Summary

The episode opens with the popularity of Claude Code and OpenClaw as the backdrop; 于文渊 says Bailian’s Token volume has essentially doubled every month since launch, with Open Cloud and Cloud Code clearly accelerating growth, especially in Agent use cases. Most of these Tokens come from high-quality frontier models; just 2 weeks after Qwen 3.5 launched, peak Tokens per Minute had reached a level no text model in the platform’s history had ever reached, pushing AI from a chat tool into the productivity stack.
The “100M Token Club” is no longer enough to measure demand; the real metrics are Token quality, peak capacity, and the ability to convert GPUs into Tokens with a better user experience and lower cost. A Token from a 0.6B small model is not equivalent to one from an Embedding model or a SOTA deep-reasoning model; tens of thousands of people join this heavy-consumption cohort every day, and 于文渊 even suggests raising the threshold to 1B while tracking time to first Token, generation speed, load smoothing, and GPU utilization.
AI will rewrite both cloud customers and cloud architecture: over the next 3 to 5 years, a large share of work done by people today will be handed to AI, while compute, storage, networking, and scheduling will look completely different. Bailian’s engineering mission is to “keep every GPU from sitting idle for even 1 second” without sacrificing stability, making GPU pools from 1K to 1M look to users like an elastic cluster accessible through a single API; the cloud’s primary users in the future may even be Agents.
于文渊’s most aggressive MaaS call is that “there is no situation that requires self-hosting,” because the cost, security, and flexibility companies think they gain from building in-house are precisely the 3 problems MaaS is more likely to solve. Per-Token optimization, GPU fill rates, and switching among multiple models all require sustained infrastructure investment; he also says confidential inference can keep the platform from seeing model files or requests, with customers holding the keys. But he explicitly framed this as a personal “hot take” representing Bailian alone, not Alibaba Cloud.
The danger in AI Coding is not that models cannot write code, but that companies set “share of code generated by AI” as a target and skip the process knowledge engineers need to develop judgment. 于文渊 endorses prototyping but believes AI is “not ready yet” for mission-critical code; if newcomers hand problems to AI from day 1, they may settle for 99% correctness and never find the 1% of problems.
The work most susceptible to a major AI productivity boost may not be front-end work close to the product and user, but systems work in operating systems, databases, and file systems, where objectives can be formalized and tests are clear. A FAST paper from last year showed that a 32B model could write a file system when the specification was sufficiently precise; 于文渊 is more convinced by “back coding”—write clear requirements or specs first, then have AI implement them. The closer a product problem is to people and the more open-ended the answer, the harder it becomes; MaaS engineering also remains an open problem as models and underlying resources keep changing.
China’s hard constraint in AI is total compute supply, not just the performance of individual chips; 于文渊 compares it with “cars already running on the highway, but there isn’t enough fuel.” He believes China can achieve autonomous, controllable compute and build the world’s leading infrastructure, but the current supply gap will seriously affect AI development; the end state may resemble “water, electricity, gas, and models” as utilities, without a powerful monopoly around any single model, but with infrastructure built from AIs offering different speeds, performance, and functions.

Deep dive

1. The Token Curve Keeps Climbing; Agents Are Just the Latest Accelerator

At the opening, the host attributed the Token shortage to a continuing lack of sufficiently cost-effective compute; 于文渊 said Bailian’s Token demand has not let up since day 1, with growth running at roughly 2x month over month. The host brought up Claude Code and OpenClaw; when pressed later, 于文渊 said Open Cloud and Cloud Code had indeed accelerated growth, particularly in Agent use cases.
Just 2 weeks after Qwen 3.5 launched, peak Tokens per Minute had reached a level no text model in Bailian’s history had ever reached. Even with an Alibaba CEO who is unusually aggressive on compute spending, the combination of model R&D and customer service still leaves capacity insufficient.
Video generation once showed the same pattern: once a model crossed the usability threshold and moved from Demo to a product that could genuinely be sold, demand surged suddenly. He therefore finds it difficult to predict the next breakout use case in the near term, but is certain that this is “just the beginning” (仅仅是一个开始).
Token count itself is misleading: a Token from a 0.6B small model, an Embedding model, and a SOTA deep-reasoning model are not equivalent in compute, intelligence, or capability. Tens of thousands of people join the heavy-user cohort consuming 100M Tokens per day every day; he did not disclose the cohort’s current size, only that it keeps growing, and said the club’s threshold should be raised to 1B.

