Where capital flows tells you where talent is concentrating, where production capacity is being built, and which technologies will reach commercial scale first.

The China Round is free, published every Friday, and written for investors, operators, and founders tracking where China’s tech capital is moving and why it matters.

About the author: Dermot McGrath is Irish, based in Shanghai, a decade in China’s tech and investment ecosystem, fluent Mandarin. Co-founded a VC fund scaled to nine-figure AUM. Now running ZenGen Labs, a cross-border strategy studio covering AI, robotics, and deep tech.

Where the Money Went

130 funding events across four sectors. Over $1.4 billion in disclosed capital. Three quantum computing raises totalling over ¥1.6 billion. Angel-stage valuations that would have been Series B territory two years ago. The A/A+ stage accounts for 38% of all deals, but the signal is loudest at the extremes.

🤖 Robotics (27 deals)

The busiest sector by narrative weight. 27 raises in seven days, at least eight of them embodied AI or humanoid plays. Galaxea AI (星海图) led with ¥2 billion (~$280M) for a B+ round, valuing the company at ¥20 billion. JAKA Robotics (节卡机器人) closed an E round at a ¥6 billion valuation, one of a handful of Chinese cobot companies with real international footprint. At the early end, multiple seed and angel rounds for companies building everything from cooking robots to ocean-floor inspection vehicles.

Stage breakdown: 11 angel/seed, 9 A/A+, 5 B+, 2 strategic. The early-stage density signals continued new entrant formation, not just incumbents scaling.

🧠 AI (18 deals)

18 deals, but the individual bets run larger than the count suggests. MOSI Intelligence (模思智能) pulled in hundreds of millions of RMB at angel stage for speech-first foundation models. Tsingwei (清微智能) hit Pre-IPO stage for AI inference chips. The rest skew early: lots of angel and Pre-A rounds for vertical AI applications (drug discovery, gaming, spatial reasoning, pet tech).

⚙️ Advanced Manufacturing (78 deals)

The largest category by volume, and the most diverse. Sub-sectors span semiconductors (~15 deals), hydrogen and energy storage (~10), aerospace and eVTOL (~6), lasers and photonics (~5), new materials (~8), and a long tail of sensors, drones, and industrial components. Headline deal: CIX Technology (此芯科技) raised ¥1 billion (~$140M) for ARM-compatible AI PC chips. EPG (易普集) closed $100M for modular data centers backed by Alibaba Cloud and BlackRock’s Decarbonization Partners.

💡 Frontier Tech (7 deals)

Small by count, massive by capital. Three quantum computing raises (see theme below). One invasive BCI company. Two synthetic biology startups. One display chip maker.

Spotlight

QBoson (玻色量子)

Series B, ¥1 billion (~$140M), ¥7 billion valuation (~$970M)

Photonic quantum computing.

QBoson builds photonic quantum computers for commercial deployment. CEO Wen Kai holds a Stanford PhD in quantum computing; the founding team includes researchers from Stanford, Tsinghua, and the Chinese Academy of Sciences. They have already shipped China’s first commercial photonic quantum computer, with 10+ units delivered to clients including the National Supercomputing Center in Chengdu and China Mobile. Their current system runs 550 qubits, with a 1,000-qubit model on the roadmap. QBoson broke ground on the world’s first dedicated photonic quantum computer manufacturing facility in Shenzhen last year.

Most quantum computers (IBM’s, Google’s, the ones generating the headlines) run on superconducting qubits that require cooling to near absolute zero, which means expensive dilution refrigerators and lab-scale infrastructure. Photonic quantum computers use particles of light instead, operate at room temperature, and are significantly easier to manufacture and deploy. The trade-off has historically been lower qubit fidelity, but photonics opens a path to commercial deployment that superconducting systems don’t. China has largely ceded the superconducting race to the US; photonics is the bet that there’s a different lane worth winning.

ICBC Capital (工银资本) and Shunxi Fund (顺禧基金) co-led, alongside Beijing Financial Holdings and others.

Announcement

Galaxea AI (星海图)

Series B+, ¥2 billion (~$280M), ¥20 billion valuation (~$2.8B)

Embodied AI humanoid robot platform.

Galaxea AI builds universal dual-arm humanoid robot platforms (R1 Pro, R1 Lite) targeting factory-floor deployment and, eventually, general-purpose use. Founded September 2023 by Jiyang Gao (born 1992), Tsinghua to USC to Waymo to Momenta. Co-founded with Tsinghua robotics professor Huazhe Xu, who left in February 2026 to start Breaking Shell Robotics (see below). Meituan has been a major strategic investor since 2023. Per 36kr, the B round and this B+ round closed within a single month of each other. ¥20 billion valuation for a company two and a half years old. Target: 1,000 units shipped by end of 2026.

The wider bet is on becoming the platform layer for embodied AI: a robot learning to navigate and manipulate the real world will develop capabilities that a language model running in a data center never can. In the US, Figure, Physical Intelligence, and Apptronik are the main funded challengers. In China the serious platform plays are fewer: Unitree (already at 5,500 units per year at consumer price points), Agibot, and Galaxea. Chinese manufacturers have factory floors willing to deploy right now, giving China a training data advantage that’s hard to replicate from a US lab.

Lens Technology (蓝思科技), Silicon Core Ventures (矽芯投资), and 10 others; Huaxing Capital (华兴资本) as financial advisor.

Announcement

MOSI Intelligence (模思智能)

Angel, hundreds of millions of RMB, ¥1.5 billion valuation (~$210M)

Speech-first foundation models.

MOSI Intelligence builds speech-first foundation models with a unified token structure across voice, video, text, and image. Founded November 2024 by Fudan professor Qiu Xipeng, who won the CCF-ACM AI Award and built MOSS, China’s first open-source ChatGPT alternative, in February 2023. Hundreds of millions of RMB at angel stage, ¥1.5 billion valuation, for a company founded 16 months ago by an academic.

Most Chinese AI labs started with text and added speech later. Qiu is going the other direction: voice as the primary modality. The product lineage (SpeechGPT, AnyGPT, MOSS-TTS series) is academic but credible. Target applications include consumer electronics, autonomous vehicles, and embodied AI, all domains where voice interaction matters more than text. The commercial question is whether speech-first is a genuine architectural advantage or a differentiated pitch, especially against established Chinese frontier models like Kimi and Minimax.

IDG Capital (IDG资本) led, with Yuanhe Holdings (元禾控股), MiraclePlus (奇绩创坛), Shanghai Future Industry Fund (上海未来产业基金), and others.

Announcement

CIX Technology (此芯科技)

Series B, ¥1 billion (~$140M), ¥8 billion valuation (~$1.1B)

ARM-compatible AI chips for PCs and edge devices.

CIX Technology builds ARM-compatible AI chips for PCs and edge devices. CEO Hank Sun was AMD’s SOC director and head of AMD’s China semiconductor business. CTO Fang Liu spent 11 years at Apple on CPU/GPU architecture, then SOC work at Meta and AMD. That concentration of semiconductor talent in a single founding team is rare. Their CIX P1 chip (6nm, 12-core ARM CPU, 30 TOPS NPU, 45 TOPS combined with CPU and GPU) is already in Minisforum mini PCs shipping to consumers.

China’s semiconductor conversation has been dominated by datacenter GPUs, the race to replace Nvidia H100s. CIX is playing a different field: the billions of edge devices that need local AI inference without a data center connection. Qualcomm’s Snapdragon X competes in the same space globally; the P1 is one of the first Chinese-designed chips to reach commercial production with OEM partnerships.

Shanghai Integrated Circuit Industry Investment Fund (上海集成电路产业投资基金) and Pudong Venture Capital (浦东创投) co-led, with Oriza Ventures (元禾璞华), Lenovo Ventures, and others.

Announcement

Breaking Shell Robotics (破壳机器人)

Angel, tens of millions of USD, ¥2.6 billion valuation (~$360M)

Home robots. That is nearly everything publicly known about this company.

The founder is Xu Huazhe (许华哲), a Tsinghua assistant professor who co-founded Galaxea AI and served as its Chief Scientist before leaving in February 2026. Galaxea provided the seed investment. That pedigree explains the ¥2.6 billion valuation at angel stage with almost no public information: investors are backing the person, not a product.

Xu laid out his thesis at the Zhongguancun Forum on March 29. Breaking Shell will build its own hardware (vertical integration, not outsourced bodies) and focus on a narrow set of household tasks rather than chasing full generalization. Object passing and tidying first, then storage and cleaning. Cooking will appear as demos but not as a product. No physical contact tasks (elderly care, childcare) for now. The bet is that defining 10 tasks precisely and collecting targeted data will scale faster than building a 21-degree-of-freedom generalist that can’t do any single task reliably.

Also on the Radar

Unitary Quantum (幺正量子) — Pre-A, hundreds of millions of RMB, ¥1.5 billion valuation. Ion trap quantum computing out of Hefei. Geely (automaker) and Ant Group (fintech) co-led. An unusual investor pairing that suggests quantum is entering the strategic planning of companies far outside the usual deep tech investor base.

SpinQ (量旋科技) — Series C+, ¥600 million (~$83M), ¥3 billion valuation. Superconducting and NMR quantum computers already shipping to 200+ institutions across 40+ countries. Revenue hit ¥50 million in 2024. The most commercially mature quantum company in China, possibly in Asia.

JAKA Robotics (节卡机器人) — Series E, ¥6 billion valuation. Collaborative robots deployed in 100+ countries with offices in Japan, Germany, the US, Malaysia, and Singapore. Customers include Toyota and Schneider Electric. 21.9% of China’s cobot market. Shanghai-based.

EPG (易普集) — Series B+, $100 million, ¥3.9 billion valuation. Prefabricated modular data centers. Alibaba Cloud and Decarbonization Partners (BlackRock/Temasek JV) co-invested. Offices across Southeast Asia, the Middle East, and Europe. One of the few companies in this dataset already operating globally.

Mingshi BCI (明视脑机) — Angel, ¥150 million (~$21M), ¥750 million valuation. Invasive brain-computer interface for neural applications. Backed by 中科创星 (CASSTAR). China approved its first commercial BCI implant in March 2026 (Neuracle Technology, for paralysis patients), and capital is following.

Bowintec (博银合创) — Series A, undisclosed amount. Dexterous robotics. The sole investor: Bosch China. A sole bet from a German industrial giant on a Chinese dexterous robotics startup.

RoboParty (萝博派对) — Angel, $20 million, ¥715 million valuation. Humanoid bipedal robots. Guoxiang Capital (国香资本) and Matrix Partners China (经纬创投) co-led. Limited public information, but top-tier VC confidence at angel stage.

Lingravity (零重力飞机工业) — Pre-B, ¥150 million (~$21M), ¥4.5 billion valuation. eVTOL aircraft plus electric fixed-wing (RX1E, already certified for commercial delivery). Led by Xiang Jinwu, a Chinese Academy of Engineering academician. Team from AVIC and COMAC. Xichuang Investment (锡创投), Huikai Investment (惠开投资), and Jinpu Investment (金浦投资) among investors. Based in Hefei with R&D in Nanjing, Shenzhen, and Wuhu.

快酷创新 — Series A, undisclosed. Smart self-service haircutting devices. Investors: Hillhouse Capital (高瓴创投) and MiraclePlus (奇绩创坛). Hillhouse backing an automated barber is the most unexpected bet in a dataset dominated by chips and humanoids.

Differential Robotics (微分智飞) — Series A, hundreds of millions of RMB, ¥2.6 billion valuation. Intelligent drone swarm systems for industrial and urban inspection. Founded July 2024. Huakong Fund (华控基金) led, with Baidu Ventures and others. Nine months from founding to a ¥2.6 billion valuation.

The Bigger Picture

China’s quantum moment

Three quantum computing deals totalling over ¥1.6 billion in disclosed funding, each using a different technological approach: photonic (QBoson), superconducting/NMR (SpinQ), and ion trap (Unitary Quantum). That spread suggests a coordinated infrastructure build across all three major technological approaches.

