Nvidia Corp. has made a “significant investment” in artificial intelligence developer Thinking Machines Lab Inc. as part of a partnership that the companis announced today.

The size of the cash infusion was not disclosed. However, the companies did reveal that Thinking Machines will purchase billions of dollars’ worth of hardware from Nvidia as part of the deal.

The startup previously raised $2 billion in seed funding from a consortium that included Nvidia, Advanced Micro Devices Inc., ServiceNow Inc. and other tech firms. That investment valued it at $12 billion. It’s unclear whether Nvidia’s latest investment increased its valuation.

Thinking Machines was founded last February by Chief Executive Mira Murati (pictured, right, with Nvidia CEO Jensen Huang), the former chief technology officer of OpenAI Group PBC. During her time there, she oversaw the development of ChatGPT, the Sora video generator and several other products.

Thinking Machines launched its first offering, a cloud service called Tinker, last year. It enables developers to create fine-tuned, or customized, versions of open-source large language models. The service supports more than a dozen LLMs, including several from Meta Platforms Inc.’s Llama series.

Tinker uses a technology called LoRA to power customers’ fine-tuned models. LoRA performs fine-tuning by attaching a small number of customized model weights to an open-source LLM. That arrangement removes the need to modify the LLM’s existing weights, which lowers training costs.

As part of its newly announced partnership with Nvidia, Thinking Machines will purchase at least 1 gigawatt’s worth of computing hardware from the chipmaker. Last year, Nvidia CEO Jensen Huang estimated that 1 gigawatt of AI computing capacity costs about $50 billion to build. Graphics processing units account for about two-thirds of the sum.

Thinking Machines plans to use Nvidia’s latest-generation Rubin GPUs in the project. The chip series includes two accelerators. The first, Rubin CPX, is optimized for a specific subset of the calculations involved in running inference workloads. The other chip, which is called simply Rubin, supports a broader range of use cases and features 336 billion transistors.

Thinking Machines will also use Nvidia’s Vera central processing units. Each CPU includes 88 cores that use the Armv9.2 instruction set and can run 176 threats. According to the companies, Thinking Machines plans to start deploying the hardware early next year.

Nvidia’s partnership with the AI developer also extends to other areas. In particular, the companies plan to develop “training and serving systems for Nvidia architectures.” They didn’t specify whether those systems will be of the software or hardware variety.

The term “serving system” usually refers to the software stack that developers use to power their inference workloads. Such software performs tasks such as distributing inference-related calculations across GPUs. Nvidia already has an open-source serving system called Dynamo.

“This partnership accelerates our capacity to build AI that people can shape and make their own, as it shapes human potential in turn,” Murati said.

Thinking Machines may be planning to use the large amount of hardware that it has commissioned from Nvidia to power new products. Two recent job postings indicate that the company is working on AI models optimized for audio processing and visual reasoning. It’s also developing custom implementations of AI model building blocks such as the Transformer architecture’s attention module.

Photo: Nvidia