When Openi, the White House, Oracle, Softbank and MGX announced that they will invest in the Stargate project – a fresh company to drive the development of artificial intelligence in the US – questions about implications for infrastructure of data centers, energy consumption and generation, and development and more immediately.
It is “like a space race with the Soviets in the 1960s,” explained Kuba Stolarski, IDC vice president and the main global research in the field of infrastructure trends and compulsory service providers.
“America is now in the AI race with the Chinese for the rest of the 1920s, starting the largest AI infrastructure project in history”, in which the general GDP of the country will be influenced by the sold goods and services ai report, Stargate Project launches a fresh age of development of artificial intelligence which followed the project announcement January 21.
However, the fresh gigantic language models, such as GPT 3.0 chat, were over five years ancient. Questioned by the GPU memory capacity limiting the speed and level of computational operations required for LLM trainingIt turned out to be high-priced. Thus, the race for achievement of performance like artificial intelligence, which achieve or exceed human cognitive abilities, means that Huge Ai will probably adapt its haals, according to carpentry.
“Armaments Race” for GPU
After preliminary efforts to build an ambitious Stargate data infrastructure – $ 500 billion in Ailene, Abilene, Texas is expected in Ailene, Ailene, Ailene in Ailene, Ailene, Openai, in the next four years, Opennai said that the first $ 100 billion will provide American leadership at AI.
However, “the development of large language models will require enough commercial use cases to justify investments” in the largest AI infrastructure project in history, according to carpentry.
In addition to LLM, competitive technologies of accelerated servers are used in various cases of exploit, he noted in report.
He emphasized that although the Stargate scale was not determined from the point of view of computing infrastructure, “the demand for GPU servers from Stargate may cause a significant load on the GPU supply, which has just begun to make it easier in 2024 after a very limited delivery of 2023. ”
“While the supply will continue to grow every year, we have observed that demand is still ahead of the supply in this extremely hot market segment,” he added.
In addition, while the involvement of President Donald Trump “can generally refer to the removal of barriers” required to carry out the project on this scale, joinery said that there are many uncertainty related to energy policy.
“An interesting phrase is, however, how some of the elements of the project, which may be necessary because of its long-term success, such as clean energy, seemingly conflict with the direction of new administration,” he said.
Finally, it is also debatable that GPU farms are useful investments, and it is not clear whether artificial general intelligence or agi “can be achieved thanks to raw computing power and existing LLM methods”.
“For example, with comparable financial support, quantum calculations can provide much more revolutionary breakthroughs than LLM,” said Stolarski.
Deepseek and narrower models
Counteracting the expansion of the LLM approach as the wise action of the Chinese company Deepseek claimed that it was able to train a comparable AI model at 11 -a reduction in the GPU capacity.
“I think that the Deepseek hit the nerve, it was the claim that they could do it for such a fraction of costs, but it is not apples for apples, because when the basic models in the US, which were used to build Deepseek.
“The levels of optimization we get will probably improve,” he said.
Opeli cooperates with technological giants such as ARM, Microsoft, Nvidia and Oracle about the development of technologies related to Stargate. Oracle promised earlier to resolve challenges related to health data, as well as an ambitious commitment to the National Database of Electronic Health Documentation and a recent initiative in the field of cybersecurity in the cloud.
“There are all those methods that they try to work in the background in the back to try to increase the efficiency of these things,” said Stolarski.
The cost must decrease to AI proliferation in the case of exploit – along with the scale of the model.
“Smaller models that are much more adapted and adapted to highly specific use will take over,” he said in Stargate position paper.
“These small language models (unlike large language models) do not require such huge infrastructure environments. Rare models, narrow models, low precise models – currently many studies are carried out to radically reduce the infrastructure needs of creating AI models while maintaining their AI models, maintaining your AI models accuracy indicators. ”
For example, Stolarski explained that IBM is building smaller models to generate its clients’ revenues.
“If we manage to be more proficient [fewer GPUs]Then the demand will boost – he said.
“I think that everyone who follows this market expected that we will get this increase in performance, and it is a matter of time, and I think that even now the knee reaction seems to be disappearing.”