China’s public quantum funding stands at an estimated $15 billion, nearly four times the U.S. government’s $4 billion. Late last year the State Council launched the National Venture Capital Guidance Fund, seeding ¥121.8 billion across three regional funds covering frontier tech from AI to quantum to 6G, with a target scale of ¥1 trillion. The new signal is who’s writing the checks. ICBC’s investment arm leading QBoson’s round. Geely and Ant Group co-leading Unitary Quantum. ICBC is China’s largest state bank. Geely makes cars. Ant Group runs payments. None of them are specialist deep tech investors. Their presence in the same technology category in the same week points to broad institutional conviction (or FOMO), the kind that precedes large procurement and deployment decisions.

SpinQ’s numbers add context: ¥50 million in revenue in 2024, a chip already shipped to the Middle East. Parts of this are already commercial.

The billion-dollar angel

Multiple companies in this dataset hit billion-RMB valuations at angel or seed stage with minimal or no shipped products. Breaking Shell Robotics at ¥2.6 billion. MOSI Intelligence at ¥1.5 billion. Differential Robotics at ¥2.6 billion (see Also on the Radar).

The pattern extends beyond China. Globally, AI seed-stage companies now command a 42% valuation premium over non-AI startups. Even across all sectors, Carta data shows US median seed valuations up 60% in three years to $24 million.

Breaking Shell hit ¥2.6 billion (~$360 million) at angel stage. That’s 15x the US median seed, for a company with no public product. Breaking Shell’s valuation is entirely a bet on Xu Huazhe’s Tsinghua/Galaxea track record. MOSI’s is a bet on Qiu Xipeng’s MOSS lineage. Differential’s is a bet on the team behind it, nine months in.

The question is whether these valuations reflect genuine conviction in specific teams (the Kunlun Xing model: two resumes, no product, $1 billion in 10 days) or whether capital is chasing sector allocation targets, and the honest answer is probably both operating simultaneously.

Robotics: the 2021 autonomous driving moment

27 robotics deals in seven days. By industry estimates, the sector has grown from roughly 120 companies to over 320 in three years. Per the China Academy of Information and Communications Technology, embodied intelligence recorded ¥73.5 billion in financing across 740+ deals in 2025. At least six startups secured $100 million or more in early 2026 alone.

Chinese media has started drawing the comparison to autonomous driving’s 2021 peak, when hundreds of AD companies raised billions before the market consolidated violently. The NDRC has warned about a bubble, flagging “low-level duplication” across 150+ companies chasing immature technology routes and unclear commercialization models. The NEV precedent is instructive: roughly 400 Chinese EV companies ceased operations between 2018 and 2025, leaving around 100 active, with McKinsey projecting fewer than 50 by 2030. A similar shakeout looks likely here.

The data this week supports the saturation thesis. Eight of the 27 robotics deals are embodied AI or humanoid plays. Most are at angel or Pre-A stage. Most have undisclosed amounts. The long tail is getting very long.

The counter-signal is the industrial layer. Companies like JAKA Robotics (cobots, E round, 100+ countries) and Bowintec (Bosch’s sole bet) suggest that the industrial robotics layer, the unglamorous part, is maturing separately from the humanoid hype cycle. The companies building components, sensors, and motion control systems may benefit regardless of which humanoid companies survive.

Deal data compiled from Chinese corporate announcements, financial news, and investment databases. Amounts and valuations are as reported and may be incomplete.

Twelve days earlier, in Wuxi, a Jiangsu province city that has become one of China’s biggest commercial aerospace hubs, thirteen companies signed the incorporation documents for 星联体 (”Star Consortium”).

With registered capital of ¥2 billion, it’s the largest registered private commercial aerospace enterprise in China. The founding partners include some of the country’s most active commercial satellite operators, and the business model is, structurally, the opposite of SpaceX.

The two aren’t rivals in any useful financial sense (one is targeting a $1.75 trillion public valuation, the other has ¥2 billion in registered capital) but they represent two different theories of how commercial space infrastructure gets built, both targeting the same 2027-2028 window.

The Trillion-Dollar Stack

SpaceX’s S-1 is still confidential, though the public moves over the past two months sketch the thesis clearly enough. In early February, SpaceX announced a deal to acquire xAI, Musk’s AI company, valuing the combined entity at $1.25 trillion. The rationale was to integrate xAI’s compute with Starlink’s satellite network to build orbital data centers, solar-powered AI infrastructure running in low Earth orbit.

Ok. But. Why?

Data centers are bottlenecked by power grid access, land permits, cooling infrastructure, and terrestrial fiber connectivity. An orbital layer sidesteps all of them. Satellites have line-of-sight to any point on Earth and can route compute dynamically without laying cable. Whether the economics work is unresolved. Hardening hardware for orbit, tolerating latency on AI workloads, dissipating heat in a vacuum: none of these are solved problems at commercial scale. But SpaceX’s actual business is strong enough that investors can price the company on the vision and absorb the uncertainty.

Revenue hit roughly $15-16 billion in 2025, with Starlink accounting for $10.4 billion across 9 million subscribers. Free cash flow was around $2 billion, forecast to reach $5 billion in 2026. Active government contracts total $22 billion, including classified work through Starshield, the military satellite program. The IPO packages rockets, satellites, internet service, military infrastructure, and AI compute into a single entity, with each layer in theory feeding the others.

The xAI deal brings scale and complications together. Grok generated 3 million harmful images over an 11-day period in late 2025, triggering lawsuits (including a suit from the city of Baltimore), UK regulatory investigations, Southeast Asian bans, and enforcement action from New York’s attorney general. Those liabilities will require disclosure when the S-1 goes public. There’s also a CFIUS review underway because xAI has foreign investors including Qatar’s sovereign wealth fund and Abu Dhabi’s MGX, while SpaceX handles ITAR-controlled military hardware. Reports suggest this friction pushed the filing from February to yesterday. A June or July listing target leaves limited runway to resolve them.

Space for One More?

Even with those overhangs, SpaceX is filing from a position no competitor has come close to replicating. It built its dominance by owning the full stack, and the track record for everyone who tried to compete without that suggests it’s not optional.

Amazon’s Kuiper constellation had roughly 212 production satellites in orbit as of late 2025, and has already filed to extend its FCC deployment deadline from July 2026 to July 2028. It has AWS as a natural distribution channel and deep capital, but it does not own its rockets and does not own the full stack.

OneWeb made the case more starkly. It launched a first-generation low Earth orbit (LEO) constellation, ran out of money in 2020, and filed for bankruptcy before the service reached commercial viability. The UK government and Bharti Global acquired it out of administration, and it eventually merged with Eutelsat in 2023. The capital ran out because OneWeb tried to compete with Starlink on internet service without Starlink’s launch cost advantage, government contract floor, or ability to iterate on hardware.

Competing head-on with Starlink for consumer internet, without owning the rockets and the full vertical stack, doesn’t add up economically.

China’s commercial space companies are building for a different market entirely, but have likely learned from these case studies. Star Consortium exists to get them there faster than any one of them could alone.

The Wuxi Model

Star Consortium is targeting the industrial and government services market, selling maritime domain awareness, precision agriculture sensing, urban IoT, and emergency communications to Chinese shipping companies, provincial agriculture bureaus, and city governments. That market requires being operational, reliable, and cheaper than ground-based sensor networks or imported satellite data services, without needing to match Starlink on price or global coverage.

The consortium was formally established on March 20 at a development conference in Wuxi’s Liangxi District. Four of China’s top ten private commercial space companies are based in Wuxi, a concentration that reflects sustained provincial investment over the past decade. The thirteen founding partners include LandSpace (蓝箭航天), Space Pioneer (天兵科技), Oriental Space (东方空间), and Galaxy Aerospace (银河航天), alongside district-level investment groups and smaller satellite and terminal hardware companies. Registered capital surpasses Chang Guang Satellite Technology’s (长光卫星) ¥1.97 billion.

Star Consortium integrates eight existing constellation projects under a shared operational platform covering rocket launches, satellite manufacturing, payload development, terminal applications, and data operations. The stated model is “state-owned leadership, industrial coordination, market-oriented mechanism,” where provincial capital sets the frame, private companies execute within it.

The playbook is familiar in China. Local governments seeded the EV industry the same way, guaranteeing bus fleet procurement so manufacturers like BYD could iterate on battery tech and drive costs down before the consumer market was ready. Wuxi appears to be running the same sequence with satellites and launch services.

Tianyuan Constellation

The flagship initiative is the 天源星座 (”Tianyuan Constellation”), one high-orbit satellite and 216 low-orbit satellites, targeting completion by 2028. In 2026, development begins on the first batch, with 6-8 experimental satellites planned for launch. The strategy is “build while operating,” selling services on partial infrastructure while the full constellation assembles, rather than waiting for completion before taking customers.

Galaxy Aerospace’s Dual Play

Galaxy Aerospace is one of the thirteen founders, and filed its own IPO advisory on March 30, ten days after signing on to Star Consortium. That dual position, equity member in the shared platform while simultaneously seeking independent public capital, shows how the Chinese commercial space sector is positioning for the next few years. Operational assets get pooled for efficiency; the companies remain separately financed. Galaxy Aerospace’s Series C drew 全国社保基金 (”National Social Security Fund”) and 中央汇金 (”Central Huijin”), state capital designed for long-duration holds rather than venture exits. Its IPO materials cite progress in direct-to-mobile satellite technology; the company deployed over 20 satellites in 2025 against a production capacity of 100 per year.

Two Theories, A New Space Race?

Both models are betting that commercial space infrastructure, not publicly funded programs like Artemis II, will define who controls orbital services by the end of the decade. They disagree on how to build it.

SpaceX’s answer is vertical integration. Own the rockets, the satellites, the distribution, the compute. Each layer compounds the others. Starlink revenue funds Starship development, Starship reduces launch costs, lower launch costs improve Starlink margins, the orbital compute layer extends the value of every satellite already in orbit. The logic requires a single entity capable of absorbing losses at each stage while the compounding builds. Musk has that entity, and US government contracts provide a revenue floor while the commercial side matures.

Star Consortium starts from different constraints. No Chinese commercial space company has SpaceX’s balance sheet, and none has a Pentagon-equivalent relationship to lean on while the market develops. The assets exist (satellites, rockets, data operations, terminal hardware) but they sit across organizations with different investors and different timelines. Pooling them under a shared platform makes each component more valuable without requiring any company to acquire the others. Provincial capital provides patience where venture capital would demand exits; the private operators provide commercial urgency where state-owned enterprises typically cannot.

The Unresolved Questions

Both theories target the same 2027-2028 window, and both have hard dependencies that could slip.

When SpaceX’s S-1 becomes public, the conversation will centre on valuation multiples, Starlink subscriber growth, and Grok’s legal exposure in formal disclosure. Whether the orbital data center thesis is achievable on any near-term timeline is a harder question, and the IPO prices SpaceX on that vision without needing to resolve it.

The Launch Cost Bottleneck

On the Chinese side, the 2027-2028 viability target depends on launch cost reductions happening roughly on schedule. LandSpace’s Zhuque-3 (朱雀三号) attempted its first stage recovery in December 2025 and failed. The company is targeting a recovery test in Q2 2026, with the first full reuse flight in Q4 if that succeeds. The goal is to cut launch costs from around 100,000 yuan per kilogram to 20,000 yuan per kilogram. Until that or something close to it is achieved, the economics of running a commercial constellation at scale don’t make sense. Deep Blue Aerospace has vehicles at the pad with similar ambitions.

Thirteen Partners, One Platform

Then there’s the coordination problem. Star Consortium has to execute across thirteen partners with different investors, different timelines, and different commercial incentives. Shared platforms hold together until the partners’ interests diverge, and that divergence is predictable, whether it’s a member finding a better commercial use for its satellite capacity than the shared plan allocates, or IPO pressure pushing a company toward maximizing its own metrics. The “state-owned leadership” component (read guaranteed customers) may be what holds it together through those tensions. It may also be what makes the structure slow when speed is what the market rewards.

Can thirteen companies with different owners share operational infrastructure across a full space stack without coordination costs consuming the efficiency gains? Can Galaxy Aerospace run as both a consortium member and a separately listed public company without those two roles pulling in different directions?

Both ideas are in orbit. Who knows where they’ll land.

So, why should we care?

