Prologue: The Night Nvidia Bled

January 27, 2025: Nvidia fell 17% on 3.5× the prior day's volume. That gap-down candle is the night this story opens with. Data: Sina Finance.

January 27, 2025. Monday, 9:30 a.m. Eastern.

Seven minutes after the opening bell, traders stared at their screens, sure it was a glitch. Nvidia's stock looked like someone had pulled out the floor — no earnings miss, no rate hike, no downgrade from any bank. The panic arrived out of nowhere, like a stone dropped into a still lake: the first ripple hadn't even settled before the second one capsized the boat.

By the close, Nvidia had fallen 17%, erasing roughly $589 billion in market value — at the time, the largest one-day loss in U.S. stock market history, more than two AMDs gone overnight. In a Silicon Valley group chat, someone posted a cold line: "Have we just been hit by a higher civilization?"

Who did it? The answer hid in an open-source release from a week earlier — from Hangzhou, by a company most Americans had never heard of.

But don't rush. The suspense isn't in that single release — it's in the three years behind it: the company that made the world's most valuable chipmaker shed half a TSMC in a single session had fought its way out of a Chinese charnel yard known as the "Hundred-Model Wars" — and that charnel yard is the real negative this night was printed from.

In 2023, 244 Chinese LLMs fought one another on the same stage. Three years later, far fewer remain — but the table is far from empty: the five giants are still there, five independent foundation-model startups still stand, and on the device side a dark horse like Xiaomi has broken out, with a thousand-plus vertical models each holding their own patch below.

So don't rush to ask "how many are left." The sharper question is: in this charnel yard, who survived — and how? What happened in between, that let a few kids from county towns tear a hole in Silicon Valley's compute hegemony?

By the end of this piece, you'll see: the survivors were not the richest, nor the most famous. They were the ones nobody had tied down.

I. Twilight of the First Movers

Both arrived first. ERNIE packed PPT and image generation into a single input box; Pangu states plainly that it serves 5 foundation models for enterprises. One failed to hold the consumer side; the other never aimed at it. Screenshots: Nov 3, 2026

ERNIE: A perfect hand, six losing moves

Any honest复盘 of Chinese LLMs must confront one awkward fact: the early birds all missed the morning. The person who said it first was Robin Li himself: "Why did we set out early yet arrive late? Because we lacked focus."

ERNIE's starting hand was enviable. Launched in 2019 — three years before ChatGPT. On March 16, 2023, Baidu became the first global giant to ship a ChatGPT challenger. A decade of R&D spending in the hundreds of billions of yuan, 27,000 AI patents. No other Chinese company had that war chest.

Then came the slide (QuestMobile data): 12.24 million monthly active users in December 2024 (No. 3 in China); 5.31 million by September 2025 (No. 9); out of the top ten by 2026 — while ByteDance's Doubao sailed past 400 million.

Six losing moves, ranked by weight:

First: closed weights plus paid tiers — switching off the data flywheel. In November 2023, ERNIE became the first Chinese assistant to charge consumers (¥59.9/month). Who was the template? OpenAI. GPT-4's closed, paid model was minting money, and Li's line — "open-source models are an IQ tax" — drew its confidence from exactly that playbook.

Cut back to America: same playbook, different continent, opposite outcome. Li mispriced two variables. One: China has no OpenAI-grade capital subsidy to sustain a closed, cash-burning model. Two: Chinese open-source rivals would come in like a tide. Zhou Hongyi publicly fired back, the open-source camp piled on, and the verdict was written in market share two years later — ERNIE went free in February 2025 and open-sourced in June, but the users were long gone. Hours in the kitchen; dinner still delivered.

Second: iteration gap. Over a year passed after ERNIE 4.0 with no major release — the roadmap only appeared once DeepSeek's shockwave arrived.

Third: brand whiplash. Wenxin Yiyan → Wenxiaoyan → Wenxin. Three renames in two years, plus leadership churn.

Fourth: outspent by two orders of magnitude. In 2025 ERNIE spent roughly ¥115 million on user acquisition. Tencent's Yuanbao spent ¥14.7 billion.

Fifth: self-cannibalization. Search ads are Baidu's profit base; an AI that answers directly is a sledgehammer to that base. Textbook innovator's dilemma.

Sixth: an organization that couldn't focus. Early on, model weights were locked away from all but a handful of internal researchers — your own people weren't using your model, and the data flywheel never spun. Li's word, "focus," translates as: ERNIE was search's child, born with "serve the search box" in its bones.

Is Baidu dead? No. In 2026, AI accounted for half of Baidu's core revenue; AI cloud grew 79% year over year, GPU cloud 283%. It lost the model layer's narrative and paid for it by selling compute. Robin Li now promises to "return to the first tier."

"Return" — that one word is the whole post-mortem.

Pangu: a road it never took

If ERNIE is "tried to win and didn't," Pangu's story is more tangled — it never really took the "general-purpose model" road at all. That's not something a one-liner like "Huawei couldn't do it" can capture. It has to be peeled back layer by layer.

Pangu was greenlit in November 2020 and launched in April 2021 — more than two years before Alibaba's Qwen or DeepSeek. Why did it never make its mark on the general-model table? Not one reason, but several stacked together.

First, and hardest: no chips. In October 2022 an export control landed, cutting off Nvidia's top GPUs from Huawei entirely — while Silicon Valley was counting cards, Huawei's entry chips for the general-model table were confiscated. You can't cook a feast without rice; that's the premise behind everything else. What it forced instead was the in-house Ascend chip: 812,000 units shipped in 2025, nearly half the domestic market, and Eric Xu says it has passed Nvidia in China. The shovel-seller isn't necessarily unwilling to pan for gold — his shovels were only ever enough for his own patch of ground.

