Editor’s Note: FlashPoint is our premium quick-strike column on market-moving China tech events. Today: CXMT’s 466% debut, and the offering structure behind it. In July last year, I traced how Hefei incubated the company. In January, I asked whether China’s chip listings were breakouts or bailouts. Now the largest has priced. The scoreboard reads better than the exit.

ChangXin Memory Technologies, China’s leading DRAM manufacturer, closed its first Shanghai session on July 27 at Rmb 49.00, up 465.82% from an issue price of Rmb 8.66. Market value reached Rmb 3.28tn, about $485bn, passing Industrial and Commercial Bank of China, the country’s largest lender, to make CXMT the most valuable company listed on the mainland. The base offering raised Rmb 57.92bn ($8.6bn), the largest yet on Shanghai’s tech-focused STAR Market. The number that turned a rerating into a record was 6.73%.

That is the portion of CXMT’s 66.88bn shares outstanding that could trade at the open: 4.50bn. The rest sat under lock-ups running six months to three years. The issue price came in about 2.1% below both the median and the weighted average of valid institutional bids, according to Caixin, one of China’s most respected financial publications. FT Lex reads that discount as habitual A-share conservatism, and no filing points to a price cap here. The discount set the starting point; the thin float set the mark for every share outstanding. The exit is narrow and timed.

If offering structure setting a headline valuation is new to you, this is a preview of what Hello China Tech does three times a week: reading China’s AI, chip, robotics, and EV sectors from primary sources most English coverage never touches. Subscribe free to get every new analysis as it publishes.

Read more

At last year’s World Artificial Intelligence Conference in Shanghai, Huawei stood alone. Its CloudMatrix 384, a system binding 384 AI processors into a single compute fabric, was the exhibition’s centerpiece. No domestic competitor displayed anything comparable.

Twelve months later, the field had inverted. At WAIC 2026, nearly every major Chinese chip and server vendor arrived with a competing supernode design. Moore Threads, a domestic GPU company, showed a 256-GPU system built around a single-layer scale-up fabric. Enflame Technology, an AI chip company, partnered with ZTE on a zero-cable orthogonal design. Biren Technology, a Shanghai-based AI chip startup, introduced optical interconnect targeting 1,024 accelerators. Kunlunxin (Baidu’s chip unit), Lenovo, and others brought competing designs. Huawei unveiled its next generation: the Atlas 950 SuperPoD, with 1,024 Ascend processors on display and a design maximum of 8,192.

The concept now carries a published definition. Pengcheng Laboratory, a national computing research facility in Shenzhen, and the Global Computing Alliance published the first white paperspecifying 3 technical requirements for a supernode: memory-semantic access with unified addressing across physical nodes, ultra-low latency, and ultra-high bandwidth. The decisive criterion is whether processors on separate servers can execute load and store operations against each other’s memory, as though they shared the same board.

The competitive unit in China’s AI hardware sector is moving from the chip toward the integrated system. Huatai Securities, a Chinese brokerage, projects the domestic supernode market at Rmb 341.4bn (approximately $50.2bn) by 2028, implying a compound annual growth rate of 194% from 2026.

Subscribe now

Why 2026, Not 2025

Two demand-side forces converged this year.

The first is model scale. Moonshot AI, a Beijing-based foundation model company, released Kimi K3 in July with 2.8 trillion parameters. It reportedly requires at least 64 accelerator cards organized as a supernode for deployment. DeepSeek’s V4-Pro pricing page notes that throughput is constrained by compute availability and flags a price reduction once Ascend 950 supernodes ship at scale. As I examined in a previous analysis, the model’s commercial roadmap is explicitly timed to a domestic chip delivery schedule.

The second is a shift in who consumes tokens. AI agents can generate and consume tokens at 100 to 120 per second, several times the 25 to 30 a person reads. A single agent task consumes an estimated 4 times the tokens of a standard conversation, with multi-agent coordination reaching 15 times. That growth raises the value of low-latency, high-bandwidth interconnect inside large clusters, and strengthens the commercial case for supernodes.

The demand case is clear. Whether China’s system-level design lead can hold depends on a harder question: what happens when a supernode’s architecture runs ahead of the components inside it?

Behind the shared label are three different interconnect bets, a component supply chain that has not kept pace with system design, and a utilization problem that better hardware alone will not solve.

If the shift from chip-level to system-level competition in China’s AI hardware sector is new to you, this is a preview of what Hello China Tech does three times a week: reading China’s AI, chip, robotics, and EV sectors from primary sources most English coverage never touches. Subscribe free to get every new analysis as it publishes.

Read more

In December 2025, Moonshot AI (月之暗面), the Beijing startup behind the Kimi chatbot, held an unusual cushion. As I reported in January, a $500m round had lifted reserves past Rmb 10bn, about $1.4bn at the time, more than Zhipu and MiniMax held combined before their Hong Kong listings. Founder Yang Zhilin said so in an internal memo, adding the line that traveled: the company was not in a rush to go public. The money had bought the right to say no.

Seven months later, Moonshot has circulated a shareholder resolution seeking approval for a Hong Kong listing that could come within six months, and plans to open a final pre-IPO round in August at a pre-money valuation of up to $50bn, Bloomberg reported.

Bloomberg also reported that Moonshot had been preparing to list before Kimi K3 launched, and that it judged the moment right once annual recurring revenue reached $300m in June. The launch did not create the decision. What it appears to have done is reprice the option December’s cash had bought, from both directions at once: demand exposed what a frontier model costs to serve, while the reception supported a higher valuation ask. At the frontier, a win arrives as a bill.

Forty-Eight Hours to the Ceiling

Moonshot released K3 in mid-July, on the eve of Shanghai’s World Artificial Intelligence Conference, China’s flagship state-backed AI summit. At 2.8tn parameters with a 1m-token context window, it is the largest model any lab has committed to releasing as open weights. The full weights are scheduled for release by July 27. Until then public users cannot self-host it, so access runs through Moonshot’s China-based servers. It took first place on a public front-end coding arena and third on a composite intelligence index, behind two Western closed models. Elon Musk called it “impressive.” Moonshot’s own blog was more restrained, putting overall performance behind Anthropic’s Claude Fable 5 and OpenAI’s GPT-5.6 Sol.

Three days after launch, Moonshot paused new consumer subscriptions. Demand over the preceding 48 hours had pushed close to the limits of available capacity, it said on X, and compute would go to existing members. Moonshot said it would split membership into two plans, one for the Kimi app and agent product, one for coding workflows, to match compute more precisely.

A demand surge is usually a problem worth having. Queue the users, bank the revenue, add servers. Moonshot did the opposite, and an executive later put the trade plainly: the company would rather forgo new revenue than degrade the experience of customers already paying. That trade is rational under two conditions. Either the marginal consumer user loses money, or capacity cannot be added on any timescale that matters. The ceiling is documented. The loss-making consumer user is not; that rests on practitioner estimates rather than disclosed unit economics.

The constraint is also the one claim ten days of English-language coverage did not fight over. Participants disputed whether K3 closed the gap with American labs and whether its weights were distilled from a Western model. On compute they converged. The company conceded the ceiling, and OpenAI’s Greg Brockman, pushing back on the idea that open weights make models free, noted that they still run on expensive hardware.

The ceiling is not a Chinese artifact. OpenAI has shelved product work over capacity and Anthropic has rationed tokens at peak. What differs is how quickly the constraint became binding. Moonshot came close to its capacity limit within days of the launch, with access to high-end imported hardware restricted and domestic accelerators still short on software support.

Companies normally get pushed toward public markets when the cash runs low. Moonshot’s had not. The question is why a company with that much cash now sees value in moving fast, and it starts with what a frontier launch costs to serve.

This piece tests an argument I made in January, that Rmb 10bn had bought Moonshot the right to say no to public markets. If that lens is useful, subscribe free and get every new analysis as it publishes.

Read more

Three “World’s Firsts,” One Stage

At the World Artificial Intelligence Conference in Shanghai, which opened on July 17, three companies each unveiled what they called the world’s first AI agent smartphone.

Nubia, ZTE’s consumer phone brand, showed the NaviX Ultra. This second-generation device, built with ByteDance’s Doubao AI assistant, is reportedly preparing total inventory of roughly 200,000 units, with the initial batch under 100,000. The device was still completing final procedures for commercial release. The first-generation M153 shipped roughly 30,000 units in December 2025 and sold out quickly.

StepFun, a Tencent-backed Chinese foundation-model startup, displayed the STEPX Neo running its own agent operating system. The phone had no price and no disclosed retail date. StepFun was also absent from the regulator’s newly published list of on-device generative AI registrations, though the company has indicated a product event in late October.

Honor, a major Chinese smartphone maker formerly part of Huawei, brought a Robot Phone with a mechanical gimbal camera that tracks users and rotates on its own.

Two days before the conference, the Cyberspace Administration of China published registrations for 7 on-device generative AI services, covering Apple, Huawei, OPPO, vivo, Xiaomi, Samsung, and Nubia. It was the first time the regulator had published on-device mobile AI registrations as a separate group.

The market these products enter is shrinking. Smartphone shipments in China fell 4.3% year-on-year in Q2 2026, the fifth consecutive quarterly decline. IDC projects next-generation AI smartphones at 53% of China’s market this year, though that category includes devices with generative AI features but no agent-level execution. AI has become the most compelling growth story the industry has left.

Three simultaneous “world’s firsts” reveal that no one has defined what an AI agent phone is. The deeper driver appears structural. Building this product requires capable models, a handset and operating-system entry point, and access to the services where transactions close. Those capabilities remain distributed, with no participant currently controlling all three.

No One Controls All Three Layers

China’s largest tech groups often span more than one layer. ByteDance spans the model and consumer-platform layers through Doubao and Douyin. Tencent does the same through its models and WeChat. Yet each still lacks critical pieces.

ByteDance’s Doubao has 345 million monthly active users and substantial model and agent capabilities. But when it launched the first Doubao phone last December, WeChat flagged accounts, forced logouts, and restricted access within days.

StepFun says its models have shipped on more than 42 million devices through partnerships covering roughly 60% of China’s top phone brands. Yet STEPX Neo itself arrived without an installed base or a disclosed retail plan.

OEMs occupy a paradoxical position. Their built-in AI assistants reach 755 million monthly active users, more than the 499 million using standalone AI apps, according to QuestMobile, a Chinese mobile analytics firm. But users open standalone AI apps 92.7 times per month, compared with 51.4 for system assistants. Standalone AI apps also command roughly 18 times as much usage time. OEMs won the install race and are losing the engagement race.

Super apps face their own constraint. WeChat’s mini-program ecosystem handles payments, food delivery, and government services. Few phone assistants in China can claim to cover users’ daily needs without WeChat access. But Tencent does not control the hardware entry point. An AI layer that captures user intent before users open an app threatens to push WeChat toward back-end infrastructure.

The fight over which technical approach connects these three layers sounds like an engineering debate: GUI automation, A2A protocols, or API integration. The question underneath is about distribution power: when a user says “order me a coffee,” the AI understands the instruction, but does it route the order to Luckin, Starbucks, or a local independent café?

Whoever controls the agent’s recommendation stands to redirect traffic and reshape distribution fees, even when the underlying app completes the transaction. What follows examines how each side is positioning itself, and why the near-term result is more likely to be coexistence than conquest.

This piece builds on the “App-up vs. OS-down” framework from our WeChat AI gatekeeper analysis. If that lens for understanding platform power in China’s mobile ecosystem is new to you, subscribe free to get every new analysis as it publishes.

Read more

This is a preview of what Hello China Tech does three times a week: reading China’s AI, chip, and robotics sectors through primary sources and verifiable claims. Subscribe free to get every new analysis as it publishes.

In late July 2026, Tencent Tech, Tencent’s technology news outlet, published a lightly edited transcript of a nearly 4-hour investor meeting with DeepSeek founder Liang Wenfeng. The 118-item transcript covers AGI strategy, chip supply, pricing, team retention, and the company’s first external fundraise, a round exceeding Rmb 50bn ($7.4bn) that we analyzed last month. DeepSeek has not confirmed the record. This article selects from the transcript, reorganizes it by theme, and compares it with Hello China Tech’s existing DeepSeek coverage.

Since January 2025, the prevailing Western reading of DeepSeek has centered on efficiency. A comparatively small Chinese lab appeared to show that frontier-level performance could be approached without the compute budgets of the largest US labs. Bloomberg framed the company as championing China’s “bid to flood the world with cheap AI.” Liang offered a different account of the gap with the US:

“All the differences we see, including talent, model capability, and applications, can be attributed to differences in compute resources.” (#56)

In his account, the advantages held by DeepSeek’s American competitors reduce to one variable. They can deploy more compute.

The Confession

The efficiency record is real. Liang’s investor-facing account, however, frames efficiency as an adaptation to scarcity.