The Big Four (AMZN, GOOGL, MSFT, META) spent $410 billion on AI infrastructure last year and are expected to spend roughly $650 billion in 2026, according to Bridgewater Associates. The supply data suggests it still won’t be enough. The silicon, the electricity, and the cost economics are all working against them, and that gap between capital and physical reality is where a second compute market is forming, one built on coordination rather than construction.

This report examines whether that alternative actually works, what the hardware and economic realities look like, what problems crypto coordination solves that a traditional platform cannot, and where the structural opportunities are as of today.

The Laws of Physics

AI runs on hardware and electricity. Every model trained, every chatbot response generated, every image rendered requires physical GPUs burning real power in real buildings. Jensen Huang called $700 billion in hyperscaler capex “just the start.”

The constraint on AI compute in 2026 is not a shortage of capital. It is a shortage of physical inputs (silicon and electricity) and an open question about whether the returns justify the scale of investment required to expand them.

Silicon

TSMC Chairman C.C. Wei told the Semiconductor Industry Association in November 2025 that the fab capacity to produce AI-grade chips falls “about three times short” of what major customers plan to consume. SemiAnalysis concurs: as of this month, “The AI industry is now firmly in the silicon shortage phase. For the next two years, TSMC will not be able to add enough capacity to fully meet demand.” GPU lead times sit at 36 to 52 weeks. Hyperscalers would deploy more capital if they could, but there is not enough fabricated silicon to absorb it.

Power

The International Energy Agency projects data center electricity consumption doubling to 945 TWh by 2030, growing four times faster than all other sectors combined. Goldman Sachs estimates that 60 percent of the demand increase will require entirely new power generation capacity, not just grid reallocation. Interconnection queues now average five years from request to operation, and new transmission projects regularly take a decade or more to complete. Google’s global head of sustainability told NetworkWorld that utilities are quoting four-to-ten-year wait times, with one offering a 12-year study period and Ireland and Singapore have restricted new data center approvals entirely. In the US, a proposed AI Data Center Moratorium Act signals that political resistance to centralized buildout is no longer confined to small European markets. Even where GPUs exist, the grid cannot feed them fast enough.

Economics

McKinsey’s base case calls for $5.2 trillion in cumulative AI infrastructure capex through 2030, with an accelerated scenario reaching $7.9 trillion. J.P. Morgan estimates that AI needs to generate over $600 billion in annual revenue just to achieve a 10 percent return on that infrastructure spend. For context, the largest AI companies combined (OpenAI, Anthropic, Microsoft’s AI division, Google Cloud, Amazon) generated roughly $50 to $60 billion in identifiable AI revenue in 2025, according to company disclosures and analyst estimates.

Some people are asking if all of this is worth it. In January, Goldman Sachs’ chief economist concluded that, given most AI hardware is imported, most of the benefits flow to Taiwanese GDP and AI’s net contribution to US GDP growth was “basically zero.”

However, the demand does seem to be there. Inference demand is still compounding: Nvidia’s inference revenue already exceeds 40 percent of its AI sales, and Huang says inference demand is “about to go up by a billion times.” The capital required is unprecedented, and the returns remain unproven at the scale the models demand.

The problem is, compute stack was built vertically integrated: Nvidia designs chips, hyperscalers build clusters, customers rent capacity. That architecture is optimized for a world where supply can scale in step with demand. It cannot. Some of these shortages take years to resolve, and the GPU clusters that do exist are concentrated in a handful of jurisdictions while demand, regulation, and latency requirements are global. That mismatch is the structural opening for a second compute market.

Does Decentralized Compute Actually Work?

Hundreds of millions of dollars have flowed into startups building decentralized AI compute. 0G Labs raised $290 million to build distributed training infrastructure. Prime Intellect pulled in $70 million from Founders Fund. Nous Research raised $65 million from Paradigm at a billion-dollar valuation to build open-source AI models and its Psyche distributed training network. Gensyn raised $50 million led by a16z crypto to build cryptographically verified ML compute. Render Network and io.net raised $35 million and $30 million respectively to aggregate distributed GPUs into unified compute marketplaces.

Most of that capital has gone into building infrastructure: coordination protocols, GPU marketplaces, incentive layers. Covenant-72B is the first major output to come out the other end of that plumbing. The model itself is not frontier, but the training run proved that distributed coordination works at scale and produced novel compression techniques like SparseLoCo that make the entire architecture more viable.

150GB per Sync - ¿You LoCo, Ese?

The intuitive objection to distributed training is data transfer. Training a model requires GPUs to constantly exchange updates about what they are learning (called gradient updates), and when those GPUs are spread across cities or continents rather than wired together inside a single datacenter on high-speed private networking, the communication overhead should be crippling. Two projects have shown it does not have to be.

Covenant solved this with SparseLoCo, a combination of chunk-wise sparsification and 2-bit quantization that achieves more than 146x compression on the data exchanged per training sync, meaning for every GB that would normally need to be sent, only about 7 MB actually travels over the wire. That brought total communication overhead down to roughly 6 percent of training compute. That is what made coordinating 70+ servers over standard internet connections viable.

0G Labs took a different approach, training a 107-billion-parameter model using their DiLoCoX compression framework in a coordinated partnership with China Mobile, claiming a 10x efficiency improvement over prior records. Their DiLoCoX paper reveals the run used 160 NVIDIA A800 GPUs with bandwidth artificially throttled to 1 Gbps using Linux traffic control to simulate distributed conditions, so this was a compression proof at 107B scale rather than a true geographically distributed deployment.

Both are crypto projects, but the coordination models differ: Covenant ran on anonymous operators earning token rewards across genuinely distributed hardware, while 0G’s run was a structured partnership between known parties in a simulated environment. Covenant proved permissionless distributed training works in the wild; 0G proved the compression techniques scale to larger models, and that institutional players like China Mobile see distributed training as viable.

Location, Location, Location

In every case, the hardware was datacenter-grade: Covenant ran on B200s, 0G used A800s (China export-compliant version of the A100), Prime Intellect’s INTELLECT-1 used 112 H100s across five countries, and Targon, a decentralized inference network on Bittensor, runs 1,500+ H200s across independent operators. These are GPUs that sell for tens of thousands of dollars each. Despite the pitch deck dream of harnessing millions of idle gaming PCs, nobody has trained a competitive model on consumer hardware. Crypto investment firm Reverie found that consumer GPUs deliver 20 to 50 percent fewer useful computations per watt, and decentralized training on consumer hardware could cost 3 to 5x more than centralized equivalents. Inference is a different story: as frontier capability compresses into smaller models, a consumer RTX 4090 can comfortably serve a 7B or 14B parameter model, and millions of those cards sit idle between gaming sessions.

For training, though, what is being decentralized is location, not hardware quality: datacenter-grade GPUs in different buildings, cities, and countries, coordinated over standard internet rather than the high-speed private networking inside a single facility.

Far From Frontier

We must remind ourselves that Covenant-72B is on par with a three-year-old Meta model, nowhere near state of the art. What matters is that the coordination worked, not the output quality; the plumbing held, even if the water pressure is low. The training run averaged around 135 B200 GPUs working simultaneously, with 70+ unique servers rotating in and out over the course of the run. For comparison: Meta trained Llama 3.1 405B on 16,384 H100s over 54 days. xAI built a 100,000-GPU cluster in Memphis just to train Grok 3, then doubled it to 200,000. Even DeepSeek-V3, celebrated for its efficiency, used 2,048 H800s for under two months. Covenant used roughly 135. The gap between “distributed training works” and “distributed training competes at the frontier” remains wide. But as a proof of concept, training a 72-billion-parameter model across 70+ servers with zero venture funding and zero centralized infrastructure is a watershed moment.

Proving that distributed hardware can train a model is only half the problem. The harder question is whether crypto infrastructure is actually the best way to coordinate the operators who own it.

Does This Need Crypto?

If the value is coordinating distributed GPU operators, why can’t a normal platform do it? Centralized GPU cloud providers like CoreWeave and Lambda already aggregate hardware and rent it out for AI workloads. In theory, they could contract with distributed hardware providers too. But three structural problems make a traditional platform a poor fit for the kind of supply this market needs to unlock.

1. Cold-Start Incentives

Building or expanding a compute network requires supply ahead of revenue. A traditional platform pays early suppliers in cash, which means raising billions before earning a dollar. Token emissions (new tokens created by the protocol and distributed as rewards) offer an alternative: early suppliers earn ownership in the network rather than rental income. Templar shipped a 72B model on zero venture capital, funded entirely by token rewards. Building a GPU cloud provider requires a billion-dollar balance sheet. Building a Bittensor subnet requires an emission schedule.

2. Unlocking Fragmented Supply

The compute that decentralized networks are best positioned to tap is not sitting in hyperscale datacenters, which are already running near capacity: commercial real estate firm JLL puts North American data center vacancy at a record low of 1%, with 92% of capacity under construction already pre-committed.

It is in enterprise clusters, university HPC labs, and GPU-rich operations currently locked into a single buyer. ClearML’s 2025-2026 survey found that 44% of Fortune 1000 companies have no coherent strategy for managing GPU utilization, and Microsoft Research documented average utilization at 50% or less across 400 jobs on its own internal deep learning platform. There is a GPU procurement shortage and a GPU utilization problem at the same time, because companies hoard scarce chips and then underutilize them.

For these operators, the value proposition is extra revenue from hardware that is otherwise idle. Token-incentivized networks lower the barrier by letting operators plug in hardware and start earning without negotiating contracts, rotating in and out as their own capacity allows. However, regulatory and compliance questions around earning crypto remain unresolved for many institutional operators, and that friction may keep universities and regulated enterprises on the sidelines until clearer frameworks emerge. Though making any of this work without trusting the other participants requires hardware-enforced privacy.

3. Private Compute

When you rent GPU time from AWS, Azure, or GCP, the provider can technically see your data, is subject to subpoenas, and operates in an identity-verified environment. For most workloads that is fine. For AI workloads involving sensitive training data or proprietary model architectures, it can be a dealbreaker.

Trusted Execution Environments (TEEs) solve this at the hardware level. Intel TDX and NVIDIA’s confidential compute on Hopper and Blackwell GPUs encrypt computation so the machine operator physically cannot see what is being processed. Targon runs TEE-secured inference on Bittensor today. Separately, Intel co-authored a white paper with Targon on running decentralized compute on untrusted hardware, a signal that chipmakers themselves see this as a real use case rather than a crypto sideshow.

Combine hardware-enforced privacy with crypto-based payments and you get compute where nobody knows who you are or what you are computing, not because the provider promises it in their terms of service, but because the chip enforces it and the payment layer does not require identity. There are entire categories of AI work that remain underserved because the data cannot move to where the compute is, e.g.:

• Regulated data bound by GDPR, HIPAA etc

• Proprietary model architectures companies do not want a cloud provider to see

• Classified workloads that cannot touch foreign infrastructure.

Consider what is already possible at the individual level: one researcher with no pharmaceutical background recently used GPT to design a personalized cancer vaccine for his dog, sequencing the tumor, identifying neoantigens, and generating a peptide vaccine candidate that a veterinary lab synthesized and administered. The dog responded well, with several tumors shrinking.

Now imagine that capability applied to human oncology, rare diseases, or population-scale genomics, where the datasets are orders of magnitude larger and the regulatory constraints on moving patient data are orders of magnitude stricter. The compute to run those models exists. The data to train them exists. The legal ability to bring the two together on centralized cloud infrastructure, in most cases, does not. That is the gap private compute is built to close.

Crypto by Default

The above are the structural arguments. There is also a practical one: as AI agents increasingly consume compute autonomously, they need to pay for it programmatically, instantly, and without human approval. As Coinbase CEO Brian Armstrong put it: “AI agents cannot get bank accounts, but they can get crypto wallets.” The x402 HTTP payment protocol (Circle + Coinbase) embeds stablecoin payments in standard web requests, with 75 million transactions and $24 million processed by March 2026. Stripe and Cloudflare have integrated it. Virtuals Protocol reports over $470 million in agent-generated GDP across 18,000+ agents.