Second, the positioning was different from day one. Pangu was born "AI for industries" — mining, weather, power. "No consumer apps" was a boundary written out plainly — Huawei wanted AI woven into its own "hardware + industry" system, not to fight a pack of cash-burning startups over a chatbot's front door. Whether that road works is its own debate; but it was never the "general model" road, and you can't measure one with the other's ruler.

Third, the 2025 open-source storm. In June, Pangu went open-source as a goodwill gesture — and drew trouble. A few researchers did something obsessive: they used "model fingerprinting" to compare Pangu's attention parameters against other models, and measured a 0.927 similarity to Qwen-2.5 14B, while independently trained models typically score below 0.7; Alibaba copyright strings were also found in the repo. Then a former employee's essay, "The Sorrow of Pangu," pushed the affair to a boil. Huawei denied the accusations but never released training logs — the silence was the loudest answer. Whatever the truth, the cut landed squarely on Pangu's technical credibility.

Fourth, the organizational-fit problem. In August 2025 the Pangu division was dissolved and folded in; its chief departed to start a company in March 2026. Huawei Cloud's model marketplace now lists 160+ third-party models, and even Huawei's own assistant runs "Pangu + DeepSeek" side by side. Li Xiang's verdict is sharp: Huawei-type companies are superb at hardware, weaker at internet and AI — organizational fit is the shortest plank in the barrel. Hardware can be bought with money and discipline; general intelligence doesn't come out of the same organizational logic.

So Pangu's quiet was less a story of "failure" than one of "trade-offs and misalignment": the chip ban blocked its road to a general model, and the inertia of its positioning and organization meant it never meant to walk that road in the first place. Yu Chengdong's June 2026 admission — "We didn't do it well. We shouldn't have." — is rare self-reflection from Huawei, but its weight shouldn't be flattened into a verdict of "couldn't build it."

Qwen: the one big-tech model that worked — how

Wait — the big players weren't wiped out. Alibaba's Qwen succeeded, and became the global open-source benchmark. Same "big-tech building models" game — why did ERNIE, Hunyuan and Pangu all break, and only Qwen survive?

Three choices, and it needed all three.

First, the 2022 "wall-breaking" was the real decisive move. In 2022, Alibaba pulled its machine-intelligence and NLP teams out of DAMO Academy and merged them into Alibaba Cloud's Tongyi Lab, with Alibaba Cloud CTO Zhou Jingren doubling as DAMO's vice dean. It looked like a reorg; it was actually a cure for a century-old disease — the Xerox PARC disease: the lab invented the mouse and the GUI, while parent Xerox guarded its copier and watched Apple and Microsoft pick the fruit. By "breaking the wall," Alibaba welded the researchers and the commercially-accountable people into one body, making scientists answer for monetization directly. Neither ERNIE nor Pangu pulled this off — their models stayed appendages of legacy businesses (search, selling hardware), never the masters of a new business.

Second, it knew what it was actually selling: not the model, but the cloud. Qwen was never positioned as "the strongest model" — it aimed to be a dependency in everyone else's software stack, like HTTP, like Linux: the better it works, the lower its profile. On August 3, 2023, Alibaba dropped the Qwen-7B weights on ModelScope and Hugging Face, the first Chinese giant to fully open-source its flagship model. Size ladder from 0.5B to 72B-plus, Apache 2.0 license, multilingual from day one. Why dare to give it away free? Because Alibaba was doing a deeper math: when developers worldwide build on Qwen, the moment they deploy, the compute almost certainly lands on Alibaba Cloud. Free models are the funnel; compute monetization is the harvest. The result: AI-related ARR broke ¥49.5 billion in 2026, 35% of Alibaba Cloud's external revenue.

Third, it hedged the chip embargo with an efficiency route. With Nvidia's high-end cards banned, Alibaba — like DeepSeek — pivoted to training efficiency, sparse architectures, and extreme engineering. Qwen even did the counterintuitive "smaller but better" (Qwen-Image 2.0 cut from 20B to 7B parameters and improved quality).

So Qwen drew the map for big tech: it's not that big companies can't build models — it's that those who don't treat the model as a new business, and don't give it a closed-loop flywheel, can't. ERNIE lacked the flywheel (search and AI fought each other); Pangu was boxed in by chips and positioning (it never took the general-model road); Hunyuan lacked time — it started early but couldn't bring itself to commit, and by the time it woke up, the track was already crowded. Qwen had all three — so it won.

(But saying "Hunyuan lost" is still premature — we'll tell its story in Part IV.)

II. The Hundred-Model Wars and the Retreat of the Celebrities

The story begins on November 30, 2022, when ChatGPT went live — the "first fire" that eventually burned through the entire industry.

China's internet took four months to react: ERNIE shipped on March 16, 2023; the "Hundred-Model Wars" officially erupted in June; over 200 models appeared within three months, peaking at 244 by September.

Then came the familiar cull: by May 2024, only 19 models were still being actively updated.

The most theatrical act was the parade of celebrities — in the same year as America's, but with the opposite instinct.

Wang Huiwen, Meituan co-founder, announced in February 2023 with $50 million of his own money. Four months later he stepped down for health reasons; his team was absorbed into Meituan.

Wang Xiaochuan, Sogou founder, started Baichuan in April 2023 — a star of the first season. He later moved his main battlefield to healthcare.

Kai-Fu Lee started 01.AI in 2023. By 2025, its pretraining team was reported to have folded into Alibaba Cloud; its Yi API later went quietly offline.