2. The Cloud’s Next Customers May Not Be People but Agents

于文渊’s 3-to-5-year forecast is sweeping: a large share of work done by people today will be handled by AI, while the shape of data centers and the compute, storage, networking, and scheduling systems inside them will be “completely different” from today.
When the host asked whether Neocloud would rewrite the existing cloud landscape, he declined to take a definitive position, confirming only that the contours of a reshuffle are already visible. Every cloud provider has to answer the same question: will the cloud’s users in the future all be Agents rather than humans?
While maintaining reliability and security, Bailian wants to fully utilize GPU pools ranging from 1K to 10K, 100K, and ultimately 1M GPUs—to “keep every GPU from sitting idle for even 1 second”—so users can access extreme elasticity and stability through a single API.

3. Global Expansion Can Flatten Load Curves—but Time-Zone Lags Limit the Upside

The host proposed using time-zone differences to keep GPUs running around the clock; 于文渊 agreed that “Token going global” is critical, but said domestic and overseas businesses cannot move in lockstep. Domestic and international growth might be only 2 months apart, yet international business could still account for just a single-digit percentage, remaining small relative to the overall market.
The shared platform ultimately has to serve both domestic and international markets, but it must overcome geopolitical and compliance hurdles. He still sees Chinese companies taking AI overseas as “the broader trend and impossible to stop.”
Bailian’s real-world deployments go beyond Coding: one beverage company placed a bot in distributor group chats, where distributors describe the product category and number of cases in natural language and the bot uses historical purchase data to replenish inventory directly. The interface is no longer a new system, but “like talking to a real person.”

4. MaaS’s Moat Is End-to-End Coordination, Not Simple Model Resale

于文渊 attributes Bailian’s differentiation to end-to-end coordination: Alibaba Cloud’s accumulated infrastructure, scale, and product lines; models supplied by sibling teams in the same lab; and the T-Head chip team all work together to tune the models, inference-serving frameworks, and hardware. The system is continuously debugged before launch rather than treating the model as a black box.
His quality promise is specific: the scores listed on the Model Card for every Qwen model “can definitely be achieved” through the Bailian API. When customers deploy Qwen or other open-source models themselves, performance, quality, and speed may not be reproducible.
Asked how thick the platform layer is between the model and the API, 于文渊’s answer was that the Qwen API is itself the Bailian API, while the real value lies in experience, cost, model performance, and capacity: “Whoever can convert compute into Tokens more efficiently has more compute.”
Bailian is not a single-model gateway either. Beyond hosting open-source models from China, users can use a Bailian API key to call MiniMax, Kimi, SiliconFlow’s DeepSeek API, and even models from inference providers. The platform serves customer demand and does not require customers to use only Qwen.

5. “There Is No Case for Self-Hosting” Reverses the Three Standard Arguments

Asked when companies should deploy privately, 于文渊 first admitted he was stumped, then offered a personal “hot take” that represents Bailian, not Alibaba Cloud: “I don’t think there is any situation that requires self-hosting.” Companies typically buy GPUs for cost control, security, and flexibility; he believes MaaS is better positioned to deliver all 3.
Cost cannot be judged by purchase price alone: companies must continuously keep up with changes in models and algorithms, optimize per-Token cost, fill GPU capacity, and balance service quality across multiple models. Every company would need infrastructure engineers working on these optimizations continuously, which is “actually very difficult.”
On security, he said cloud providers cannot see and will not inspect user data. Bailian is also promoting confidential inference, which would keep the platform from seeing model files and any requests, with end-to-end keys held by the customer. This, he said, is the guarantee provided by cryptography.
Flexibility also favors MaaS, because “the greatest certainty today is actually uncertainty about what AI will need tomorrow.” Model architectures and inference efficiency continue to change, and fixed GPU purchases are not as flexible as companies imagine.

