In the seventh lesson, “Scaling failure” is listed as the biggest systemic risk. This judgment is very important and worth unpacking on its own. Because if scaling really saturates, the capital expenditure narrative across the entire AI industry chain will collapse — GPU demand growth will slow, HBM will be oversupplied, and power projects will be cut.
So what signals should we watch for here?
First, let’s talk about what scaling laws are.
Over the past five years, the core logic of AI progress has been: double the model parameters, double the data, double the compute, and model capability will improve in a predictable way. This pattern has been repeatedly validated in the jumps from GPT-2 to GPT-3 and from GPT-3 to GPT-4.
The capital expenditure decisions of all the large model companies — OpenAI, Anthropic, Google, Meta — are fundamentally built on this belief: as long as we keep piling on compute, capability will keep improving. The trillion-dollar investment in the entire AI infrastructure buildout is based on this assumption.
But starting in the second half of 2025, cracks began to appear in this belief.
The first batch of signals came from the releases of the new-generation flagship models — GPT-5, Claude 4.5, Gemini 3. Compared with the previous generation, they did make progress, but the magnitude was not the kind of “generational leap” seen from GPT-3 to GPT-4.
A common observation in the industry is that even if you spend ten times as much compute, the capability improvement is far below what it used to be. Some benchmarks are already close to perfect scores, making further gains difficult; on others, such as reasoning, the improvement comes more from new methods like “Reasoning Effort” than from pure scaling.
To judge whether scaling has really saturated, you can look at a few signals.
First, the benchmark gains of the next generation of models.
If, after GPT-6, Claude 6, and Gemini 4 are released, improvements on mainstream benchmarks continue to slow — for example, if the gain over the previous generation is less than 10% — and this pattern appears for two consecutive generations, that is a very strong signal.
Second, the capital expenditure narrative of the top labs.
The core story OpenAI and Anthropic use when raising funds has always been, “I need more compute to train the next generation of models.” If these companies start talking more about “efficiency,” “synthetic data,” and “inference-time compute” as alternative paths, rather than “bigger and stronger,” that is a signal that scaling is no longer the main line.
Third, the rise of new paradigms.
In the past two years, many of AI’s important advances have come less from scaling and more from test-time thinking, synthetic data, multi-agent systems, and fine-tuning specialized models. If the returns from these directions far exceed those from “just adding another round of compute,” that means the compute demand curve will be reshaped rather than simply continuing to rise.
Fourth, the capex guidance of the hyperscalers.
In 2026, the combined capital expenditure of the four hyperscalers is expected to reach about $725 billion, up 77% from 2025. This is the largest infrastructure investment in history, and it is based on the assumption that the next generation of models will be even larger. If that assumption wavers, hyperscaler capex will slow first — and that will show up in the market before scaling failure becomes obvious.
It needs to be said that scaling “failure” does not mean “completely useless,” but rather that “marginal returns have fallen to a level that can no longer support trillions in capital expenditure.” Even if scaling slows, inference demand is still exploding, so the infrastructure story will not end immediately. But it will shift from being “training-led” to “inference-led,” and the main stock-market narrative will shift with it.
So for investors, the important question is not whether scaling will come to an end, but where scarcity is migrating between training and inference. That is where the framework of “scarcity migration paths” in the course really comes into play.