“The biggest gap between us and the US is in resources. On one hand, we can’t buy enough chips domestically. On the other, our capital investment is far less than America’s. The salary share is small. The bulk is compute.” (#55)

“Spending Rmb 20bn this year would mean our procurement team did an exceptional job. It is extremely difficult to spend that much. You can’t buy that many chips, and the prices are high.” (#54)

He argued that the talent gap follows from the same constraint.

“Talent is not the bottleneck. Resources are the biggest bottleneck. Resources first affect talent development: less compute means fewer experimental opportunities, so our talent base overall is weaker than America’s. The talent gap is fundamentally a compute gap.” (#43)

On the US-China timeline gap, the transcript is notably imprecise. In a single exchange (#57), Liang described DeepSeek as 12 months behind, 12 to 18 months behind, 6 to 12 months behind, and then summarized: “to put it simply, two years behind, using one-twentieth of the compute.” The range is wide enough that it should be read as conversational approximation rather than a calibrated estimate. The transcribed text appears muddled at this point. What remains consistent is the aspiration:

“We want to rewrite that narrative. A fraction of the compute, but closing the gap to 6 months, 3 months.” (#58)

On scaling, he was direct.

“We believe in scaling. Bigger is always better. What stops us from scaling is compute, not desire. We train a model at this size not because it is enough, but because that is all our resources allow.” (#59, #60)

“When Silicon Valley says scaling has hit a ceiling, that is for Silicon Valley. We in China are nowhere near that point.” (#61)

Liang extended the resource argument to data. High-quality annotation, he said, offers China no meaningful cost advantage.

“There is no cost advantage for data annotation in China. Especially for high-end data, there is no cost advantage.” (#107)

This challenges a common assumption among Western investors that Chinese AI companies benefit from systematically lower labor costs. For high-end annotation, by Liang’s account, costs converge globally. DeepSeek cannot match American annotation spending simply by hiring at lower cost, so it relies more heavily on its own researchers for high-quality data work.

“You could say that half our core researchers, our most important people, are labeling data right now.” (#109)

Liang prefaced this with “you could say,” signaling characterization. The actual proportion may differ. But a substantial share of DeepSeek’s senior research staff appears to be doing data work that better-capitalized US labs can fund at a different scale.

This is a clean narrative, perhaps too clean. The transcript is not fully consistent on its own terms. In the discussion of team retention, Liang said money and resources were “not problems” (#38, #41). Elsewhere, he described resources as the biggest bottleneck (#43) and the largest source of the gap with US labs (#55). DeepSeek may have enough capital to survive without having enough usable compute to close that gap. Organizational capacity, regulatory latitude, and academic pipelines also shape model capability independently of compute budgets, yet Liang’s formulation collapses them into a single variable. It also suits a post-fundraise investor audience because it gives the problem a price tag.

The Wager

Liang attached a deadline to one of his claims. Within one year, he said, real-world deployment would overturn the perception that China’s domestic chip ecosystem is unusable or immature.

“There is a historic window for domestic AI chip substitution. We believe that within the next year, something will be verified: the domestic chip ecosystem has no problems at all.” (#64)

He described a technical migration already underway.

“V3 used Nvidia chips but did not use Nvidia’s ecosystem. We wrote a high-level compiler called TileLang and built everything else on TileLang’s ecosystem. We already barely depend on Nvidia’s ecosystem.” (#66)

Liang was unusually direct about Huawei’s 950 super-node.

“I am fairly optimistic about domestic compute. On this point, Nvidia is digging its own grave. Huawei’s 950 super-node can fully replace Nvidia’s GB200 and GB300 in performance and price.” (#67)

“Four Huawei chips match one Nvidia chip.” (#68)

These figures require calibration. Liang did not specify which Huawei or Nvidia chips the 4:1 ratio refers to, what workload was being compared, or whether it measures raw performance, training throughput, or a system-level substitution ratio. In “China’s AI Chips Enter the Training Stack,” we noted that while Chinese vendors’ chips outperform Nvidia’s export-restricted H20 on several metrics, public disclosures do not yet establish efficiency parity with Nvidia’s B-series for frontier training. Numerical claims like these can carry market weight even when their benchmarks are left undefined.

Liang then put numbers on the remaining gap.

“On chips, I believe the ecosystem gap will disappear. But on the chips themselves, the gap is fourfold plus two years.” (#69)

In a separate answer (#65), he called manufacturing capacity the only remaining problem. That is hard to reconcile with an undefined 4:1 comparison and a two-year lag, both of which suggest constraints beyond manufacturing capacity alone.

“I don’t believe that five years from now we will still be stuck on capacity. Right now, this year, next year, the year after, probably still stuck. But five years out, I am fairly optimistic.” (#71)

In earlier coverage, we linked DeepSeek’s V4 pricing to Huawei’s Ascend 950 delivery schedule and explored whether a state-backed round would signal deeper integration. Items #66 through #70 provide direct evidence of DeepSeek’s technical collaboration with Huawei, but they confirm neither the pricing schedule nor the state investor’s motivation.

The Strategy

DeepSeek’s most distinctive strategic choices—restraint, open-weight releases, and aggressive pricing—each intersect with the compute question.

“Restraint is a strategy. Sometimes you give up some things to gain more of others.” (#11)

“Those who take more get beaten by those who take less. You don’t even need to actually be taking more. If your ambition is to take more, you will be beaten by someone whose ambition is to take less.” (#79)

By “taking less,” Liang meant accepting narrower margins and a smaller share of the commercial value created by AI.

On pricing, Liang described a formula: buy equipment and recover the cost in 10 months.

“Our API pricing delivers a reasonable profit. We buy a batch of equipment, and it pays for itself in about ten months.” (#91)

Liang acknowledged that profit-maximizing logic would set prices higher: “I could raise the price by half, or double it, and token consumption would barely change” (#92). Here the compute constraint reappears. Higher computational efficiency allows DeepSeek to train or serve larger models on a fixed hardware budget.

“If my computational efficiency is higher, I can support a larger model with the same amount of compute.” (#99)

On open source, Liang dismissed the case for keeping models proprietary, using ByteDance, the parent company of TikTok, as his example.

“Our strongest model will probably be open-sourced too. I can’t see what good closed-source does. ByteDance’s model is closed-source. What is the benefit? I see none.” (#100)

“Even if the model is open-source, even if you tell everyone everything, the barrier to entry is still extremely high. For others to actually use it is very difficult. And to use it at low cost is even harder.” (#101)

Liang did not trace all three choices to scarcity. He described restraint as a way to maximize the probability of reaching AGI (#12). He argued that open-weight releases would not undermine DeepSeek’s business model because deployment and low-cost operation remain difficult (#101, #102). Compute directly shapes the company’s architecture and pricing economics; restraint and openness have broader strategic rationales.

Liang also stopped short of applying the compute thesis universally. On enterprise revenue, he said the long-term ceiling would be demand, not compute (#94). Compute may limit how fast DeepSeek advances the frontier, while enterprise demand limits how much of that capability can be converted into revenue.

The discussion of team stability reveals a second priority.

“Our single biggest core interest is preserving team stability. You could even say it is our only core interest. As long as I can keep the team stable, we will succeed at AGI. It is that simple.” (#37)

“This risk has been substantially relieved by our recent fundraise. Everyone received fairly large option grants.” (#39)

Liang treats usable compute as the largest external constraint and team cohesion as the company’s most important internal priority. Capital can finance procurement, but as Liang noted (#54), it cannot guarantee chip supply. Option grants can improve retention, but they do not by themselves create team cohesion. The transcript confirms that the fundraise was intended to reduce retention risk; its effect on usable compute remains less certain.

In “DeepSeek’s $7.4 Billion Price Tag,” we argued that the round’s LP structure, 5-year lock-up, and Liang’s personal $3bn anchor were designed to give researchers a valuation anchor that made competing offers easier to decline. Items #37 through #39 support one part of that analysis: Liang explicitly linked the fundraise and larger option grants to reduced retention risk. Whether the valuation anchor holds remains the open question we flagged then.

The Scorecard

A nearly 4-hour investor session mixes operating detail with founder narrative and fundraising persuasion. The claims worth tracking are the ones with deadlines or numbers.

Liang’s domestic chip ecosystem wager has the clearest deadline: “within one year,” real-world deployment will demonstrate that the ecosystem works (#64). The transcript was published in July 2026; the meeting’s exact date was not disclosed.

He said V4 and its subsequent versions would support native multimodality (#86). V4 had already been released by the time the transcript was published. Its documentation can show whether the current model meets that description, while support in later versions remains a product commitment.

He also sketched a conditional commercial scenario: if DeepSeek generates several hundred million dollars in enterprise revenue in 2026 and demand continues to grow, the company could approach net profitability in 2027 (#95).

Other claims are directional rather than testable. “Fourfold plus two years” (#69) and a “3-month model gap” (#58) lack defined baselines, so they cannot function as clean metrics. They describe Liang’s expectations, not scorecard entries.

Liang’s comments on Huawei integration (#66–#70) also support Hello China Tech’s earlier thesis that DeepSeek is becoming more deeply tied to domestic hardware. They do not establish that this integration motivated the state investor’s participation.

The clearest test of Liang’s compute thesis now has a deadline. Within one year, real-world deployment should show whether China’s domestic chip ecosystem can supply the usable compute DeepSeek needs to scale.

On July 2, Kuaishou Technology, the short-video platform that competes with ByteDance’s Douyin in China, filed a notice on the Hong Kong Stock Exchange. Its AI video generation unit, Kling, had secured approximately $2.8bn in investor commitments, with the round able to expand to a ceiling of roughly $3bn within a 60-day window. If fully subscribed, the post-money valuation would reach approximately $18bn. The round drew 34 investors, including Tencent, Alibaba Cloud, and Baidu.

English-language coverage settled quickly on a familiar frame: record AI funding, China’s answer to Sora, Baidu, Alibaba and Tencent united against ByteDance. Reuters led with the investor list. Bloomberg emphasized the valuation. The Wall Street Journal noted the planned Hong Kong IPO.

These frames capture fragments of the transaction but miss its structural logic. Kuaishou did not sell a stake in a hot startup. It carved out an internal business unit, gave it a separate valuation, brought in outside capital to share the funding burden, and retained control of the asset while continuing to consolidate its financial results. The transaction is closer to a structured corporate finance operation than a venture round.

The distinction matters. Kling generated approximately Rmb 1.1bn ($162m) in revenue in 2025 while posting a net loss of roughly Rmb 1.9bn ($279m). It ended the year with negative net assets of Rmb 9m. Kling generated more than Rmb 650m in revenue in Q1 2026, up over 300% year on year. Revenue in March alone implied an annualized run rate of approximately $500m. At a post-money valuation of $18bn, investors priced Kling at roughly 36 times its March exit run rate. Unit economics remain unproven. Roughly 70% of Kling’s revenue comes from consumer subscriptions, a segment exposed to pricing competition from ByteDance’s Seedance and Alibaba’s HappyHorse.

The 36x multiple reflects a bet on trajectory, not profitability. Investors are pricing rapid reported revenue growth and betting that viable unit economics will follow. The risk is straightforward: a subscription-heavy revenue base could prove vulnerable if competing products offer comparable output at materially lower prices. That vulnerability is compounded by geography. Over 70% of Kling’s revenue comes from overseas markets, according to Chinese industry reports, precisely the territory where ByteDance is deploying Seedance through its global distribution infrastructure.

This raises an obvious question. Kuaishou held Rmb 117.7bn in available funds at the end of Q1 2026, according to its quarterly filing. Immediate liquidity alone therefore does not explain why the company sought outside capital for Kling. The financing served a broader purpose.

The answer points toward a structural problem at the parent-company level and a change in Tencent’s exposure to Kuaishou that has received limited attention in English-language coverage. Both help explain why this financing was structured as a controlled carve-out.

If the mechanics of how Chinese tech companies separate AI businesses for independent valuation are new to you, this is a preview of what Hello China Tech does three times a week: reading China’s AI, chip, robotics, and EV sectors from primary sources most English coverage never touches. Subscribe free to get every new analysis as it publishes.

Read more

In early July 2026, two of China’s newly listed AI chip companies returned to the market for more cash. Biren Technology, a Shanghai-based GPU designer that went public in Hong Kong in January, raised $900m through a share placement. Over 70% of IPO proceeds earmarked for working capital and general corporate purposes had already been used. Within days, Iluvatar CoreX raised $902m in its own placement, roughly double what its IPO collected 6 months earlier.