Cut back to America: celebrities dove in on both shores the same year — but American celebrities trained models themselves. Musk left and built xAI with his own hands on the model; Dario left with eleven researchers and built Anthropic. Chinese celebrities managed the people who trained models — manifestos, hiring, org charts, no hands on the furnace. The fork in the road was right there: training a foundation model is 24 months of staring at loss curves. Fame buys headlines and term sheets; it does not buy intuition for the loss curve.

By 2024 the league table had condensed into the "Six Tigers": Zhipu, Moonshot, MiniMax, StepFun, Baichuan, 01.AI. The two with celebrity founders are both out of the foundation race today.

III. Same Schools, One Outsider

Two lineages plus one outsider: Shaolin scattered disciples everywhere, while that lone lamp hangs from no branch.

The rest of the story starts with one question: who trained these people?

The American answer is simple: one lab — OpenAI — scattered into four empires. The Chinese answer is more interesting: two sources, plus one outsider.

Source one: Tsinghua — the Shaolin of the jianghu, with disciples everywhere. Jie Tang's KEG lab is the "library" of Chinese LLMs: its disciples scatter and each founds a sect. Zhipu's chief Zhang Peng is the steadiest of Tang's students — an engineer from Tsinghua's AMiner, chasing no trend and betting no life. He delivered GLM-130B in August 2022, four months before ChatGPT, and later released ChatGLM-6B, the "first lesson" downloaded over ten million times. The prospectus reads: government and enterprise private deployment contributes over 80% of revenue at nearly 60% gross margin — unsexy, but it pays the bills. That's the orthodox-sect way: no flashy swordplay, only a fire kept burning for a hundred years.

And from another branch of Tang's line came the "prodigy" Yang Zhilin — in the group by sophomore year, publishing at top venues as an undergrad, the one everyone saw coming. Behind BitDist (ModelBest) stands Tsinghua NLP's Liu Zhiyuan; Wang Xiaochuan is Tsinghua too. Of the surviving foundation players, two and a half carry Tsinghua blood. Shaolin's seventy-two arts, each with its inheritor.

Source two: SenseTime — a Wudang disciple who tempers his internal power before descending the mountain. Yan Junjie rose from intern to vice president at SenseTime, then founded MiniMax in December 2021. What SenseTime taught him was a hard "engineering" discipline: build products first, use them to feed the model — in the jianghu's words, "temper the body before training the inner power."

MiniMax survived in a way no one else did: instead of fighting the C-end traffic war at home, it moved the battlefield overseas and collected money from "emotional companionship" — a niche nobody else respected. Its core product, Talkie, grew paying users from under 500,000 in 2024 to 1.77 million in 2025, with overseas revenue at 73% of the total. The most counterintuitive part: while Kimi and others burned hundreds of millions on user acquisition, MiniMax's selling expenses actually fell 40.3% — companion products grow on emotional retention, not ads.

It's the only one of the five with a complete, auditable set of financials (00100.HK): 2025 revenue of $79.04 million, +158.9% year over year, gross margin up from 12.2% to 25.4%. Don't underestimate this "feed the foundation with overseas C-end cash flow" model — it made MiniMax the most cycle-resistant of them all: however brutal the domestic price war got, its Talkie subscriptions kept collecting money overseas. Others gambled their lives at the table; MiniMax opened a pawnshop next door — a Wudang sword, not easily drawn, but when it is, the softness hides a needle.

Source three (really, a fault line): Microsoft. MSRA was the academy of the PC era; in the LLM era it produced just one survivor, StepFun (founder Jiang Daxin is ex-Microsoft). A generation of AI talent stayed trapped inside organizations — that gap is itself part of why China's incumbents dropped the first baton.

The outsider: Liang Wenfeng — the sweep-monk with no sect. A machine-vision master's from Zhejiang University and founder of quant fund High-Flyer. He built a 1,100-GPU cluster in 2019 and stockpiled over ten thousand A100s by 2021 — a year before the export controls. He is neither NLP gentry nor venture-backed. Ask his lineage, and he can't name one; ask his backing, and the answer is two words — High-Flyer — managing hundreds of billions.

In a jianghu obsessed with sects and pedigree, Liang Wenfeng is the one who belongs to no mountain yet holds the deepest inner power. Let Shaolin and Wudang preen all they like — the sweep-monk never steps onto the dueling stage; he stays in the library, wiping his sword, over and over.

But plenty of people reduce DeepSeek's success to "having the cards" — and that's the biggest misread.The cards were only the admission ticket. What actually won it was a "money-saving chain," from architecture down to hardware, that wrung every drop out of every unit of compute.

Others spent "more compute for a stronger model"; Liang spent "more precise algorithms to make the same compute go twice as far." V2's Multi-head Latent Attention (MLA) compressed memory usage by 93%; DeepSeekMoE let each token activate only 37B parameters, cutting training cost 42.5%; by V3, 671B total parameters, 2,048 H800s, two months, $5.58 million — while closed models of comparable ability cost ten times that. R1's pure reinforcement-learning stage cost just $294,000.

Liang "voted" for this technical route with his signature: across 11 papers in three years, he never put his name on the hundred-author flagship reports — he signed, stroke by stroke, the deepest atomic technologies: MLA, DeepSeekMoE, NSA, DSec. He wasn't betting on "I have a big model"; he was betting on "I rewrote the rules of compute."