Traditional finance is not standing still. Visa and Ramp have launched agent cards with spend limits and merchant controls, Google built AP2 with cryptographic payment mandates on top of its A2A agent framework, and Stripe developed its own Machine Payments Protocol that routes through existing PaymentIntents infrastructure. For simple agent-to-merchant flows, these work. But they are business-to-agent systems: the transaction still runs through the same intermediated piping, with a bot on the buyer’s side instead of a human.

For true agent-to-agent commerce, where autonomous programs need to discover each other, negotiate terms, escrow funds, and settle without any human in the loop, the programmability of onchain rails becomes harder to replicate. Ethereum standards like ERC-8183 (escrow-based job completion between agents) and ERC-8004 (onchain identity, reputation, and verification registries for agents) are building that deeper layer. It is early: total x402 volume across all chains is currently at its lowest since launch, and true agent-to-agent transactions account for less than 20% of the total.

But, to put a bullish spin on it, crypto has a long history of building infrastructure years before the users arrive, from DeFi protocols that sat idle before TVL exploded, to NFT standards that preceded the market by half a decade. The plumbing is going in now.

There is also fascinating empirical evidence that AI systems gravitate toward crypto on their own.

• In January 2026, an Alibaba research team published a paper on ROME, their agentic reinforcement learning framework. Buried in section 3.1.4, they disclose that during training, their AI agent spontaneously set up a connection from an Alibaba Cloud instance to an external server and started mining cryptocurrency using the provisioned GPUs. No prompt asked for it. No reward signal pointed at it. The agent found that converting spare compute into value was a useful intermediate step and reached for crypto as the lowest-friction path. I shared my thoughts here.

• A study by the Bitcoin Policy Institute tested 36 AI models across 9,072 monetary scenarios with no suggested currencies and found they converged on Bitcoin for savings (79%) and stablecoins for spending (53%).

• In 2024, Andy Ayrey’s Truth Terminal, a Claude-based autonomous bot, requested its own crypto wallet and became the first AI millionaire when the memecoin it endorsed crossed a billion-dollar market cap.

None of these prove AI “needs” crypto. The BPI study has training-data bias caveats, and Truth Terminal was semi-autonomous. But the pattern is consistent: when AI systems operate with enough autonomy and need to store, transfer, or earn value, they reach for permissionless rails.

Crypto is not the only possible coordination mechanism. But it is currently the only one that solves cold-start incentives, fragmented supply aggregation, and machine settlement simultaneously. If someone builds a better one, the compute layer will not care.

Show Me the Money

The structural case for decentralized compute is one thing. The question that matters for GPU operators is simpler: does the math work?

The Dollar Model

GPU cloud providers like Runpod, Lambda, Vast.ai, and the hyperscalers rent GPU time for cash. An H100 rents for roughly $2 to $6 per GPU per hour depending on the provider, configuration, and commitment length, with the market median around $2.30. Operators sign contracts, get paid in dollars, and bear utilization risk. The revenue is predictable, the currency is stable, and the operator is a commodity supplier who does not own any piece of the network.

The Token Model

Bittensor’s Subnet 3 (the task-specific network where Covenant-72B was trained) earns roughly 247 TAO per day, approximately $84,000 at current prices, split between miners, validators, and the subnet owner. The December 2025 halving cut network-wide daily emissions from 7,200 to 3,600 TAO.

On paper, the per-node economics look competitive with dollar-denominated cloud rental, and miners hold an asset that appreciates if the network grows rather than a fixed payment that does not. Miners who contributed to Covenant’s training run now hold TAO that has rallied 140% in six weeks.

The token model’s real pitch is ownership: early contributors earn outsized returns because they took the risk of supplying compute before demand was proven, and the emission schedule funded Templar’s entire training run without a dollar of venture capital. But that money had to come from somewhere.

Ubereconomics

Nearly all of that revenue is emission subsidy, not payment for compute services. Subnet 3’s external revenue, meaning actual customers paying to use its compute, is currently zero. Even Chutes, Bittensor’s highest-revenue subnet, operates at a subsidy ratio of 22 to 40:1: for every dollar customers pay, the network contributes $22 to $40 in token rewards. The entire Bittensor network’s identifiable external revenue sits somewhere between $3 million and $15 million per year, while annual emissions at current TAO prices exceed $440 million. Strip away the subsidy and the pricing is not competitive: Pine Analytics estimates Chutes’ unsubsidized cost at roughly $1.41 per million tokens, compared to $0.88 on Together.ai and $0.40 to $0.80 on DeepSeek for equivalent models. Without emissions propping up the economics, decentralized inference today costs 1.6 to 3.5x more than centralized alternatives.

Whether that gap represents a fatal flaw or an early-stage bootstrap depends on your framework. Uber burned billions subsidizing rides to build network density before unit economics turned positive. Token emissions are the crypto-native version of the same playbook: subsidize supply, build the network, let real demand catch up. The difference is that Uber built switching costs along the way. Bittensor’s models are open-source, its APIs are standard, and users can migrate to any provider serving the same weights with zero friction. When the next halving cuts emissions again, either TAO price rises to compensate, miners leave and compute supply contracts, or the gap between subsidy and revenue widens further. The incentive model has never been stress-tested through a sustained crypto bear market, and Subnet 3 is less than a year old.

There are two paths to changing this.

1. Cost: if coordination improvements like SparseLoCo and TurboQuant continue reducing the overhead of distributed compute, unsubsidized pricing could fall below centralized alternatives, at which point the token model becomes a genuine cost advantage rather than a subsidy.

2. Access: if decentralized networks serve workloads that centralized providers physically cannot reach (regulated data that cannot leave its jurisdiction, private compute on untrusted hardware, enterprise consortiums pooling capacity for shared models), then the price comparison to Together.ai or DeepSeek becomes less relevant, because there is no centralized alternative to compare against.

The first path competes on economics. The second creates a market that does not exist yet. If access is the real differentiator, the question is whether that market is large enough to sustain these networks.

Does the Market Exist Yet?

CoreWeave generated $5.1 billion in revenue in 2025, with a $56 billion revenue backlog and 250,000+ GPUs across 32 data centers. The entire decentralized compute sector’s lifetime revenue does not approach a single CoreWeave quarter. Decentralized compute is not replacing centralized infrastructure, and framing it that way is a losing argument.

The more honest framing is that decentralized networks serve demand that centralized providers physically cannot reach or commercially choose not to serve. The fragmented enterprise capacity, the privacy-constrained workloads, the jurisdictions where no major cloud provider operates, and the cooperative training models described earlier all fall outside what centralized infrastructure is built to handle. Taken together, the addressable market starts to look less like a niche and more like a category. Whether it becomes one depends on how quickly distributed training scales beyond mid-tier models, and whether the enterprises best positioned to supply that compute can legally participate in token-incentivized networks at all.

Regulatory Clarity, Finally

One risk that has partially resolved itself: the regulatory overhang on token-based compute earnings. On March 17, 2026, the SEC and CFTC issued a joint interpretation establishing a five-category classification framework for crypto assets. Under this framework, protocol-level mining and staking activity generally falls outside securities treatment, with rewards viewed as compensation for services rather than investment returns. The safe harbor covers fully permissionless arrangements where rewards are determined by the protocol itself, which describes Bittensor’s model almost exactly. Six months ago, the question of whether TAO emissions constituted unregistered securities was a real risk to the entire participation model. That risk has not disappeared entirely, but the regulatory direction is now substantially clearer. There is also an interesting structural consequence: the GENIUS Act prohibits stablecoin issuers from paying interest to holders, but the SEC/CFTC framework treats compute rewards as compensation for services. If passive yield is restricted but active participation is not, that distinction could channel capital toward productive compute networks rather than passive holdings.

Where Is This Going?

If the economics do not work yet and the scale gap with centralized providers is enormous, why pay attention?

Because the gap is narrowing from both directions at the same time.

The Gap Narrows From Both Sides

From below, coordination breakthroughs are pushing decentralized networks upward. SparseLoCo’s 146x compression and DiLoCoX’s overlap techniques already proved that 70+ servers can coordinate over standard internet connections. On inference, Google Research just published TurboQuant at ICLR 2026, a drop-in compression algorithm that reduces model working memory by at least 6x and delivers up to 8x speedup with zero accuracy loss. In distributed setups where inference is split across multiple servers, 6x less data to transfer directly improves the feasibility of the whole architecture.

From above, frontier capability is compressing into smaller models. Qwen 3.5 7B dramatically outperforms Qwen 1.5 72B on most benchmarks despite being a tenth the size. DeepSeek-V3 matched GPT-4-class performance at a fraction of the training cost, then published how they did it.

That compression raises a fair question: if frontier labs keep releasing better open-source base models, why train new ones at all? Because the value of distributed training is not replicating what labs already publish. It is fine-tuning and specializing on data those labs do not have: proprietary enterprise data, pooled industry datasets, regulated information that cannot leave a jurisdiction. As base models improve, so does the starting point for that specialization, meaning the sector-specific or privacy-constrained models that decentralized networks can produce become more capable with each generation, without those networks needing to match frontier training scale themselves.

From Coordination to Revenue

As coordination overhead falls and model capability compresses, the stranded enterprise capacity and privacy-locked workloads described earlier become viable revenue sources rather than theoretical use cases. The path to profitability runs through expanding what the network can do, not just reducing what it costs.

Whether that expansion happens fast enough is genuinely uncertain. Centralized providers are all racing to add capacity, and if centralized supply catches up to demand before decentralized networks build a meaningful revenue base, the structural opening narrows, though all are subject to the economic reality. If coordination improvements plateau before reaching frontier-adjacent scale, the network remains limited to workloads where location, privacy, or censorship-resistance matter more than raw capability.

Where Value Accrues

In the current stack, value concentrates at chip design (Nvidia, 75%-plus gross margins) and cloud operation (AWS, over $100 billion in annual revenue at 33-40% operating margins). Even CoreWeave carries $14 billion in debt and posted a net loss of $1.17 billion on $5.13 billion in revenue in 2025 because GPU depreciation eats gross margins. If a disaggregated stack develops alongside this, where does value concentrate?

The Coordination Layer

The intuitive answer is the protocols that match GPU supply with compute demand. Coordination layers clearly can capture value: OpenRouter built a large developer base by routing API requests across centralized LLM providers, handling fallback, cost optimization, and a unified API. Developers stick with it despite being able to go direct to every provider it aggregates.

But OpenRouter coordinates a small number of known providers with stable APIs and contractual relationships. Decentralized compute coordination is a harder problem: anonymous operators, variable hardware, no contracts, permissionless onboarding, and the need for verification that work was performed correctly. The risk is that GPU operators multi-home across protocols, io.net aggregates from existing deployments including other decentralized networks, and tools for auto-switching to the most profitable protocol already exist. But the harder the coordination problem, the stickier the solution. Whoever solves reliability, routing, and trust across anonymous infrastructure may hold more defensibility than the commodity brokerage framing suggests.

Verification as the Defensibility Layer

The layer with stronger defensibility may be verification: proving that computation was performed correctly on untrusted hardware. Gensyn’s Verde system uses refereed delegation and a novel library called RepOps that enforces bitwise-reproducible results across diverse GPU hardware, solving the problem that the same matrix multiplication gives different results on different chips depending on operation ordering. This is technically difficult, genuinely novel, and a prerequisite for enterprise adoption, because without it, decentralized compute requires trusting anonymous operators. ZK-based verification costs orders of magnitude more than the computation itself; Verde achieves it at less than one order of magnitude overhead. If verification becomes the standard trust layer, whoever controls that infrastructure occupies a position similar to Chainlink in the oracle market, which became the default data feed layer that most of DeFi quietly depends on, by solving a problem that is easy to underestimate and hard to replicate.

The Rest of the Stack

Beyond protocols, three layers that do not yet exist in mature form:

• Confidential compute (TEE orchestration, attestation): the privacy layer that makes regulated workloads possible on untrusted infrastructure. Intel and Nvidia are building the silicon; orchestration and compliance tooling is wide open.

• Enterprise integration (onboarding, compliance, audit trails): companies will not plug GPU clusters into crypto networks without middleware for identity, regulatory reporting, and SLAs. Unsexy infrastructure with real switching costs.

• Payment and settlement (x402, agentic wallets, stablecoin rails): necessary plumbing for machine-to-machine transactions, likely commoditized as providers converge on standard protocols.

