Follow BigDATAwire:

September 30, 2024

OctoAI Snapped Up by Nvidia

(sdecoret/Shutterstock)

OctoAI, a Seattle AI startup that allowed users to run a variety of foundation models while shielding them from the complexity of the underlying infrastructure, has been acquired by GPU giant Nvidia.

Rumors have swirled about a possible acquisition since mid-September, when The Information broke the story that the two companies were in “advanced talks” regarding an acquisition by Nvidia. Last week, OctoAI Founder and CEO Luis Ceze confirmed the deal in a LinkedIn post.

“I’m very excited to share that I’m joining NVIDIA,” Ceze wrote. “Building OctoAI (now NVIDIA) has been a fantastic journey going from pioneering research on ML compilers at the University of Washington, to driving Apache TVM adoption to building products valued by customers in a fast moving industry.”

There are lots of unanswered questions regarding this acquisition, including what role Ceze will have at Nvidia. Ceze, who shared the title of VP AI systems software at Nvidia, says in his LinkedIn post that he’s “excited to join NVIDIA and contribute to expanding efforts in ML compilers and cloud infrastructure for AI.”

It’s also unclear how much of the OctoAI team will make the transition over the Nvidia, or what Nvidia will do with the OctoAI software, if anything at all. It appears OctoAI customers will need to start looking for a new platform to host their AI workloads, according to letters OctoAI is writing to customers.

“We have made the strategic decision to initiate the wind down of the commercial availability of our services,”  read one letter an OctoAI customers posted to LinkedIn. “As such, effective October 31, 20224, we are terminating your access to all OctoAI services and deactivating your account as permitted under the OctoAI services Terms of Use.”

Nvidia has not publicly announced the deal, which The Information reported is worth about $165 million. With debt and other factors included, the deal could exceed $250 million, according to a report by GeekWire.

OctoAI was co-founded in 2019 by Ceze and several individuals from his research group at the University of Washington, including Jared Roesch, Jason Knight, Thierry Moreau, and Tianqi Chen. The company, which initially was named OctoML, was founded to commercialize Apache TVM, an open source software project led by Ceze that provides an abstraction layer between machine learning models and the hardware that it runs on.

OctoAI co-founder and CEO Luis Ceze is now VP of AI systems software at Nvidia

“Think of it as a middle layer between frameworks and hardware,” Ceze, who was a BigDATAwire Person to Watch for 2024, told the publication back in 2021. “We offer a clean abstraction across a wide variety of hardware.”

On the software side, Apache TVM supported all the latest frameworks, including TensorFlow, PyTorch, Keras, MXnet, Core ML, and ONNX. On the hardware side, Apache TVM supported a wide range of processors, including Intel and AMD CPUs, Nvidia GPUs, ARM CPUs, FPGAs, and even mobile chips from Qualcomm. An on-prem version dubbed the Octomizer was also available.

OctoML developed a commercial version of Apache TVM that allowed customers to bring their ML models to a variety of hardware. Customers could pick and choose which processors to run ML training and inference workloads on, thereby helping to maximize performance and minimize cost.

As the generative AI craze kicked in, OctoML pivoted its business slightly. While it always supported computer vision and natural language processing (NLP) workloads, the company sought to emphasize not only the performance and cost benefits of its platform, but also the flexibility it brings to customers and the capability to move where various components of GenAI workloads run to avoid lock-in.

OctoAI had raised about $132 million over four rounds. The latest was a Series C round led by Tiger Global Management in November 2021 that raised $85 million. The company was reportedly valued at about $900 million at that time.

Related Items:

Birds Aren’t Real. And Neither Is MLOps

Coming to Grips with Prompt Lock-In

‘Octomize’ Your ML Code

 

 

BigDATAwire