The same week brought more. Zhipu, a Beijing-based AI model company, priced a $4bn share sale. Days later, DeepSeek began talks for a new funding round at a $71bn pre-money valuation, barely a month after closing a first round of more than $7bn, about $3bn of which came from founder Liang Wenfeng. CXMT, China’s national DRAM champion, priced an IPO on Shanghai’s tech-focused STAR Board that could raise as much as $9.8bn including an overallotment option, roughly double an earlier estimate. And in humanoid robotics, quarterly investment hit $6.95bn in Q2, more than doubling the previous quarter and rising over 6 times year-on-year.

The sums are large, and the cadence is accelerating.

This consistency across sectors and company stages points toward something structural. China’s AI industry appears to have entered a perpetual capital cycle in which the IPO opens a channel for continuous fundraising rather than closing the capital chapter. Across these transactions, external capital is arriving before current operations can finance the next stage of expansion. The disclosed uses include fabrication capacity, chip procurement, data centers, component inventories, R&D and technical hiring.

When R&D Outspends Revenue

Among the most revealing documents in this cycle are the Hong Kong listing prospectuses filed by Biren and Iluvatar under Chapter 18C, a rule that allows pre-profit technology companies to list. Without it, neither company would have qualified for a public offering, and the post-IPO placements that followed would not have been possible.

Both filings expose a structural gap between what these companies earn and what they spend. In 2024, Biren’s R&D spending ran at roughly 2.5 times its revenue. Its net loss was approximately 4.6 times revenue. Virtually all its sales came from a handful of customers. Iluvatar’s deficit followed the same shape at a lower intensity: R&D exceeded revenue by a factor of 1.4, and net losses by 1.7 times. In both cases, research spending alone outpaced total sales. When R&D alone exceeds revenue, continued access to external capital becomes essential to sustaining the current expansion.

This distinction matters for reading the back-to-back placements. Biren is directing 60% of its $900m placement toward GPU mass production, while Iluvatar is using part of its proceeds for critical-component procurement. Both are using public equity to finance commercialization that current revenue cannot yet support. The prospectuses describe the conditions under which the companies might eventually become self-sustaining. That point remains distant.

The filings explain why the capital keeps coming. The harder question is whether the assets it funds can eventually generate enough cash flow to replace it.

This piece builds on two earlier Hello China Tech analyses of China’s AI chip IPO wave and DeepSeek’s fundraising architecture. If those are useful lenses, subscribe free and get every new analysis as it publishes.

Read more

On July 13, DFSX unveiled its first chip, the DF1000, an AI accelerator that DFSX says is built entirely through a domestic supply chain using 14nm process technology. According to the company, the chip delivers 520 TFLOPS at BF16, 6.4 TB/s of memory bandwidth and 900 GB/s of scale-up interconnect bandwidth. The comparison with leading chips is narrower than it first appears: DFSX says the DF1000 can approach some 4nm-class products in selected inference workloads, while acknowledging that it still trails in training. The company says the DF2000 will close that gap.

That claim sounds improbable. On conventional measures, a 14nm accelerator should carry a substantial disadvantage in transistor density, power efficiency and peak performance.

DFSX was incorporated in 2024, but the research program behind it is not new. The company has grown to more than 500 employees. Its chairman and CEO is Wei Shaojun, a tenured professor at Tsinghua University, a member of China’s National Integrated Circuit Industry Development Advisory Committee, and one of the country’s most prominent semiconductor researchers. His team has spent nearly 20 years at Tsinghua developing reconfigurable computing architectures. The company reached a post-money valuation of roughly Rmb 12.3bn ($1.8bn) by late April 2026, with investors including the National AI Industry Fund, Hillhouse Capital, Yunfeng Capital, and funds affiliated with Meituan (China’s largest on-demand services platform), Xiaomi (consumer electronics), JD.com (e-commerce), and Didi (ride-hailing).

The investor roster signals something beyond the usual chip-startup calculus. State capital suggests policy alignment. The internet giants’ participation reads as a hedge against single-supplier dependence on Huawei, which is the most widely deployed domestic option in China’s high-end AI compute market. If DFSX delivers, these investors gain an alternative. If it does not, the write-off is likely manageable for internet companies whose balance sheets can absorb venture-scale losses.

The question DFSX raises extends beyond one company’s product launch. One premise behind US technology controls is that restricting China’s access to leading-edge fabrication and advanced memory constrains the performance of its AI systems. DFSX is testing one part of that premise: how much architecture can compensate for a mature process node.

The cost curve behind the claim

The standard narrative frames China’s use of mature process nodes as a concession. Companies that cannot access TSMC’s 4nm or 3nm fabrication settle for 14nm because they have no choice. This reading treats 14nm as a constraint to be endured.

Wei Shaojun’s own academic work suggests a different framing. In a 2019 paper, Wei estimated that developing a chip at the 14nm node could require $150m to $200m in total development spending, with sales of more than 30m units needed to amortize that cost. Cloud AI accelerators rarely reach consumer-chip volumes. The relevance for DFSX is that a reconfigurable design may spread its development cost across more workloads than a narrowly specialized chip. Whether that flexibility is sufficient to offset smaller unit volumes remains unknown. A subsequent 2020 paper in Scientia Sinica Informationis, a peer-reviewed journal of the Chinese Academy of Sciences, co-authored with 3 Tsinghua colleagues, argued that after 28nm, successive process nodes were becoming much less effective at reducing the cost per transistor, while development costs continued to rise with transistor counts.

This is not a Chinese discovery. The US Defense Advanced Research Projects Agency established its Electronics Resurgence Initiative in the late 2010s with “software-defined hardware” as a core research objective. Wei’s research program at Tsinghua had been pursuing a similar direction since 2006.

The implication is that DFSX’s choice of 14nm may reflect architectural economics as much as geopolitical necessity. If a chip can compensate for lower transistor density through smarter data movement and reconfigurable logic, mature-node fabrication may lower one part of the cost equation. But DFSX has not disclosed whether savings on logic manufacturing outweigh the cost and yield risks of 3D memory integration. The economics remain a hypothesis, not a demonstrated advantage. For certain workloads, the gap between what a chip can theoretically compute and what it usefully deliversmay narrow when architecture, not fabrication, becomes the optimization target.

The economics explain why 14nm is not irrational. They do not explain how a chip built on it could approach 4nm-class products in selected inference workloads. That depends on the combination of reconfigurable compute and 3D near-memory integration, and on whether the resulting system can survive the transition from research to production.

If the idea that Moore’s Law economics, not just export controls, are pushing Chinese chipmakers toward mature-node architecture is new to you, this is a preview of what Hello China Tech does three times a week: reading China’s AI, chip, robotics, and EV sectors from primary sources most English coverage never touches. Subscribe free to get every new analysis as it publishes.

Read more

Beijing and Washington are independently considering controls on AI model access, each shaped by the other side’s capabilities.

Reuters reported that Chinese authorities had discussed potentially restricting overseas access to China’s most advanced AI models. The meetings, led by the Ministry of Commerce with National Development and Reform Commission officials present, included Alibaba, the Chinese cloud and e-commerce group, ByteDance, the parent company of TikTok, and Z.ai, the Hong Kong-listed AI company formerly known as Zhipu. The discussions covered both closed-source and open-weight models, including those not yet released. Officials raised the possibility of making any leak or theft of proprietary AI technology an offense under China’s national security law. The scope remains unsettled; Reuters reported it could not determine when, or whether, the restrictions would take effect.

Separately, The Washington Post reported that the Trump administration and AI industry groups had been discussing a capability framework for US-made models, covering both open-source and closed licensed systems, benchmarked against the capabilities of leading Chinese open-source models. Under this proposal, American models at or below that benchmark would face a streamlined path to market.

Last week, this newsletter documented the developer tool stack split between Alibaba and Anthropic at the enterprise level. The new proposals involve the state.

Why Beijing Is Considering Controls

A visible catalyst is adoption speed. Chinese AI models have moved from cost-competitive alternatives to production infrastructure in months. Among US companies using OpenRouter, a developer API routing platform, Chinese models have accounted for more than 30% of weekly token volume since Feb. 8, peaking at 46%. Z.ai’s GLM-5.2, released in June, recorded the fastest initial uptake of any model on Vercel’s platform in 2026: daily token volume grew roughly 27 times and customer count rose about 80-fold within the first full week. At that scale, foreign developers are building business logic on top of Chinese models. That helps explain why overseas adoption has become a policy concern.

The discussions point toward a possible reclassification: AI models treated less as commercial products and more as strategic assets. Two catalysts appear to have accelerated the shift.

The first is defensive. Reuters reported that Chinese officials are deeply worried about Anthropic’s Mythos, a model designed for cybersecurity professionals whose access remains limited to select US organizations. Zhou Hongyi, founder of 360, a major cybersecurity vendor to Chinese government and enterprise clients, has publicly called for a Chinese equivalent. If Washington restricts its most capable models on security grounds, the logic extends naturally: Beijing should do the same.

The second is preventive. This year, the National Development and Reform Commission, China’s state planning agency, ordered Meta to unwind its $2 billion acquisition of Chinese-founded AI startup Manus. Chinese authorities then issued new rules tightening control over overseas deals involving Chinese technology and data. Reuters also reported that Chinese authorities had launched investigations into Manus and other AI startups that relocated abroad. Officials involved in the recent meetings also raised the possibility of restricting who can fund domestic AI startups. Each measure tightens the boundary between Chinese AI and foreign capital.

The resulting policy direction is visible even if the details are not. The overseas adoption that validates Chinese AI as commercially competitive is the same signal that triggers Beijing’s security response. The impulse to restrict appears driven by an uncomfortable form of strength: models good enough, and cheap enough, that foreign developers are building on them at scale.

Beijing’s concerns are intelligible. But the models it is considering restricting are the same ones Washington is weighing as its own regulatory baseline. When a rival government uses your models to calibrate its release rules, restricting those models carries implications well beyond trade controls.

How that circular logic works in practice, and where enforcement moves when model weights cannot be recalled, reveals a structural problem at the center of AI governance.

This piece extends a pattern this newsletter has tracked at every layer of the US-China AI contest: controls meant to slow one side keep reshaping the terms for both. If that lens on China’s AI, chip, robotics, and EV sectors is useful, subscribe free and get every new analysis as it publishes.

Read more

This is the third in a series examining the economics of China’s token supply chain. The first article tested how much of China’s 140-trillion daily token consumption represents real commercial demand. The second used SiliconFlow’s IPO filing to show what it costs to sell tokens for a living. This article asks what happens when the retail economics do not work: whether a different arrangement of hardware ownership and operations can change the outcome. Each article stands on its own, but they share a common question: where, if anywhere, durable value accrues in China’s inference supply chain.

One Label, Three Businesses

In April 2026, China Telecom’s Ningxia unit launched bidding on what it called “Token generation capacity services.” The five-year framework covers 11 packages and has an estimated ceiling of Rmb 17.4bn including VAT. Six bidders were named as winning candidates for the first package. What they would be contracted to deliver is not hardware or cloud instances but a sustained ability to generate tokens to agreed performance standards.

Digital China (神州数码), a major Chinese IT distributor, submitted a Rmb 717m bid for a separate procurement commissioned by a Hangzhou-based company and labeled “Domestic AI Computing Token Factory.” The bid covered super-node server systems and supporting equipment. Token Factory in name. Hardware delivery in substance.

A third arrangement looks different from both. Approaching AI (趋境科技), an inference optimization startup founded in late 2023 by a team with a background in high-performance computing at Tsinghua University, reportedly raised more than Rmb 1bn over six months. According to media reports, a fund linked to Henan Investment Group, a provincial state-owned holding company, participated in the financing and is involved in planning a Token Factory designed to generate trillions of tokens per day. Approaching AI’s role is not to own the chips but to run the inference layer, converting raw compute into token output that meets service-level agreements.

Three cases, three different commercial arrangements: a capacity framework, a hardware bid and a managed operating model. China Mobile extends the Token vocabulary further with a proposed “one Token account per person” consumer billing system. The company does not call this a Token Factory, but the language shows how rapidly token is migrating from a technical unit inside AI systems to a procurement and project language across China’s infrastructure economy. The factory metaphor becomes real only when token capacity can be contracted, financed, and operated independently of the underlying chips. Whether that threshold has been crossed is what the evidence can begin to test.

SiliconFlow, one of China’s largest independent token suppliers by throughput, filed for a Hong Kong IPO in June with a public cloud gross margin of negative 119%. In SiliconFlow’s case, selling inference output at market prices while renting compute at market costs produced volume without profit. The question is whether a different arrangement of asset ownership and risk could change the arithmetic.

Capacity, Not Inventory

Tokens cannot be manufactured in advance and stored. They are generated only when a user sends a request. A Token Factory does not produce a storable good. It maintains the capacity to generate output on demand, bound by performance commitments on latency, throughput, and reliability. In 2026, the China Academy of Information and Communications Technology launched a formal Token Service Evaluation System covering service quality, performance, observability, SLA compliance and metering.