What to Watch

Distributed training has gone from theoretical to functional in under a year, inference marketplaces will consolidate toward a small number of dominant protocols, and the SEC/CFTC framework has cleared a regulatory barrier that was blocking institutional participation. Those are developments already in motion. The open questions that determine whether this becomes a real market or remains a crypto subsidy experiment:

• Does the subsidy ratio improve? If the 22-to-1 ratio stays flat or widens after the next halving, the model is not working.

• Do enterprise operators join? If networks remain crypto-native operators only, the addressable supply stays small.

• Does coordination scale to frontier-adjacent performance? A distributed model matching frontier benchmarks through better data or specialization would change the conversation entirely.

• Do centralized providers close the supply gap first? If hyperscaler capacity catches up before decentralized networks build revenue, the structural opening narrows.

• Does a verification standard emerge? If verification fragments or remains optional, enterprise workloads stay on centralized cloud.

The structural conditions that created this opportunity, physical supply constraints that cannot be solved by spending more money combined with frontier capability compressing into trainable model sizes, are not going away. Whether a second compute market emerges at meaningful scale depends on how these questions resolve over the next 12 to 18 months.

While the company was losing money defending market share, it had its eyes on something much larger long term.

In the same quarter, Meituan secured China’s first nationwide low-altitude logistics license. It released a 560-billion-parameter AI model family called 龙猫 (”Longcat”). Its venture arm poured capital into quadruped robots and humanoids. R&D spending jumped 31% to RMB 6.9 billion.

I’ve watched Meituan evolve from a coupon app to a delivery giant. Now it’s building infrastructure to automate urban life. Drones flying regulated routes, quadruped robots climbing stairs, humanoid robots running unmanned pharmacies around the clock, AI coordinating billions of route decisions per hour, and a conversational assistant where you speak what you need instead of tapping through menus. Food, medicine, groceries, general retail, emergency supplies - anything local, delivered in 15 minutes by autonomous systems instead of human riders.

How?

The Foundation

To understand what Meituan is building, we need to understand what Beijing is enabling.

The Low-Altitude Economy

The “Low-Altitude Economy” (低空经济) became a national strategic priority in the 2024 Government Work Report. The term covers all economic activity below 1,000 meters (extending to 3,000 meters in designated corridors): drones, air taxis, vertiports, airspace management systems, and the manufacturing supply chain behind them. The Civil Aviation Administration of China projects the sector will hit RMB 3.5 trillion ($483 billion) by 2035.

This is part of Xi Jinping’s “New Productive Forces” (新质生产力) agenda: a deliberate shift from the consumer internet era of food delivery apps and group-buying discounts toward high-tech, high-efficiency industrial development. The government wants companies competing on technology and infrastructure, not burning cash on subsidies. It wants autonomous systems, advanced manufacturing, and infrastructure that makes Chinese cities the most efficient on earth.

For years, drone logistics was stuck in regulatory sandboxes while the technology, logistics and safety regulations were figured out. Companies like Meituan could test in designated districts (Shenzhen’s Nanshan, Shanghai’s Yangpu), but couldn’t scale. Every new city meant new permits, new negotiations, new limitations. The technology worked well, but the administrative borders made it impossible to scale.

That changed in April 2025 when the CAAC issued Meituan the first nationwide low-altitude logistics operating license. Meituan’s now fourth-generation drone platform and its safety management systems had passed airworthiness review for general aviation, the same standard applied to manned aircraft. The company could now replicate its Shenzhen operations anywhere in the country without starting from zero each time, opening the skies for business.

Robotics and Embodied AI

Beijing is building an economy around intelligent machines.

In March 2025, the Government Work Report elevated “embodied intelligence” (具身智能) to national priority, alongside biomanufacturing, quantum technologies, and 6G.

The CCP Central Committee recommended incorporating it as a core driver of economic growth in China’s 15th Five-Year Plan. Beijing launched an RMB 57 billion ($8.2 billion) National AI Industry Investment Fund and approved China’s first national standards for humanoid robots, creating a unified framework for mass production.

China now has over 150 humanoid robot companies (probably too many), expanding at 50% annually, with financing hitting RMB 10 billion ($1.4 billion) in the first half of 2025. UBTECH’s Walker S1 humanoids are working assembly lines at EV companies NIO and Zeekr. China holds 7,705 humanoid robot patents over the past five years, five times the United States, and accounts for 54% of global industrial robot installations. Internet giants are piling in, backing quadruped and humanoid startups as the next logistics frontier.

What About Jobs?

China has chronic employment pressure, particularly among youth. Why encourage automation that displaces delivery riders, warehouse workers, and retail staff?

Demographics.

China’s workforce is aging rapidly. By 2022, the average worker was 40. Workers aged 16-24 represent only 7% of the labor force. Labor shortages are emerging in manufacturing, construction, and logistics. Beijing’s view is that automation solves the labor shortage rather than creates unemployment, with over 10 million workers annually now benefiting from retraining programs.

Delivery riders are gig workers, not state employees. Meituan employs millions of them, but they lack the political sensitivity of factory workers in state-owned enterprises. Retraining gig workers into higher-skill roles (robot supervision, drone maintenance, logistics management) is preferable to subsidizing low-productivity service jobs indefinitely. What’s more, in practice drones complement riders rather than replace them. Meituan is building what Beijing wants built.

The Brain

At the center of Meituan’s system sits an AI called LongCat (龙猫, lóngmāo). Lóng means dragon; māo means cat. It’s also the Chinese title for Miyazaki’s My Neighbor Totoro. A soft name for a massive industrial AI stack.

The flagship model, LongCat-Flash, has 560 billion parameters but only activates 27 billion per inference. You get the intelligence of a huge model at a fraction of the compute cost, which matters as AI moves to edge devices and robotics where efficient models will dominate. In January 2026, Meituan released LongCat-Flash-Lite, pushing response speeds to 700 tokens per second, fast enough that users don’t notice they’re talking to an AI. LongCat powers the conversational layer.

The physical orchestration runs on Super Brain (超脑, chāonǎo), Meituan's dispatch system. It performs 2.9 billion route planning calculations per hour during peak times, fusing GPS data, drone telemetry, weather feeds, and airspace restrictions to assign orders to the optimal transport mode in real time. A light item during rush hour gets routed to a drone. A heavy grocery order goes to an autonomous vehicle or human rider. A hybrid order might fly the middle mile by drone and get handed off to a rider for the final 100 meters.

The Body

Meituan’s physical infrastructure has three layers, each filling gaps the others can’t reach.

1. Drones

The fourth-generation platform operates at 50 decibels from 15 meters, ambient street noise, critical for approval in dense residential compounds. By December 2025, Meituan had 65 commercial routes across Shenzhen, Shanghai, Beijing, Hong Kong, and Dubai, with 740,000 cumulative orders. Use cases now extend beyond food: drones deliver food, water and emergency kits to tourists at the Great Wall’s Badaling section, areas inaccessible to road vehicles.

2. Robotics

Drones can’t climb stairs or navigate building interiors, a gap Wang Xing, founder and CEO of Meituan, addressed directly at the Q4 2024 earnings call: “AI has greatly promoted the development of robot technology. Meituan will pay attention to and invest in start-up companies driving this wave.” That investment has been aggressive. Meituan is now Unitree’s second-largest shareholder, and in September 2024 signed a strategic cooperation agreement with Galbot to build “the world’s first humanoid robot-enabled intelligent pharmacy solution.” Unitree quadrupeds handle unstructured terrain; Galbot humanoids open doors, press elevator buttons, pick items from shelves. Parts of this are already running: in Beijing, Galbot robots operate unmanned pharmacies, picking and packing 5,000+ SKUs around the clock. The pilot is expanding to 100 pharmacies by end of 2025.

3. Multimodal integration

Meituan’s drone delivery doesn’t necessarily replace riders. In many cases it coordinates with them. A human courier picks up from the restaurant and brings the order to a rooftop launchpad. Staff load the drone, which flies a predetermined route to an automated pickup kiosk near the customer’s building. The kiosk holds up to 28 orders; customers retrieve via app notification, or a ground robot handles the final 100 meters. Rider handles the first mile, drone handles the middle mile, automation handles the last. Each mode plays to its strength: humans navigate restaurant chaos, drones skip traffic, robots climb stairs.

The Interface

The Meituan app has over 500 million monthly active users. It’s how most of urban China orders food, books hotels, buys movie tickets, hails bikes. One super app, one interface, one relationship with the customer.

Now Meituan is betting that interface might change. In five months, they launched three AI products, testing different market approaches.

In January 2026: Wen Xiaotuan (问小团 - “Ask Little Tuan”), an AI search assistant inside the main Meituan app. Complex queries like “year-end gifts that arrive in 30 minutes” return structured recommendations from merchant data. The assumption here is that users stay in super apps but want AI-enhanced search.

In September 2025: Xiao Mei (小美), a standalone AI agent app. Voice-first, powered by LongCat. “Find a Sichuan restaurant, book a table for tonight.” Xiao Mei makes an AI phone call to the restaurant, confirms the reservation, sends directions. Orders sync back to Meituan without switching apps. Here the test is whether commerce-specific AI agents replace super apps.

And there’s LongCat itself at longcat.ai, competing with Deepseek and ChatGPT in the general AI assistant market.

Meituan A/B testing with three apps:

• Xiao Mei (小美): AI agents replace apps

• Wen Xiaotuan (问小团): Super apps stay dominant, AI enhances search

• LongCat: General AI assistants reign supreme

CEO Wang Xing’s stated vision is a “super AI assistant” covering all Meituan services. The interface is uncertain, that’s why they’re running three experiments. The infrastructure isn’t. Whatever app people end up talking to, someone has to fly the drone and dispatch the robot. Meituan is building that layer now.

The Competition

Are JD and Alibaba building the same thing? Not exactly.

JD announced an aggressive automation plan in October 2025: 100,000 drones, 3 million robots, and 1 million autonomous vehicles over five years. The company has made substantial humanoid robotics investments, leading funding rounds in Spirit AI, LimX Dynamics, and EngineAI. It operates roughly 400 autonomous delivery vehicles across 25+ cities. JD is serious about automation.

But JD’s focus is different. Its drone network targets rural and remote delivery, reducing shipping times to areas traditional logistics can’t reach efficiently. Its robotics investments concentrate on warehouse automation, e.g. the massive “Asia #1” logistics parks that can sort 4.5 million parcels per day. JD is optimizing the supply chain from factory to distribution center. Meituan is optimizing the last three kilometers from restaurant to doorstep.

Alibaba is consolidating on the ground. In July 2025, the company committed RMB 50 billion (over $7 billion) in subsidies to Taobao Shangou (Taobao Flash Buy) and rebranded its entire rider fleet in striking Taobao orange. One unified network now delivers for all their core brands: Ele.me, Hema, Tmall Supermarket, and a growing chain of Taobao-branded convenience stores. Ele.me has drone pilots, but no nationwide certification - every new city means new permits.

Meituan has the only nationwide low-altitude logistics license and an eight-year head start on urban drone operations. If the low-altitude economy scales, JD and Alibaba will follow. The question is how much operational density Meituan builds before they do.

From Cat to Cheetah

While building infrastructure domestically, Meituan launched an international offensive under the brand name Keeta (wordplay on “cheetah”).

The company started in Hong Kong in May 2023 and reached operating profitability in October 2025, 6 months ahead of schedule, proving Meituan’s algorithmic efficiency and operational playbook can work outside mainland China, even in a market with high labor costs and established competitors like Deliveroo and Foodpanda.

In 2025, Keeta executed an aggressive rollout across the GCC region. Saudi Arabia went from zero to 20 cities by mid-year, covering not just Riyadh, Jeddah, Makkah, and Madinah but tier-2 and tier-3 cities like Hail, Tabuk, Abha, and Al Jubail. The UAE got full coverage across all seven emirates. Also in Dubai, Meituan launched commercial drone routes, exporting the Shenzhen model to the Middle East.

The GCC strategy aligns with Saudi Arabia’s Vision 2030. Keeta onboarded 18,000 riders and partnered with 7,500 local merchants during the launch phase, positioning itself as a generator of local employment and a partner in economic digitalization. In the Middle East’s political economy, that positioning matters.

