, real customers, revenue, profits—many of those can’t be examined properly. That’s where the industry is now. In the next three to five years, whether it’s small robots, factory robots, or home robots, real landable products will be the test.
Journalist: How will this shakeout happen? Which type of company will get eliminated first? Big tech is entering—do startups still have a chance?
Zhang Haixing: The shakeout is already quietly starting, and it’ll accelerate in the coming years.
First, natural elimination. Second, a lot of M&A will happen. For example, companies that raised several hundred million, burned cash on lab research, but can’t deliver mass-produced products with clear applications—they’ll burn out. Another type: companies that stack tons of unrelated low-end SKUs just to inflate sales volume—low-quality growth can’t hide lack of core technology.
M&A will be big. Hardware-good players bought by brain-good players. Some with cash go public and then buy a brain team. That’s also a capital exit path.
Big tech will come—a full-front assault. Many are already in. If a startup doesn’t have real barriers in core hardware, system reliability, or vertical scenarios, it’ll quickly become an appendage of the big ecosystems.
Competition is already saturated and intense. But startups still have a window. New eras create new brands. Users won’t think they need to buy robots from traditional old brands—they’ll want something newer. Startups still have a beginner’s protection period.
Those that stay on the table will be the ones with the most data, highest intelligence, largest user base, and greatest shipment volume.
Journalist: What’s the capital strategy for Matrix Ultra Intelligence going forward?
Zhang Haixing: We’re not anxious about it. Cash on hand is decent. We’re using government support, fundraising, and bank tools. We’ve always been disciplined about spending—tight on every cost. Our philosophy is to spend little to achieve big. We haven’t been first to burn cash on algorithms.
As for IPO, the earliest plan is around 2028–2029. We have state-backed LP investors and listed company shareholders who also need liquidity—they need a timeline. But that timing also aligns with our own judgment on technology maturity and business closure.
Journalist: Back to you personally—the label “designer” has always stuck. Some even question whether a consumer electronics designer has enough technical depth for building robots. After all this talk about hardware architecture, costs, algorithms—are you deliberately trying to shake that label?
Zhang Haixing: Not deliberately—and it’s hard to shake anyway. The label started when I won the Red Dot award and Professor Li Kaifu put it on me. The award was in the company’s name, and I led the design. Later, Musk invited me to set up Tesla’s China Design Research Center, so that image got reinforced.
Inside the company, everyone knows I haven’t touched design software in years. Notice that in our conversation, I haven’t brought up design once. It’s not intentional avoidance—it’s because what I actually do is this: system architecture, mechanical engineering, supply chain cost optimization.
I define myself as a Product Manager, or an Inventor.
Of course, wearing that label isn’t all bad—at least it means I have the best aesthetic in the industry. Taste and style are there. So with the same performance, we can sell at a higher price, with better margins. I figure—don’t sweat it.
Journalist: I saw on your company wall: “Vision: Build the largest AI robot company on Earth.” What does “largest” mean? Where do you see Matrix Ultra Intelligence in five years?
Zhang Haixing: In the five-year window, I hope we’re a well-known AI brand in both B2B and consumer markets. Customers using our products are very happy, and we keep iterating more advanced stuff.
In terms of revenue structure, hardware body should be less than half—more from compute subscriptions and scenario services. Because in the future, compute is productivity. Buying a robot is essentially buying compute that can turn into physical productivity. That’s a much healthier business model than just selling robots—and probably the inevitable future for every embodied intelligence company.
Building the world’s largest AI robot company—that’s our long-term vision. Goal is TOP 1 in scale and market cap.
