Stratechery overturns the AI bubble. What do we do with AI

Original title: Agent Over Bubbles
Original by Ben Thompson, Stratechery
Photo by Peggy Block Beats

EDITOR: AGAINST THE BACKDROP OF THE CONTINUED WARMING OF AI INVESTMENT AND INDUSTRY NARRATIVES, THE EXISTENCE OF BUBBLES HAS BECOME A CENTRAL ISSUE OF THE MARKET'S DISCUSSIONS. ON THE ONE HAND, EXTREME RISK NARRATIVES HAVE REINFORCED FEARS OF TECHNOLOGY LOSING CONTROL; ON THE OTHER HAND, RAPIDLY EXPANDING CAPITAL SPENDING AND VALUATION LEVELS HAVE KEPT THE BUBBLE THEORY ALIVE. UNDER THIS DIVERGENCE, MARKET JUDGEMENTS SHOW OBVIOUS UNCERTAINTY。

Ben Thompson, the founder of the Science and Technology Analysis Platform Stratechery, has long focused on the evolution of technology industry structures and business models. On the occasion of the GTC 2026 meeting, he revised his earlier judgment that "AI is in a bubble": instead of looking at it as a bubble, he understood it as a round of structural growth driven by technological paradigm changes。

This judgement is based on observations of the LLM three key leaps. Since 2022, when ChatGPT first demonstrated its ability to model large languages to the market, LLM has evolved from “available but unreliable” to “compellable” to “capable of carrying out its tasks”. In particular, at the end of 2025, with the release of Anthropic Opus 4.5 and OpenAI GPT-5.2-Codex, the agentic workload began to move from concept to reality。

The key is not the model itself, but the emergence of the "entharness". Agent decorates the user with the model, is responsible for the movement of the model, calls the tool and validates the results, and transforms AI from a tool requiring continuous human intervention to an implementation system that can assign tasks. This change not only increases reliability, but also extends the AI application boundary。

Based on this paradigm shift, the authors further note that the expansion of AI demand no longer depends on the size of the user, but more on the ability of the unit user to move; at the same time, the presence of a "winner-for-all" load on the agglomeration load will continue to boost the demand for high performance algorithms and provide structural opportunities for chip manufacturers and cloud service providers。

UNDER THIS FRAMEWORK, CURRENT LARGE-SCALE CAPITAL SPENDING IS NO LONGER MERELY A SPECULATIVE BET FOR THE FUTURE, BUT MORE LIKELY A PRECURSOR TO REAL NEEDS. AS AI MOVES FROM "AID TOOLS" TO "IMPLEMENTATION INFRASTRUCTURE", ITS ECONOMIC IMPACT MAY HAVE JUST BEGUN TO APPEAR。

The following is the original text:

In the past, I preferred the latter, and even thought that foams were not necessarily bad at certain stages。

BUT AT THIS MOMENT, AS I WAS AT THE OPENING OF THE GTC IN MARCH 2026, IN INVERDA, MY JUDGMENT CHANGED: IT WASN'T NECESSARILY A BUBBLE. (THE IRONY IS THAT THE JUDGMENT ITSELF MAY BE EXACTLY THE SIGNAL OF THE BUBBLE

LLM THREE PARADIGM LEAPS

IN THE LAST FEW WEEKS, WHEN I WAS TALKING ABOUT INVERDA AND ORACLE, I REPEATEDLY MENTIONED THAT LLM HAD UNDERGONE THREE KEY LEAPS。

Phase 1: ChatGPT

The first point was the release of ChatGPT in November 2022, which hardly needs to be repeated. Although a large-language model based on Transformer had emerged as early as 2017 and continued to improve its capabilities, it had been underestimated for a long time. Even in October 2022, in an interview with Stratechery, I argued that this technology, while amazing, lacked productization and entrepreneurial motivation。

But a few weeks later, everything turned upside down. ChatGPT makes the world truly aware for the first time of the capabilities of the LLM。

However, the earlier version was also impressed by two observations, in particular by the "foam theorists":

First, models often go wrong, even when they don't know the answer. It makes it more like a "dazzling tool" that's amazing but unreliable。

Secondly, it is still very useful even then, provided that you know how to use it and that you constantly verify the output and correct the error。

phase 2: o1

The second turning point is the o1 model published in September 2024 by OpenAI. At that time, LLM had made significant progress with stronger basic models and post-training techniques, with more accurate outputs and fewer hallucinations。

but the key breakthrough is that it "think" and then answers。

THE TRADITIONAL LLM DEPENDS ON THE PATH, AND ONCE IT GOES WRONG IN THE REASONING, IT GOES WRONG. THIS IS A FUNDAMENTAL WEAKNESS OF THE SELF-RETURN MODEL. THE REASONING MODEL, ON THE OTHER HAND, SELF-ASSESSSSSSSSS THE ANSWERS, AND IT PROVIDES AN ANSWER, JUDGING WHETHER IT IS CORRECT OR NOT, AND IF NECESSARY, TRY ANOTHER PATH。