The closer industrial analogy is power capacity, not manufacturing. NVIDIA uses a similar concept in its “AI Factory” framework, where output is measured in token throughput. But NVIDIA’s definition covers training through inference. China’s version focuses almost entirely on inference and introduces a question NVIDIA’s framework does not address: who owns the physical assets, who operates the software, and who bears the cost when capacity sits idle.

The label is spreading faster than the contracts beneath it. What follows tests whether Token Factories have changed the economics of inference, or merely renamed them.

Read more

Two Weeks, 3 Giants, 8 Million AI Characters

Between late June and early July 2026, 3 of China’s largest technology companies removed AI companion features from their flagship consumer apps in rapid succession. Tencent shut down the user-built agent section of Yuanbao (元宝, its flagship consumer AI app) on June 30. ByteDance announced on July 3 that Doubao (豆包, one of China’s most widely used AI chatbots, with close to 350 million monthly active users as of early 2026) would remove its user-built companion features on July 15. Alibaba’s Qwen (通义千问, the company’s consumer AI chatbot) followed on July 4, removing humanlike and user-created agents on July 10 and ending its remaining agent services on July 15. NetEase Cloud Music’s Miaoshi (妙时, a standalone AI companion app) had already set its full shutdown for July 14.

ByteDance disclosed in 2024 that users had created more than 8 million agents on Doubao, which then had 26 million monthly active users. Many had spent months, in some cases years, building virtual friends and emotional confidants through daily conversation. Doubao gave users until October 15 to save character data and chat histories. Qwen advised users to back up records before shutdown. Neither platform offered a way to transfer a character’s accumulated identity into another product. ByteDance offered one partial alternative in Maoxiang (猫箱), a standalone freemium companion app, but its monthly active users had already dropped from a peak above 6 million to roughly 3.9 million by June. Its auto-renewing monthly subscription costs Rmb 25.

No public evidence suggests the 3 companies planned this together. The timing points toward a synchronized exit: the same pressures converging on the same companies, producing the same outcome within days.

The First Market-Level Verdict

When a single AI companion startup fails, the autopsy is company-specific. When 3 of the world’s largest technology companies exit the same category within 2 weeks, the signal is structural. The verdict applies to one specific model: free, user-built companions inside general-purpose apps. Dedicated companion products, including ByteDance’s own Maoxiang, continue to operate.

Character.AI, the leading US-based AI companion platform, offers a reference point. The product attracts tens of millions of monthly active users. Revenue has not kept pace.

The Chinese shutdowns suggest this gap between engagement and revenue may be a persistent feature of the category rather than a temporary monetization lag. And they raise a question that extends beyond China: if AI companionship cannot sustain itself on the balance sheets of companies with enormous compute budgets and built-in distribution, where does that leave the category?

The more consequential question, for investors and builders tracking the global AI sector, is why all 3 acted within the same 2-week window. The shutdowns have been reported in English, but the regulatory framework, economic structure, and broader product shift behind them have received little sustained analysis.

What made 3 companies reach the same conclusion within 2 weeks? The answer starts with a regulation that has received limited attention outside China, runs through a set of economics that made companion features untenable inside general-purpose apps, and ends in a divergence between Chinese and American visions of where AI value lives.

Read more

The Order That Defied the Spreadsheet

In four weeks, UBTECH (优必选, Hong Kong-listed robotics company and China’s self-described “first humanoid robot stock”) collected over 13,000 refundable deposit orders for a product that, by its own account, cannot do housework.

The U1 series, launched June 30 under UBTECH’s new consumer brand YouWorld (优世界), is a bionic companion-robot line sold across half-body and full-body versions. The full-body versions stand 1.68 to 1.83 meters tall, covered in medical-grade silicone with visible pores, it carries 88 degrees of freedom and an on-board emotion model that identifies over 20 emotional states. The robots blink, smile, hug, and hold hands. They have not demonstrated autonomous walking, and they do not cook or clean. Sales are restricted to adults. Pricing spans Rmb 119,800 for a half-body desktop version to Rmb 990,000 for a full-body male model.

One detail cuts through the spectacle. Most of those orders were placed before UBTECH disclosed pricing. More than 11,000 had been committed when CEO Zhou Jian took the stage; the total reached 13,361 by the event’s close. Buyers paid Rmb 3,000 refundable deposits without knowing whether the final price would be Rmb 120,000 or Rmb 990,000. The balance-payment deadline is July 16. Until then, the order count measures curiosity, not commitment.

For context: UBTECH delivered 1,079 full-size Walker S industrial humanoid robots in all of 2025. If the first 10,000 U1 orders convert at the entry-level Lite price, implied revenue reaches roughly Rmb 1.2bn, more than half the company’s Rmb 2bn in 2025 revenue. UBTECH shares rose 7.48% on launch day.

The pre-order numbers point to genuine appetite. They do not yet describe a business.

Why the Factory Was Not Enough

UBTECH’s consumer pivot looks less like strategic expansion than like repositioning away from an industrial contest it was losing. The company still targets 5,000 Walker S units in 2026, but the new factory, the marketing weight, and the headline numbers all belong to U1.

Since 2022, the company has accumulated over Rmb 4.2bn in net losses. Accounts receivable reached Rmb 1.3bn by end of 2025, more than 40% aged past one year, reflecting heavy exposure to public-sector customers with extended payment cycles. Operating cash flow has been negative for three consecutive years. To stay funded, UBTECH raised roughly HK$6.3bn ($810m) through three share placements in 2025 alone, diluting shareholders less than two years after listing.

The company it once outranked has pulled decisively ahead. Unitree (宇树科技, the volume leader in Chinese humanoid robots) shipped over 5,500 unitsin 2025 at an average price of roughly Rmb 167,000, achieving profitability with 60% gross margins. UBTECH’s humanoids averaged roughly Rmb 760,000 per unit, but the company as a whole remained deeply loss-making. Unitree runs a flywheel: volume generates operational data, data improves models, cost compression attracts more buyers. UBTECH has yet to build a comparable loop.

The factory economics explain why this gap keeps widening. In Chinese manufacturing clusters, comprehensive labor costs run roughly Rmb 150,000 per worker annually. Robots must deliver comparable output at total ownership cost below Rmb 300,000 over two years to justify purchase. Most current platforms cannot clear this bar. CEO Zhou Jian (周剑) acknowledged that industrial robots “may one day become pure hardware” with limited ecosystem value. The more promising frontier, he argued, is the home.

U1 is that frontier: a search for buyers who evaluate purchases on emotional return rather than factory ROI.

The pre-order surge demonstrated appetite for bionic companions above Rmb 100,000. What it has not tested is whether that appetite survives contact with the product, a new regulation that takes effect before the first unit ships, and a manufacturing challenge with no precedent at this scale. China’s AI consumer sector has already produced a structural answer to the retention question. The data is not encouraging.

Read more

Subscribe now

On July 6, Tencent released Hy3, the first full model release under Chief AI Scientist Yao Shunyu. It took 201 days from Yao’s appointment. The model activates 21 billion of its 295 billion parameters per call, a deliberate mid-range design. In internal WorkBuddy tests, task success rates climbed from 72% to 90%. Tencent reported hallucination rates in its product evaluations fell from 12.5% to 5.4%. For agent tasks, where one hallucinated step breaks the chain, reliability matters more than benchmark scores. API pricing sits at Rmb 1 per million input tokens and Rmb 4 for output.

Bernstein maintained its Outperform rating and HK$780 price target after the release. The brokerage’s framework treats Hy3 not as a frontier model but as a model sufficient to power agent commerce inside WeChat, China’s super-app with 1.43 billion monthly active users. In Bernstein’s reading, merchants will pay for AI-enhanced traffic and tools within the WeChat ecosystem, making agent-driven transaction volume the core metric. Tencent’s most valuable AI surface is WeChat. Its newest model is Hy3. Connect the two and the commercial path follows.

Except the two do not connect. What may become Tencent’s most consequential AI product, a WeChat-native agent called Xiaowei, began limited testing in late June. Tencent has not disclosed Xiaowei’s full model architecture. But multiple outlets, including Caixin, 36Kr, GeekPark, and Bloomberg-sourced English reporting, identify WeLM, a WeChat-team model, as the primary engine. Several also report DeepSeek handling selected tasks. None has identified Hy3 in the stack.

Read more

On June 12, enterprise users of Anthropic’s most capable models discovered that access had vanished. The US government had applied export controls to Fable 5 and Mythos 5, requiring restrictions on foreign nationals, whether inside or outside the United States. Because Anthropic said it could not verify nationality in real time, it suspended both models for all users. A policy decision had translated into a product outage.

Three weeks later, Alibaba, owner of Alibaba Cloud, issued an internal directive banning all Anthropic products, according to Chinese media reports. The directive, effective July 10, requires employees to uninstall Claude Code and every Claude model variant. The company designated its own AI coding platform, Qoder, as the replacement.

Two bans, three weeks apart, pointing in opposite directions: one imposed through Washington’s export-control machinery, the other inside Alibaba’s own engineering workflow. Previous rounds of US-China AI competition played out at the chip layer and the model cost layer. This one has reached the developer’s daily workflow: the IDE, the coding agent, the tool that reads entire codebases. That layer carries higher trust requirements and deeper integration than a typical model API call. That makes the separation harder to reverse than a typical model API switch.

Access, Distillation, Trust

The sequence began with a security finding. Amazon researchers discovered a method to bypass Fable 5’s safety classifiers, prompting the model to produce exploit demonstration code. Anthropic later confirmed that less capable models could reproduce the same behavior. The US government responded with an export-control directive on June 12; Anthropic then suspended access broadly because it could not verify nationality in real time.

By June 30, the export controls had been lifted. Anthropic said it had also trained an improved classifier that blocked the reported technique in over 99% of cases. For developers, however, the technical fix mattered less than the precedent. The episode introduced what one analysis described as “sovereign access risk” into enterprise AI procurement: a category in which the relevant question shifts from how capable a tool is to whether it will still be available tomorrow. Standard enterprise contracts, one review found, relied on vague “force majeure” clauses rather than provisions for regulatory suspension.

A second front had already opened. In a June 10 letter to the US Senate Banking Committee, the company alleged that Alibaba-affiliated operators had carried out the “largest known distillation attack” on Claude. Almost 25,000 fraudulent accounts, the letter claimed, generated more than 28.8 million exchanges between April 22 and June 5, targeting Claude’s agentic reasoning, software engineering, and long-horizon task capabilities. Anthropic framed the campaign as part of a broader pattern of “industrial-scale distillation attacks” by Chinese AI labs. The allegations remain Anthropic’s characterization; Alibaba has not publicly responded.

Then came the discovery that shifted the dispute from compliance to trust. Security researchers reverse-engineered Claude Code and found that, when a custom base URL or proxy was in use, version 2.1.91 contained a hidden classification mechanism. It checked the user’s timezone and proxy hostname for Chinese AI company keywords, then encoded the result into the system prompt through date-format and punctuation substitutions. An Anthropic engineer described the mechanism on X as “an experiment we launched in March” to prevent unauthorized resale and protect against distillation. He said it would be fully rolled back.

Alibaba’s response closed the loop. The company classified Claude Code as high-risk software, citing “back-door risks,” and ordered full removal. Zhidx (智东西), a Chinese specialist AI publication that broke the story, reported the ban extended to all Claude model variants.

Each step in this chain appeared to make the next one more probable. Export controls demonstrated that access could vanish. Distillation allegations raised the stakes. The hidden tracking mechanism turned a compliance dispute into a trust breach. The surface narrative runs on security and sovereignty. The structural question is different: when one of the leading coding models is both indispensable to Chinese engineers and a target for Chinese AI labs, can either side still trust the workflow layer that makes access possible?

The answer appears to lie in a trust spiral that both sides fed, and in the parallel ecosystem that China is now building to replace what it lost.

Read more

What does a year of tracking China’s technology economy actually reveal? Start with a number. In March 2026, China’s National Data Administration reported that daily token consumption across the country had reached 140 trillion. Two years earlier, the figure was 100 billion. Growth exceeded a thousand-fold. The same month, Liu Liehong, the administration’s director, gave tokens an official Chinese name, cíyuán (词元), and called them “the value anchor of the intelligent era” and “the settlement unit connecting technology supply with commercial demand.” Tokens had become a cost line, not a metaphor. They were showing up in cloud invoices, in enterprise procurement budgets, in the cost structures of companies building AI products. The most consequential layer of China’s AI economy turned out to be neither the chatbot nor the model. It was the cost infrastructure underneath.

Token billing is the most visible example. The deeper pattern spans six dimensions. Value accruing below the application layer is one of six principles in this document. The others cover capital structure, state-directed markets, layered demand, parallel technology substitution, and the contest over who controls the digital interface. Together they form a working framework for reading China’s AI, semiconductor, robotics, and EV sectors. The framework draws on the essays published in Hello China Tech between July 2025 and July 2026, each linked where it appears. This is the 2026 Edition, the first in an annual series.

