The price index of that rental, called
Sdh100rtLaunched today. Every day, it will crunch 3.5 million data points from more than 30 sources around the world, so that the average spot rent can be given to use one. Nvidia h100 GPU for one hour. (“Spot Price” is the one that is delivered a commodity immediately.)

“I really believe that the calculation will be the greatest resource for humanity over the next few years,” Li says. “If my thesis is correct, it will require more sophisticated risk management.”

According to Lee, such an index will have more opportunities for a broad set of players to join the AI ​​Tool and AI industry. How do you get from an index? The basic story of silicone data helps explain it.

By the beginning of last year, Lee was in charge of global data integration BloombergIn that case she met with several small companies that were trying to distribute AI-fuel data products, and many of them were struggling with the same problem. They could only offer their products at a certain rate, but the cost of the GPU time was unpredictable to them. Therefore, so their benefits were margins.

With specific items such as energy, companies can plan for these swings by knowing historical trends and hedging with financial products such as futures contracts. But this was not present for the main object of AI: time on a GPU. Hence Lee has been determined to build the foundation for those products, and the result is the SDH100RT price index.

He chose the index Nvidia H100, because it is the most widely deployed GPU, and is used to train new AI modelHowever, a price index for Nvidia A100s, which deal with a lot of estimate tasks, works as well. And he has developed a method that will determine when it makes sense to the sequencing prices for others Ai chipsLike
AMD And Navidia’s blackwell series.

Carmen Lee established silicone data after a term Bloomberg,Silicon data

Armed with data, Startups And other people who create new AI products will be able to better understand their potential costs, so they can set their services to a profitable value. And the new AI infrastructure will be able to set a benchmark for their own revenue. But just as, in Lee’s opinion, it is that the new sources of capital AI may join the space.

BanksFor example, a relatively inexpensive supplier of capital, took note. But because they have strict risk control and do not have enough GPU price data, they are not in a position to fund AI projects. Lee hopes that SDH100RT will allow banks to lend to a broad set of players in the AI ​​industry and allow them to come up with financial products that already reduce the risk for those.

Insight and inequalities from data

Although it has been launched today, silicone data has monitored the GPU fare prices for months. As you can expect, some interesting insights have been unveiled by having a window in the price of AI training. There are some things that Lee has discovered. (She is publishing
This analysis Regularly since last September.)

Eastern coast rules! West Coast Drols: H100 rental price is very stable United StatesBut there is one Constant east coast profitIn March, you can find one hour work from H100 to US $ 5.76 on the East Coast. But the price of the same hour will be $ 6.80 on the West Coast.

Help Hyperscaler Chips: Heroic In web services Design your own chips And Server Is Low prices For cloud giant customers. According to silicon data, at around $ 4.80 per hour, the GPU per average unit price for Trainium 2 of AWS is less than half the price to use NVidia H100. Its first generation chips infantia and trainium both come in less than $ 1.50 per hour, which is less than half the price of today’s estimates, NVDia A100. However, H100s is considered the only option for state -of -the -art model training, so their performance can justify extra scratches.

Deepsek’s minor effects: January lampsake shock Very little Spot for the price of rent. You may remember that performance and Reported less training Hangzo-based Deepsac’s LLM was surprised and sent AI-related shares into a patch of Pearl Clutching. “When Deepsek came out, [stock] Gone mad in the market, “Lee says.” But the price of the spot did not change much. “At the beginning of the lampsac, the price of H100 increased to $ 2.50 per hour, but it was still at $ 2.40 per hour at $ 2.30 per hour before the range of $ 2.60 per hour.

Intel Is more posh than AMD, GPUS Always are always under the control of CPU, usually in 4: 1 ratio. And for that CPU spot, elections have been contested between the market Intel and AMD. (Nvidia also makes its CPU, called Grace.) But it seems that customers are willing to pay a little premium for the Intel-Integrated System. For NVIDIA A100 System, people with Intel CPUS get about 40 percent more than those with AMD. ProcessorFor H100, the effect involves the interconnect technique. If a computer used SXM or PCIE as its link, Intel received a high price. But for those using NVLINK interconnect scheme of Nvidia, AMD received a premium.

Commodity of AI

Can you really boil the price of AI at the same number? After all, the performance of the computer and its utility for a particular customer include a lot of factors. For example, a customer can undergo training with data, which, for legal reasons, cannot cross international borders. So why should they care about the price in another country? And, anyone who has examined the major benchmark results of machine learning,
Mlperf, Can see, the performance of the same Nvidia GPU can vary widely depending on the system that is in it and it is running software.

According to Lee, the commodity view can work. The index of silicon data normalizes all these differences and gives different weight to those such as a data center participates in the rental market, its location, its data sources and many, many other things.

Perhaps the greatest support for AI’s idea as an object
Nvidia CEO Jensen Huang she herself. In the company’s big developer event, GTCHe pushed to think Data center As “AI factories”, whose output will be measured in how many tokens, the smallest unit of LLM information, they can produce per second.