The EV Déjà Vu: Why China's Humanoid Robot Blitz May Be Winning the Battle — but Losing the War

China controls 80% of global humanoid robot shipments, has crushed hardware prices, and supplies the components Tesla cannot live without. The EV script is playing out — almost pixel-perfectly. But humanoid robotics is not an EV. The three ways this war is fundamentally different will determine which side ultimately wins.

SECTION 1

The Déjà Vu Is Real

Anyone who tracked China's electric vehicle industry between 2018 and 2024 will feel an unnerving familiarity when they look at the humanoid robotics landscape today. The storylines rhyme with almost suspicious precision.

In EVs, Tesla opened a gigafactory in Shanghai, turbocharging China's battery and drivetrain supply chain. BYD, NIO, Li Auto, and a hundred others followed — cheaper, faster-iterating, and ruthlessly price-competitive. The result: China captured the global volume game while Tesla captured the global profit game, taking an outsized share of industry earnings through its Full Self-Driving software.

In humanoid robotics, the same opening act is underway. Tesla's Optimus program has become the single most powerful catalyst for China's precision-components sector — orders flowing to Sanhua Intelligent Control, Topgroup, Green Harmonic, and scores of actuator and sensor makers across the Yangtze River Delta. Meanwhile, Chinese OEMs — Unitree, AgiBot, Zhiyuan, UBTECH — have taken the same playbook: ship fast, price aggressively, iterate relentlessly.

The numbers tell the story bluntly.

Sources: Omdia, IDC, Morgan Stanley, McKinsey, SemiAnalysis, company filings (2025–2026)

Morgan Stanley, in its June 2026 sector update, raised its China shipment forecast for the year from 28,000 to 50,000 units — a figure that reflects not hype but actual purchase orders and supply-chain capacity bookings. The bank describes the industry as having officially entered its 'early commercialisation phase.'

SECTION 2

EV vs. Humanoid: A Side-by-Side Anatomy

Before unpacking the differences, it is worth cataloguing just how structurally similar the two competitions look at this stage.

Sources: RobotToday analysis, McKinsey, SemiAnalysis, company disclosures

The surface similarities are striking enough to generate a confident prediction: as in EVs, China will dominate on hardware volume and cost, while the US captures the high-value software layer. The EV script, replayed.

That prediction is probably correct in its broad outline. But the deeper you look at the structural differences between the two industries, the more you realise that 'probably correct' conceals three profound ways in which the humanoid competition will be harder, longer, and more treacherous than the EV race — for both sides.

SECTION 3

Three Ways This War Is Fundamentally Different

Difference 1: EVs Sell Energy. Humanoid Robots Sell Behaviour.

This is the most fundamental distinction, and it is almost entirely absent from mainstream analysis of the sector.

An electric vehicle's core value proposition is fixed at the moment of manufacture. Range, top speed, charging time — these can be precisely measured, precisely compared, and precisely replicated. Once Chinese manufacturers mastered the lithium-iron-phosphate battery cell, the differentiation between a $25,000 BYD and a $40,000 Tesla became, in practical terms for the average consumer, a question of software and brand rather than hardware capability. The hardware had been democratised. Price war followed inevitably.

A humanoid robot's core value proposition does not exist at the moment of manufacture. It is created afterwards — through deployment, through failure, through the gradual accumulation of embodied knowledge. Two identical Unitree G1 robots, fresh off the same production line, represent the same product. The same two robots after six months — one working on a BYD wiring-harness line in Shenzhen, the other sitting in a warehouse — are fundamentally different machines. One has learned ten thousand ways that wiring harnesses can go wrong. The other has learned nothing.

This is not a trivial distinction. It transforms the economics of the entire industry:

• EV purchase decisions are one-time, easily benchmarked, and price-sensitive. Industrial buyers can compare range charts and make rational choices.

• Humanoid robot purchase decisions are the beginning of a long-term relationship. Buyers are not just purchasing a machine — they are purchasing access to a software and data ecosystem that will determine whether the machine becomes genuinely useful or permanently mediocre.

• A cheap, underperforming humanoid robot is not like a cheap EV that gets you from A to B with less range. It is like a cheap enterprise software platform that never quite does what it promises — except it also physically occupies your factory floor.

