It has been years since Zhang Yiming last stood in front of a large ByteDance audience. The founder, who handed the CEO title to Liang Rubo in 2021 and retreated from public view, is now 43 years old and — with a personal fortune estimated at ¥670 billion — the richest person in China. He spends his time on the thing he has decided matters most: the company's AI strategy.
So when he walked into a recent all-hands meeting for ByteDance's AI organization, the roughly one thousand researchers and engineers in the room knew the words would carry weight. What they heard was not a rallying cry. According to The Paper, which cited multiple people with knowledge of the meeting, Zhang told his team that ByteDance has the ability to compete with OpenAI and Anthropic. Then he delivered the sentence that has since echoed through China's technology industry: truly catching up with the frontier, he said, will take until 2029 or 2030 at the earliest.
Not this year. Not next year. Not after one more training run. A four-year runway, announced inside a company famous for moving at breakneck speed — and announced the same week that domestic rivals were busy shipping models, courts were debating data rights, and the global AI market was re-pricing itself.
The most interesting part of the speech was not the timeline. It was what Zhang blamed for the gap, and the method he ruled out for closing it.
An Unusually Honest All-Hands
Corporate all-hands meetings in China's tech industry tend to follow a script: affirm the mission, celebrate the metrics, hint at the breakthrough. Zhang's November address reportedly did none of that. Instead, it functioned as an internal calibration.
The message had two halves. The first was reassurance: ByteDance can compete with the world's best AI labs. The second was discipline: do not confuse consumer product dominance with frontier model capability, and do not chase the appearance of parity. Zhang explicitly set the company's benchmark at the international first tier — not at domestic rivals, and not at app-store download charts.
This was a pointed framing. Doubao, ByteDance's AI assistant, is the most-used AI-native application in China, with third-party estimates putting its monthly active users around 380 million. The company's video generation model Seedance leads global benchmarks. By every consumer metric, ByteDance AI is winning. Yet its founder chose the moment of that victory to tell his team they are behind.
The candor follows a pattern. In an August internal meeting, Zhang had already told the Seed team to practice "long-termism and delayed gratification" rather than "using someone else's output to trade for temporary leaderboard rankings." CEO Liang Rubo reinforced the message at ByteDance's mid-year all-hands, acknowledging that while Doubao remains competitive in consumer scenarios and Seedance leads in video generation, the gap between ByteDance's large language models and overseas leaders had widened.
*Inside China's tech giants, AI strategy meetings have replaced product launches as the industry's main stage. (Image: Unsplash)*
The Machine Behind the Meeting
The team Zhang addressed is the centerpiece of ByteDance's AI architecture. Seed, the company's foundation-model unit, was assembled through one of the most aggressive talent and organization builds in the industry.
| Element | Detail | Source/Context |
|---|---|---|
| Team size | ~1,000 researchers and engineers | Consolidated since 2025 |
| Head of research | Wu Yonghui | Joined Feb 2025 from Google DeepMind |
| Internal structure | Edge, Focus, Base teams | Reorganized under Wu after his arrival |
| Current flagship | Seed 2.1 (released June 2026) | Multimodal LLM; no open frontier weights |
| In training | Reported model up to 10 trillion parameters | Overseas media reports; not officially confirmed |
| Product surface | Doubao assistant, Seedance video model | Doubao ~380M MAU (third-party estimates) |
Wu Yonghui's arrival marked a turning point. Within three months of the former DeepMind research vice president joining, Seed's published paper output exceeded all of 2024 combined, and the unit's reporting lines were redrawn so that core algorithm teams reported directly to him. The structure — Edge for applied capability, Focus for concentrated research bets, Base for pretraining foundations — was designed to give ByteDance what it had never had: a permanent frontier-model organization rather than a rotating cast of project teams.
That organization is now training what overseas media have described as a model of up to ten trillion parameters — a scale, if confirmed, in the same class as Anthropic's largest systems. ByteDance has not officially commented on the report, and Zhang's own framing suggests the company does not believe scale alone closes the gap.
