The Financialization of AI: When Compute Becomes a Market Before It Becomes Mature
There are moments when a news item is not just a news item. It is a signal.
The recent report that China is considering the creation of futures contracts on AI tokens through the Shanghai Futures Exchange may look, at first glance, like just another episode in the global artificial intelligence race. According to Reuters, the plan is still at an early stage and concerns futures contracts linked to “AI tokens” — units of information processed by AI models and already used as a way to measure consumption and pricing in AI services.
On the surface, this does not necessarily mean that China is creating an “AI stock exchange” in the traditional sense. We are not talking about a new market where shares of AI companies are traded. We are talking about something more subtle, more technical, and possibly more important: the transformation of AI computational consumption into a financial product.
And that is where the real question begins.
Not whether artificial intelligence matters. It does.
Not whether compute is real infrastructure. It is.
Not whether companies need tools to forecast, manage, and hedge future costs. They do.
The question is different:
Is it wise to create financial market products on top of a technology infrastructure that is still so new, so volatile, and so poorly understood?
My answer is: not without serious caution.
Not because every derivative is bad. Not because every futures market is dangerous by nature. Futures markets can serve a useful function. They can help companies manage price volatility. They can improve price discovery. They can give producers and consumers a way to plan ahead in uncertain markets.
But we have seen before what happens when markets take something real, package it into financial products, multiply it through leverage, wrap it in complex models, and then convince everyone that risk has disappeared simply because it has been priced.
The 2008 financial crisis did not happen because mortgages existed. Mortgages were real. Behind them were houses, people, incomes, banks, contracts, and collateral. The problem began when those mortgages were transformed into securities, those securities into complex products, those products into supposedly safe assets, and the real risk disappeared inside layers of financial engineering.
The lesson of 2008 is not that derivatives are always evil.
The lesson is deeper:
Complexity becomes dangerous when markets confuse pricing with understanding.
And this is exactly what worries me about artificial intelligence.
AI is still in a phase of intense formation. We do not yet have a stable understanding of the long-term cost of models, the future economics of inference, the evolution of GPUs, the real demand for compute, the burden of regulation, the pressure on energy infrastructure, the geopolitical fragility of chip supply chains, or even which model architectures will dominate in the coming years.
And yet, before we fully understand the infrastructure, we are already beginning to build financial markets on top of it.
In the United States, CME Group has announced plans for futures based on GPU rental-rate benchmarks, aiming to help manage volatility and price risk in the compute market. ICE has also announced plans for GPU compute futures based on indices covering several major GPU types.
These moves are not irrational. A large AI company, a data center operator, a cloud provider, or an enterprise heavily dependent on model usage may indeed want to lock in future compute costs. Just as an airline may hedge fuel prices, an AI company may want to hedge compute prices.
That is the healthy version.
The dangerous version begins when hedging turns into speculation.
When the market no longer primarily serves those who have a real need for compute, but those who want to bet on the future price of compute. When the product is no longer used to protect productive activity, but to create new layers of financial profit on top of an infrastructure that only a small number of people truly understand.
That is when the story begins to sound familiar.
In 2008, the system was not simply selling “air.” It was selling certainty about something it did not truly understand. It was selling the illusion that because risk had been distributed, it had disappeared. It was selling the idea that models knew better than reality.
In AI, the risk is similar, but not identical.
Here, the underlying asset is not a house. It is the ability to process information. It is tokens. It is GPU hours. It is inference. It is latency. It is energy consumption. It is data center capacity. It is access to advanced chips. Ultimately, it is the ability of a company or a state to produce, train, and use intelligence as a service.
This is not “air.” It is real infrastructure.
But it can very easily become “air” when it is disconnected from real use.
If a company buys compute futures to protect itself from rising costs, that can make sense. But if funds, banks, and traders begin building complex products on top of compute indices without any relationship to actual AI usage, then we are no longer talking only about technological infrastructure.
We are talking about the financialization of intelligence itself.
And that is a major shift.
Because artificial intelligence is not just another technology sector. It is horizontal infrastructure. It will affect production, defense, healthcare, education, public administration, logistics, finance, energy, and research. If the cost of access to this infrastructure becomes the object of intense financial speculation, then the risk does not concern only investors. It concerns the economic function of businesses and, at a deeper level, the strategic power of states.
This is where the geopolitical dimension becomes unavoidable.
The battle over AI is no longer just a battle between companies. It is not simply OpenAI versus Google, Anthropic versus Meta, Nvidia versus whoever tries to challenge it. It is a battle between states, infrastructures, energy systems, supply chains, chips, data centers, and regulatory frameworks.
If oil was the strategic commodity of the 20th century, compute may become one of the strategic commodities of the 21st.
Not in the same physical form. Not with the same logistics. But with comparable strategic importance.
Whoever controls compute controls, to a significant degree, the speed at which they can develop, train, and deploy artificial intelligence.
China seems to understand this not merely as a technological issue, but as a matter of national and financial strategy. The possible creation of AI token futures is not only about market innovation. It is also about positioning. It is about creating mechanisms to manage cost, scarcity, volatility, and access in a world where compute is becoming a foundation of economic and political power.
But this is exactly where caution is needed.
When something becomes strategic for states and attractive to markets at the same time, the probability of excess increases. Governments want power. Companies want access. Investors want returns. Markets want products. And somewhere inside this convergence, the most important question is often lost:
Do we really understand what we are pricing?
What exactly is an AI token as a financial underlying?
How does it relate to the real cost of inference?
How is it affected by improvements in model efficiency?
What happens if tomorrow’s models become dramatically cheaper to run?
What happens if architectures change?
What happens if demand shifts from large centralized models to smaller specialized ones?
What happens if compute costs fall sharply?
What happens if access to chips becomes more restricted for geopolitical reasons?
In mature markets, these questions are supported by historical depth, data, cycles, experience, and control mechanisms. In AI, many of these factors are still being formed.
That is what makes the creation of financial products so risky.
Not because there is no real asset. There is.
But because the asset is not yet maturely understood.
And markets have a dangerous tendency: when they do not fully understand something, they often turn it into an index. Once it becomes an index, it becomes a product. Once it becomes a product, it becomes an object of leverage. And once leverage enters the picture, the distance between real value and financial imagination can become enormous.
This does not mean that financial innovation should stop.
It means that financial innovation should follow understanding, not precede it.
In the case of AI, there is a real danger that we may do the opposite.
We may create the products first and then try to understand what they actually measure.
And then the problem will not be artificial intelligence. The problem will be human certainty around artificial intelligence.
That was also the deeper lesson of 2008. It was not that mathematics failed. It was not that innovation itself was wrong. It was that excessive confidence, when combined with opacity, leverage, and institutional weakness, can turn a useful economic tool into a systemic threat.
AI is already surrounded by an atmosphere of exaggerated expectation. Every company wants to appear AI-first. Every state wants a national AI strategy. Every investor is looking for the next Nvidia. Every market is looking for the next product. In such an environment, the creation of futures on compute or AI tokens may be presented as a natural evolution.
Perhaps it is.
But natural evolution does not mean safe evolution.
The critical issue is not whether AI will eventually have financial markets around it. It probably will. The critical issue is whether those markets will serve real productive use or whether they will create a new layer of speculation on top of an infrastructure that is still being built.
Because in the end, the greatest risk is not that we will sell “air.”
The greatest risk is that we will sell as certainty something we have not yet understood.
And unfortunately, we have seen that before.