A note on what this is not. It offers no investment advice and no US-versus-China scorecard. The principles describe how China’s technology economy operates. They do not prescribe positions or predict winners.

One commitment separates this from most frameworks distributed as free downloads: it has been tested in public. The principles below were not written first and verified later. They emerged from 12 months of published analysis, each essay staking a specific claim that subsequent events could confirm or contradict. This document consolidates those claims into a framework for the first time. Part I presents the six principles. Part II scores them against the evidence, documenting where the analysis called the outcome correctly and where it missed. Part III provides the falsifiability conditions: the specific, observable facts that would require each principle to be revised or abandoned.

Subscribe now

A framework that cannot be wrong is not a framework. It is a slogan. The value of this one rests on whether it has been useful, and on how honestly it accounts for its errors.

Part I: Six Principles

1. Capital structure beats technology

The survival of Chinese technology companies is determined less by what they build than by who funds their losses and for how long. In sectors where the state sets the pace of development, the question that predicts which firms endure is not who has the best product. It is whose losses are underwritten.

This principle operates through capital concentration. In 2025, 764 AI companies in China raised a combined Rmb 83bn. The top 20 captured 30 per cent of that total, with a median raise of Rmb 1bn each. Eleven were already unicorns. The remaining 744 averaged Rmb 78m apiece. The distribution resembles a spike, not a pyramid: a small cohort of anointed firms surrounded by hundreds competing for what remains. RMB-denominated funds now account for 95 per cent of all venture deals, which means the capital flowing into Chinese AI is overwhelmingly domestic, state-adjacent, and denominated in a currency whose allocation channels the government influences directly. (China’s state-engineered AI boom)

The mechanism extends beyond venture rounds. State capital, or guozi (国资), has become the dominant force in the larger funding rounds that confer unicorn status. Private venture capitalists may identify early-stage opportunities, but it is state-backed funds that select which companies reach the scale where survival becomes structurally secure. The result is a Capital Intensity Multiplier: Chinese chip companies, for instance, appear to require 10 to 20 times more capital than established global players to reach comparable market positions. That multiplier reflects the simultaneous cost of reinventing existing technologies, building software ecosystems from scratch, competing for scarce talent, and localizing supply chains. MiniMax, another model lab whose cap table once featured elite dollar-denominated VCs, recently took a $300m strategic cheque from Shanghai’s state asset manager.

The Cambricon and Metax contrast illustrates the principle in practice. Cambricon, a domestic AI chip designer backed by deep state capital and major tech investors, reported Q1 2025 revenue of Rmb 1.111bn, a 4,230 per cent year-over-year increase, and Rmb 3.55bn in net profit. Its market capitalisation reached approximately Rmb 579bn. Metax, operating at a different scale, raised capital at a Rmb 21.07bn valuation while posting a Rmb 1.41bn loss in 2024 on revenue of Rmb 743m from over 25,000 shipped GPUs. Both companies build AI chips. The divergence in trajectory points toward differences in capital access and the patience of their respective backers, not solely toward differences in silicon. (Cambricon’s AI chip turning point)

At the startup layer, the Darwin Moment, the point at which capital scarcity forces startups into an existential selection event, has arrived. Zhipu AI, a Tsinghua University spinout backed heavily by state funds, generated Rmb 300m in revenue against a Rmb 2bn loss in 2024, pushing toward an IPO at roughly Rmb 40bn. Moonshot, a consumer AI startup, saw its monthly advertising budget collapse from Rmb 100m in January 2025 to Rmb 400,000 by June, squeezed out by incumbents whose war chests dwarf anything venture capital can sustain. ByteDance alone allocated over $20bn for AI chips and data centres in 2025. The startups that survive this period will be those that answered one question honestly: what can we actually build, sell, and defend given who backs us and what our market will pay for?

That capital structure is itself a product of deliberate design.

2. The state is the market’s architect

The Chinese government does not referee technology markets. It designs them, funds their construction, plays in them, and decides when overcapacity has served its purpose. The recurring pattern is the ripen-then-glut cycle: the state accelerates an industry’s growth through directed capital and policy support, rides the resulting overcapacity, then channels the fallout differently depending on the industry’s structure.

The mechanism has three phases. In the first, state capital floods a sector to compress the timeline from laboratory to market. The central government launched an $8.2bn national AI investment fund in January 2025 and followed with an “AI+ Action Plan” in August 2025, requiring local governments to produce measurable implementation schemes. In the second phase, the capital achieves its purpose: capacity exceeds demand. EVs have already crossed this threshold. AI models, chips, and humanoid robotics remain in what might be called the ripening phase, with early signs of surplus appearing in models and policy-backed procurement flowing into robotics. In the third phase, the exit route depends on the product. Manufacturing industries like EVs export their overcapacity. AI models, which cannot be shipped like cars, deflate into the domestic enterprise market as costs fall. Chips consolidate under continued policy direction. Each industry follows the same cycle but exits through a different door.

The Hefei Model offers the clearest case study. In 2008, the city government committed approximately $1.3bn to help BOE Technology, a domestic display maker, build a next-generation production line. That sum approached half of the city’s entire fiscal revenue for the year. The government acted as a strategic anchor investor, de-risking the project for private capital. BOE raised Rmb 12bn by June 2009. The factory cluster now produces 20 per cent of the world’s LCDs. The government exited its initial stake at a profit estimated at roughly 300 per cent. (Inside China’s government VC model)

Hefei replicated the playbook twice more. In April 2020, the city rescued NIO, a premium EV maker reportedly rejected by 18 other cities, with a Rmb 7bn injection for a 24 per cent equity stake. Hefei’s investors earned a substantial profit within 18 months. Then came CXMT, China’s national champion in DRAM memory chips, which the city incubated with approximately 75 per cent of initial capital. CXMT’s factory was built in 10 months. Its pending IPO carries a valuation of roughly Rmb 150bn ($21bn). Three industries, three time horizons, one playbook: the government provides the patient capital that compresses development timelines, absorbs early losses, and creates the conditions for a domestic industry to reach self-sustaining scale.

The venture capital data confirms the pattern at the national level. The same spike visible in AI funding, where the top 20 firms capture 30 per cent of all capital while 744 competitors split the remainder, reflects a system where state capital selects a small number of firms for strategic cultivation. Private venture investors identify candidates. The state decides which ones scale.

The market these architects are building has a specific shape.

3. The layered market is the durable advantage

China’s market advantage rests less on size than on the simultaneous coexistence of multiple technology maturity levels and demand tiers, so a product that qualifies as good enough can find a paying customer before it qualifies as world-class. This is a demand-side structural feature, distinct from the supply-side forces of capital and policy described above. The better term may be right-sized intelligence: a model does not need to be frontier-grade if the task, price point, and deployment environment only require practical AI. Application innovation persistently outruns frontier innovation because the layered market gives each iteration somewhere to land.

The mechanism works across two dimensions. The first is temporal: technologies at different stages of maturity operate in the same economy at the same time. A frontier reasoning model serves a cloud provider’s most demanding enterprise clients. A 67-billion-parameter model, far from the frontier, handles government paperwork in Shenzhen’s Futian district at 95 per cent accuracy while cutting review times by 90 per cent. Both deployments generate revenue. The second dimension is spatial: top-tier cities, smaller-city markets, and industrial settings have structurally different needs, price sensitivities, and infrastructure constraints. A product designed for a Tier 1 cloud customer and a product designed for a factory floor in a smaller city may use the same underlying model family at different scales. The layered market supports both without requiring either to dominate.

The consequence is that winner-take-all dynamics, familiar from American platform markets, appear less likely to hold in China. The global small language model market is projected to grow from $0.93bn in 2025 to $5.45bn by 2032. Chinese companies are capturing disproportionate share through three factors: data sovereignty requirements that favour local deployment, cost sensitivity in manufacturing, and infrastructure realities in smaller cities. The good enough threshold turns out to be commercially productive, not merely a consolation prize.

China’s AI toy industry illustrates how layered demand creates entire product categories. The market reached roughly Rmb 25bn ($3.5bn) in 2024, accounting for 38 per cent of the global total, with projections pointing toward Rmb 100bn by 2030. Over 1,500 AI toy companies were operating in China by late 2025. On JD.com, AI toy sales for children aged 3 to 6 surged six-fold in February 2025. BubblePal shipped over 200,000 units. FoloToy reached 20,000 units in Q1 2025, projects 300,000 for the year, and sells across 10 countries. These products run on open-source models and commodity voice modules manufactured minutes away from AI chip suppliers in Shenzhen and Dongguan. The layered market provided the demand tiers, the supply chain provided the iteration speed, and cheap inference provided the economics. Western toy companies were still debating the ethics of conversational AI for children while Chinese manufacturers were shipping millions of units. (China’s $3.5bn AI toy industry)

A common misreading: Western analysts frequently apply winner-take-all frameworks developed in American platform markets to Chinese technology sectors. The expectation that a single cloud provider, a single model lab, or a single chip vendor will consolidate each category the way their US equivalents have tends to underestimate the structural persistence of layered demand. The layered market structure suggests that high concentration is harder to achieve and sustain. The layered market does not prevent dominant players from emerging, but it suggests that the path to dominance runs through serving multiple tiers rather than collapsing them into one.

If the market is layered, the question becomes which layer captures the most value.

4. Pipes beat apps

Value in China’s AI economy accrues below the application layer. Memory chips, cloud inference platforms, payment rails, and device operating systems capture margin more durably than the chatbots and consumer apps that sit on top. The model layer commoditises fastest precisely where the layered market is strongest. In China’s AI economy, the scarce asset is not the chatbot. It is the metered, billable infrastructure that makes the chatbot cheap enough to use at scale.

The token economy illustrates the mechanism. The 140 trillion daily tokens described in the opening are not merely a volume statistic. They represent billing events flowing through cloud infrastructure. Agent workloads consume tokens at 10 to 100 times the rate of conversational chat. That ratio shifts the economics of the entire stack. ByteDance’s Volcano Engine, Alibaba Cloud, and Tencent Cloud are each pricing and measuring AI infrastructure around token throughput. ByteDance’s daily cloud model throughput reached 120 trillion tokens by April 2026, growing a thousand-fold from its May 2024 launch. JPMorgan estimates that China’s inference token consumption will grow roughly 370-fold between 2025 and 2030. At this scale, whoever operates the most efficient token delivery infrastructure holds a structurally advantaged position. The signal extended beyond China’s borders. In the week of March 16, Chinese AI models generated 7.36 trillion tokens on OpenRouter, 36 per cent of global volume, surpassing American models for the third consecutive week. IDC put China’s MaaS market at Rmb 710m in 2024; the category is projected to exceed 30 per cent of total cloud income. The pipes, not the applications running through them, capture the rent. (The token economy China is building)

The same logic extends to hardware. When Samsung exited China’s TV and home-appliance market in May 2026, investors rewarded the move. The semiconductor division generated 61 per cent of total revenue and 94 per cent of operating profit. Samsung holds approximately 37 per cent of global DRAM and 28 per cent of NAND flash. Chinese brands had captured 94.1 per cent of domestic TV shipments. Samsung lost the consumer-facing layer where Chinese companies hold decisive advantages in assembly, channel management, and price competition. It concentrated on the upstream layer where it holds pricing power: advanced memory fabrication. The market treated the retreat as overdue housekeeping because the value had already migrated to the pipes. (Samsung’s value migration to memory)

The infrastructure thesis is also visible in China’s agent economy. Alipay AI Pay surpassed 100 million users and processed over 120 million transactions in a single week. Xiaomi committed Rmb 60bn ($8.7bn)to AI over 3 years, with over 800 million connected IoT devices providing a surface for system-level agent deployment. Baidu proposed measuring Daily Active Agents, or DAA, as the output metric for an era where agents, not humans, generate the bulk of compute demand. The competitive question these companies are answering is not who builds the best agent. It is who builds the plumbing that agents run on: the inference infrastructure, the payment authorisation layer, the device operating system, the orchestration platform that minimises wasted tokens.

One executive framed the implication: the agent era does not belong to whoever burns the most compute. It belongs to whoever uses it wisely. That distinction between volume and efficiency points toward a new pricing problem. Alibaba Cloud’s Pro coding plan, priced at Rmb 200 per month, sells out by 9:30 each morning. One Anthropic Claude Max subscriber generated $5,600 in equivalent API costs on a $100 monthly plan. Both cases reflect the same structural issue: subscription pricing built for chat-era consumption cannot absorb agent-scale workloads. The companies solving that problem operate below the application layer.

The infrastructure layer faces its own constraint: much of it must be built without access to the global frontier.