How ByteDance Arrived at the Slow Lane
The 2029 timeline only makes sense against the company's three-year AI reorganization, one of the most consequential — and least publicly understood — strategic arcs in Chinese technology.
| Period | Move | Significance |
|---|---|---|
| 2023 | AI Lab dissolved; research redistributed | Ended the academic-lab model |
| 2024 | Seed and Flow established as parallel units | Split foundation models from application products |
| Feb 2025 | Wu Yonghui joins from DeepMind | Talent anchor for frontier research |
| Feb 2026 | Seed 2.0 ships; ranks #9 on LMArena | Only Chinese model in the global top ten at the time |
| June 2026 | Seed 2.1 ships | Admits limits on open-ended and research-level tasks |
| Aug 2026 | Feishu split: product team into Doubao, GTM into Volcano Engine | Consolidated the application layer under the AI assistant |
| Nov 2026 | Zhang's all-hands: catch-up "earliest 2029–2030" | Public acknowledgment of the gap |
The February high point is worth pausing on. Seed 2.0's ninth-place finish on the LMArena chatbot arena was a genuine milestone — the first time a Chinese model had cracked the global top ten on the industry's most-watched crowdsourced benchmark. It did not last. By mid-year, DeepSeek's V4, Moonshot's Kimi K3, and Zhipu's GLM-5.2 had all posted results that pushed Seed out of the domestic lead, and none of those rivals needed ByteDance's budget to do it.
Then came the application-layer consolidation. In August, ByteDance dissolved Feishu's standalone structure, folding its product engineers into Doubao and its go-to-market organization into the Volcano Engine cloud unit. The message: one super-app entry point, one cloud distribution arm, and no redundant empires. Feishu had reportedly passed ¥3 billion in annual recurring revenue in 2025 — profitable, growing, and still subordinated to the AI thesis.
*Benchmark tables and user metrics now shape capital allocation across China's AI labs. (Image: Unsplash)*
The Numbers Behind the Confidence
Why does Zhang believe 2029 is reachable at all? The honest answer is resources. ByteDance is one of the few companies on earth that can fund a decade-long frontier-model program without asking anyone's permission.
| Metric | Figure | Notes |
|---|---|---|
| 2026 capex discussion | Up to $70 billion | Bloomberg, May 2026; preliminary, focused on AI data centers |
| LLM business ARR | ~$4 billion | Industry estimates, mid-2026 |
| Doubao MAU | ~382 million | Third-party estimates, Aug 2026 |
| Doubao ranking | #1 AI-native app in China | QuestMobile, Sept 2026 |
| Feishu ARR (2025) | ¥3+ billion | Before its split into Doubao/Volcano Engine |
| Founder wealth | ¥670 billion | Zhang Yiming, China's richest person |
The $70 billion figure deserves care. It originated in a Bloomberg report this past May describing internal discussions to raise 2026 capital expenditure to as much as that number — not a finalized budget. Even at a fraction of it, however, ByteDance's infrastructure commitment dwarfs that of every Chinese AI startup combined, and it is matched in China only by Alibaba.
Revenue is the other pillar. Industry estimates put ByteDance's large-model business at roughly $4 billion in annual recurring revenue — an unusually strong commercial position for a frontier lab, built on the Doubao subscription, Volcano Engine model APIs, and the Seedance video-generation business. OpenAI and Anthropic remain far larger, but ByteDance is closer than any Chinese peer to the lab-plus-cloud revenue model that funds OpenAI's compute appetite.
The Data Wall
Zhang's diagnosis of the gap was blunt: Seed's pretraining dataset is "still too small." In an era when scaling laws still govern model evolution, the size, quality, and knowledge density of training corpora set a hard ceiling on capability — and this is the one input that cannot simply be purchased.