And beneath that route was the parent company's confidence. High-Flyer, a quant fund managing hundreds of billions, in Liang's own words: "Money has never been our problem; export controls on advanced chips are." Precisely because there was no "recoup the model's cost" line on the ledger, he could price the API at rock bottom and release the weights outright — open-sourcing destroys other people's pricing, not his own valuation narrative.

So DeepSeek's "restraint" (no super-app, no parameter race, no fundraising) was, at bottom, a luxury — only someone who doesn't need the money and isn't afraid to lose can fight a war of "only competing against myself." "Untied to capital and academic orthodoxy" turned out to be the single most explanatory variable in the entire industry.

One detail worth savoring: the three chiefs who made it — Liang Wenfeng from Wuchuan, Yang Zhilin from Shantou, Yan Junjie from Xiayi — are all kids from small county towns. No family learning, no capital-city pedigree, just hard skill, fought all the way from the jianghu's frontier to center stage.

Cut back to America: the Chinese founders and the American crowd were already academically related. Yang Zhilin's CMU advisor was Ruslan Salakhutdinov, Apple's first AI director; Transformer-XL and XLNet were co-authored with Google. Kimi's obsession with long context was never a trend-chase — long context is literally his doctoral birthright. Think of a swordsman: every other move he only half-practiced, but this one — the "Long-River Setting-Sun Blade" — is the technique he has soaked in since childhood, fused into his very blood. That's why even beaten to the ground, that one move remained.

And the feuds broke out inside this same small circle — not open commercial warfare, but shadow combat between schoolmates and colleagues.

First, something that sent a chill down the industry's spine.

October 2024: a ByteDance internal notice that no outsider noticed — but that detonated inside the circle. An intern had quietly injected malicious code into a shared LLM training run. In plain words: someone poisoned the furnace. Thousands of cards, dozens of hours of compute — possibly burned into a vat of contaminated scrap. A year later, arbitration documents showed ByteDance suing for ¥8 million. The industry's first model-poisoning case — and the first time the field realized it had spent years worrying about "aligning the model with humanity" while failing to align its own people.

Then the poaching war. In 2025, multiple outlets reported ByteDance offering doubled salaries to poach DeepSeek researchers. This was daylight robbery — of brains, not customers. Liang Wenfeng's answer was sharper than any press release: he didn't retaliate; he "grew his own." The first author of the industry-shaking NSA paper was an intern; at DeepSeek, requesting compute needs no approval, and seniority impresses no one. You poach my people — I'll train more than you can steal.

The last landmine was the loudest, and the most personal — it tangled mentor, schoolmate, investor and self-interest into a knot that refuses to come undone.

The feud began in 2016. That year, Yang Zhilin, with his senior schoolmates Zhang Yutao and Chen Qicong, founded a company called Recurrent AI in Tsinghua's circle, building enterprise conversational AI. Among its investors was GSR Ventures — one of the earliest to put money in.

The turn came at the end of 2022. ChatGPT's fire convinced Yang that general-purpose LLMs were the road, and he struck out anew. In 2023, Moonshot AI was born. Under the spin-off arrangement, old employer Recurrent would receive 9.5% of the new company's shares for free — a nod to old debts. But the problem was procedural: GSR insists the shareholder-resolution signatures the spin-off required weren't completed until January 2024, and the consent waivers from five old shareholders never arrived — meaning the old shareholders were left out in the cold, without knowing.

The real powder keg was someone else. Zhang Yutong, GSR's second-ever managing partner — the one who backed Xiaohongshu. She'd sat on Recurrent's board, left GSR in April 2024, and turned up as Moonshot's co-founder — eventually landing in the "president" seat, running strategy, fundraising and commercialization.

On December 5, 2024, Zhu Shenghu fired off a string of WeChat Moments posts, dragging into daylight a secret the circle had whispered about for half a year. His words, in plain terms, were three hammer-blows:

One, Moonshot was a project Recurrent had incubated internally for over two years, and the spin-off procedure still isn't finished. Two, Zhang Yutong, as a Recurrent director, hid from everyone that she'd received 9 million shares in the new company for free — about 14% of the initial shares, more than the 9.5% the parent Recurrent received. Three, this is fiduciary duty — the high-voltage line an investor may not cross — so GSR fired her "to set the record straight and protect the fund's reputation."

Zhu even named her husband, Wang Zhen: the 9 million shares were held through a BVI vehicle under Wang's name, and she later received another 3 million shares from Yang — 12 million in all, making her Moonshot's second-largest individual shareholder at founding, behind only Yang himself.

Yang's counterattack came within a night. Late on December 6, he posted a long response, item by item: the spin-off was agreed with Recurrent's CEO in February 2023, with a written record; the board resolution was signed by every director; Zhang's shares were granted as a co-founder on multi-year vesting — compensation for years of future service, not a free gift. And his final stance was the sharpest: he would defend Zhang to the end, no severance. Internally, people even said — "14% is still too little for her."

Why defend her so hard? Because the math ran like this: Alibaba's $1 billion lead investment, the round that jumped the valuation from $300 million to $2.5 billion — Zhang was the one who made that connection. Kimi hitting 36 million MAU, that had her fingerprints too. In a jianghu that lives on fundraising, someone who can find money — and knows how to find it — is worth more than anything.

Zhu's side had its own unspoken grievance. In March 2023 he'd said publicly he had "no interest in investing in Chinese foundation models"; a month later, Alibaba led Moonshot. He'd sentenced the track to death with his own mouth, and its valuation ran from $3 billion to $50 billion — while he and the five old shareholders behind him were exactly the ones left off the train. So public opinion turned the question back on him: back when we asked if you'd invest, you publicly said no; now that it's about to ring the bell, you arrive with a lawyer's letter, timed to the IPO — is that "holding someone to account," or a roadside shakedown?

