Silicon Data Launches First GPU Rental Price Index

This rental price index, called SDH100RT, was launched today. Every day, you will collect 3.5 million databases from more than 30 sources around the world to provide the average immediate rental price to use the NVIDIA H100 graphics processing unit for an hour. (“Immediate Price” is what the commodity is delivered immediately now.)

“I really think that compute will be the largest human resource in the next few years,” tells me. “If my thesis is correct, you will need to manage more advanced risks.”

According to Li, this index will lead to AI tools cheaper and more opportunities for a broader group of players to participate in the artificial intelligence industry. How do you get from an index to all of this? The story of the origin of silicone data helps explain it.

$ 1.04: The rental price feature for NVIDIA H100 graphics processing units on the eastern coast of the United States against those on the western coast.

Even early last year, I had a responsible for the integration of global data in Bloomberg. In this situation, she met many small companies that were trying to provide data products fed by artificial intelligence, many of which suffer from the same problem. They were unable to offer their products at a fixed price, but the cost of GPU that they need is unpredictable. Therefore, it was the margins of its profit.

With typical goods such as energy, companies can plan these fluctuations by knowing historical trends and hedging with financial products such as future contracts. But this was not present for the main artificial intelligence commodity, and the time on the graphics processing unit. So Li started to create the basis for these products, and the result is the SDH100RT price index.

The NVIDIA H100 index chose, because the most widely spread GPU, used to train new AI models. However, the NVIDIA A100s price index, which treats a lot of inference tasks, is also in business. Matthew has developed a logical index of artificial intelligence chips, such as those in the AMD and NVIDIA series in Blackweell.

Carmen Lee founded silicone data long after Bloomberg.Silicon data

Armed with data, startups, and others that build new products for Amnesty International will be able to better understand their potential costs, so that they can set their services at a profitable price. Those who build the new Amnesty International infrastructure will be able to set a standard for their own revenues. But as important, in my opinion, new capital sources can participate in the area of ​​artificial intelligence.

Banks, for example, are a relatively inexpensive resource, notice me. But since they have strict risks controls and there was not enough GPU price data, they were not in a position that allowed them to finance artificial intelligence projects. He hopes that SDH100RT will allow banks to lend a wide range of players in the artificial intelligence industry and allow them to reach financial products that reduce the risks of those already in them.

Vision and abnormalities of data

Despite its launch today, silicone data tracks GPU rental prices for several months. As you may expect, the presence of a window in the price of artificial intelligence training has unveiled some interesting ideas. The following are some of the things he discovered to me. (These analyzes have been publishing regularly since last September.)

The rules of the eastern coast! He asked the western coast saliva: H100 rental price is very stable in the United States, but there is a The continuous eastern coast feature. In March, you can get an hour of work from HE100 on the eastern coast for $ 5.76. But that same watch will cost you $ 6.80 on the western coast.

Easter chips: Amazon Web Service Design its special chips And his servers Low prices For cloud giant customers. According to silicone data, at about $ 4.80 per hour, the average unit price per unit of graphics processing for AWS is less than half of the price to use NVIDIA H100. Each of the first generation chips comes from the inference and trainees less than 1.50 dollars per hour, which is less than half of the price of the Outference Workhorm today, NVIDIA A100. However, H100s is believed to be the only option to train in advanced models, so its performance may be justified.

The humble impact January Dibsic shock I did little To the instant rental price. You may remember that the performance and reporting of low -cost training for Deepseek’s Llms, headquartered in Hangzhou, surprised a lot and sent stocks related to lack of intelligence into a piece of pearl capacity. “When Dibsic came out, [stock] “The market has gone, but the immediate price has not changed much,” he says. In the appearance of Deepseek for the first time, the H100 price rose moderately to $ 2.50 per hour, but that was still in $ 2.40 per hour to $ 2.60 per hour from the previous months.

Intel is more luxurious than AMD: Graphics processing units are always under the control of the CPU, usually 4: 1. The market is stabbed for the CPU and AMD. (Nafidia also makes its central processing unit, called beauty) But it seems that customers are ready to pay a little bonus for powered systems. For NVIDIA A100, those with Intel CPUs have achieved about 40 percent of the AMD processor. For H100, the effect depends on the interested intercourse technique. If a SXM or PCIE computer is used as its links, it has brought Intel a higher price. But for those who use the NVIDIA Nvlink Interconnect chart, AMD got premium.

AI

Can you really boil the price of artificial intelligence to one? After all, there are many factors that have a computer’s performance and benefit to a specific customer. For example, the customer may be trained with data that, for legal reasons, can cross international borders. Why do they care about the price in another country? Just as anyone who has examined the leading standard results in Machine Learning, Mlperf, can see, the performance of the NVIDIA graphics processing unit itself can vary depending on the system that is operating and the program it works on.

According to Li, the commodity view can work. Silicon data index normalizes all these differences and gives different weights to things such as the amount of data center sharing in the rental market, location, data sources and many other things.

Perhaps the greatest support for the idea of ​​artificial intelligence as a commodity is from the CEO of Nvidia Jensen Huang himself. At the Grand Developer of the company, GTC, he pushed to think about data centers as “Amnesty International factories” whose output will be measured in the number of symbols, the smallest unit of information used by LLM, and they can produce them in a second.