The constraint has a specific structure for ByteDance. The company sits on one of the world's richest content ecosystems — short video, news, literature, livestream transcripts. But frontier pretraining demands something different: global, multilingual, high-knowledge-density text at a scale measured in tens of trillions of tokens, much of it locked behind copyright regimes and language barriers that no amount of domestic content can substitute for.
| Training Input | ByteDance Position | Frontier Requirement |
|---|---|---|
| Compute | Effectively unlimited capex | Hundreds of thousands of accelerators |
| Engineering talent | ~1,000-person Seed org, DeepMind pedigree | Top-tier pretraining expertise |
| Chinese-language content | Among the world's deepest | A fraction of ideal pretraining mix |
| Global multilingual corpora | Structurally constrained | Dominant share of the token budget |
| Licensed/proprietary text | Limited vs. US labs | Growing share as public web data saturates |
| Unique modality data | Exceptional video corpus | Video increasingly central to multimodal training |
The last row is where ByteDance's strategy gets interesting. If text data is the wall, video may be the door. No Western lab owns a comparable corpus of billions of short videos with engagement metadata, and multimodal models are only beginning to absorb that kind of signal. Zhang's investment in Seedance and video-first products is not a detour around the data wall — it may be his plan to route through territory where ByteDance starts with a structural advantage.
The Anti-Distillation Doctrine
The most consequential instruction from the November meeting was not about dates. It was about method. Zhang reiterated that ByteDance must not rely on distillation — training on the outputs of stronger models — and the company has reportedly operationalized the ban: internal systems now run API anomaly detection and filter open-source model data to catch researchers quietly feeding third-party outputs into training sets.
| Lab | Route to Capability | Distillation Stance | Cost Profile |
|---|---|---|---|
| ByteDance | Self-built data pipeline, 10T-param training run | Explicitly banned as a dependency | Capital-heavy, slow |
| DeepSeek | Architecture efficiency, RL-led post-training | Publicly denied distillation claims | Compute-frugal, fast iteration |
| Moonshot AI | Open-weight frontier releases | Emphasizes independent training | VC-funded, research-led |
| Alibaba (Qwen) | Open ecosystem + cloud distribution | Open weights, permissive derivatives | Ecosystem-leveraged |
The doctrine cuts against the grain of an industry where distillation — in its gray-zone form — is widely suspected of powering fast benchmark gains. Several Chinese labs have faced pointed questions this year about whether their models learned from restricted Western APIs. Zhang's position, articulated since August, is that such gains are counterfeit: they inflate evaluation scores while hollowing out the underlying capability, and they leave the company legally and reputationally exposed.
Whether the ban is fully enforceable inside a thousand-person research organization is an open question; whistleblower-style leaks and API forensics suggest no lab can police its training data perfectly. But as a strategic signal, the meaning is unmistakable. ByteDance has chosen the slowest, most expensive path — and its founder wants that choice on the record.
Why 2029 Matters — and Why It Might Not
Set the timeline against the backdrop it was announced into. Global AI valuations are wobbling: media reports this week put Anthropic's IPO valuation ceiling near $1.02 trillion while OpenAI's latest round reportedly settled at $1.25 trillion — down from earlier heights. China's own AI equity market has spent the autumn arguing with itself about bubbles. Into that anxiety, Zhang delivered a four-year admission.
| Player | Position (Nov 2026) | Trajectory |
|---|---|---|
| OpenAI | ~$1.25T valuation; frontier models + consumer scale | Defending the frontier |
| Anthropic | ~$1.02T IPO ceiling; enterprise/code strength | Ascending |
| ByteDance | ~380M Doubao MAU; Seed mid-tier among Chinese LLMs | Investing into the gap |
| Domestic rivals | DeepSeek, Kimi, GLM benchmark leads at various points | Fragmented lead |
There are two readings of the 2029 date. The generous one: Zhang is doing what serious lab leaders do, refusing to let product metrics launder research shortfalls, and buying his team the one thing frontier research needs — permission to be boring for four years. The skeptical one: a moving-target deadline is unfalsifiable, and announcing one conveniently lowers expectations for a model program that has slipped from February's podium. OpenAI and Anthropic will not pause while ByteDance catches up; the frontier in 2029 will be wherever they are then, not where they are now.