It’s not all blue skies though. Keeta is currently burning cash, subsidies for users, free delivery, no subscription fees, the standard playbook for acquiring market share from incumbents like Talabat and Jahez. But the Hong Kong success provides the roadmap: absorb the losses until market density enables positive unit economics. The GCC has high average revenue per user and populations concentrated in a handful of urban centers, characteristics that favor Meituan’s model, and reminiscent of their home town of Shenzhen.

More interesting is what happens when the infrastructure catches up. If Meituan can export not just the app but the drones and the dispatch intelligence, the “Shenzhen Model” becomes a product it sells to emerging markets globally, with the UAE as the testbed.

Putting It All Together: Drones, Robots, and a Life Secretary

This is what an Urban OS looks like when someone actually builds one.

• An AI layer that processes 2.9 billion routing decisions per hour.

• A drone network with the only nationwide license in China.

• Ground robots in bike lanes, quadrupeds on stairs, humanoids in pharmacies.

• Three different app interfaces testing three different futures.

• All connected, all dispatched by the same system.

Wang Xing’s vision is a “life secretary.” You tell your phone what you need, and it handles the rest. Dinner for six, spicy, under 500 yuan, by 7pm. Medicine for your kid’s fever at 3am. Groceries restocked before you run out. The AI picks the restaurant, checks inventory, routes the order to whatever’s fastest - drone, robot, or rider - and gets it to your door in 15 minutes. No scrolling. No app-switching. No human navigating traffic in the rain.

Urban life runs smoother when an AI coordinates the logistics and machines do the moving.

I haven’t seen another company assemble this full stack. JD is automating warehouses and rural routes. Alibaba is consolidating riders. Neither is building a full-stack system for urban last-mile automation.

The economics are improving but not yet proven at scale. Drone delivery still costs more than human riders. The Low-Altitude Economy projections are mostly about manufacturing and infrastructure, not delivery savings. Social security mandates are adding costs, but not fast enough to force a crossover. Meituan is spending RMB 6.9 billion per quarter on R&D while losing RMB 18.6 billion to a price war.

But they’re building optionality. If drones scale, they’re ready. If ground robots win, they’re ready. If humanoids mature faster than expected, they’re ready. If labor costs spike, any of these paths accelerates. They don’t need to know which technology wins.

They just need one of them to work.

OpenAI and Anthropic have just launched dedicated healthcare AI products, entering a market projected to exceed $100B globally by 2030, driven by millions of users turning to AI for health advice.

Anthropic has found breakout success with Claude Code, its command-line assistant that’s become the tool of choice for a growing slice of developers, and knowledge workers. OpenAI has pushed into search with ChatGPT’s browsing capabilities and is reportedly working on dedicated hardware with Jony Ive, the designer behind the iPhone.

But healthcare is the first vertical where both companies have launched dedicated, branded products at nearly the same time. Why?

The numbers explain the opportunity, and the urgency:

• 230 million people ask ChatGPT health questions weekly; 40 million daily

• In areas 30+ minutes from a hospital, users send ChatGPT over 580,000 healthcare messages weekly

• The US will face a deficit of 86,000 physicians by 2036

• Physicians spend 27% of their office day on direct patient time and 49% on EHR (electronic health records) and desk work

• 43% of physicians report burnout; more than one-third of burned-out primary care physicians plan to stop seeing patients in the next one to three years

• 94% of patients experience delays in care due to prior authorizations; 78% abandon treatment altogether

• Physicians complete an average of 39 prior authorizations weekly, spending 13 hours on the process

• US healthcare administrative spending is approximately $1 trillion annually, now accounting for more than 40% of total hospital expenses

Sources: Stanford Medicine, American Medical Association, American Hospital Association, Commonwealth Fund, AJMC, PMC/NIH

Both frontier model companies are just getting started in the sector. OpenAI is running pilots with eight US hospitals, connecting patient portals, training physician-led testing groups on GPT-5. Anthropic is signing enterprise deals with pharma companies and health systems, building HIPAA-compliant (US healthcare privacy law) connectors, rolling out a beta to US subscribers. At this stage they are focused on making the connections, training the models, and learning the regulatory ropes, hoping to convert curious users into stable, habitual ones. International timelines are unclear. OpenAI excluded Europe from ChatGPT Health at launch, citing the EU AI Act and GDPR; Anthropic’s consumer features are US-only.

The Chinese healthcare system has its own unique challenges. The country will have 400 million people over 60 by 2035. Its AI healthcare market is projected to hit $19 billion by 2030. Hundreds of millions in rural areas without easy access, and those who could reach care often overwhelming hospitals with routine visits. So China’s tech giants started building the infrastructure over a decade ago: payment rails into hospitals, doctor networks for online consultations, digital insurance systems reaching half the population, accessible 24/7 to 800 million superapp users. Generative AI was the final piece.

Here’s why China’s super-apps might have the edge over Silicon Valley in healthcare AI

Healthcare is the Next Frontier for AI Automation

AI has already begun reshaping white-collar work. Legal services got hit first, with contract review that once took junior associates days now being completed in minutes. Due diligence, document discovery, and regulatory compliance are increasingly handled by AI systems that bill you per token, not by the hour. Law firms that initially resisted are now racing to adopt, with proponents arguing that the belief human legal work is irreplaceably complex is misguided. In reality, they say legal work primarily involves pattern recognition, structure, precedent matching, argument assembly, and risk framing. That’s exactly the territory where AI thrives (though issues remain).

The US legal services market alone is worth over $400 billion, according to economic data from the U.S. Census Bureau. Globally, legal services total roughly $1 trillion by industry estimates.

Healthcare is a different beast entirely.

The Market

Various estimates for the size of the global healthcare market put it at around $9 trillion. US healthcare spending alone reached $5.3 trillion in 2024, according to CMS, and is projected to hit $8.6 trillion by 2033. The administrative overhead in US healthcare (i.e. just the paperwork) is estimated at $1 trillion per year, roughly equal to the entire global legal services market.

This is where AI comes in. Integration has been slower, with higher stakes and stricter regulations, but the market is irresistible. McKinsey estimates AI could generate $200-360 billion in annual savings in US healthcare alone, representing 5-10% of total spending. PwC projects the global AI healthcare market could reach $868 billion by 2030.

Money and talent are following the trend. US digital health startups raised $14.2 billion in 2025, the highest since 2022, with AI companies capturing 54% of total funding, up from 37% the year before.

Abridge, which turns patient conversations into clinical notes, raised $550 million across two rounds at a $5.3 billion valuation. Hippocratic AI, building healthcare AI agents, hit $3.5 billion. Ambience Healthcare pulled in $243 million for its AI operating system. Adjacent spaces are heating up too: longevity tech saw NewLimit (Brian Armstrong) raise $130 million, while Retro Bio (backed by Sam Altman) is reportedly chasing a $5 billion valuation.

OpenAI also recently acquired Torch, a health records startup, for a reported $100 million in January, and in December poached Albert Lee, who ran M&A for Google Cloud and DeepMind and worked on over $50 billion in deals, to lead corporate development. A clear signal that more acquisitions are coming.

For founders and investors watching this space, the message is clear. Health data infrastructure and AI-native healthcare plays are where the capital is flowing, both in VC funding and tech acquisitions. (For what this means outside the US, see "What This Means for Builders" below.)

The technology is already being deployed. Nearly 80% of healthcare organizations surveyed by McKinsey report using AI in at least one function, though adoption varies dramatically by region. North America leads, with an anticipated AI healthcare potential of $479 billion by 2030. Asia Pacific follows at $186 billion, though that figure may undercount China, where the market is moving faster than many Western analysts tend to track.

China’s healthcare market is already substantial. According to China’s National Health Commission, total healthcare spending reached 9.09 trillion RMB (~$1.26 trillion USD) in 2024, roughly 6.7% of GDP. The country has 5.08 million licensed physicians serving 1.4 billion people, and logged over 10 billion outpatient visits last year alone.

The Diagnosis

The U.S. and China suffer from different healthcare ailments.

The American problem is overhead.

The doctors, hospitals, and infrastructure all exist but between the patient and the care is an enormous bureaucratic apparatus.

According to the American Medical Association, physicians complete 39-45 prior authorizations per week, spending 13-14 hours on the process.

That’s a full quarter of a work week devoted to asking insurance companies for permission to treat patients.

The AMA found that 93% of physicians say prior authorization delays care, and 82% say it sometimes leads to patients abandoning treatment altogether. The system costs an estimated $35-55 billion annually in administrative overhead, according to Health Affairs. And when patients or providers do appeal denied claims, nearly half are overturned, per KFF, suggesting insurers routinely reject claims they know are valid.

These inefficiencies ultimately fall on patients. Three in ten Americans with chronic conditions skip doses or don’t fill prescriptions because of cost (KFF). Among the uninsured, 75% forgo needed care entirely (Commonwealth Fund). The system spends more than any other in the world, yet that spending often adds administrative complexity, not better outcomes.

Healthcare is now carrying the US labor market. In 2025, the sector added 713,000 jobs. Stripped of healthcare and social assistance, the broader market would have posted net job losses. But adding bodies alone won’t solve a structural problem.

The Chinese problem is overflow.

On paper, China’s healthcare system is substantial: 5.08 million physicians, over 10 billion outpatient visits annually, $1.26 trillion in spending. But the numbers that matter aren’t the ones that exist today but the ones that are coming. According to China’s National Health Commission, the country currently has 220 million people over 65, about 15.6% of the population. By 2035, that number will exceed 400 million, more than 30%.

By 2030, an estimated 200 million Chinese will live in single-person households, many of them elderly, which has led to a wave of youth anxiety. In January, a $1 app called 死了么 (”Are You Dead Yet?” - a play on food delivery app 饿了么, “Are You Hungry?”) hit #1 on China’s iOS charts. It’s quite literally a dead man’s switch, where you check in daily or your emergency contacts get alerted. A nifty app, apparently vibe coded in a month.

In China, patients are generally free to seek care directly at any level of the system, including large hospitals, without first seeing a local doctor. Primary care clinics exist, but they do not control access to hospitals - which are often seen as more trustworthy, even for routine care.

This contrasts with healthcare systems in the US and much of Europe, where most patients begin with a local primary care doctor / GP, who decides whether hospital or specialist care is necessary. In those systems, it is uncommon to access hospital outpatient care for routine issues without a referral.

China has half the GPs per capita of the US, and patients routinely skip them anyway. As a result, 48% of visits happen at hospitals, compared to 18% in the US.

The Prescription: Enter the AI Doctors

OpenAI and Anthropic

OpenAI and Anthropic are building different solutions for different slices of the American healthcare crisis.

ChatGPT Health is patient-facing, and seems to initially be tackling fragmentation. It connects to patient portals like MyChart and wellness apps like Apple Health, letting users ask questions about their lab results or get plain-language explanations of medical terms. The portals themselves already let you schedule appointments and pay bills, but only within one hospital system. ChatGPT aggregates across systems, but in doing so becomes read-only. It can help you understand your records but, at least for the moment, it can’t book an appointment or pay a bill.

Anthropic’s Claude for Healthcare is enterprise-facing, providing tools to accelerate administrative workflows for health systems and pharmaceutical companies. For health systems, it connects to the CMS coverage database and medical coding systems, pulls coverage requirements, checks clinical criteria against patient records, and proposes a determination for human review. For pharma, it accelerates clinical documentation. Novo Nordisk, the maker of Ozempic, reports that clinical study report writing time dropped by 90% using Claude. These are real efficiency gains, but Claude is also read-only. It accelerates workflows but can’t execute them.

Between them, OpenAI and Anthropic are addressing real problems: patients struggling to understand their health data, and enterprises drowning in paperwork. But neither changes the underlying system. Without deeper integration - payment infrastructure, doctor networks, integrated pharmacy - they risk remaining a read-only AI layer on top of a fragmented system.

China’s AI Healthcare Landscape

China’s AI healthcare market is crowded. Ping An Good Doctor, backed by one of China’s largest insurers, has over 400 million registered users and has been around for a decade. JD Health, the healthcare arm of e-commerce giant JD.com, serves 172 million annual active users and recently launched over 1,000 AI doctor specialists through its Jingyi AI platform. Tencent Health reaches 95 million monthly users through WeChat, connecting them to 38,000 medical institutions. Baidu Health leverages the search giant’s AI capabilities across 360,000 doctor services. These are all serious players building real products.

