China’s AI Startup DeepSeek surprised the world with the R1 model release. This surprised the world, almost the same as the major models of Google and Openai, despite the claims to train with relatively modest GPUs.
Deepseek’s relative efficiency is that some experts and investors are wondering if AI really has a large -scale hardware expenditure. And it can change the demand for data centers -and the energy required to give them power.
The company claims that it has run 2,048 NVIDIA H800 GPU for two months to train a slight old model.
Few companies are exposed as NVIDIA, and their stock prices decreased 16 % at the time of publication. Perhaps the startups and power producers who bet on new nuclear power and natural gas abilities are probably more vulnerable.
In particular, nuclear power is at the top of the Renaissance for many years, and has promised to build and operate a new generation of power plants safely and cheaply with the advances in the design of fuel and reactor design. Until now, there was almost no reason to burn first. The nucleus is still more expensive than wind, solar, and natural gas. Furthermore, the next -generation nuclear power has not been tested on a commercial scale yet.
The rapid increase in power demand from AI has changed the equation. The data center is expected to consume 12 % of all US power, so the share will be more than tripled in 2023. Dollar in question. Google has promised to buy 500 megawatts from Nuclear Startup Kairos. Amazon has led an investment of $ 500 million to the X-Energy, a nuclear power startup, and Microsoft has cooperated with the Contellation Energy in a 3-miles of $ 1.6 billion renovation.
But what happens if the problem is exaggerated?
There is no hard and fast rule that suggests that the only way to improve AI performance is to use more calculations. For a while, the tactics worked well, but recently more computing has not obtained the same consequences. AI researchers have cast for solutions, and DeepSeek may have found one of the R1 models.
Of course, not everyone is convinced.
“Deepseek’s achievements may be groundbreaking, but we have questioned the concept that the feat was performed without using advanced GPUs,” says Citigroup analyst atif Malik. I am.
Nevertheless, history will find a way to make AI cheaper and more efficient, even if Deepseek is hiding something. After all, it is easier and potentially faster to impose a Ph.D. that develops a better model than building a new power plant.
The current wave of the new reactor is scheduled to be online until 2030, and the new natural gas power plant will not be available until the end of 10 years. In that context, the power investment of a high -tech company looks like a hedge in case the software bet does not bread out.
If so, expect high -tech companies to reduce their ambitions. High -tech companies have almost always chosen the latter, if the physical assets or software were given billions of dollars.
Where does it leave nuclear power startups and energy companies? it depends. Some people can generate electricity at a low cost that does not matter if the power of AI needs to decline. The world was impressive, and even before the AI bubble began to swell, it was expected that electricity demand would grow.
However, since there is no demand from AI, these cost pressure will probably increase. Wind power, solar light, and battery are inexpensive and cheaper, and are essentially mass -produced in modules. Developers can gradually develop new renewable plants, provide electricity (and revenue) before the entire project is completed, and at the same time managing uncertain demand to manage the future. The same is not possible about the reactor or gas turbine. High -tech companies know this quietly in renewable energy to supply data centers.
Few people have predicted the current AI boom, but it is unlikely that everyone will know how they will develop in the next five years. As a result, more secure bets on energy flows into a proven technology that can be quickly developed and expanded according to the rapidly evolving market. Today, renewable energy complies with the bill.