This means that the model begins to proactively manage errors and reduces the burden of user intervention. The results are also significant. If ChatGPT's breakthrough is "letting LLM work," then O1's breakthrough is "make LLM reliable"。

Phase 3: Agent (Opus 4.5 / Codex)

At the end of 2025, a third leap was made。

In November 2025, Anthropic released Opus 4.5, initially resonating flat. But in December, Claude Code, who carried the model, suddenly showed unprecedented capability; almost simultaneously, OpenAI released GPT-5.2-Codex, showing a similar level。

People have been talking about “Agent” before, but at this moment they have finally begun to really accomplish their tasks, even a complex task that takes hours, and is done correctly。

The key is not in the model itself, but in the control layer (harness), which is the software layer of the movement model, the call tool, the execution process. In other words, the user no longer directly operates the model but delivers the target, using the Agent schedule model, the call tool, the execution process and the validation of the results。

Take programming as an example:

Phase 1: Model generation code

Phase II: Models are reasoned during generation

Phase three: Agent Generation Code → Performs testing → Auto-run testing → Replays if wrong, without continuous intervention by the user。

This means that the core flaws of the ChatGPT era are being systematically addressed, with higher correctness, greater reasoning and automatic validation mechanisms。

The only remaining question is: What should we do with it

The threshold of initiative is falling

I have repeatedly emphasized these three points in order to explain why the entire industry is suffering from a serious deficit of capacity and why excessive capital spending is justified。

There are three paradigms in which the need for computing is completely different:

• Phase I: calculator-intensive training, with lower reasoning costs

phase ii: surging costs of reasoning (more token + more frequency of use)

Phase 3 (Agent): Multiple calls to reasoning models, Agent itself consumes calculus (or even favours CPU), uses frequency for further explosions

But more importantly, the third point is that changes in the structure of demand are seriously underestimated。

Currently, there are far more people using chat robots than there are people using Agent, and many people are not really using AI. This is because the use of AI requires "activeness". LLM is a tool that has no purpose, no will, but is called on its own initiative。

But Agent changed that, and it reduced the demand for human initiative. In the future, a person can command multiple Agents。

This means that even a small number of people are “active” enough to generate huge computing needs and economic output。

AI STILL NEEDS TO BE "MAN-DRIVEN" BUT NO MORE "MAN-PEOPLE"。

Pay drivers for enterprises

THE LIMITED WILLINGNESS OF THE CONSUMER SIDE TO PAY AI HAS BECOME CLEAR. IT IS BUSINESS THAT IS REALLY WILLING TO PAY FOR PRODUCTIVITY。

THE MOST EXCITING THING FOR BUSINESSES IS NOT JUST AI TO IMPROVE EFFICIENCY, BUT AI TO REPLACE MANPOWER AND BE MORE EFFICIENT。

The reality is that it is often a few in large companies that really push business forward, but the organization is large, with significant coordination costs. Agent's role is to magnify the influence of "value-driven people" while reducing organizational friction。

The result is "a smaller number of people, higher output, lower cost." That is why future layoffs are likely to be more than "cyclical adjustments" but structural changes。

THE COMPANY WILL RETHINK NOT ONLY WHETHER THERE ARE MANY PEOPLE IN THE EPIDEMIC, BUT ALSO WHETHER WE DON'T NEED SO MANY IN THE AI ERA

Why isn't this foam

From this perspective, the logic of "not foam" is clearer:

1. THE CORE WEAKNESSES OF LLM ARE BEING CONTINUOUSLY ADDRESSED BY THE ALGORITHM AND ARCHITECTURE

2. Number of people driving demand is falling

Agent's gain is not just a drop, but a gain

It is therefore not difficult to understand why all cloud manufacturers are saying that the capacity to calculate is short of demand and that capital spending has continued to increase significantly。

Agent Reconstructing Value Chains

Another key question is, if models are eventually commodified, can OpenAI and Anthropic make money

The traditional view is not, but Agent changed that. The point is that the real value is not in the model itself, but in the integration of the Model + Control System。

Profits tend to flow to the "integration layer" rather than alternative modules. Like apples, its hardware is not commodified because it is integrated with software. Similarly, Agent needs the depth of synergy between models and Harness, which makes OpenAI and Anthropic key integraters in the value chain rather than an alternative。

Microsoft's shift was a sign that it had emphasized "models can be replaced" but had to give it up when the real Agent product was launched。

This means that models are not necessarily fully commodified, because Agent needs integration capabilities。

The final paradox

I have to go back to that paradox。

I always thought that as long as everyone was worried about bubbles, they were not bubbles; the real bubbles were when no one questioned them。

And now my conclusion is that this is not a bubble。

But if "I said it wasn't bubble" itself, it proved to be foam, that's all。

[ Chuckles ]Original Link]