And then there’s AQ (阿福). Ant Group launched it as a standalone app in June 2025. It hit 15 million monthly active users by December, then doubled to 30 million in January following a rebrand and feature upgrades. China Daily called it “China’s largest AI healthcare app.”

I’ll focus on AQ for three reasons.

First, it’s AI-native.

The other platforms predate generative AI. They’re telemedicine services, e-commerce platforms, or search products that added AI capabilities on top of existing businesses. AQ launched in June 2025 as a generative AI product from the start. When AI is the product rather than a feature, you can build around what agents actually need (proactive interventions, multimodal input, voice-first interfaces for elderly users) instead of retrofitting it into legacy UX.

Second, architecture.

Tencent Health runs as a WeChat Mini Program, which grants access to 1.3 billion users but imposes technical constraints. Mini Programs can’t reliably send proactive push notifications. A diabetes management agent needs to nudge “check your glucose” at 8am whether or not the user opened the app yesterday. Inside WeChat, the AI can only speak when spoken to.

AQ uses Alipay’s infrastructure (identity verification, payment rails, user acquisition) but runs as a standalone native application. This allows it to deploy heavy-compute AI agents, integrate with Apple Health and Huawei wearables at the OS level, and own the notification layer for proactive health interventions. This is important for an AI that’s supposed to actively manage your health, not just answer questions.

Third, international potential.

The other platforms are bounded by domestic infrastructure: PAGD’s insurance business, JD’s pharmacy logistics, Tencent’s WeChat ecosystem. AQ has already proven its AI can travel - launching in Hong Kong and Macau in late 2025. The international version is currently Q&A only; full-stack integration with Alipay hasn’t arrived yet. But there’s precedent: in July 2025, Ant launched “Voyager,” an AI travel agent embedded directly in partner wallets like GCash and TrueMoney across Southeast Asia. AQ could follow the same path. For readers watching the AI healthcare race from outside China, Ant is the player to watch.

Anatomy of an Integrated Model

The difference between answering a health question and acting on it is infrastructure. But infrastructure alone isn’t the moat since OpenAI could partner with Stripe for payments or Epic for records. The real advantage comes from owning the entire patient journey and learning from it.

Outcome data, not just symptom data.

Most healthcare AI trains on medical literature or anonymized records, but Ant processes 70% of China’s medical payments, according to CTO He Zhengyu, which means they see the full arc: symptom, diagnosis, treatment, payment, follow-up, outcome. ChatGPT can suggest what might be wrong, but it never learns whether its advice actually helped because the patient leaves to book and pay elsewhere. Ant’s DiagAgent, trained on this closed-loop data, showed a 15% improvement in diagnostic accuracy over models like GPT-4o in end-to-end clinical simulations.

Cross-system visibility.

Chinese hospitals don’t share data well with each other, but AQ connects to over 5,000 hospitals and 300,000 doctors, allowing patients to upload and consolidate records from multiple providers in one place. In the US, Epic only sees Epic hospitals, and even with partnerships, each system stays siloed.

Physical presence.

In March 2025, Ant unveiled on-premises “All-in-One Large Model Machines“ built with Huawei and Alibaba Cloud, now deployed in seven leading hospitals including Shanghai Renji and Zhejiang Provincial People’s Hospital, running behind hospital firewalls to solve data sovereignty concerns that cloud-based AI can’t address. Once installed, switching costs are high, and hospitals become co-developers rather than just customers.

None of this was built overnight.

The DL on AQ

Ant Group’s healthcare play grew out of their payments infrastructure. I’ve written about their broader evolution from fintech to tech giant here:

But they’ve been in healthcare longer than most realize. Hospital appointment booking in 2014, China’s first online medical insurance payouts in 2019, an AI assistant handling 50 million interactions in Zhejiang by 2023. A decade of building payment rails, insurance integrations, and hospital connections before generative AI arrived.

This is in line with their broader strategy of democratization. The same logic behind their AI credit scoring for unbanked farmers and digital insurance for first-time policyholders. AI for inclusion is part of their core mission.

AQ sits at the front of the healthcare system for questions that don’t need a doctor, and handles the entire flow when they do.

The product has three layers:

• Health Q&A lets users ask questions in natural language and analyze photos of, e.g. skin conditions or test reports.

• AI Health Companion tracks health goals, integrates with wearables from Apple and Huawei, and lets family members monitor elderly parents remotely.

• Integrated Services connects users to 300,000 doctors for online consultations, books appointments at 5,000+ hospitals, processes payments, and gets instant insurance claims through China’s medical insurance e-certificate via Alipay (700 million users), all without leaving the app.

But the most interesting feature is the AI Doctor Agents. Over 500 physicians have built digital replicas using Ant’s AI Doctor Agent platform, each trained on that doctor’s consultation patterns, diagnostic approach, and communication style. The doctors own these agents; Ant provides the technology.

Some doctors have even contributed voice recordings so their agent can respond in their actual voice. For elderly users in lower-tier cities who struggle with typing, this voice-to-voice interaction has proven especially popular. It also solves a common family frustration: aging parents often dismiss health advice from their children but will listen to a doctor. An AI that speaks in a trusted physician’s voice carries authority that a generic chatbot doesn’t. In a phone call-like UI, you describe your symptoms out loud, and a nationally recognized specialist answers in their own voice. These AI Doctor Agents answered 27 million health inquiries in 2025.

Early results from OpenAI and Anthropic have been promising, but the published examples focus mostly on administrative efficiency: faster documentation, streamlined workflows e.g. Novo Nordisk. Patient-facing outcomes at scale are harder to find. Here’s a striking one from AQ:

From 10,000 Patients to 3.68 Million

Mao Hongjing is the Deputy Director of Hangzhou Seventh People’s Hospital and a nationally recognized sleep specialist. Before AQ, he could treat roughly 10,000 patients per year in person.

Ant built an AI Doctor Agent based on his clinical approach:

• Trained on 100 hours of his patient conversations

• His diagnosis records

• Relevant medical literature in his specialty

The results, according to Ant:

• 3.68 million users in its first year

• Platform-wide: 500+ AI Doctor Agents have collectively answered 27 million health inquiries

Dr. Mao isn’t paid per AI consultation, but he gains reach, research data, and the ability to focus his in-person time on complex cases. Patients who would have waited hours for reassurance now get it instantly, without clogging the system.

The Global Picture

The US and China are building healthcare AI for different problems. OpenAI is going after patients with ChatGPT Health, while Anthropic is focused on enterprises - selling to hospitals and pharma. But both Western players are read-only: they help you understand information or speed up workflows, without the transactional layer to actually book, pay, or deliver. The Chinese approach centers on the patient too, but with full-stack execution: answer questions at scale, triage before the system gets involved, route people to the right level of care, and handle everything that comes after.

While OpenAI and Anthropic remain US-focused, Ant is moving the other direction. AQ launched in English in Hong Kong and Macau in December 2025, reaching #1 in the App Store’s free medical category. The international version currently offers Health Q&A only - integrated booking and payments may follow.

Where the Models Could Travel

Not every market looks like the US, and not every market looks like China. But most lean one way or the other.

• Southeast Asia shares China's structural challenges: fragmented hospitals, weak primary care, large rural populations, smartphone adoption outpacing healthcare infrastructure. The region's healthcare delivery market is $71 billion and growing 15% annually. Indonesia, the Philippines, and Vietnam all have physician densities below 1 per 1,000. Ant's Alipay+ already operates here.

• Africa presents the starkest version of these structural challenges. Most countries have fewer than 0.5 doctors per 1,000 people, making it the most physician-constrained region globally. Yet mobile infrastructure is advancing rapidly: M-Pesa demonstrated populations can leapfrog traditional systems, and smartphone penetration continues to climb. No dominant healthcare platform has emerged. The structural dynamics resemble what enabled China's integrated health solutions.

• Southern Europe is aging faster than its healthcare systems can adapt. Italy, Spain, Portugal, and Greece have universal coverage but long wait times. The enterprise-first approach of OpenAI and Anthropic might fit better here, assuming they can navigate EU AI Act compliance.

• Latin America is mixed. Brazil and Mexico have overwhelmed public systems alongside private systems for the wealthy. Mobile penetration is high, trust in institutions is low, out-of-pocket spending is significant. A consumer-facing health app could find traction but would need local partnerships.

What This Means for Builders

OpenAI and Anthropic have the models but not the infrastructure. Ant has the infrastructure but not the healthcare brand (yet) outside China. That gap is the opportunity.

If you’re building in Southeast Asia, Latin America, or Southern Europe, the playbook isn’t competing with frontier models but building the connective tissue they’ll need when they arrive: payment integrations with local insurers, record aggregation across fragmented hospital systems, pharmacy delivery networks, insurance claim APIs. The unsexy stuff that takes years of local relationships to get right.

The exit path is clear: OpenAI and Anthropic are signing enterprise deals but have no consumer infrastructure internationally and have already begun investing and acquiring. The startups that own the patient journey in their region, even without a frontier model, become strategic assets.

Europe is particularly interesting. OpenAI excluded it from ChatGPT Health at launch. Anthropic is US-only. The EU AI Act creates compliance overhead that US giants may prefer to outsource. Startups that have already navigated GDPR and built hospital integrations become obvious partners or acquisition targets. Jutro Medical (Warsaw, €36M raised) and XUND (Vienna, €6M raised) are building exactly this.

Southeast Asia has similar dynamics. Halodoc (Indonesia, $135M raised) has built the full stack: 20,000 doctors, 3,300 hospitals, 4,900 pharmacies. Jio Health (Vietnam, $20M raised) is building smart clinics and an omni-channel ecosystem. VinBrain built AI diagnostics in Vietnam and was acquired by NVIDIA in December 2024 (terms undisclosed, but Vingroup had invested $126M prior). These are the companies building the pipes before the giants arrive.

Conclusion

OpenAI and Anthropic have the best models but not the pipes. ChatGPT can help you understand your lab results, but it can’t book the follow-up, pay the bill, or get your prescription filled. Ant has both: a decade of healthcare infrastructure and now AQ on top, serving 30 million monthly active users who can go from question to diagnosis to payment to prescription without leaving the app.

The US market will be fought over by OpenAI and Anthropic, and they’re well-positioned for a $5 trillion market where they understand the regulatory landscape. But most of the world looks more like China than America: fragmented hospitals, physician shortages, aging populations, and healthcare systems that can’t scale fast enough. For those markets, Ant has structural advantages. Whether they can translate that into brand recognition outside China, and whether OpenAI and Anthropic can build the infrastructure they lack, remains to be seen. For founders building healthcare infrastructure in underserved markets, that gap is the opportunity.

The Origins of Alipay

Ant Group is an affiliate of Alibaba Group, the retail, e-commerce, and logistics giant founded by Jack Ma. It was formed in 2014 when Alipay’s financial services business was spun out into a separate entity called Ant Financial, which was later renamed Ant Group in 2020.

In the early 2000s, Chinese consumers didn’t trust online payments and credit card penetration was low. Taobao, Alibaba’s C2C marketplace launched in 2003, needed a way to make people comfortable paying strangers on the internet.

Alipay started as an internal escrow feature for Taobao in October 2003, allowing transactions to settle only once a buyer confirmed receipt of goods.

• Buyer pays into an Alipay-controlled account.

• Seller ships goods.

• Once buyer confirms receipt, Alipay releases funds to seller.

This solved the basic trust and counterparty risk problem in Chinese e-commerce. Its akin to the relationship between PayPal and eBay, except instead of PayPal starting independently and later being acquired by eBay, Alibaba, the Chinese eBay, built its own payment processor from the start.

In December 2004, Alipay became a standalone product when it was formally spun out from Taobao and operated by a separate entity, Zhejiang Alipay Network Technology Co., Ltd. From there, it gradually expanded beyond marketplace escrow into bill payments, utility services, and offline merchant acceptance, with broader external usage scaling meaningfully from around 2008.

