Meet the $10,000 Nvidia chip powering the race for A.I.

Hundreds of GPUs are required to train artificial intelligence models, like large language models. The chips need to be powerful enough to crunch terabytes of data quickly to recognize patterns. After that, GPUs like the A100 are also needed for “inference,” or using the model to generate text, make predictions, or identify objects inside photos.

This means that AI companies need access to a lot of A100s. Some entrepreneurs in the space even see the number of A100s they have access to as a sign of progress.

“A year ago we had 32 A100s,” Stability AI CEO Emad Mostaque wrote on Twitter in January. “Dream big and stack moar GPUs kids. Brrr.” Stability AI is the company that helped develop Stable Diffusion, an image generator that drew attention last fall, and reportedly has a valuation of over $1 billion.

Now, Stability AI has access to over 5,400 A100 GPUs, according to one estimate from the State of AI report, which charts and tracks which companies and universities have the largest collection of A100 GPUs — although it doesn’t include cloud providers, which don’t publish their numbers publicly.

Nvidia stands to benefit from the AI hype cycle. During Wednesday’s fiscal fourth-quarter earnings report, although overall sales declined 21%, investors pushed the stock up about 14% on Thursday, mainly because the company’s AI chip business — reported as data centers — rose by 11% to more than $3.6 billion in sales during the quarter, showing continued growth.

Nvidia shares are up 65% so far in 2023, outpacing the S&P 500 and other semiconductor stocks alike.

Nvidia CEO Jensen Huang couldn’t stop talking about AI on a call with analysts on Wednesday, suggesting that the recent boom in artificial intelligence is at the center of the company’s strategy.

“The activity around the AI infrastructure that we built, and the activity around inferencing using Hopper and Ampere to influence large language models has just gone through the roof in the last 60 days,” Huang said. “There’s no question that whatever our views are of this year as we enter the year has been fairly dramatically changed as a result of the last 60, 90 days.”

Ampere is Nvidia’s code name for the A100 generation of chips. Hopper is the code name for the new generation, including H100, which recently started shipping.

Compared to other kinds of software, like serving a webpage, which uses processing power occasionally in bursts for microseconds, machine learning tasks can take up the whole computer’s processing power, sometimes for hours or days.

This means companies that find themselves with a hit AI product often need to acquire more GPUs to handle peak periods or improve their models.

These GPUs aren’t cheap. In addition to a single A100 on a card that can be slotted into an existing server, many data centers use a system that includes eight A100 GPUs working together.

This system, Nvidia’s DGX A100, has a suggested price of nearly $200,000, although it comes with the chips needed. On Wednesday, Nvidia said it would sell cloud access to DGX systems directly, which will likely reduce the entry cost for tinkerers and researchers.

It’s easy to see how the cost of A100s can add up.

For example, an estimate from New Street Research found that the OpenAI-based ChatGPT model inside Bing’s search could require 8 GPUs to deliver a response to a question in less than one second.

At that rate, Microsoft would need over 20,000 8-GPU servers just to deploy the model in Bing to everyone, suggesting Microsoft’s feature could cost $4 billion in infrastructure spending.

“If you’re from Microsoft, and you want to scale that, at the scale of Bing, that’s maybe $4 billion. If you want to scale at the scale of Google, which serves 8 or 9 billion queries every day, you actually need to spend $80 billion on DGXs.” said Antoine Chkaiban, a technology analyst at New Street Research. “The numbers we came up with are huge. But they’re simply the reflection of the fact that every single user taking to such a large language model requires a massive supercomputer while they’re using it.”

The latest version of Stable Diffusion, an image generator, was trained on 256 A100 GPUs, or 32 machines with 8 A100s each, according to information online posted by Stability AI, totaling 200,000 compute hours.

At the market price, training the model alone cost $600,000, Stability AI CEO Mostaque said on Twitter, suggesting in a tweet exchange the price was unusually inexpensive compared to rivals. That doesn’t count the cost of “inference,” or deploying the model.

Huang, Nvidia’s CEO, said in an interview with CNBC’s Katie Tarasov that the company’s products are actually inexpensive for the amount of computation that these kinds of models need.

“We took what otherwise would be a $1 billion data center running CPUs, and we shrunk it down into a data center of $100 million,” Huang said. “Now, $100 million, when you put that in the cloud and shared by 100 companies, is almost nothing.”

Huang said that Nvidia’s GPUs allow startups to train models for a much lower cost than if they used a traditional computer processor.

“Now you could build something like a large language model, like a GPT, for something like $10, $20 million,” Huang said. “That’s really, really affordable.”

Nvidia isn’t the only company making GPUs for artificial intelligence uses. AMD and Intel have competing graphics processors, and big cloud companies like Google and Amazon are developing and deploying their own chips specially designed for AI workloads.

Still, “AI hardware remains strongly consolidated to NVIDIA,” according to the State of AI compute report. As of December, more than 21,000 open-source AI papers said they used Nvidia chips.

Most researchers included in the State of AI Compute Index used the V100, Nvidia’s chip that came out in 2017, but A100 grew fast in 2022 to be the third-most used Nvidia chip, just behind a $1500-or-less consumer graphics chip originally intended for gaming.

The A100 also has the distinction of being one of only a few chips to have export controls placed on it because of national defense reasons. Last fall, Nvidia said in an SEC filing that the U.S. government imposed a license requirement barring the export of the A100 and the H100 to China, Hong Kong, and Russia.

“The USG indicated that the new license requirement will address the risk that the covered products may be used in, or diverted to, a ‘military end use’ or ‘military end user’ in China and Russia,” Nvidia said in its filing. Nvidia previously said it adapted some of its chips for the Chinese market to comply with U.S. export restrictions.

The fiercest competition for the A100 may be its successor. The A100 was first introduced in 2020, an eternity ago in chip cycles. The H100, introduced in 2022, is starting to be produced in volume — in fact, Nvidia recorded more revenue from H100 chips in the quarter ending in January than the A100, it said on Wednesday, although the H100 is more expensive per unit.

The H100, Nvidia says, is the first one of its data center GPUs to be optimized for transformers, an increasingly important technique that many of the latest and top AI applications use. Nvidia said on Wednesday that it wants to make AI training over 1 million percent faster. That could mean that, eventually, AI companies wouldn’t need so many Nvidia chips.