Newsletter — week of August 31, 2026

Market performance

The S&P 500 ends the week nearly flat, with a gain of about 0.1%, in a week shortened by the Labor Day holiday on Monday and marked by large daily swings. The index first suffered at the very start of the week from a renewed military escalation between the United States and Iran, before rebounding strongly on Thursday thanks to remarks deemed reassuring from a Federal Reserve official on the path of rates. Friday, however, closed the week on a more cautious note after a U.S. jobs report came in much stronger than expected.

Market news

This week’s newsletter will be a little special, focused on the announcement of ChatGPT-6 and its implications for the AI market.

OpenAI launched GPT-6 Astra on September 3, 2026, its new flagship model. Here are the key takeaways:

It’s a real generational shift (not just an update like GPT-5.5 or 5.6).

It’s built for « agentic » AI: able to work on its own across multiple steps (navigating software, coding, correcting itself), not just answering questions.

It was trained on more than 100,000 chips (GPUs) at the Stargate site in Texas, the heaviest training run ever done by OpenAI.

API pricing: $10 per million input tokens, $50 for output—2.5 times more expensive than the previous model (GPT-5.6 Sol).

The Jevons paradox and AI

In the 19th century, an English economist named Jevons noticed something surprising: when steam engines became more efficient (and therefore used less coal per unit of work), total coal consumption rose instead of falling. Why? Because the efficiency made the steam engine profitable for many new uses, so it was used far more.

That’s exactly what’s happening with artificial intelligence today, and GPT-6 Astra is a fine demonstration of it.

The unit cost of a token collapses, while demand for compute grows faster still

A token is the basic unit processed by a ChatGPT-style language model, corresponding roughly to a fragment of a word. For a given quality level, its cost has fallen dramatically: a GPT-4-generation model (2023) cost about $20 per million tokens at the time, versus less than $1 today for an equivalent level of performance.

Over the same interval, however, the real infrastructure needs grew at a much faster pace. The capacity of the data centers operated by OpenAI went from 0.2 gigawatt (GW, a unit of electrical power) in 2023 to 1.9 GW in 2025—a nearly tenfold increase in two years.

This is because the number of tokens generated has exploded, as you can see below (source: tokenperday.com):

Agentic AI as the structural driver of this rise

GPT-6 Astra isn’t limited to answering one-off queries. It’s designed to carry out multi-step tasks autonomously: navigating an interface, producing code, checking the result, starting over if needed. Each task of this kind triggers between 10 and 20 successive calls to the model, versus a single one for a classic conversational query. According to a Gartner analysis, agentic workflows thus consume between 5 and 30 times more tokens than a standard generative query.

This is precisely the Jevons paradox mechanism applied to AI: the fall in unit cost does not reduce total spending—it broadens the range of tasks entrusted to the model, which fuels an increase in demand for compute greater than the price decline itself.

Why the market’s top-performing segment stays expensive for the long run

While the market’s average cost declines over time, the price of the best-performing model of the moment never falls durably. GPT-6 Astra is priced 2.5 times higher than its predecessor, precisely because it requires more compute to produce each token. Each new generation of cutting-edge model thus raises the price ceiling of the high-end segment.

Below is the chart showing the average price of the most powerful AI models, which climbed from about $1.3 to more than $2.3 per million tokens between January and August 2026, with a marked peak in July—illustrating the continuous rise of the price ceiling at the top of the market:

This chart is a direct, quantified confirmation of the Jevons phenomenon for AI. While the market’s average cost declines over time, the cutting-edge segment—where both GPT-6 Astra and Claude Fable 5.1 sit (both at $10/$50 per million tokens)—keeps rising.

Conclusion

Two opposing dynamics combine, in line with the mechanism Jevons described:

The unit cost falls, which continuously broadens the field of profitable uses for AI.

The volume consumed grows faster still, under the combined effect of new uses and agentic workflows that are structurally more token-hungry.

The cutting-edge segment stays expensive, with each new benchmark model, such as GPT-6 Astra, redefining the price ceiling upward.

The trajectory taking shape is therefore not that of an artificial intelligence that is « less and less costly » in absolute terms. It is that of an AI that is less and less costly at constant quality, but whose total spending—along with the chip and electricity needs underpinning it—keeps growing, precisely because the demand it generates keeps expanding.

This phenomenon is fundamental to understand if you want to invest in AI. And that’s why we remain invested in the companies we consider the best positioned to benefit from these ever-growing compute needs.

Our subscribers already know it: this conviction shows up in the numbers. Since January 1, our portfolio has posted a return of 45.9%. To access the full composition of our positions, our trades, and our weekly analysis, join our subscribers now.