The case has dragged on since November 2024, unresolved at the Hong Kong International Arbitration Centre (HKIAC). And its strangest feature is that it grows more expensive the longer it drags: the claim is priced as a "percentage of initial shares," and as the valuation rolled from $300 million to $50 billion, the same 10.5% swelled from tens of millions to $700 million (about ¥4.7 billion). In October 2026, Reuters broke that the two sides had begun confidential settlement talks on the eve of the IPO — because a Hong Kong prospectus cannot dodge a "material pending arbitration," and until that landmine is defused, sponsors won't sign and cornerstone investors won't order.

This game has no winner. Yang carries the "procedural flaw + data ownership" accusations; Zhang carries the moral debt of "fiduciary duty"; Zhu carries the question of "returning with a vengeance after missing the wave." And the most telling part: they argue furiously, yet they're fighting over seats at the same table of capital — while the retail investors who pay the bill were never at the table at all.

Same schools don't guarantee same paths — that's a jianghu. When the same small circle starts poisoning, poaching, and flipping tables on itself — that's a jianghu, no qualifier needed.

IV. The Giants' Second Half: ByteDance and Tencent

Doubao puts PPT and image generation right in the input box, chasing engagement; Hunyuan's brand page shows nothing but "Hy” — the series name behind Hy3 and Hy4. One trades traffic for time, the other throws money to chase it. Screenshots: Nov 3, 2026

The earlier chapters covered the early birds and the celebrity exits. But the real war wasn't fought among startups — it was fought between two giants holding traffic and cash, betting on the same compute table. One played "traffic for time." The other played "money to chase time."

ByteDance: king of traffic, and its monetization bind

Doubao is a monster. When it launched in May 2024, daily token volume was about 4 trillion. Two years later it hit 180 trillion — a 1,500-fold jump. Monthly active users: 419 million, a runaway No. 1, 2.6× the runner-up Qwen, leaving DeepSeek and Yuanbao far behind.

How do you grow a monster like that? ByteDance's old craft: pour the entire group's traffic into a single funnel. Douyin opened two entry points — long-press the like icon to ask a question, a Doubao assistant pinned atop the message list — with Jianying, Volcano Engine, and e-commerce all wired in. ByteDance's internal phrasing: "bring Doubao to the user, not the user to Doubao." By November 2024, Doubao hit 59.98 million MAU, second globally behind only ChatGPT.

But the traffic grows exactly as fast as the bill. This is the cruel difference between LLMs and WeChat or Douyin: each added WeChat user costs near-zero in servers; each added LLM query, image, or video is GPU being burned in real time. One cost breakdown: hardware depreciation is 58% of a single Doubao inference, power 29% — nearly 90% real resource consumption, not amortized infrastructure.

Hence the most absurd pair of numbers in Chinese AI: a national-scale app with over 200 million DAU earning under ¥1 million a day. Over 200 million people use it daily; it makes less than a million yuan a day, mostly from e-commerce commissions — while burning tens of millions in compute. Third-party estimates put Doubao's monthly net loss at ¥210–270 million.

In June 2026, Doubao did something that poked a hornet's nest: it posted the first price list. Three tiers — ¥68, ¥200, ¥500 a month, the top tier ¥5,088 a year, endlessly compared to ChatGPT Pro's $200. Social media split into "charge me and I uninstall" versus "there's no free lunch." In a market where almost every Chinese model still shouted "free," the first one to charge was always going to take the most abuse.

But Doubao's real problem isn't whether to charge — it's that the capability can't carry the traffic. This is the shortfall ByteDance least wants to face, and the one getting harder to hide.

ByteDance's strengths were all bet on multimodal and the product loop — video, image, voice, meshed seamlessly with Douyin and Jianying. But the part of a language model that matters most — deep reasoning — is exactly its weakest link. The smoking gun: ByteDance's RL team only became a first-tier division (Horizon RL) in August 2026. Before that, RL was a post-training add-on, not a main thrust — whereas DeepSeek treated RL as the lifeblood from day one. Third-party testing (SQBench, 220 tasks each) puts it plainly: Doubao shows "noticeably fewer fully-correct passes in long-context reasoning, trailing Qwen and DeepSeek on complex business and math-audit tasks" — yet it tops the board on "engineering execution," software-engineering tasks.

In plain words: Doubao is a superb "doer," not a deep "thinker." It writes code, fixes bugs, generates video — but on layered reasoning it tends to "skip steps," simplifying key conditions and eliding the middle of the derivation.

And the official word on the gap is already out. In August 2026, ByteDance CEO Liang Rubo admitted at an all-hands: the gap between Seed's LLM and the leading overseas models is "widening," and asked the team to "accept being behind for a while." A giant holding 400 million MAU and ¥200 billion in capex, publicly saying "accept being behind" — almost unheard of in ByteDance's history.

Root causes? Two. First, closed weights. Doubao releases no open weights, so its capability is judged only by self-evaluation and leaderboards — no "ten-thousand-eyes" external scrutiny from the open-source community, so problems surface late. Second, path inertia. ByteDance's DNA is "product, traffic, monetization" — unbeatable in multimodal, but a burden in "a decade to sharpen one sword" fundamental intelligence. Running research at the cadence of product makes things that look good but aren't deep.

So ByteDance's choice is explicit: in 2026 it raised the AI infrastructure budget from ¥160 billion to ¥200 billion, and is reportedly training a model with up to 10 trillion parameters (over three times Kimi K3's 2.8 trillion). This is buying time with compute — betting that stacking cards can make up the "depth" class it missed. But the training stability and data quality of a 10-trillion-parameter model are precisely the unsolved problems of the moment. Whether this gamble pays off is one of the biggest suspense points to watch into 2027.