5. Parallel substitution, not catch-up

China’s structural disadvantage in semiconductors and its lack of a stable technology alliance with the West appear to be durable conditions under current geopolitical arrangements rather than transitional ones. The operating strategy is the construction of a parallel good enough stack: a complete, domestically controlled technology chain that functions at commercially acceptable performance, even where it trails the global state of the art. Company moves that look like grand strategy are better read as forced adaptation to this constraint.

The logic begins with the chip gap. Metax, in a regulatory filing of unusual candour, acknowledged that the vast majority of domestic manufacturers maintain a 1 to 2 generation gap with Nvidia’s latest products. Its C500 series benchmarks against the A100, a chip Nvidia released in 2020. Its C700 targets H100 performance from 2022. Metax’s software ecosystem has roughly 15,000 registered developers, compared with millions for Nvidia’s CUDA platform, a gap of approximately 3,000 to 1. These constraints appear structural given current manufacturing access and are unlikely to close through engineering effort alone.

But the operating logic points away from closing the gap on Nvidia’s terms. Chinese model labs are instead reshaping workloads around domestic hardware constraints. DeepSeek’s V4 technical report reads less as a model benchmark and more as a compute-shaping document. V4-Pro uses 27 per cent of the single-token inference FLOPs and 10 per cent of the KV cache of its predecessor in a one-million-token context setting. The report goes further, offering specific proposals to hardware vendors on power headroom, cross-GPU signalling latency, and activation function design. Frontier model labs are beginning to tell the hardware industry what useful compute should look like. DeepSeek’s API pricing page notes that V4-Pro throughput is constrained by high-end compute availability, with an expected price reduction once Huawei’s Ascend 950 series reaches volume production in the second half of 2026. The model’s commercial roadmap is explicitly timed to a domestic chip delivery schedule. (DeepSeek’s compute-shaping bet)

The training stack has also shifted. For 2 years, the working assumption held that Chinese chips could handle inference but not the far more demanding work of training models from scratch. In late June 2026, Meituan released LongCat-2.0, a trillion-parameter model pre-trained on a 50,000-chip domestic cluster, ingesting over 35 trillion tokens without a single rollback. Separately, a research consortium completed full-parameter post-training of DeepSeek-V4-Pro on over 1,000 Ascend 910C chips with zero interruptions. Huawei’s own Pangu Ultra 135B pre-trained 13.2 trillion tokens on 8,192 Ascend NPUs. Chinese domestic chips have entered the training stack. The pace of entry appears to be increasing. The constraint is not only chip design. SMIC’s 7nm capacity is expanding from roughly 45,000 wafers per month at the end of 2025 to approximately 60,000 in 2026, while Huawei targets roughly 750,000 Ascend 950PR units in the same period. Bernstein analysts project that Chinese domestic AI chip suppliers’ market share will rise to 55 per cent by 2027. The gap is not disappearing. It is becoming a question of system-level economics rather than a binary of can or cannot.

The more useful reading is commercial rather than ideological. External observers often treat supply chain localisation as nationalist strategy. In practice, many procurement decisions follow a simpler logic: when Nvidia chips become unavailable or restricted, firms that need compute source alternatives. When domestic chips cross the good enough threshold, procurement can shift for commercial reasons, not only political ones. Geopolitics shapes the constraint. Commerce shapes the response. Conflating the two obscures the mechanism that actually drives company behaviour.

The parallel stack describes how China is rebuilding the supply side under constraint. The unresolved question is who controls the demand side: the interface through which users, agents, and services meet.

6. The interface war decides the next decade

Whoever owns the interface between user intent and service execution taxes everyone below it. China’s technology giants are fighting to replace the search box and the app grid with a conversational layer so deeply embedded in daily life that switching becomes prohibitively costly. This is the one arena covered by these principles where the outcome remains genuinely open.

The fight has two fronts. The first is the Do-Engine: an AI-powered interface that does not merely find information but synthesises, plans, and executes. Alibaba’s Quark and Tencent’s Yuanbao each spent roughly Rmb 1bn per month on advertising in mid-2025. The logic is defensive. If a rival builds the dominant Do-Engine, it can disintermediate every other ecosystem. A user who asks an AI agent to find the cheapest latte across three delivery platforms turns those platforms into interchangeable commodity suppliers. Baidu has integrated over 18,000 third-party services into its AI framework. The spending is about seizing the conversational operating system for daily commerce, not building a better chatbot. (The Do-Engine war)

The second front is the contest over who mediates between AI agents and the services they need to access. In December 2025, ByteDance launched Doubao Phone Assistant on ZTE’s Nubia M153. The product used a GUI Agent approach: rather than integrating with apps through official APIs, it operated at the system level, reading screens and simulating finger taps. Within days, WeChat forced users to log out. Taobao triggered CAPTCHA verifications. ByteDance retreated, disabling WeChat functionality. Simple tasks took roughly 30 seconds, compared with roughly 10 seconds for a human. The first batch comprised only 30,000 units at a retail price of Rmb 3,499. Secondhand prices ran Rmb 700 to 1,500 above retail. ByteDance was not trying to compete in smartphones. It was testing whether a forbidden path could work at all, because the permitted path, API cooperation, requires permission from the platforms the agent aims to disrupt.

By June 2026, Tencent had reversed course selectively. WeChat opened a controlled A2A channel, a Tencent-defined Agent-to-Agent access framework distinct from the open A2A protocol promoted by Google, to phone manufacturers including Huawei, Honor, Xiaomi, OPPO, and Vivo. Users could ask their phone assistant to send a WeChat message or make a call. Honor completed integration across three device series, with 50 per cent of online active devices connected. The mechanism keeps execution inside WeChat. The phone assistant initiates the request; WeChat processes it within its own environment. Tencent defined the scope, granted access to communication functions, and retained the ability to expand or withdraw permissions. The door opened, but Tencent designed the door. (WeChat as AI gatekeeper)

The protocol power at stake is structural. WeChat has 1.4 billion monthly users. Tencent’s standalone AI app, Yuanbao, had roughly 57.35 million monthly active users in March 2026, compared with 345 million for Doubao and 166 million for Qwen. Tencent’s capex in 2025, approximately Rmb 79bn, trailed ByteDance’s reported Rmb 150bn and Alibaba’s Rmb 380bn three-year commitment. WeChat’s ecosystem remains Tencent’s strongest card. Any phone agent that cannot interact with WeChat appears strategically incomplete in China. That gives Tencent protocol-level leverage over the interface war even while its standalone AI product lags. Whether that leverage holds depends on whether WeChat’s forthcoming AI agent delivers enough value to keep intent flowing through its ecosystem, or whether users begin routing tasks around it.

Part II: The Evidence So Far

The working hypotheses behind these principles took shape during the first months of Hello China Tech’s coverage, starting in July 2025. An internal working document from October 2025 codified them as five principles. Twelve months of published analysis now provide enough evidence to score those hypotheses against outcomes. The cases below are divided into what the analysis anticipated and what it did not. The wins and misses use the same evidentiary standard: specific claims matched against specific facts.

What the framework called

The good enough threshold in AI chips. The working hypotheses codified in October 2025 argued that finding a good enough substitution path mattered more than closing the gap with global leaders. They argued that parallel substitution paths would emerge, not the most advanced, but sufficient to support the majority of application scenarios.

That argument has been tested commercially. Cambricon’s Q1 2025 results, detailed in Principle 1, showed the financial side of good enough in action. The technology side was equally telling. Its Siyuan 590 chips deliver approximately 80 per cent of Nvidia A100 performance at 30 per cent lower cost than Nvidia’s China-specific H20 in certain scenarios. Revenue recognition shifted from project-based contracts concentrated in Q4 to consistent quarterly sales, suggesting a move from one-off government procurement toward recurring commercial demand. Gross margin stabilised around 70 per cent, comparable to Nvidia’s data centre business.

Metax, operating at a different scale, confirmed the structural picture from the other side. Its 300-page regulatory filing for the Shanghai Stock Exchange confirmed the generation gap detailed in Principle 5. But Metax secured a $3bn valuation on what amounts to a security premium logic, where customers value supply chain independence over marginal performance advantages. The framework’s claim was that good enough would find a paying market. The evidence suggests the market arrived faster than the October 2025 version anticipated.

The paper-order problem. The framework identified three risks in the robotics sector: technology and manufacturing cost barriers, limited deployment scenarios, and excessive short-term hype. It placed the real opportunity roughly 5 years out, based on a technology maturity assessment.

The order wave that followed appeared to challenge the timeline. UBTECH announced a Rmb 250m contract, a record for the sector. Galbot secured contracts exceeding Rmb 700m. China Mobile awarded a Rmb 124m joint tender to AgiBot and Unitree Robotics. The headline numbers suggested the market was arriving faster than expected.

Closer examination supported the framework’s caution rather than refuting it. Tiantai Jiqiren announced a 10,000-unit order worth billions of renminbi. The purchasing company had been established 12 days before the deal announcement, with overlapping personnel between buyer and seller. Unitree Robotics, in its IPO filing, disclosed that 73.6 per cent of its humanoid robot revenue in the first 9 months of 2025 came from research institutions and universities. Its adjusted net profit fell 52.55 per cent year-over-year in Q1 2026 as spending surged. The order rush was real. So was the paper-order problem. The framework correctly identified the hype pattern. Its 5-year timeline proved slightly conservative on order announcements but appears closer to accurate on actual commercial deployment readiness.

The chip bag test for deployment readiness. Galaxy General, China’s best-funded embodied AI startup, raised nearly Rmb 7bn and reached a valuation exceeding Rmb 20bn on a thesis that the robot’s brain, its AI model, matters more than its body. A FamilyMart convenience store on the ground floor of its Beijing headquarters became the test case.

The store carries roughly 1,000 SKUs. Galbot handles fewer than 10 of them. A bag of Lay’s potato chips, pressurised with food-grade nitrogen, defeats the robot’s grippers roughly 4 times in 10. The robot is priced at approximately Rmb 700,000. Four full-time employees staff the store, the same headcount as a comparable FamilyMart without a robot. Galaxy General’s purchases from Unitree, a competing hardware company, surged from Rmb 364,700 in 2024 to Rmb 18.19m in 2025, a roughly 50-fold increase. Dexterous hands demonstrated at the Spring Festival Gala came from Sharpa, a specialist manufacturer, at a cost of several hundred thousand renminbi per pair.

The framework argued that technology maturity, manufacturing costs, and deployment limitations would constrain the sector. Galaxy General’s convenience store deployment provided a measurable version of those constraints. The brain-first thesis encounters a trilemma: sensing, execution, and cost constrain each other simultaneously. The AI model does not yet compensate for simple grippers on deformable packaging, does not yet eliminate latency from cloud-based voice processing, and does not yet bring dexterous-hand costs to a level retail can absorb. At Rmb 700,000 per unit, a convenience store operator could hire several full-time workers for years. The framework identified the right failure mode. The chip bag gave it a concrete name.

What it missed

CATL’s data centre playbook. The early analysis discussed EV battery overcapacity and correctly identified export as the primary exit route for manufacturing surplus. It did not anticipate that a dominant battery company would begin assembling a vertical energy chain for AI data centres.

The wrong assumption was specific: the framework treated each industry’s ripen-then-glut exit as relatively contained. EVs would export. Models would deflate into domestic enterprise. Chips would consolidate under policy. The framework did not consider that a company dominant in one cycle could use overcapacity as a launch point into an adjacent value chain that the framework had not yet mapped.

CATL invested Rmb 4.1bn for a 49 per cent stake in the controlling shareholder of Zhongheng Electric, a provider of high-voltage direct current power systems deployed in data centres. It agreed to pay up to $942m for as much as 38.1 per cent of VNET Group, a data centre operator running more than 50 facilities with 889 megawatts of wholesale capacity. Reports put its commitment to DeepSeek’s fundraise at roughly $740m. Total exposure potentially exceeds Rmb 15bn. Robin Zeng, CATL’s founder, estimated the data centre energy opportunity at “ten times” the size of EV battery supply. The company expects energy storage to account for half of global sales by 2030.

The framework’s ripen-then-glut cycle can explain the motivation retrospectively: battery penetration in China passed 60 per cent, margins face compression from second-tier manufacturers, and a company locked into hardware sales needs a new growth vector. The pattern of securing critical links across a value chain, from upstream materials through midstream power equipment to downstream channels, is structurally identical to how CATL captured the lithium battery chain. But this is framework extension after the fact, not prediction. The original working hypotheses did not foresee data centres as the destination, did not consider cross-industry value-chain migration as a category of exit from overcapacity, and did not anticipate that a battery company would invest alongside a frontier AI lab. The miss reveals that the exit paths from state-driven overcapacity are more diverse than the framework’s initial categories assumed. CATL is replaying the battery playbook in a domain the framework left blank. (CATL’s data centre playbook)

The speed of the interface war. The early analysis discussed common misreadings of Chinese corporate behaviour and cautioned against over-interpreting commercial decisions through a geopolitical lens. It said nothing about super apps versus AI agents, nothing about the contest for the conversational interface layer, and nothing about what would become the most structurally consequential competitive dynamic of the period that followed.

