2026 Is the Year of Return: Models Have No Secrets, Applications Have Moats
Model capability progress is slowing, Chinese open-source models have compressed the gap to 6-12 months, and the API business has no monopolistic barrier; what is truly valuable is the layer of context, tools and environment outside the model.
The video won't play here. Listen to the audio instead:
The argument · tap a timestamp to hear it
The AI commercialization chain was validated this year
Late last year Yusen's analogy was: AI is still learning fast, it's a very smart elementary school student, but an elementary school student going to work probably can't earn money yet, so we first need to see technological progress, then that progress landing in products, and only then commercialization. In 2025 this chain was validated: ChatGPT revenue grew very fast, Anthropic's API grew very fast, coding applications like Cursor and Cloud Code reached close to the $1 billion ARR level, and agent applications like Manus and GenSpark are heading toward $100 million ARR. At the start of the year he judged that reasoning ability, coding ability and Computer Use together would unlock the year of the Agent, and this year we did indeed see L3-level coding agents that can program on their own for more than an hour, as well as general-purpose agent products.
— Dai YusenChinese models have compressed the gap to 6 to 12 months
At the start of the year everyone worried that Chinese companies lacked cards and lacked talent and would lag for a long time. After DeepSeek burst onto the scene, everyone found you could catch up to near the world frontier at one-tenth the cost or even less; later Chain One, Kimi K2, GLM, MiniMax and a series of ByteDance models could all, a few months after a US SOTA release, put out open-source models with close performance at one-fifth to one-tenth the cost. At the start of the year the open-source ecosystem was still dominated by Llama; now it is basically Chinese models' world, and even Cursor's Composer 1 is widely believed to be trained on Chinese open-source models. Yusen thinks completely surpassing them in the short term is very hard, but narrowing the gap to 6 to 12 months is fairly realistic.
— Dai YusenModels have no secrets, and API is not a monopoly business
Yusen says information and talent flow very fast in this industry, especially since Silicon Valley has no non-competes, so a lot of know-how cannot be fully kept secret, and on top of that everyone has hundreds of thousands of cards in hand and the data is roughly similar, so it is hard to open up a generational gap. He gives the example of Manus: at launch it was originally all Claude 3.7 APIs, because at the time only that could run agentic products, but it soon found Gemini also ran well and was cheaper, so it switched a large volume to Gemini; later Codex launched and its coding ability was also strong. For an API user, switching between models is easy, and application companies can instead draw on the strengths of many and use the best model for different tasks — which precisely shows that selling APIs itself has a low barrier and is not a monopolistic business.
— Dai YusenApplications only have a moat when they go from sashimi to a full banquet
Most of the first batch of AI applications in 2023 just made an external UI for the model; users interacted with the model through the UI and what they got was still the model's output. Yusen's analogy was ‘sashimi, which is fish plus wasabi and soy sauce’. Complex applications now are different: they need the user's own context, industry-specific data, and the history of the user's exchanges with the application (memory), which the model itself does not have; at the same time agents can use more tools and change their environment, for example writing a piece of code for their own use. So outside the model layer there is an extra context layer, and an environment-and-tools layer, and the analogy goes from sashimi to a full Manchu-Han banquet — ingredient quality still matters, but service, environment and cooking skill matter too. Manus beat all advanced models on the Remote Labor Index test released by Scale, precisely because it really does more on top of the model.
— Dai Yusen2026 is the Year of R, and the first R is Return
Over the past few years everyone bet on AI investment continuing to exceed expectations, drawn by the potential for large returns. But now the investment is getting bigger and bigger, and roughly 50% of the investment in data centers is compute; these cards may become obsolete after 4 to 6 years, so investment needs to see results within a shorter return cycle. At the same time marginal returns on the model side are declining: the progress of GPT-5 over GPT-4 is not as great as GPT-4 over GPT-3, while the investment is much larger, and it still cannot stop Chinese open-source models from approaching quickly at low cost. On the application side last year was low expectations and high growth; now everyone is very optimistic about 26 and 27, and expectations are already very high. The $20 billion question raised by Sequoia's David later became the $60 billion question, and now this account is getting harder and harder to square.
— Dai YusenThe second R is Research, the third R is Remember
Yusen says AI history has always been an alternating process of scaling and research, and now it is again time for research to improve things; Demis also believes one to two research breakthroughs are still needed to reach AGI. Silicon Valley has therefore seen a new trend of investing in New Labs — opportunities centered on researchers and exploring differentiation from leading model companies, such as Thinking Machines, Ilya's SSI, Reflections and Human End, new research labs that start out at the $1 billion valuation level. The third R is Remember/Memory: without memory, the result you get asking AI and the result I get asking AI are the same; now ChatGPT's understanding of him is already much higher than a new chatbot's, because there are three years of chat history. But memory now is basically still retrieval-based, like a friend relying on a very thick notebook to write down every sentence, rather than truly understanding you at the model weight level.