Both readings can be true at once, and the market's reaction suggests investors heard the second one loudest: commentary around the speech immediately turned to valuation discipline for concept-driven AI stocks, with analysts advising a shift toward companies with real cash flow — a category in which, not coincidentally, ByteDance's application business qualifies.
What Comes Next
The near-term milestones are concrete. Seed's next major model release will be read as the first referendum on the slow-lane doctrine — specifically whether it can reclaim ground against DeepSeek V4, Kimi K3, and GLM-5.2 on open-ended reasoning and agentic tasks rather than on consumer-style evaluations. Any credible detail on the reported ten-trillion-parameter training run would shift the entire industry's compute calculus. And Doubao's commercial trajectory — subscription attach rates, Volcano Engine API growth, Seedance's video-generation revenue — will determine how much patience the P&L can fund.
Beyond that, watch three things. First, data: any ByteDance move into licensed content, synthetic data at frontier scale, or multimodal data infrastructure will show how it plans to breach the data wall it admits to. Second, people: Seed's hiring under Wu Yonghui has been aggressive, and departures — or arrivals — from rival labs are the leading indicator of whether the strategy is working. Third, doctrine: if other Chinese labs publicly commit to no-distillation training, Zhang will have dragged an entire industry onto his slow lane; if they do not, he has handed rivals a four-year head start on the cheap shortcut.
Zhang Yiming built the world's most valuable startup by being faster than everyone. His bet now is that the AI race rewards the opposite — that in a market drunk on monthly leaderboards, the only durable advantage is the discipline to ignore them. 2029 is far away. The timer, as he made sure his team understood, has already started.
Social Voices
@量子位实习生 on Zhihu
"一个人敢在内部说'我们还要四年',比一百个发布会都有含金量。字节这是把DeepSeek式的坦诚学过来了。"
>
"A man who dares to say 'we still need four years' internally carries more weight than a hundred launch events. ByteDance has learned DeepSeek-style honesty."
@Momo on Douban
"翻译一下:榜单先不冲了,反正追不上。但别真信2029,那是给投资人画的终点线,移动靶。"
>
"Translation: we're done chasing leaderboards since we can't win. And don't take 2029 literally — it's a finish line painted for investors, and the target moves."
@AI_watcher_kevin on X
"The most underrated detail in Zhang Yiming's talk: ByteDance runs API anomaly detection to stop its own researchers from secretly distilling rival models. That's an internal compliance regime for training data. Almost no lab admits it needs one."
@互联网小助手 on Weibo
"豆包月活3.8亿,Seed却承认落后,这才是清醒。应用赢不代表模型赢,某些厂商该听听。"
>
"Doubao has 380M MAU while Seed admits it's behind — that's clarity. Winning in apps doesn't mean winning in models. Certain companies should listen."
@36Kr commenter (top thread)
"10万亿参数如果是真的,说明字节在按第一梯队的入场券做准备。但参数不等于数据,张一鸣自己把数据列为瓶颈,这个顺序说明他真懂行。"
>
"If the 10-trillion-parameter report is real, ByteDance is buying a first-tier ticket. But parameters aren't data — Zhang himself named data as the bottleneck. That ordering shows he actually understands the problem."
@silent_chain on GitHub discussion
"Everyone in the comments is debating 2029. The actual signal is the distillation ban. If the biggest Chinese lab publicly commits to clean training data, it raises the compliance floor for every other lab. That's the part that changes the industry."
*Related reading: Doubao's 12 Trillion Token Explosion · DeepSeek V5 and Huawei Ascend 950: The Sovereignty Test · Kimi K3.1 Pro and the Open-Weights Frontier Gap · Two Walls: US Chip Controls and China's Mirror Regime*