This shifts the competitive dynamic in a direction that heavily favours trust, software quality, and ecosystem depth over pure price. China can win on Day 1 of the sale. Whether it wins on Day 180 of the deployment depends on something supply chains cannot manufacture: demonstrable intelligence.

Difference 2: EV Data Was Passive. Robot Data Must Be Earned.

Tesla's most formidable competitive weapon in the EV era was its data flywheel. Sell millions of cars, equip them with cameras, collect billions of miles of real-world driving data, feed it back into the neural network, make the car smarter, sell more cars. This loop is elegant because it is passive: the data comes to Tesla; Tesla does not have to go to the data.

In humanoid robotics, the equivalent loop is broken at a critical point. Elon Musk acknowledged this himself on Tesla's Q1 2026 earnings call: Optimus units in the Fremont factory are 'primarily for learning, not productive tasks.' The robots are present. The work is not yet happening.

Why? Because the data that trains a useful robot is not camera footage of a factory. It is the proprioceptive record of a robotic hand grasping a wire bundle incorrectly 10,000 times, adjusting grip pressure 10,000 times, and eventually developing a reliable grip policy. This data can only be generated by robots that are actually doing the work — struggling, failing, recovering, and recording every interaction.

Tesla has cars on roads generating FSD data automatically. Tesla does not yet have enough Optimus units in enough factories doing enough real work to generate equivalent robot-behaviour data at scale.

China's OEMs, by contrast, shipped over 10,000 units into industrial settings in 2025. These robots are not performing as well as an Optimus might theoretically perform. But they are working, failing, and learning in the environments that matter most: the automotive assembly lines, 3C electronics factories, and logistics operations that represent the first commercial wave of humanoid deployment.

The analogy that captures this most precisely is not Tesla vs. BYD. It is early-era Google Maps vs. local taxi drivers. Google had superior algorithms. The taxi drivers had twenty years of intimate knowledge of every alley, bottleneck, and shortcut in their city. Over time, data won. But 'over time' took longer than expected, cost more than expected, and required genuine street-level deployment — not simulation.

The critical question for the humanoid industry is therefore not 'whose AI is better today?' but 'who will have accumulated more deployable robot-hours by 2028?' On current trajectories, the answer still favours China — but Tesla's Fremont conversion, if it succeeds in ramping to tens of thousands of units by late 2026, begins to close the gap at speed.

Difference 3: EVs Had One Battlefield. Humanoid Robots Have Three.

The EV competition, for all its complexity, ultimately resolved into a single competitive dimension: who can deliver the best vehicle at the best price. Hardware, software, and brand premium all mattered, but they could be collapsed into a single consumer decision.

Humanoid robotics cannot be collapsed. It is simultaneously three separate wars, operating on different timescales, with different leaders, and — crucially — different rules.

Sources: RobotToday analysis, McKinsey, SemiAnalysis, Morgan Stanley

Layer One: Hardware (China Has Won)

McKinsey's April 2026 supply chain report confirmed what the industry already knew: removing Chinese suppliers from Tesla's Optimus BOM would triple the unit cost, from approximately $46,000 to $131,000. China controls around 69% of rare-earth mining and 90% of NdFeB magnet processing. Musk has publicly acknowledged that magnet supply constraints have directly affected Optimus production timelines.

This advantage is structural, not cyclical. It took China twenty years of manufacturing investment to build. It cannot be onshored in three. Mexico and Thailand 'geopolitical firewall' strategies — routing Chinese components through third-country factories — are workarounds, not solutions. The hardware layer belongs to China, and this will remain true for the foreseeable competitive horizon.

Layer Two: Data (Currently Contested)

China leads on deployment volume. Its OEMs have more robots in more factories accumulating more physical-world interaction data. This advantage is real and should not be dismissed.

But it is not yet decisive, for two reasons. First, volume without quality is a weak advantage in AI training. Ten thousand robots performing simple pick-and-place tasks in controlled environments generate less useful training signal than one hundred robots performing complex multi-step assembly tasks with full sensory instrumentation. Second, Tesla's 'Digital Optimus' initiative — coupling the physical robot with a cloud-based world simulator and the Cortex 2.0 supercomputing cluster — creates a synthetic data generation capability that can partially compensate for limited physical deployment.

