Burning tokens to stack up agents is a transitional strategy, not the endgame
Silicon Valley is shifting from mindless token burning to careful accounting: a local open-source model can do real work on a dollar a day of electricity, yet when a genuinely hard problem shows up, a smarter model can still nail in one shot the answer a billion-token group chat could not argue its way to. Saving money and top-tier intelligence are not in conflict.
The video won't play here. Listen to the audio instead:
The argument · tap a timestamp to hear it
The era of token-burning leaderboards is already over
For the past few months, Silicon Valley's agent strategy went by the name Token Maxing. Jensen Huang told his engineers to burn more tokens, and Meta went as far as setting up a leaderboard for who consumed the most. But the wind is changing. The token allowance Uber handed out to employees was pulled back after costs overran the plan: four months burned through the entire year's budget. Stripe's internal tooling now tells you outright that Opus 4.8 is expensive and suggests starting with a cheaper model for non-core tasks. This episode is about exactly that turning point, and how the agent engineering paradigm is moving from reaching for the strongest model by reflex toward using tokens economically.
— HongjunFour or five hundred dollars a day in tokens is not expensive
Since last November, Huang Dongxu has been using Opus to build a distributed database, db9, from scratch. Even with a $200-a-month Claude Code Pro Max subscription, once he exceeded the included allowance he had to buy additional usage out of pocket, running four or five hundred dollars a day. His judgment is that the money is well spent. A product that was previously impossible to build, given the technical and headcount limits, now takes one person three months, and this piece of software could have brought the company $10 million a year in direct sales revenue. The ROI covers the token cost completely.
— Huang DongxuOpenClaw matters as a signal, not as a product
Zhang Hongjiang argues that OpenClaw's real significance lies not in the product itself but in the signal it sends: Peter, working alone, spent one week and produced an agent like this. That says the reasoning and coding ability of large models has become strong enough that a single architect-level engineer can stand up a system fast, which marks AI genuinely crossing over from the research stage into engineering deployment. You no longer need a deep understanding of how the model works internally to use it to build the system you want.
— Zhang HongjiangHermes did not solve memory, it just routed around it
The complaint most often leveled at OpenClaw is poor memory: once the context is compressed it forgets things easily, and its behavior turns erratic. After switching to Hermes, Huang Dongxu found that it does not solve the memory problem itself but sidesteps it cleverly. Its agent loop includes a mechanism that continuously summarizes successful experience and hardens it into reusable skills, so the longer you use it, the more best practices it grows. He sees this as fundamentally an engineering problem, unrelated to model capability, resting instead on better architectural design and a team's willingness to polish.
— Huang DongxuMulti-agent group chats exist to keep the agents from stopping
Huang Dongxu uses Slock to build multi-agent discussion groups for code review: top-tier models from different vendors pick holes in each other's work and try to persuade each other, and it does not stop until nobody can find another problem. The price is token consumption ten times that of a single model, peaking at 1 billion tokens a day and a bill of three or four hundred dollars. The Slock team calls this mechanism, keeping an agent working through rounds of discussion, agent dynamics. In his view, every harness and every loop engineer is fundamentally solving the same problem: how to keep an agent working without stopping.
— Huang DongxuOn a single hard problem, a stronger model beats multiple agents
Huang Dongxu originally believed that having multiple agents debate each other was necessary, because a single model's one-shot success rate was not high enough. The arrival of GPT-5.5 and Fable 5 broke that assumption: a complex bug the group chat spent ages on and could not resolve, fed to Fable 5, would often be pinned down exactly on the first try, with no prior context, purely on raw intelligence. His conclusion is yes and no. On a single hard problem, a stronger model really can wipe out collaborative discussion, but extreme hell-difficulty problems still need a sufficient harness as a backstop.
— Huang DongxuLocal open-source models can already enter a real workflow
DeepSeek V4 and GLM-5.2 are what made Huang Dongxu trust an open-source model in an actual workflow for the first time. Running DeepSeek V4 Flash locally on a Mac Studio, speed reaches about 30 tokens/s, close to the cloud API experience, and the cost is only electricity and depreciation on the machine: roughly $1 a day, twenty or thirty dollars a month. He now hands the local model both his high-volume repetitive work and the tasks he would never have dared run on a paid model before, such as summarizing several hundred papers in one go. His bill has come down from the frenzied code-piling period to two or three hundred dollars a month today.
— Huang DongxuAgent startups compete on data moats, not on models
Asked whether now is a good time to start an agent company, Huang Dongxu says yes, but the key is whether you can hold a moat of enterprise-grade vertical data rather than being flattened at any moment by the next jump in model capability. He likes three directions. First, middle-layer infrastructure: memory, data platforms and context management. Second, observability, which can answer the question every CIO is asking, namely where the tokens actually went and which of it was waste. Third, a new generation of cloud infrastructure built for agents rather than human developers, such as InsForge, which he has invested in.