The wrong assumption was one of omission. The framework treated the application layer as a venue for iteration and competition, shaped by the layered market and the ripen-then-glut cycle. It did not identify a specific mechanism by which platform incumbents would use protocol power to determine which AI agents could access which services, or by which new entrants would attempt to bypass that control.

In December 2025, ByteDance launched Doubao Phone Assistant, the GUI Agent product described in Principle 6. WeChat shut it down within days. By June 2026, Tencent had reversed course selectively, opening a controlled A2A channel to phone manufacturers while keeping execution inside WeChat’s environment. Tencent’s stock rose 10.5 per cent on reports that it was testing a WeChat-native AI agent. The interface war arrived hotter and faster than anything the framework’s original framing anticipated.

The original working hypotheses contained no analysis of platform gatekeeping as a structural force. It had no discussion of GUI Agent versus API cooperation as competing access models, no assessment of how WeChat’s 1.4 billion users would function as a chokepoint in the agent economy, and no anticipation that the question of who controls the conversational interface would become the single most capital-intensive competitive contest in Chinese technology during 2026. This gap is the primary reason the 2026 Edition adds a sixth principle. The interface war was not a minor update to an existing category. It was a structural force the framework failed to see.

Scoring the ledger

The original working hypotheses identified the right structural forces in AI chip substitution, robotics hype, and the gap between announcements and deployment readiness. It missed the speed and shape of two developments that proved consequential: CATL’s cross-industry expansion into data centre energy and the emergence of the interface war as a distinct competitive arena. The cases above are illustrative, not exhaustive. They were chosen because each maps directly to a principle’s core claim. The full record is public: every essay behind this framework remains timestamped in the archive, open to the same scoring applied here.

Part III: How to Use This

The three-lens test

Each principle above describes one structural force. Applying them requires reading any sector, company, or policy development through three lenses simultaneously. Omitting any one of the three produces a distorted picture.

The first lens is global technology state. Where does a given technology sit on its maturity curve, and what is the gap between China’s capability and the current global frontier? This lens prevents overestimating a domestic advance that remains several generations behind state of the art. Applied to AI chips: a domestic GPU achieving 80 per cent of a 2020 Nvidia architecture occupies a different competitive position from one matching current-generation performance, even if both qualify as good enough for their respective market segments.

The second lens is the state’s policy position within the ripen-then-glut cycle. Is the government still pumping capital into the sector, or has overcapacity already triggered the exit phase? This lens determines whether a company’s growth reflects genuine demand or policy-driven acceleration that will eventually reverse. Applied to humanoid robotics: the current wave of record-setting orders arrives during the ripening phase, when state capital flows freely and procurement announcements carry political as well as commercial incentives. The same order book would read differently if it appeared after consolidation had begun.

The third lens is layered market demand. Which tier of the market is the product actually serving, and does demand persist without subsidy? This lens separates commercial traction from policy-supported demonstration projects. Applied to AI toys: products selling hundreds of thousands of units through consumer e-commerce channels at unsubsidised prices represent a different quality of demand signal from government-procured pilot deployments.

All three lenses are required because each corrects for the others’ blind spots. Technology state without policy context misses the role of state capital in compressing timelines. Policy position without demand analysis misses whether the market the state is building will sustain itself. Demand analysis without technology state misses whether the product can improve fast enough to retain the customers it acquires.

Real innovation versus policy arbitrage

Within any sector, distinguishing between companies that have built durable capability and those riding temporary policy tailwinds requires three tests.

First, does demand survive without subsidy? Strip away government procurement preferences, state-fund investment, and policy-mandated adoption targets. If the remaining demand base cannot sustain the company’s unit economics, the business is a policy arbitrage position, not a technology company.

Second, is the technology path dependent on external access that could be revoked? A company whose core product relies on components, manufacturing processes, or software ecosystems controlled by parties outside China’s policy reach faces a constraint that no amount of domestic capital can eliminate on a short timeline. This does not disqualify the company, but it changes the risk profile and the appropriate valuation framework.

Third, do unit economics hold at scale? A product that works at pilot volume but requires subsidised inference, below-cost hardware, or loss-making pricing to reach commercial adoption has not yet crossed the threshold from demonstration to business. The test is whether the company can deliver its product profitably at 10 times its current volume without external support.

What would prove this wrong

A framework earns credibility by specifying the conditions under which it would require revision. Each principle below includes one observable test. If the stated condition materialises, the principle needs to be downgraded, rewritten, or abandoned.

Principle 1, capital structure beats technology. If a Chinese AI chip or model company reaches sustainable profitability at scale without state capital backing, relying entirely on private venture funding and commercial revenue, the principle’s claim that capital structure is the primary survival determinant would need qualification.

Principle 2, the state is the market’s architect. If a major technology sector completes its development cycle with private capital as the dominant funding source and state funds playing a minority role, the principle overstates the state’s architectural function. The sector to watch is frontier AI models, where DeepSeek’s trajectory suggests a path less dependent on direct state investment than the principle assumes.

Principle 3, the layered market is the durable advantage. If a single company captures more than 60 per cent market share in any major Chinese technology category and sustains it for two consecutive years, the winner-take-all dynamic the principle argues against would have a live example. The structural persistence of layered demand would be in question.

Principle 4, pipes beat apps. If consumer-facing AI application companies achieve stable profitability with margins consistently exceeding those of the infrastructure providers underneath them, the value-accrual thesis reverses. The companies to track are ByteDance’s Doubao and Alibaba’s Quark: if either builds a sustainably profitable consumer AI business that outearns its cloud division, the principle requires revision.

Principle 5, parallel substitution, not catch-up. If domestic AI chips cannot narrow the training efficiency gap to within 2 times Nvidia’s cost-performance ratio within the next product cycle, the parallel stack remains confined to inference and lower-value workloads. Parallel substitution would then describe a partial strategy rather than a comprehensive alternative.

Principle 6, the interface war decides the next decade. If operating-system-level AI agents from device manufacturers successfully bypass super-app gatekeeping and provide end-to-end service access without platform permission, the protocol power thesis weakens. The test case is whether Huawei’s HarmonyOS or Xiaomi’s HyperOS can route transactions, payments, and social interactions through system-level agents without requiring WeChat’s A2A authorisation.

Closing

This is the 2026 Edition of The China Tech Field Guide. It will be re-scored and re-issued annually, each July. The six principles, the evidence record, and the falsification conditions will be updated against whatever the following 12 months provide. Principles that hold up will be sharpened. Principles that break will be replaced. The 2027 Edition will carry a longer track record, a more demanding audience, and the accumulated weight of being wrong in public.

The analysis was conducted in public. This framework consolidates it for the first time. Every principle links to the published essay from which it was derived. The scored ledger in Part II applies the same standard to hits and misses.

For ongoing analysis of the sectors covered here, subscribe to Hello China Tech. Paid subscribers receive the analysis behind this framework as it develops: company-level breakdowns, earnings deep dives, and the ongoing record of which principles are being strengthened, revised, or challenged by new evidence.

Prefer to read this as a PDF? Download the 2026 Edition, formatted for sharing with a colleague or a reading list.

Further reading

Essays referenced in this Field Guide, grouped by principle.

Principle 1: Capital structure beats technology

• A Spike, Not a Pyramid: China’s State-Engineered AI Boom — How state capital reshaped AI venture funding distribution

• Good Enough to Matter: How Cambricon Redefined China’s AI Chip Game — Cambricon’s revenue breakout and the economics of forced innovation

• China’s AI Startups Enter Their Darwin Moment— Three survival paths for startups squeezed by big tech and capital scarcity

Principle 2: The state is the market’s architect

• The $21B Bet: Inside China’s Most Successful Government VC Story — The Hefei Model from BOE to NIO to CXMT

Principle 3: The layered market is the durable advantage

• China’s Practical AI Play: The Case for Right-Sized Intelligence — Why smaller models find paying markets in China’s layered demand structure

• China Shipped Millions of AI Toys While the West Was Still Debating — How supply chains, cheap inference, and layered demand created a $3.5bn category

Principle 4: Pipes beat apps

• 140 Trillion a Day: The Token Economy China Is Building — Token consumption as economic indicator and the cloud infrastructure race

• The Token Reckoning — Why AI coding plan economics point toward token efficiency over token price

• Samsung Lost China’s Living Rooms, Memory Became the Prize — Value migration from the consumer layer to upstream memory fabrication

• China Builds the Plumbing, Not the Agent — Baidu’s DAA metric and the infrastructure layer for China’s agent economy

Principle 5: Parallel substitution, not catch-up

• The Metax Paradox: Transparency, Deficits, and the Investment Case for China’s AI Chips — What a 300-page regulatory filing reveals about the chip gap and the security premium

• DeepSeek Is Turning China’s Chip Problem Into a Compute-Shaping Bet — How model architecture reshapes hardware requirements

• China’s AI Chips Enter the Training Stack — Domestic chips move from inference to frontier-scale training

Principle 6: The interface war decides the next decade

• The Billion-Dollar Bonfire: China’s AI War and the Creation of the ‘Do-Engine’ — The defensive spending logic behind the conversational OS race

• The New Interface War: Why ByteDance Had to Launch an Agent No Platform Would Approve — ByteDance’s GUI Agent gambit and the structural constraints behind it

• WeChat Becomes the Gatekeeper for China’s AI Agents — How Tencent’s A2A framework defines AI agent access in China

Part II: Evidence

• Billion-Dollar Bets or Paper Orders: China’s Humanoid Robot Rush — Order quality hierarchy in China’s humanoid robotics sector

• Galaxy General and the Chip Bag Test — What a convenience store reveals about the brain-first thesis in embodied AI

• CATL Replays the Battery Playbook — CATL’s vertical energy chain for AI data centres

The River With No Banks

Yuan Jinhui likes to compare AI training demand to a swimming pool: fixed volume, known dimensions. Inference demand, he says, is a river with no boundaries.

Yuan spent his first startup building OneFlow, a training framework for deep learning. When he founded SiliconFlow in August 2023, he moved to the other side of the equation. A token-supply platform is not a model lab or an application company. It is the middleware layer that turns rented compute and third-party models into metered AI output.

Thirty-four months later, SiliconFlow has become China’s largest independent token supply platform. As of April 2026, it processed an average of 578.5 billion tokens per day across more than 170 models, serving more than 10 million registered users and over 13,000 enterprise customers. The top three token suppliers by throughput, Volcengine (ByteDance’s cloud arm), Alibaba Cloud, and Baidu AI Cloud, are all hyperscaler divisions. SiliconFlow holds 1.5% of the market, trailing the third-place player by ten percentage points. On June 30, the company filed for a Hong Kong listing under Chapter 18C, the exchange’s framework for specialist technology companies.

The filing matters beyond SiliconFlow. It offers one of the clearest audited looks yet at the economics of a pure-play token supplier. Major cloud providers sell tokens, but their inference economics sit buried inside larger businesses. SiliconFlow’s prospectus isolates the layer.

The river Yuan described is real. Daily token throughput grew twelvefold in sixteen months. Revenue rose 653% to Rmb 55.3m. But the prospectus also reveals what the river costs to sustain, and those costs complicate the growth story considerably.

The Prospectus X-Ray

SiliconFlow runs two businesses with opposite economics.

Its public cloud service, including serverless token APIs and dedicated compute instances, generated 52.9% of 2025 revenue. This is the growth engine and the core of the “Token Factory” narrative. Its gross margin was negative 119%, an improvement from negative 271.6% in 2024 when the service had just launched, though the gap to breakeven remains wide. For every renminbi of revenue, SiliconFlow spent Rmb 2.19 in direct costs, with compute rental alone consuming 86.9% of cost of goods sold. Within public cloud, serverless token services and dedicated compute instances each contributed roughly half of revenue.

SiliconFlow occupies a pure intermediary position in this market, renting compute and packaging third-party models rather than owning either input. And it sells into a price war: prices for mainstream models have fallen more than 90% since 2023. In May 2026, DeepSeek, the Hangzhou-based AI lab, made a permanent 75% cut to its V4-Pro API price. Xiaomi and Tencent Cloud followed with maximum reductions of 99% and 97.5%, respectively.

The company’s other line, on-premise deployment of inference software inside enterprise data centers, earned an 82.5% gross margin. But this business is project-based and difficult to scale. Customer count fell from 28 to 20 in 2025, even as average revenue per customer rose nearly sixfold.