But don't write ByteDance's losses off yet. Its real money machine is the video model Seedance. Launched February 2026, it runs at about $2 billion annualized revenue, over ¥1 billion a month, roughly 70% gross margin, 80% market share — some in the industry call it "the most profitable single product in Chinese AI commercialization history." And the flywheel is vicious: a 100-minute drama costs ¥500k–1M to shoot with real actors, ¥50k–100k with AI; the saved production budget flows straight back into Douyin ad spend. Others sell shovels and the ore belongs to the customer; the ore ByteDance's shovel digs has to be poured into its own smelter — one cut on model calls, one on traffic, one on in-show ads.

Cut back to America: OpenAI killed Sora because it bled without earning. ByteDance turned Seedance into a cash cow — and reportedly pushed OpenAI to consider "reviving" Sora. Same track, one burned it shut, the other harvests in silence. On video, ByteDance built the global case study.

ByteDance's bet is explicit: ¥200 billion in capex (or more), ¥85 billion on chips, trading ad profits for a ticket into the compute era. It is betting that — the closed loop outruns the cash burn.

Tencent: slow, but catching up — too early to call it a loss

Tencent's move came only after its back was to the wall.

It was once the most "Buddhist" player: Hunyuan shipped early but never pushed hard, sitting on WeChat's traffic and gaming's cash cow, seemingly unhurried. Then, in February 2026, Tencent did something ruthless — it tore down and rebuilt its pretraining and RL infrastructure. It also poached OpenAI's chief researcher Yao Shunyu to be chief AI scientist, with a one-line philosophy: don't stack parameters; optimize for usefulness.

The rebuilt Hunyuan, rebranded Hy, sprinted: Hy3 preview in April (295B total, 21B active), Hy3 official in July, Hy4 preview in August (770B total, 49B active, 1M context) — two years' road covered in six months. The payoff was immediate: Hy4 preview shot to No. 1 on OpenRouter at 14.7 trillion tokens, leaving even its own Hy3 at No. 7. Blind tests had it edging GLM-5.3 and Kimi K3.

But the bill came, and it was brutal. In Q2 2026, Tencent's capex hit ¥52.8 billion, up 176% year over year; free cash flow turned negative at -¥13.8 billion — the first negative quarter since listing. This is the price of tearing off the "Buddhist" label: a company that generated ~¥100 billion in annual free cash flow for a decade had to borrow to buy compute for the first time. It raised over ¥73 billion across two bond sales in half a year, all earmarked for AI infrastructure. UBS estimates Tencent's AI business loses about ¥15 billion a quarter, dragging profit growth by roughly five points.

Cut back to America: ByteDance and Tencent are answering the same exam question — the internet-era formula of "subsidize for scale, monetize scale with ads" breaks down in LLMs, because marginal cost no longer trends to zero but rises linearly with usage. ByteDance's answer is to charge first and let video fund the language model; Tencent's is to charge nothing, lean on gaming and ad cash cows, and paper the gap with debt. Neither route is proven — but both are betting the same thing: whoever closes the loop first wins the second half.

One detail worth savoring: Tencent's Yuanbao is still free, and it runs a "DeepSeek + Hunyuan" dual-model switcher — it chases the leader while loading the leader's model into its own app. "Use them until you can beat them" — the most ruthless and the most pragmatic move a big-tech pursuer can make.

But keep the word "Hunyuan" far away from "failure." Hunyuan didn't fail to build — it simply hadn't tried hard yet. ERNIE's disease was a legacy business dragging it down; Pangu's fate was that neither its chips nor its positioning sat on the general-model road. Hunyuan has neither: it sits on WeChat's 1.4 billion MAU and a gaming-and-ads cash cow, so it never lacked money or scenarios — it lacked the "we have no choice" moment. February 2026's teardown-and-rebuild was that moment landing; Hy4 topping OpenRouter in six months proves that when Tencent gets serious, it can catch anyone.

So Tencent's real question isn't "will it fail" — it's "can time amortize the price." The negative free cash flow is tuition; whether that tuition buys a fight-capable foundation model or just an expensive side-trip won't be known until 2027, when we see whether Hy can close the loop inside WeChat's ecosystem. Calling it a loss now is too early.

V. A Third Way: Xiaomi MiMo, the Device-Side Dark Horse

Beyond the giants and the startups lies a third route many people overlook — device makers training their own models. And the dark horse that broke out on this track in 2026 is Xiaomi.

Xiaomi started later than almost everyone, yet took a wilder path. Lei Jun only pushed to set up the model team in 2023, with a two-line charter: reasoning models, plus lightweight on-device deployment. While others scaled up and moved to the cloud, Xiaomi asked from day one how to cram a model into a phone, a car, a speaker. Lei Jun, known in the jianghu as "the workhorse," has scrapped in the phone game for over a decade and lives by one iron rule: don't chase No. 1 — chase the seat nearest the user.

But what turned Xiaomi from "a hardware maker dabbling" into "a player at the table" was one master it recruited in November 2025 — Luo Fuli, a former DeepSeek researcher. The appointment was itself a move on the lineage map: beyond the Tsinghua and SenseTime lineages, Xiaomi tapped the "outsider" bloodline of DeepSeek and brought in a veteran. The sweep-monk's door, it turned out, had a disciple willing to descend the mountain.