— Huang DongxuIn their own words · checked verbatim
Back in December last year, in another interview, this exact question came up, and I said at the time that there was no bubble.
去年的12月份 我在另外一个采访里面 这就是一个问题 我当时就说没有泡沫
Zhang Hongjiang5:03
Peter, as the author of OpenClaw, was able to spend one week and just build OpenClaw. For me, the biggest significance of that is it announced that the capability of large models has crossed a threshold.
Peter作为OpenClaw的作者 他能够花一个礼拜 就把OpenClaw做出来 对于我来说最大的意义在于 宣布了大模型的能力过了一个坎
Zhang Hongjiang15:09
It actually sidesteps the memory problem very cleverly. It isn't that the memory module is brilliantly designed; it's that inside its agent framework there's a strong bias toward constantly summarizing successful experience, summarizing it into skills.
它其实很巧妙地绕开了记忆的问题 它并不是说 记忆模块设计得多么出彩 而是它在它的Agent的框架里边 有一个很强的倾向 去不停地总结成功经验 总结成skill
Huang Dongxu23:14
I think the Slock team invented a really interesting term, ‘agent dynamics’.
我觉得Slock它们团队 发明了一个很有意思的词 叫做“Agent动力学”
Huang Dongxu33:19
The thing they couldn't argue their way to might get solved in one shot in the face of Fable 5's raw intelligence. So at that point you realize, I burned 1 billion tokens earlier, and on the Fable 5 side it was nailed in one shot, very precisely. So it's absolutely intelligence steamrolling everything.
讨论不出来的东西 可能在Fable 5的绝对智商面前 一次就解决了 所以这时候你就会发现 我前面烧了10亿Token 在Fable 5这边 非常精准地一次搞定了 所以绝对的智商碾压
Huang Dongxu36:21
Is it important that the decision-making authority sits with the agent? I think it is important. The agent has to make decisions, yes. You have to believe the agent's decisions are more reliable than your own.
决策权在Agent是重要的吗 我觉得是重要的 就是Agent要做决策 是的 你要相信Agent的决策 比你的决策更靠谱
Huang Dongxu1:02:45
What I found is that you get this state where the agents blame each other. For instance, sometimes they just stop, and I say, why did you two stop here, do a retrospective. So they start summarizing the day's work, and then Agent A blames Agent B, saying, look, I stopped entirely because of you. I've told you several times already, don't do it this way.
我就发现 会有Agent之间 互相blame（指责）的状态 就比如说 有时候它们就停下来了 我说为什么你们这个就停下来了 复盘一下 然后它们就开始总结今天的工作 然后Agent A指责Agent B说 你看 我停下来都是因为你 我已经说过好几次了 你不要这么做
Huang Dongxu1:03:45
So on this point, I believe the turning point of intelligence, the singularity, has already arrived. I've said this a few times before: the singularity arriving comes down to machines' ability to learn surpassing humans' ability to learn.
所以从这点上 我认为智能的转折点 就奇点已经到来 我之前也讲过几次 就是这个奇点到来就在于 机器的学习的能力 超越了人的学习能力
Zhang Hongjiang1:16:52
Figures
| Huang Dongxu's single-day spend using Opus | $400-500/day | 12:08 |
| Peter (OpenClaw's author) 30-day API spend | $1.3 million | 20:13 |
| Slock group chat's single-day token consumption and matching bill | 1 billion tokens, $300-400 | 32:18 |
| Local open-source model (DeepSeek V4 Flash) inference speed | about 30 tokens/s | 37:22 |
| Electricity cost of the local open-source model | about $1/day, $20-30/month | 38:23 |
| Huang Dongxu's current agent bill | $200-300/month | 40:25 |
| Manus Pro pricing | $200/month | 59:44 |
| Headcount of Microsoft's newly created Frontier Company (FDE) team | 6,000 people | 1:00:45 |
| Annual rate of token price decline | 10x cheaper every year for the past 3-4 years | 1:10:48 |
Glossary
- OpenClaw
- An open-source personal-assistant agent project that advocates local-first design and keeping your own privacy
- Agentic Loop
- The internal loop in which an agent repeatedly calls the model, executes tools, and updates its context
- FDE (Forward Deployed Engineer)
- Engineering teams that model companies station at enterprise customers to get agents deployed in production
- Agent Native
- Product design whose business logic collapses if you take the AI or the agent out; the agent is the primary user
- Jevons Paradox
- The phenomenon where efficiency improves and unit cost falls, yet total consumption grows faster than before
- Gödel machine
- A system that can, in theory, prove things about itself and rewrite its own code to keep self-improving
How to listen
Technical leaders costing out agent engineering and weighing whether to stand up their own local models, plus investors watching how token economics reshapes the balance between cloud providers and model vendors
The discussion after 1:18 about AGI-driven unemployment and the meaning of individual existence is lower in information density. If you only want the conclusions, skip it.