The data layer's verdict will not be clear until 2027 or 2028. By then, both the quantity and the quality of China's deployment data will be measurable against the performance of Tesla's Optimus in real factory settings. This is the decisive battleground — and it remains genuinely open.

Layer Three: The AI Brain (US Advantage, Geopolitically Walled)

On raw AI capability, the American stack is ahead. Tesla's end-to-end neural network architecture — the same system that underpins FSD v12 — applied to physical robot control represents a qualitative leap over the vision-language-action models that Chinese OEMs are currently deploying. Figure AI's partnership with OpenAI, backed by Microsoft's compute infrastructure, gives the American humanoid ecosystem access to frontier large language model capabilities that China's export-controlled chip environment cannot easily replicate.

Unitree's open-sourcing of UnifoLM-VLA-0 in March 2026 — a vision-language-action foundation model built on Qwen 2.5 VL 7B, released under CC BY-NC-SA 4.0 — is the most significant Chinese counter-move in this layer. It signals that at least one Chinese OEM understands that the endgame is not hardware but intelligence, and is willing to sacrifice short-term IP protection to build an open ecosystem. The strategic logic is unmistakable: China's OEMs are replicating the Android playbook — deploying open-source AI to build developer gravity and ecosystem reach, as a direct counter to the compute wall that US export controls have erected. The battle lines of the AI brain war are now drawn with uncomfortable clarity: Tesla's closed, vertically integrated 'iOS model' against China's open-source, community-driven 'Android model.' History suggests both can coexist — but history also shows that the open model tends to win on volume, while the closed model wins on margin. In humanoid robotics, both volume and margin matter enormously, which is precisely why neither side can afford to cede the other's ground.

Whether UnifoLM-VLA-0 is sufficient is a separate question. Against the computational resources Tesla is deploying — Cortex 2.0 at full capacity, AI5 inference chips targeting on-device training — the current Chinese AI stack remains outgunned. The geopolitical walls, meanwhile, are getting higher: H100 export restrictions, Unitree's 2025 congressional security review, and the broader US-China technology decoupling all constrain China's ability to close the AI gap through imports.

This is where the EV analogy breaks down most completely. In EVs, the 'intelligence layer' — FSD — existed on a software-only substrate. It could theoretically be replicated by any team with sufficient AI talent. The intelligence layer in humanoid robotics is physically embodied: it runs on chips, requires massive compute clusters, and depends on hardware ecosystems that are subject to export control. China's path to competitive humanoid AI runs directly through the most closely guarded bottleneck in the US-China technology competition.

SECTION 4

The Correction China's Industry Needs to Hear

For Chinese OEMs and their investors, the EV analogy is seductive precisely because it is partially true. China has the supply chain. China has the volume. China has the price advantage. In EVs, those three factors were sufficient to build a durable global industry.

In humanoid robotics, they are necessary but not sufficient — and the timeline for this becoming apparent is shorter than most Chinese players appreciate.

The danger is a specific failure mode: winning the hardware race so thoroughly, and burning through margins so completely in doing so, that the industry reaches the software inflection point without the capital or the institutional capability to compete.

Consider the arithmetic. Unitree's G1 currently generates approximately 67% gross margins at its $13,500 price point, according to SemiAnalysis's component-level analysis. This is strong. But Unitree's upcoming IPO, which targets a $610 million raise on the Shanghai exchange, implies R&D spending commitments that require sustained margins. The moment a serious price war begins among China's 140-plus humanoid OEMs — and it will begin, because that is what always happens in Chinese hardware markets — those margins will compress toward the 10–15% range that characterised the peak of EV commoditisation.

At 10% gross margin, funding a competitive foundation model programme costs more than the hardware business can sustain. At that point, the Chinese humanoid industry faces the same existential question that haunts every Chinese hardware-first company: how do you transition from selling boxes to selling intelligence, when selling boxes has consumed all your capital and margin?

Unitree's open-source VLA move can be read as an implicit acknowledgement of this risk — an attempt to build developer ecosystem gravity before the margin compression arrives. Whether it will be enough depends on execution speed that remains unproven at scale.

