The GDP of many a nation is overshadowed by the budgets currently being pumped into AI. Can those investments be justified?
Alphabet is set to invest 185 billion dollars in 2026, primarily in data centers and AI infrastructure. OpenAI wants to spend about 600 billion dollars on hardware for AI by 2030, and parties such as Softbank, Nvidia, and Amazon are happily investing tens of billions themselves to support those plans. The company was able to raise another 122 billion dollars from investors.
Anthropic, for its part, is going to buy 3.5 GW of data center capacity from Google. Financial Times estimates the cost for one gigawatt of infrastructure at thirty to fifty billion dollars, although hardware prices are currently still increasing. The billions being juggled are too abstract to be comprehensible. A comparison might help: Belgium’s GDP in 2025 was approximately 689.36 billion dollars.
High investment, low revenue
There is no large profit to offset these massive investments. On the contrary: Anthropic says it hopes for a revenue of 30 billion dollars, up from nine billion dollars in 2025. OpenAI generates two billion dollars in revenue per month and is seeing that figure rise. It seems plausible that the company will at least match Anthropic. The amounts are high, but the total revenue is only a fraction of the investments yet to come.
Are those investments justified then? Looking at the figures above, the fear of a bubble is not unfounded. It seems that tech companies are going all-in with their financial reserves. Mark Collier, Executive Director of the PyTorch Foundation, also does not shy away from the comparison to a game of chance: “Humanity is making a big bet,” he states on stage at the first PyTorch conference in Europe, which recently took place in Paris.
AI, without the hype
The PyTorch Foundation is a non-profit that manages important open-source projects for AI, including the PyTorch framework itself. Together with other projects including vLLM and Ray, PyTorch plays a crucial role in the development of AI, the training of LLMs, and hardware support. Unlike OpenAI, Anthropic, Nvidia, and the other major players, however, Collier is not sitting at a money tap. He leads an open-source community that strongly believes in AI but is less prone to hype.
We see that during the conference. When asked about their favorite topics, the developers present overwhelmingly indicate ‘inference’ (82.1%). Training appeals to 71.5 percent, while only 54 percent of the experts want to talk about agents. This stands in stark contrast to the hype surrounding AI agents created by the marketing departments of software companies worldwide.
“I think the open-source community largely consists of doers, with an aversion to hype,” Collier says, framing the results. “Agents are a hype, and ultimately mostly a system for using models.” In other words: agents are a form of nicely packaged inference, so inference itself is more interesting. The figures primarily illustrate that the PyTorch conference is a gathering of down-to-earth AI specialists. Keynotes are therefore about concrete problems, illustrated with relevant code. There is little room for blowing hot air.
Tipping Point
What does Collier think of the enormous investments in AI? Is there a bubble? “It’s like reading tea leaves,” he admits. “But I don’t think so. We have only just reached a tipping point. For about six months now, models have really been at the level where clear economic value is being created.”
For about six months now, models have really been at the level where clear economic value is being created
Mark Collier, Executive Director PyTorch Foundation
Collier points to concrete applications, such as coding. Models like Claude Code really help programmers get complex work done faster. We also see this during the conference, where programmers openly thank Claude Code for the rapid progress they have made in developing new tools. We are not seeing predictions, but concrete results.
AI in companies
“The improvements have become recursive,” says Collier. He sees another shift: “Two years ago, we saw that training and adaptation of models was mainly done by researchers. Today, there are real use cases at large companies. Uber, for example, has trained thousands of AI models itself and uses them in production because the approach was efficient. Large companies in the financial services sector are also training their own models.”
“The barrier to entry is also lowering,” Collier continues. “When a new model comes out, it takes barely a few months before new techniques are developed to make that model more efficient. We are seeing fivefold improvements in a short time. More efficient models mean that companies need smaller GPU clusters to get started with them.”
Pace of progress as a benchmark
Collier therefore sees progression in the economic story behind AI. “There is real progress, but there is no benchmark with which you can map out the expectations around AI. It is clear to me that the speed of progress is incredible. That’s why I’m less skeptical about AI than many others.”
Collier is not blind to the discrepancy between the investments in and the revenue from AI today, but given the enormous pace of tangible progress, he believes that is not a disaster. It is the promise of that pace of progress, coupled with the growing economic relevance that goes with it, that must justify the investments.
That is also the reasoning of OpenAI, which points out that revenue in 2024 was still one billion dollars per quarter, which is half of what is now realized monthly. The company’s revenue is currently growing four times faster than that of Alphabet or Meta. If this trend continues, AI will generate enough income in the relatively short term to justify the billions in spending. Collier indicates that predicting is difficult, but even without a sensitivity to hype, he joins the camp of the optimists.