6. AI Coding’s Ceiling Depends on Whether Humans Retain Judgment

于文渊 still recommends that students study computer science. He cited a judgment from the 1980s, relayed by a lab mentor: “There will be 2 kinds of people in the future: those who are used by computers, and those who use computers.” When he studied at university in the 2000s, he thought the statement still held; it still holds today.
His counterintuitive advice is that newcomers should not let AI write too much code. He used the example of medical interns to make the point: if someone accepts only AI conclusions from the moment they see their 1st patient, even a 99% accuracy rate leaves them without the experience to identify the remaining 1% of errors.
The host cited Cursor trend data showing the share of people mindlessly using Tab to autocomplete code had moved from roughly 20%-30% to 70%-80%. 于文渊 said he gets “very nervous” in Code Review when code is obviously AI-generated. Vibe Coding is already good enough for Prototypes, but production code requires understanding whether every line does what it should and whether side effects such as memory leaks and open file handles are acceptable.
He considers setting the share of AI-generated code as a corporate target “extremely dangerous,” because a company’s history, a founder’s style, and process knowledge cannot be conveyed through a few prompts. The better framing is that 1 AI can do work that previously required several engineers, not that 1 AI replaces several engineers. Using AI still presupposes that you can do the work yourself.
He is more persuaded by “back coding”: write a clear requirements document or specification first, then have AI implement it. That sets a higher bar for architects, who must be able to explain the work precisely; today, companies cannot expect complex work to be done well with only 2 or 3 prompts.

7. Formalized Work Gets Boosted First; Compute Eventually Becomes “Water, Electricity, Gas, and Models”

The host used IKEA furniture to illustrate process knowledge: having the materials and instructions is still not equivalent to having a craftsperson who has assembled the furniture many times. Even if AI eventually reaches 99.9%, programmers still need to build their capabilities around the 0.1% it cannot do; 于文渊 agreed that proficiency comes from repeatedly completing the process itself.
The host suggested that the people most likely to be replaced might be the best writers of code; 于文渊 added that front-end engineers close to the product and user often need the know-how to get users genuinely using the product, making them harder to replace. By contrast, systems work in operating systems, databases, and file systems has clearer objectives, while codebase quality, test cases, and results are easier to define—making it more susceptible to a major AI productivity boost.
A FAST paper published last year at a top storage conference showed that a 32B model can write a file system when the specification is sufficiently clear. But 于文渊 stressed that this does not mean complex work can be done with only 2 or 3 prompts; MaaS engineering remains an open problem because models, underlying resources, and compute are changing too quickly.
China’s compute problem is like oil: the question is not whether China can produce oil or how many oil fields it has, but whether daily demand matches the oil that can be supplied each day. He is confident that China can achieve autonomous, controllable domestic compute and build the world’s leading infrastructure; but the current supply gap will affect China’s AI development, and the impact of Nvidia supply—or a cutoff—would be “large, extremely large.”
Nvidia is the de facto standard. T-Head is a team he has used personally, and he considers its developer experience “extremely smooth” and its efficiency high. He has not personally used Moore Threads or MetaX, so he offered no further assessment.
Whether the oil is light or heavy, “someone will always be able to refine it”; the key remains total supply. He asked: if the business can grow 10x next year, why can’t compute grow 100x? Even if given 1,000x more compute, he believes stronger model training and making more AI applications cheaper would absorb it.
Asked about the landscape at the end of 2026, he would say only that Agent and AI generation will be among the biggest sources of growth this year, without judging which will be larger. On Neocloud, he is not particularly bullish on companies that simply resell Nvidia or similar compute; he prefers AI-native services that abstract away hardware and complexity, including companies such as Firefox and Together, as well as sandbox hosting, cloud desktops, browsers, search, and Agent observability. He also pointed to opportunities resembling Datadog.
The end state may not produce a powerful monopoly around any single model, but look more like the power grid: users do not distinguish between nuclear and hydro power; they only need 220V AC from the outlet, while the backend accommodates AI with different speeds, performance, and functions. After water, electricity, and gas, models too will become infrastructure-grade utilities.

134. A Review of Data with 谢晨: New Oil, Data Pyramid, and the Recipe Theravance's strategic review with Andy Summers $TBPH