🔴 The Sense of a Bubble Put to the Test by Infinite Capital

More than 6,000 of you read us every weekend. Thank you for your trust.

Last week, we discussed an AI infrastructure bubble.

• With a major risk: the explosion of secondary debt, the debt of suppliers who, faced with the data center building frenzy, have to borrow and carry a latent risk. That risk exists in Europe too.
• But while the valuations and the sums are large, many Silicon Valley investors keep arguing that this is not a bubble but a rational realignment of capital.

This week, we are going to explore this new industrial reconfiguration and its economic and political consequences. Taking Silicon Valley's side.

A quick dive into the current thinking on AI.

First of all, the AI market has changed profoundly since October 28, 2025. OpenAI has finally made its new agreement with Microsoft official.

• The cap on returns, which was constrained by the original structure, has been removed. Under the OpenAI Foundation, a for-profit entity has been created, the OpenAI Group PBC (Public Benefit Corporation), which now concentrates all commercial activities: model development, products (ChatGPT, API, etc.) and industrial partnerships.
• Microsoft holds roughly 27% of the equity in this structure and retains the intellectual property rights to OpenAI's models until 2032.
• Microsoft remains the main compute provider via Azure (with a commitment of roughly $250 billion over the term). From a marketing standpoint, it was important for Microsoft to be able to tell its customers that it hosts ChatGPT.
• But the agreement opens up, for the first time, the possibility for OpenAI to use other clouds and to work with third parties on non-API products, marking the end of Microsoft's total exclusivity.
• Finally, the AGI clause, which stipulates that if OpenAI reaches artificial general intelligence (validated by an independent panel) then Microsoft's rights will have to be renegotiated, still seems to be in place.

Fear of the bubble bursting was surely one of the main reasons for a quick deal. Nobody, least of all Microsoft, whose stratospheric valuation is tied to AI, had any interest in leaving OpenAI's situation wobbly.

We are now in a new world with new rules, which we are going to study in detail here.

Key takeaways

• This is not a speculative bubble: Unlike the 2000s, today's infrastructure (GPUs, data centers) is actually being used, with a positive return on investment.
• The tech giants control the entire value chain: Google, OpenAI, Nvidia and Meta have the financial, human and technical resources to lock down innovation around AI. It's going to be very hard for companies like Mistral (which wants to raise 6 billion to stay in the race).
• AI chips are becoming the new battlefield: The Google (TPU) vs. Nvidia duel structures the coming balance of power and explains Nvidia's latest circular partnerships. Google has announced it will sell its chips to other players.
• Falling gross margins are the new business model: Unlike SaaS software, AI applications consume enormous amounts of compute. Thin margins are offset by real adoption.
• Business models are going to be profoundly altered: The shift from usage-based billing to outcome-based billing will change the business model of AI.
• The robots are coming: Tesla and China are leading the humanoid race. This is the new playing field.

One thing is certain: we are going to have to learn to normalize investments of several trillion dollars and a decline in tech's historically high margins.

AI is still in a phase of massive industrial integration. Here are a few elements from a note written for one of our clients on these questions.

Structural differences from the 2000 bubble

97% of installed fiber was unused (not “lit”) at the peak of the 2000 bubble. Today, every GPU is saturated.

• Valuations are far more rational than they were back then (Nvidia at 40x earnings vs. Cisco at 150x in 2000).
• Google, Meta, Microsoft and Amazon have invested heavily in AI, buying GPUs and building data centers.
• Since this ramp-up in spending (capex), their return on invested capital (ROIC) has risen by about 10 points. That means these massive investments are profitable.
• Unlike the 2000 bubble, when infrastructure was underused, here the resources (GPUs) are fully exploited, which directly improves these groups' profitability.
• The question being asked is obviously that of the business models that keep these new infrastructures running.

Strategic implication: The idea that this is not a speculative frenzy but a change in volume and infrastructure is gaining ground in Silicon Valley.

Concentration of capital and barriers to entry

The digital giants are sitting on $500 billion in cash.

• The leading tech companies (Apple, Microsoft, Google Alphabet, Amazon, Meta, Nvidia) have collectively accumulated roughly $500 billion in available cash on their balance sheets.
• This excess liquidity gives them an “unlimited” capacity to invest in AI infrastructure (data centers, GPUs, chips, software).
• Example: Microsoft has more than $80 billion in cash. Alphabet tops $100 billion.
• To avoid being smothered, new entrants like CoreWeave, Mistral Data Center and Fluidstack are helped by Nvidia but have neither the same scale of capital nor, probably, priority access to hardware (something Elon Musk pulled off brilliantly for his Colossus data center last year).
• On top of their cash reserves, these groups generate a combined free cash flow of more than $300 billion every year.
• This lets them reinvest massively every year, without depending on debt or outside financing.