Under Luo Fuli, MiMo climbed a steep four-step curve in three years: a 7B dense model in April 2025 → a 309B MoE in December 2025 → a 1-trillion-parameter Pro in March 2026 → a 1.02-trillion model in April 2026. Every generation open-sourced, MIT license, free to use and modify. By September 2026, MiMo-V2.6 scored 46 on the Artificial Analysis Intelligence Index — above GLM-5.3 (45) and Kimi K3 (44), the top open-weight model in the world. And its per-task cost of $0.13 is one-fifteenth of GLM-5.3's or Kimi K3's.

Even sharper is how Xiaomi turned "open source" into a reputation war: not just the weights, but the RL training code, 7,000-plus training environments and the technical report all released — it even livestreamed its RL training runs. In a Chinese model scene that leans closed, that's almost a provocation.

But Xiaomi's ambition was never "selling model APIs" — its token revenue trails the giants and leading independents by an order of magnitude. What Xiaomi wants is to make MiMo the nervous system of its "Human x Car x Home" ecosystem: its first AI-native phone, the Xiaomi Miclaw (codenamed "Lobster"), launched in March 2026, with MiMo flowing into HAD assisted driving, HyperOS, speakers and smart home. Lei Jun's stake: over ¥60 billion into AI over three years, ¥16 billion in 2026 alone.

Xiaomi's calculus is clear: it won't fight for the top of the general-intelligence leaderboard; it's betting on the still-untested route of "edge-cloud synergy" — squeezing models into devices with INT4 quantization, sparsification and hardware-aware kernels to drag inference cost down to "edge-viable." In July 2026, Xiaomi's "Pengpai AI" on-device model filed for regulatory approval alongside Apple, Huawei, OPPO, vivo and Samsung — marking on-device LLMs' shift from novelty to standard.

Xiaomi completes another piece of the Chinese AI jianghu: while the giants pile compute in the cloud and the startups race on extreme value-for-money, device makers are grabbing "the inch of screen closest to the user." It may never be the king of general intelligence, but it's the only player binding a trillion-parameter model to phone shipments on the same table — and that contest, too, is nowhere near decided.

VI. The R1 Massacre

To understand why R1 was a massacre, look at the battlefield it landed on — a calm era when everyone still believed money could buy users. And its fuse had been lit a month earlier: on December 26, 2024, DeepSeek released V3 with a final training run officially costed at $5.576 million — a tenth or less of what Silicon Valley spent on comparable models. Nobody realized that number was a ticking bomb.

May 2024: the price war. ByteDance's Doubao drove inference prices into the "li era" (a tenth of a cent per thousand tokens); Alibaba cut prices on nine Qwen models by up to 97%; Baidu, Tencent, iFlytek and Zhipu followed. The premium on raw model capability was stripped on the spot.

With "selling the model" dead, consumer apps had one play left: buy traffic. The most aggressive buyer was also the best-taste team — Kimi, at up to ¥199 million in monthly ad spend. Its growth team stared at the dashboard, watching the curve climb, waiting for the "burn for scale, scale for funding, fund to burn" loop to start spinning.

Until the night of January 20, 2025.

In Hangzhou, a company that had never spent a yuan on advertising posted R1's weights online. Free, excellent, commercially usable — rivaling Silicon Valley's best closed models.

Kimi's growth team lived through the most agonizing dawn in the industry that night. The users they'd spent over a billion yuan to buy were draining away in real time — per public reports, R1 gained 100 million users in seven days, while Kimi's MAU slid from 21.01 million to 7.29 million, out of the top ten.

The average bought user lasted four days. That number was now the whole industry's sorest slap.

The cull began. The Six Tigers bled out collectively. In ByteDance's war room, the ad budget was frozen on the spot — yesterday's siege ammunition, today's toxic liability. Zhu Shenghu, who had vowed not to invest in foundation models, publicly turned bull — the internet's favorite "told you so" reversal.

Cut back to America: this is the prologue's night. The Monday seven days after R1, Nvidia fell 17% — about $589 billion, roughly two AMDs. Silicon Valley argued "Sputnik moment" for a week while Altman cut prices overnight.

From then on, Chinese LLMs were measured with a new ruler: not parameters or benchmark scores, but cost per task, agent success rates, and iteration speed.

And it's exactly here that Kimi's story turned — it's the only one of the Six Tigers to be beaten to the bottom and climb back on its own.

R1 drove it from 21.01 million MAU down to 7.29 million, out of the top ten. For anyone else, that's a dead end. But Yang Zhilin did something only a technical founder could: admit the traffic-buying route was wrong, cut the ad spend, retreat to the lab, and retrain the model. A year later, Kimi K3 dropped — 2.8 trillion parameters, the world's largest open-source model, topping Hugging Face's trending board within 30 minutes and ranking No. 1 globally in Code Arena's frontend blind tests.

How did K3 turn it around? Not by stacking parameters, but by Yang replacing three foundations of the Transformer that the field had used for a decade: the optimizer (Adam → Muon), attention (full attention → KDA hybrid linear attention), and residual connections (→ Attention Residuals). This original bottom-layer work made million-token-context decoding 6.3× faster and training 25% more efficient. In other words: others traded "more compute for capability"; Yang traded "smarter architecture for compute" — the same species as Liang Wenfeng.

Kimi's survival is also in its revenue structure: B-end API is over 70% of revenue, with ARR climbing from $100 million in March 2026 to $300 million in June and $1 billion by August. A 300-person company with no OKRs and no department walls multiplied its ARR tenfold in three months. But its fatal flaw is right here: the stronger it gets, the more people use it, the more compute it lacks — K3 closed new-user subscriptions within 48 hours. Cards are the wall Kimi has never gotten around.