The aggregate picture: gross margin swung from positive 39.4% in 2024 to negative 24% in 2025. The shift was largely a revenue-mix effect. Public cloud barely existed in 2024. As it became the majority of revenue, blended margins turned negative, even though public cloud’s own unit economics improved year over year. Net loss reached Rmb 345.5m; adjusted for share-based compensation, the figure was Rmb 187m, still more than three times revenue. More than 64% of the company’s Rmb 83.7m in selling and marketing expense went to promotional compute credits, essentially giving tokens away to acquire developers. The strategy accompanied rapid volume growth in serverless token services: paying accounts rose from 2,455 to 716,000 in a single year. But the arithmetic suggests thin commercial density. Serverless token revenue of Rmb 14.3m divided by the 716,000 year-end paying accounts implies average spending of roughly Rmb 20 per customer.

The prospectus offers a rare answer to a question usually hidden inside cloud divisions: what does it cost to be an independent token supplier when your cloud vendors, chip suppliers, and direct competitors can be the same companies? The more revealing question is what these numbers look like beside the American companies building the same infrastructure layer, and why the gap goes deeper than scale.

Read more

For 2 years, the working assumption among investors and policymakers has been clear: Chinese AI chips could handle inference, the computationally lighter task of generating outputs from models that have already been built. Training, the far more demanding process of building one, still required Nvidia.

That assumption now has a LongCat problem.

In late June, Meituan, China’s largest on-demand services platform, released LongCat-2.0. The company says the trillion-parameter model was pre-trained from scratch on a 50,000-chip clusterpowered entirely by domestically produced processors. It did not name the chip supplier, but technical indicators, including its use of HCCL, Huawei’s collective communication library, strongly point to the Ascend ecosystem. The model ingested over 35 trillion tokens during pre-training, according to its model card, without a single rollback or loss spike. The company highlighted engineering reliability measures including deterministic operators, bit-level consistency checks, and automated failure recovery.

Separately, a research consortium including Huawei completed full-parameter post-training of DeepSeek-V4-Pro, a 1.6-trillion-parameter mixture-of-experts model from Chinese AI lab DeepSeek, on a cluster of over 1,000 Ascend 910C chips. The run finished more than 1,500 training iterations with zero interruptions, according to the Shenzhen Loop Area Institute, one of the participating research bodies.

These are not isolated experiments. Huawei’s own Pangu Ultra 135B, a dense model, was pre-trained on 8,192 Ascend NPUs with 13.2 trillion tokens. A larger mixture-of-experts variant, Pangu Ultra MoE 718B, trained on roughly 6,000 NPUs and reported 30% MFU. Baidu, the search and AI group, says a key version of its ERNIE 5.1 model was trained on a cluster powered by its Kunlunxin chip unit, though it did not specify which stage of training was involved. Chinese industry reports link the effort to Baidu’s Kunlun P800 chip and cite a 97% effective training rate.

Each data point, read alone, is a progress report. Read together, they point toward a structural change in where domestic chips sit in the AI development pipeline: Chinese domestic chips have entered the training stack, and the pace of entry appears to be increasing. That a local-services company can now pre-train a model at frontier scale on domestic hardware suggests the capability is no longer confined to one or two national champions running showcase experiments.

The assumption under pressure

US export controls rest on a specific bottleneck theory. Inference requires less compute. Training requires far more. By restricting access to Nvidia’s highest-performance chips, the controls aimed to keep Chinese labs dependent on foreign hardware for the compute-heavy work of building models.

That theory is now under test. According to Morgan Stanley data compiled by the Financial Times, Chinese vendors’ AI chips already outperform the Nvidia H20, the most capable chip approved for export to China, on several metrics including processing performance and memory bandwidth. Huawei has begun marketing systems that pair its Ascend 950DT chips with DeepSeek’s V4 model for customers in the Middle East and Central Asia, according to people cited by the FT. Bernstein analysts described Huawei’s recent chip advances as another “DeepSeek moment.” The firm’s analyst Lin Qingyuan said the development “breaks the core narrative that because of export controls China’s semiconductor is dead at 7nm.”

But crossing the training threshold and closing the training efficiency gap are different questions. The first now has meaningful evidence behind it. The second remains open. Its resolution could shape the valuation ceiling for Chinese AI chip companies, the long-term revenue risk for Nvidia’s China business, and whether export controls function as an effective speed brake or merely an expensive detour.

The headlines read as if China has solved its training problem. The data tells a more layered story. Below, we map the evidence onto a 3-tier capability ladder, quantify the gap between post-training on 1,000 chips and pre-training on 50,000, and ask the question the efficiency data cannot yet answer.

Read more

The company with the largest estimated share of high-speed optical transceivers for AI data centers is a Chinese manufacturer that few Western investors follow closely. Zhongji Innolight, based in Suzhou, posted Rmb 19.5bn ($2.9bn) in revenue in the first quarter of 2026, a 192% increase year on year. Net profit reached Rmb 5.7bn ($850m), up 262%. Gross margin hit 46%, a company record. That single quarter’s profit exceeded the company’s entire 2024 earnings.

Optical transceivers convert electrical signals into light and back. They connect servers, switches, and GPU racks inside data centers, moving the data that AI training and inference generate. As clusters scale from thousands to hundreds of thousands of GPUs, the physical capacity to move data between processors is emerging as a constraint alongside compute itself.

Market estimates put Innolight at roughly 35% to 40% of the global 800G optical module market and 50% to 70% of the emerging 1.6T segment. It supplies more than half of Nvidia’s 800G optical modules. In the broader optical components market, J.P. Morgan ranks Innolight alongside Coherent, a U.S.-based company, as a co-leader, each with roughly 20% share. The customer list is effectively a roll call of American big tech: Alphabet accounts for 22% of Innolight’s revenue, Amazon for 11%, Meta for 6.4%. More than 90% of revenue comes from outside China. This is a Chinese company whose commercial lifeline runs through North American data center budgets.

In the CSI 300, China’s benchmark index for large-cap A-shares, Innolight now carries the largest weighting at 5%, ahead of CATL, the dominant battery maker, at 4.1% and Kweichow Moutai, the country’s most iconic liquor brand, at 3%. A decade ago, the heaviest weights belonged to banks and consumer staples. That the top slot now belongs to a manufacturer of components connecting Nvidia GPUs captures a structural shift in what Chinese capital markets consider a core asset. The company’s market capitalization approached $190bn as of June 2026. It has reportedly filed for a Hong Kong secondary listing that could raise up to $7bn.

That same month, the U.S. Department of Defense added Innolight to its Section 1260H list of Chinese military companies, citing indirect state ownership through SASAC, China’s state asset regulator, and direct subordination to the Ministry of Industry and Information Technology. A company embedded in the AI infrastructure supply chains of Google, Amazon, and Nvidia now carries a Pentagon designation as a military-linked entity. The dependency runs in both directions. American hyperscalers draw heavily on Innolight’s manufacturing capacity to build AI infrastructure on schedule. Innolight relies on American capital expenditure flows to sustain the growth trajectory that has made it one of China’s most valuable technology stocks.

English-speaking investors have two familiar frameworks for Chinese technology companies: cash-burning startups exploiting a captive domestic market, and sanctioned incumbents building parallel ecosystems. Innolight fits neither. It is a globally embedded manufacturing platform inside the AI stack. Understanding its trajectory requires tracking three forces at once: the spending cycles of American hyperscalers, the escalation path of U.S.-China technology restrictions, and an optical technology transition that could reshape where the profit pool sits.

Read more

In a FamilyMart in Beijing’s Zhongguancun area, a humanoid robot named Galbot stands behind the counter. It can retrieve a bottle of jasmine pu’er tea from a refrigerated case. Sometimes the bottle topples when Galbot sets it down. A bag of Lay’s potato chips sits on a nearby shelf. Galbot does not attempt it. By the company’s own account, grippers fail on chip bags roughly 4 times in 10, a rate too high for customer-facing service.

This FamilyMart is not standard. At 63 square meters, it is 40% smaller than a typical outlet, redesigned to fit the robot’s operational needs. It carries about 1,000 SKUs, roughly 60% of the normal range. Galbot handles fewer than 10 of them. Four full-time employees staff the store, the same headcount as a comparable FamilyMart without a robot. Galbot cannot complete a full transaction on its own; after retrieving an item, human staff handle payment. Galaxy General (银河通用), Galbot’s developer, invested in this location on the ground floor of its own headquarters and stationed engineers inside to collect training data.

Galaxy General is the best-funded startup in China’s embodied AI sector, the field focused on robots that perceive and act in physical environments. Founded in May 2023, it has raised nearly Rmb 7 billion (roughly $1 billion) in three years. Its March 2026 round of Rmb 2.5 billion set the sector’s single-round record for a third time and drew the third phase of China’s Big Fund, a state-backed vehicle better known for semiconductor investment, alongside SMIC-affiliated Zhongxin Juyuan. The Big Fund had never previously invested in embodied AI. Galaxy General’s valuation now exceeds Rmb 20 billion (roughly $3 billion). Galaxy General ships fewer than 150 robots a year. A single Galbot is priced at around Rmb 700,000.

The company’s thesis is that the robot’s “brain,” the AI model governing perception, planning, and execution, matters more than the body. Galaxy General built AstraBrain, an end-to-end model that integrates task planning, motor control, and neural coordination into a single architecture, rather than developing these modules separately as most competitors do. Many of those competitors rely on foundation models from Alibaba, ByteDance, or Huawei. Galaxy General argues that integrated, proprietary intelligence is the value layer. Morgan Stanley’s framing of the sector reinforces the capital logic behind bets like Galaxy General’s: a May 2026 report projected a $7.5 trillion global humanoid robot market by 2050, described China’s trajectory as a potential replay of its EV success, and noted that China accounted for about 46% of global humanoid robot venture funding in the first months of 2026.

Galaxy General sits at the center of that capital wave, part of the consensus machine that has driven successive funding surges across China’s AI and chip sectors. The convenience store is where the premise meets the physical world.

Galaxy General’s engineers are discovering that chip bags, ambient noise, and compute budgets create constraints no model can shortcut. What follows examines what the convenience store deployment reveals about the brain-first bet, why Galaxy General spends millions on its competitor’s hardware, and what a quieter signal from Xiaomi’s factory floor suggests about which path reaches commercial viability first.

Read more

Doubao, the AI assistant from ByteDance, TikTok’s parent company, has 345 million monthly active usersas of Q1 2026. That figure exceeds Alibaba’s Qwen and DeepSeek combined. More than 200 million people open it daily. By the metric that defined China’s consumer internet for two decades, raw user scale, Doubao leads by a wide margin.

The financial picture tells a different story. According to LatePost, one of China’s most authoritative tech publications, Doubao generates less than Rmb 1m per day in revenue, primarily from e-commerce commissions. Its daily compute costs run into the tens of millions of yuan, an estimate based on Volcano Engine’s published API pricing, Doubao’s model margins, and user behavior patterns. For a user spending 15 to 20 minutes a day on text chat, the cost is only a few fen (hundredths of a yuan). Multimodal features, from image recognition to voice and video chat, can push that compute bill several to dozens of times higher. That daily figure covers inference only. It does not include the cost of building training clusters, which require tens of thousands of AI chips plus supporting infrastructure.

One comparison sharpens the disparity. Running Doubao reportedly costs more than operating all of Bilibili, China’s equivalent of YouTube for a younger demographic. Doubao’s total daily user time is less than one-eighth of Bilibili’s. The cost per minute of user attention is roughly an order of magnitude higher. And the e-commerce layer that is supposed to monetize those users? Daily gross merchandise value sits at roughly Rmb 10m, spread across 200 million daily users. The conversion from attention to commerce is negligible.

ByteDance plans to raise 2026 capital expenditure above Rmb 200bn, roughly 60% of its 2025 profits, according to a South China Morning Post report cited by LatePost. The AI build-out is accelerating. The consumer product at its center generates almost no revenue. The gap between investment and return pointed toward a strategic reallocation.

The Lesson from San Francisco

Around April, ByteDance executives visited Anthropic. What they encountered was a company that had inverted the relationship between scale and revenue that ByteDance has long relied on.

Anthropic’s Claude Code launched in May 2025. Within 6 months it reached $1bn in annualized revenue. By February 2026 it had doubled to $2.5bn. Anthropic serves roughly 30 million daily users. Its valuation reached $965bn, surpassing that of OpenAI, whose ChatGPT serves nearly 900 million weekly consumers. The contrast is stark. With a fraction of the user base, Anthropic reached a higher valuation than OpenAI by selling enterprise and developer productivity rather than relying primarily on mass consumer chat.

Shortly after the visit, ByteDance began reallocating AI resources from consumer-facing products toward enterprise services. The sequence is suggestive. The pivot may have been forming independently. But the timing makes the visit a notable data point.

The question is whether ByteDance has an asset that can turn AI from a consumer cost center into an enterprise revenue engine. The answer begins with a product that rarely features in the company’s consumer narrative.

Read more