Annual examples (approximate):

• Apple: ~$100 billion
• Microsoft: ~$65 billion
• Alphabet: ~$60 billion
• Nvidia: ~$60 billion
• Meta: ~$40 billion
• Amazon: ~$30 billion

Contrary to what many hoped, AI does not upend the established order; it strengthens the position of the dominant players (Big Tech), because they have the data (Google Search, YouTube, Facebook, Instagram, etc.), the distribution (browser, OS, smartphones, connected devices, platforms), the compute resources (GPUs, in-house cloud) and the talent (researchers, AI engineers)

But above all the capital to sustain low-margin investments in AI for several years.

In a context of concentrated technological power, the digital sovereignty strategies of governments and SMEs are complicated, because they cannot absorb low-margin revenues.

Amazon and Uber only won their bet because they operate in a global market that turns minimal margins into a revenue stream consolidated on a global scale.

A national, or even European, initiative in AI will not be enough. Unless we adopt a Chinese-style model.

Falling margins are the new business model

• A drop in margins is seen positively by the market: it signals real use of AI models.
• SaaS companies that refuse to rethink their model risk being overtaken, or even disappearing.
• Microsoft and Adobe seem to show that a transition to AI is possible, even with pressure on margins.

Strategic implication: Financial markets are gradually coming to accept that tech will produce lower margins in order to stay in the race.
As we explained in a previous newsletter, it remains to be seen whether or not electricity will be available in sufficient quantity.

The 2026 infrastructure duel: Nvidia vs. Google (TPU)

• Nvidia is no longer just a chipmaker: it is a builder of complete data centers, with software, networking and racks.
• Google (via its TPUs) actually offers the main alternative to the Nvidia ecosystem, but that technology is exclusively in-house.
• Google licensing its TPUs to other platforms (via Broadcom) could destabilize the other dedicated-chip initiatives. Indeed, if Google starts selling them at scale, the other chip projects (e.g., Amazon Trainium, Groq) risk losing all relevance.

Google is already 3 to 4 generations ahead. It is almost impossible for a new chip company to catch up in the short term.

Strategic implication: Control of the hardware and software layers is becoming a lever of economic and geopolitical power.

With its OpenAI partnership, AMD plays the role of a fallback in case OpenAI's in-house chips aren't ready. These are now alternative GPUs, but they remain less visible than Nvidia's.

The battle for control of user interfaces and the feedback loop

• The major stake of 2026 is controlling post-interaction data (feedback loops).
• AI models now learn through their users, creating a flywheel effect similar to the era of social platforms. As with search, and then social networks, it's the volume of interaction that makes the difference.

That is why the battle for entry points is only just beginning: browsers, chatbots, assistants.

Shifting business models: from usage-based billing to outcome-based billing

• AI favors performance-based business models measured in human equivalents. Many metrics are being put in place, not to assess how AI replaces human work, but to bill for the human-equivalent work.
• Companies that integrate AI will have to roll out zero-margin AI products, backed by their existing business. The only upside is that the outcomes pay for that investment.
• Many listed companies are still refusing this turn, for fear of how the stock market will react.

Knowing how to monetize intelligence, not just software, will be the Holy Grail.
The question is who, between an AI-first company and a company that integrates AI, will have the better usage margin.

Robotics: the physical extension of AI

• Humanoid robots (e.g., Tesla Optimus) are progressing at an impressive pace. It is above all in China that the progress is most striking.
• Looking at the investments, it seems one of the debates about the shape of robots is settled: the humanoid wins through compatibility with the human environment. Margins will be even thinner, so there is no room for error in the implementation.
• China will be Tesla's great rival in this sector, and Europe is conspicuous by its absence.

Provisional conclusion

So there are two conversations: a secondary infrastructure bubble, coupled with a phase of strategic restructuring of global digital capital.

The goal is the acceleration of the asymmetry between those who have the infrastructure, the data and the distribution, and those who depend on them.

The question of cognitive autonomy, which occupies me through my new activities, is how to learn to navigate a world where superintelligences, passive intelligences and local intelligences will coexist. How do we keep the thread?

This question is strategic for governments and institutions.

• Many retail players, at first, didn't want to compete with Amazon, because they felt the margins were too thin compared with traditional retail.
• And yet Amazon showed it was possible to live on ultra-thin margins and to use them to create one of the most profitable businesses in history: Amazon Web Services.

AI has the same problem, but with a factor of 1,000.

The trouble is that the non-AI world is not ready for a commoditization of intelligence. In a company like Amazon, that means laying people off and replacing them with GPU purchases. Amazon understood before others that very thin margins will force AWS to invent a new, far less profitable model.

Few industrial players in Europe have understood this so far.

For paying subscribers, a few bonuses all the same!