SECTION 5

The Warning Tesla and the US Industry Must Take Seriously

For American observers, the temptation is the mirror-image error: to assume that superior AI capability will inevitably translate into commercial dominance, and that China's hardware lead is a temporary inconvenience rather than a structural constraint.

Tesla has missed every Optimus production target since the programme was announced in 2021. The 2025 target of 5,000 units resulted in delivery of 'hundreds' — a miss of greater than 90%. The official timeline targets V3 production start in late July or August 2026 at Fremont — but with zero volume commitment. High-volume output is now explicitly targeted for 2027, with the AI5 chip that powers full Optimus capability not entering mass production until mid-2027 at the earliest. Consumer sales remain targeted for end-2027 — a date that, given Tesla's record of missing every Optimus milestone since 2021, most industry observers treat as optimistic at best. As of mid-2026, this assessment has been confirmed by Tesla's own production reality.

Meanwhile, Chinese competitors are not standing still on AI. The Spring Festival Gala in January 2026 — in which Unitree robots performed complex martial arts routines, trampoline somersaults reaching three metres, and sustained 4 m/s running speeds fully autonomously — demonstrated that the gap in physical performance between Chinese and American humanoids is narrowing faster than most Western analysts expected.

More importantly, every month that Tesla's Optimus is not working in real factories at scale is a month in which Chinese robots are. The data flywheel that Tesla needs to fulfil its AI potential requires physical deployment. Physical deployment requires production. The Fremont factory conversion is underway, but the history of this programme suggests that production ramp timelines will be longer than announced.

Tesla's structural advantage in AI capability is real but fragile. It depends on sustained Optimus deployment success to generate the training data that would make the AI advantage self-reinforcing. Without that data flywheel beginning to spin at meaningful scale by mid-2027, the AI brain advantage remains theoretical rather than demonstrated.

SECTION 6

The Verdict: Winning the Body, Contesting the Soul

The EV analogy is correct in its diagnosis but incomplete in its prescription.

China has, with high confidence, already won Layer One: the hardware and supply-chain war is over, and the outcome favours Chinese manufacturers in a way that decades of investment have made nearly irreversible. The BOM advantage, the rare-earth control, the manufacturing density of the Yangtze River Delta — these cannot be replicated by American policy within any commercially relevant timeframe.

Layer Two — the data war — is live and contested. China's volume advantage is real. Tesla's quality-of-deployment potential is real. The outcome here is genuinely uncertain and will determine more about the long-term competitive structure of this industry than any other single factor.

Layer Three — the AI brain war — currently favours the United States, but it is shielded by geopolitical walls whose durability is itself uncertain. If US-China technology decoupling continues to intensify, China's AI ceiling drops and the US advantage compounds. If a geopolitical thaw occurs, or if Chinese researchers find competitive workarounds to compute constraints, the gap could narrow significantly.

The ultimate question — and it is one that no analyst can currently answer with confidence — is whether China can traverse the distance from hardware dominance to AI platform credibility before the margin compression that hardware commoditisation inevitably brings makes that journey financially impossible.

In EVs, China was allowed to run that race without an existential constraint at the software layer. BYD does not need permission from Nvidia or the US Department of Commerce to sell a car. Chinese humanoid OEMs, competing in a world where the intelligence layer runs on export-controlled chips and proprietary AI infrastructure, face a constraint that has no equivalent in the EV story.

The most probable medium-term outcome — by 2028 to 2030 — is a bifurcated global market: Chinese humanoid robots dominating volume deployments in price-sensitive industrial settings across Asia and the Global South, while American platforms — Tesla Optimus, Figure, potentially others — capture the high-margin, AI-intensive deployments in Western automotive, semiconductor, and logistics facilities.

This is not quite the EV script. In EVs, the bifurcation was about geography and price point, but both sides were selling the same kind of product. In humanoid robotics, the bifurcation may be about the very nature of the product: China selling sophisticated hardware tools; the US selling intelligent platforms. The former are cheaper. The latter are, potentially, more profitable — per robot, per year, indefinitely.

Whether that distinction makes the US the long-term winner depends on one thing above all: whether Tesla can execute a production ramp that was announced in 2021 and has not yet, as of mid-2026, resulted in a single commercial sale.

History suggests caution. The robots at the Spring Festival Gala did backflips. The factory lines in Fremont are still being installed.