— Dai YusenThe bubble has not peaked, but long-end expectations may turn pessimistic
Yusen thinks every technological revolution in human history has brought a bubble, and it is quite natural that AI may bring the largest bubble in human history. But we are not yet at the very top of the bubble, because the top of a bubble shows asset prices absurdly high, with even bad companies and fraudulent companies very high, whereas right now short-term valuations of companies like Nvidia are even very low. What he really worries about is the change in long-end expectations: right now the short end is still very optimistic about Nvidia's expectations for next year and the year after, but if the long end turns pessimistic about a longer AI payoff cycle and a slowdown in AGI's arrival, then no matter how good short-end results are they will not be the main trading factor. He judges that US stocks may see a fairly obvious pullback next year, possibly in the second half, and OpenAI's own user and revenue growth will be an important trigger.
— Dai YusenOnce intelligence gets cheap, agency and taste are what's valuable
Yusen says that throughout human history intelligence has always been a luxury; those who could mobilize smart people at scale were either top-tier enterprises or national governments. Now $20 buys a lot of intelligence, so first, execution becomes cheap, because AI can help you execute and write code; what matters is agency and taste — what you want to do, whether you have something you want to do yourself, and which of the three methods AI gives you you choose. Second, we used to do things linearly, first A then B then C, but now many things can be done in parallel; programming with AI is giving different agents different tasks and being responsible for issuing instructions yourself — you have to be a good boss to AI. Third, use the most advanced tools more to experience the future; the people who used ChatGPT earliest may for that reason have a more optimistic judgment about AI.
— Dai YusenIn their own words · checked verbatim
AI is still learning fast. It's a very smart elementary school student. But an elementary school student going to work probably can't earn money yet. So we first need to see technological progress, then that progress landing in products, and then commercialization.
AI还在快速学习 他是一个非常聪明的小学生 但是一个小学生去打工呢 可能还不能赚钱 所以我们是要先看到技术的进步 然后技术的进步在产品上落地 然后再进行商业化
Dai Yusen3:00
If you replace a lot of programmers, it doesn't mean you can earn those programmers' salaries; it means the tasks those programmers used to complete have become worthless.
如果你替代了很多的程序员 并不意味着说你能赚到这些程序员的工资 而是说这些程序员原来完成的任务 变得不值钱了
Dai Yusen1:11:35
It's like your friend's understanding of you comes entirely from a very thick notebook recording every sentence of your chats. True memory is surely that he doesn't need to carry that notebook — he already understands you more deeply inside.
好比说 你的朋友对你的理解 完全来自于他有个很厚的笔记本 记录了跟你聊天的每一句话 那真正的记忆肯定是他不用带这个笔记本 他已经在内心深处更理解你
Dai Yusen1:21:41
When everyone is very optimistic, prepare for pessimism; when everyone is very pessimistic, prepare for optimism. I think that is a contrarian way of thinking.
就是大家都非常乐观的时候 做好悲观的准备 大家都非常悲观的时候 做好乐观的准备 我觉得才是一个 反共识的一个思维方式
Dai Yusen1:44:51
Don't define yourself with the product form of the previous era. Right, I think don't call it the so-and-so of AI — so-and-so is a product form of the previous era.
不要拿上一个时代的产品形态去定义自己 对 我觉得不要叫说AI的某某某 某某某是一个上一个时代的产品形态
Dai Yusen3:14:58
Figures
| ARR level of Cursor / Cloud Code | close to $1 billion | 3:00 |
| ARR level of Manus / GenSpark | $100 million | 3:00 |
| First-round valuation of Thinking Machine Lab | $10 billion | 53:21 |
| Second-round valuation of Thinking Machine Lab | $50 billion | 53:21 |
| Latest-round valuation of Mistral | $14 billion | 54:21 |
| Share of compute in data center investment | about 50% | 1:04:29 |
Glossary
- Year of R
- Yusen's keyword for 2026; R stands for Return, Research, Remember.
- New Labs
- New companies centered on researchers, exploring paradigms differentiated from leading model companies.
- proactive agent
- An agent that can proactively discover and solve problems without a person initiating it.
- Remote Labor Index
- A test released by Scale measuring how applications perform on real remote work tasks.
- OOD
- out of distribution, referring to data or behavior outside the training data distribution.
How to listen
Investors watching the pace of AI primary-market bets, AI application founders choosing a direction, and engineers who want to understand the moat debate around model companies.
After 2:36 the chit-chat can be skipped; just listen to the four segments on OpenAI, Google, Anthropic and Manus, the rest of the commentary is scattered.