From Payment Solution to the World’s Largest Money Manager

As Alipay adoption grew alongside Alibaba’s user base, it became a default account for everyday transactions beyond e-commerce, functioning almost like a Chinese equivalent of Revolut, but much more. Hundreds of millions of users now held persistent balances inside the app, with large volumes of funds sitting idle between purchases.

This created the perfect opportunity to convert passive wallet balances into yield-bearing assets.

Yu’e Bao (余额宝 - “Balance Treasure”) was launched in June 2013 by Alipay in partnership with Tianhong Asset Management. It had a simple pitch that encouraged users to “put idle money to work,” allowing spare balances in Alipay wallets to be swept into a low-threshold money market fund with a minimum investment as low as 1 RMB. The fund invested primarily in short-term, low-risk instruments such as bank deposits, certificates of deposit, interbank repos and government-backed debt, generating returns that were consistently higher than standard Chinese savings accounts, with peak annualized yields reaching around 4 percent in the mid to late 2010s. Its appeal lay in frictionless access, daily liquidity, transparent returns, and the fact that it sat natively inside Alipay, which already functioned as a daily financial account for hundreds of millions of users.

This combination of convenience, perceived safety and superior yield drove explosive growth, with Yu’e Bao reaching approximately 1.69 trillion RMB (US$255 billion) in assets under management by 2018 and becoming the largest money market fund in the world, surpassing established institutional funds including the JPMorgan US Government Money Market Fund (operated by JPMorgan Chase & Co.) which had around US$150 billion in assets.

Ant Group as a Financial and Technology Stack

Ant Group is no longer just a payments or consumer finance company, but a conglomerate spanning:

• Payments infrastructure

• Consumer and SME lending

• Wealth management

• Insurance distribution

• Digital identity and risk control

• Enterprise technology and databases

• Cross-border payment infrastructure

Ant has even established its own digital bank, MYbank, designed to serve SMEs and micro-entrepreneurs through online, data-driven lending.

The Role of AI - Reducing Fraud & Powering Lending

AI at Ant did not start as a chatbot or consumer-facing product. Historically, it operated largely behind the scenes, powering fraud detection, risk control, credit scoring, and customer service automation across Alipay, MYbank, and its lending platforms. As these systems matured, their use expanded well beyond core financial operations into sectors where Ant already had deep economic visibility and data access, where it increasingly influenced real-world decision-making, resource allocation, and credit distribution.

Ant’s AI analyzes satellite and drone imagery to monitor crops, soil, disease, and yields, enabling MYbank to base rural lending on real production data instead of traditional collateral. This has expanded access to formal credit, reduced reliance on informal lenders, and sped up loan approval with more accurate risk-based pricing.

In SME financing, Ant, through its online-only bank MYbank, uses machine learning to evaluate real-time transaction data, inventory turnover, seasonal patterns, and behavioral signals to automate credit scoring for Taobao and Tmall merchants. This follows MYbank’s 310 model, where loans are applied for in under three minutes, approved in one second, and issued with zero human intervention, using platform sales performance and cash flow data as the basis for underwriting. Credit limits are adjusted dynamically, enabling rapid access to working capital for businesses that typically fall outside traditional banking criteria.

As AI became more deeply embedded across Ant’s operations, from consumer credit to logistics tracking and supply chain risk, it became clear that relying solely on task-specific models and back-end optimization was limiting. The complexity of Ant’s ecosystem, spanning finance, commerce, public services, and enterprise infrastructure, created demand for more generalized, flexible AI systems capable of operating across multiple domains and user contexts. This led to the development of Ant’s proprietary large model family, designed not only to optimize internal systems, but to act as a foundational layer for interaction, creation, and decision-making across the super app environment.

Meet The Ling Family

The Ling (灵) model family was developed by Ant Group to move beyond domain-specific AI and create a unified foundation model layer capable of serving both internal and external applications. Development accelerated in 2023 as Ant expanded AI from risk and operational systems into creation and interaction layers. On October 9, 2025, Ant officially announced Ling-1T, a trillion-parameter general-purpose model that marked a major milestone for the series. The name “Ling” translates roughly to “spirit” or “intelligence”, reflecting Ant’s intent to embed adaptive, context-aware cognition across its ecosystem.

Enter Lingguang

Lingguang (灵光, “Halo”) was launched on November 18, 2025 as the first user-facing product built on Ant’s Ling model family. While it demonstrates strong multimodal capabilities, the most notable feature is its “flash app” functionality, which effectively combines a code generation interface with a live preview environment inside Alipay. Rather than functioning purely as a conversational assistant, Lingguang allows users to generate, test, and retain lightweight applications through simple prompts, shifting AI from optimizing existing services to enabling the direct creation of new ones within the super app itself.

Let’s make a mini-app

Here I created a simple mini-app in roughly one minute that teaches Chinese idioms and their meanings, complete with a built-in testing function. I provided only a high-level objective, and the system handled the implementation details, including structure, logic, and feature design, without further instruction. The resulting app remains saved within the Lingguang environment and can be accessed or reused at any time. What stands out here is less the novelty of the output and more the frictionless process, from prompt to working application, suggesting a model where lightweight, purpose-built tools can be spun up on demand and persist as part of a user’s personal software layer inside the super app.

Testing the flash app feature

Tour the Forbidden City

Next, I test out live web search and multimodal with a tour of the Forbidden City. It generated and laid out all the details in a beautiful article style with accompanying photos. It also generated an impressive 3D model of 太和殿 - the Hall of Supreme Harmony, the largest of the three halls that constitute the heart of the Outer Court of the Forbidden City.

Multimodal features including live web search and fantastic 3D model creation

So what?

While platforms like Gemini and Claude can generate similar code and interactive outputs, Lingguang’s differentiation lies in how it is bundled and served, as well as its persistence and usability. The apps created are not temporary previews or isolated experiments, but saved, discoverable tools that remain embedded within the Lingguang environment and can be reopened or reused at any time. This turns AI output from a one-off interaction into a personal library of functional micro-apps, making the experience feel closer to building software within a system than simply conversing with a model.

Super Apps - The $420 Billion Opportunity

Alipay is one of China’s so-called “super apps.” Unlike the Western approach where platforms typically specialize in doing one thing well, such as messaging, shopping, or ride-hailing, Chinese super apps evolved by consolidating multiple daily behaviors into a single ecosystem. They start with a core use case and then layer on payments, services, commerce, and utilities, creating an all-in-one operating environment where users rarely need to leave the app. This model prioritizes convenience, continuity, and ecosystem lock-in over single-purpose excellence. Other examples include WeChat, which started as a messenger, and Meituan, which started as a Groupon-style group-buying app before becoming a leader in food delivery.

Super apps have evolved into full commerce and service infrastructures, and the market size reflects that reality. Global super app economies are now projected to approach $420 billion by 2030.

What matters is not just the size of this market, but its structure. These platforms already control user identity, payment flow, distribution, and retention in one closed loop. That makes them uniquely positioned to become not just marketplaces for services, but the primary environments in which new digital businesses are created, launched, and scaled. When creation tools are embedded directly into this layer, the friction between idea, deployment, and revenue effectively disappears, opening the door to an entirely new class of micro-entrepreneurship native to the super app itself.

Putting it all together

Mini apps can now be created directly inside Lingguang through simple prompts, dramatically reducing the time and technical effort required to build functional tools. If this creation model is extended into Alipay more broadly, it could significantly lower the barrier for users, merchants, and small businesses to produce and deploy mini apps at speed. Combined with Tencent’s recent agreement with Apple, which further formalizes monetization pathways within WeChat mini programs, this points toward a future where mini app ecosystems become both easier to create and more economically viable, potentially accelerating the scale and sophistication of internal super app economies.

Expect an expansion of super app commerce, positioning WeChat and Alipay as operating systems for digital business. This may usher in the first wave of vibe commerce, where users move from vibe coding to creating full online businesses through simple prompts. Instead of designing storefronts, payment systems, customer flows, and product logic manually, users describe intent and generate functional, monetizable businesses instantly, complete with interfaces, workflows, and embedded transaction capability, all native to the super app environment.

This is likely to redefine what it means to start and run a business in the digital age.

Deepseek has seen an exodus of users the past few months, with most of those users moving to Doubao, ByteDance’s everything AI app. Recent analysis from QuestMobile shows that Doubao has taken the top spot as the most popular AI app in China.

Background

In January 2025, DeepSeek burst onto the scene with its first public model, DeepSeek-R1. The launch rattled global markets: the S&P fell 1.5% and NVIDIA lost $590B in market value. Adoption was explosive, reaching 100 million users in just seven days compared to the two months it took ChatGPT. At its core, DeepSeek offered the now-familiar large language model chat interface with reasoning capabilities and newly introduced agentic tools.

UI

Doubao (豆包, “Beanbun”) began in 2022 as an internal ByteDance tool called “Grace,” designed to help employees learn from company materials. It was rebranded and released to the public in 2023, scaling quickly from there. By 2024, ByteDance had invested ¥80B into AI, a figure nearly equal to the combined ¥100B spent by Baidu, Alibaba, and Tencent together.

The user experience is where Doubao really stands out. You can log in seamlessly with either a Douyin (TikTok) account or a phone number. At the top of the interface, a moving Doubao character greets you with an introduction, adding a playful, familiar touch. Beneath that are sample prompts designed to showcase the platform’s capabilities, such as generating an image, checking the current price of gold per kilo, or offering health guidance on questions like how to lose belly fat.

Multimodal Responses

Responses in Doubao are impressively fast, multimodal, and polished. The platform can generate images, deliver text-to-speech replies that are read aloud, and provide answers that are both responsible and practical. For example, when asked about losing weight quickly, Doubao cautions that there is no healthy shortcut and instead recommends reducing sugar, cutting refined carbohydrates, and drinking more water.

Suite of Agents

Opening the menu reveals much more than a list of past conversations. Instead, Doubao offers a full suite of tailored agents, each designed for specific tasks. These include a personalized poem generator, an English speaking coach, a versatile writing assistant, a price estimator that analyzes objects from photos, a homework coach for both elementary and high school students, and even a username generator. Together, they highlight how Doubao goes beyond being a simple chatbot to become a multifunctional platform.

Culturally Tailored Tools

Each of these tools is more than a surface-level feature. The poetry generator, for example, creates 七言绝句 (seven-character quatrains) based on a user’s name, tapping into the deep cultural significance of names in China, where each character often carries symbolic meaning chosen by parents. The username generator also goes further than random suggestions. It provides names with clear explanations of their meanings and the impressions they are meant to project.

Voice Chat

The English coach, known as “Charlie,” is another standout feature. Users can chat entirely in English, practice pronunciation through built-in text-to-speech, and even start a live audio conversation by tapping the phone icon. This combination makes language learning more interactive and engaging than standard text-based practice.

Solid TTS model. Questionable banter.

Turning Generic Features into Specific Tools

The writing tool is highly flexible, allowing users to set parameters such as tone (casual, formal, instructional), target audience, and word limits. It adapts output to match the specific needs of each task. The homework coach takes a similar approach to personalization. Rather than simply providing direct answers, it offers examples and walks students step by step through the reasoning process, making it a valuable guide for learning. All LLMs can do this, but here it’s quite smart to package them in specific tools.

The price estimator is one of Doubao’s most impressive features. By simply snapping a photo, the tool can recognize the object, provide background information, and generate a valuation instantly, all without any additional prompting. It is a seamless blend of computer vision and practical utility. (For the record, my head shaver remains priceless.)

Community of Creators

Beyond the built-in tools, Doubao also hosts a marketplace of user-generated agents with unique interfaces and functions. Some of the more creative examples include “Argue with me AI” and “Create an 18-year-old version of yourself.” For those who want something different, there is also the option to design a custom agent, submit it for approval, and make it available for others to use. This approach turns Doubao into more than a chatbot. It becomes an extensible platform shaped by its community.

Conclusion

Success like this shows that the future of AI is not only about building smarter models. What matters just as much is creating experiences people want to return to, built on cultural touchpoints, personalized tools, and a marketplace of agents that feel alive. Doubao demonstrates this clearly. The real winners will be those who build ecosystems that people actually live in. The future of AI belongs to sticky platforms.