So look at how the Six Tigers' survivors each made it: MiniMax collects money overseas from emotional companionship, Kimi turned around on original bottom-layer architecture and B-end API, Zhipu feeds itself on government/enterprise private deployment, DeepSeek fights the efficiency war on its parent's blood supply. Not one of them lives on "selling the model" — because that route died in the price war.

VII. Reverse Output

After R1, the Chinese story flipped from "catching up" to "exporting."

Capturing the open-source flag. By 2026, Qwen had spawned over 150,000 derivative models — 4.7× Llama's count — and Hugging Face's official blog called Qwen "the community's default base model." The American open-source flag was pulled down by Chinese models. As of October 2026, the top five most-routed models on OpenRouter are all Chinese, together nearly half of weekly routed tokens, for 20-plus consecutive weeks.

Capitalization. Zhipu listed on HKEX on January 8, 2026 (02513.HK) — "the world's first listed foundation-model company" — followed the next day by MiniMax (00100.HK), which jumped over 100% on debut. DeepSeek broke its "never raise" vow: a first round of roughly ¥51 billion in June 2026, valuing it near ¥338 billion, with Liang Wenfeng personally contributing about ¥20 billion. Moonshot closed a $2 billion round at a $20B-plus valuation.

But to see the truth of "how many are left," you need a wider ruler. Independent foundation-model players were never just five: beyond Zhipu, MiniMax, Moonshot and DeepSeek, StepFun is still at the table (reportedly closing a ~$2.5 billion round in May 2026 and pushing toward a Hong Kong IPO), and ModelBest holds the edge-side small-model niche — the so-called "Six Tigers" today is four sprinting toward IPO and two pivoting vertical, not "only five left." Above them, all five giants (ByteDance, Alibaba, Tencent, Baidu, Huawei) remain. Outward, Xiaomi, OPPO, vivo and Honor are building a new front on the device side. Below, a thousand-plus vertical models (finance, healthcare, energy, government) each survive their own way. The table didn't shrink — it stratified, from a free-for-all anyone could enter into a four-tier board: giants, startups, device makers, verticals.

The sharpest irony: on January 15, 2025, Zhipu was placed on the U.S. Commerce Department's Entity List — the first Chinese model company to receive the "sanctions seal of approval." One year later, it rang the listing bell.

But hold the applause. No. 1 in calls is not No. 1 in cash. Chinese models carry roughly 60% of global token volume, yet the top Chinese companies' combined revenue is barely a tenth of OpenAI plus Anthropic (per Wall Street estimates). Average price on Chinese public clouds: ~¥13.5 per million tokens, versus ¥117+ overseas — an order of magnitude apart. Zhipu peaked on debut and has since fallen by nearly half; the market has begun repricing this cohort on revenue structure, not narratives.

Cut back to America: in 2026 OpenAI raised $122 billion at an $852 billion valuation; Anthropic approaches a trillion. Two continents, two capital games, an order of magnitude apart — but the criterion never changed: whoever the ropes don't bind survives.

Epilogue: Two Ways to Live One Arms Race

Burn cards over there, save cards over here — the same arms race, two ways to live.

Put the two storylines side by side and they are two ways of living through the same arms race.

The "compute is survival" obsession grew, in America, into Colossus and its million GPUs. In China it grew into something else — engineering efficiency forged by embargo: FlashMLA and NSA squeezing every drop from H800s; every tightening of the export controls deepened the efficiency route another layer.

America's way is to burn cards: survive on capital transfusions, each round more staggering than the last. China's way is to save cards: DeepSeek ran on High-Flyer's own money until the product could pay for itself — not raising money became the weapon, while the companies tied to the fundraising cycle died by the dozens.

America: one lab scattered into four empires. China: two sources plus one outsider.

And the most savor-worthy twist: China's AI pursuit was planned all along — DeepSeek and Qwen are moves in a long, deliberate industrial endgame. But R1's four-ounces-moving-a-thousand-pounds arrival was a variable nobody scheduled.

Closing Notes

Fold the board up and it comes down to this:

1. First-mover advantage has a half-life under 18 months. On this track, "early" is a liability, not an asset — ERNIE set out early and arrived late; Pangu set out earlier still but was boxed in by chips and positioning; neither held the table.
2. The incumbents shared one disease: the model served the organization; the organization served the legacy business.
3. Fame doesn't train models. Celebrity companies exited collectively — highlighting how rare it is for founders to train models themselves.
4. The giants' second half is two algorithms for one compute bill: ByteDance trades traffic for time, charges first, and lets video fund the language model; Tencent throws money to chase time, turning its free cash flow negative for the first time since listing. Both are betting the same thing — closing the loop before bleeding out.
5. Survival meant being untied: untied from capital (DeepSeek), from legacy business (ERNIE's fatal flaw), from having one's hands tied by chips and positioning (Pangu's misalignment), from the traffic-buying illusion (Kimi's lesson and pivot).
6. The embargo cuts both ways: it confiscated the chips for the table — and forged an efficiency discipline nobody else has.
7. China and America are two ways of living one arms race: one burns cards, one saves them; one survives on capital, one on not needing it.
8. The jianghu has no endgame, and there was never "only five left": five giants, four or five startups, device makers led by Xiaomi, and a thousand-plus vertical models each holding ground — the table didn't shrink, it stratified. Consumer paywalls, agent monetization, and the next paradigm are already queuing at the door.

Previous installments: The First Fire of the AI Jianghu · From the Shanghai Gigafactory to Colossus · The Ultimate Landscape of the Agent Wars