AI chip startup Recogni, having recently pivoted from edge to data center AI, is working with Juniper Networks on co-developing the rack-scale system for its second-generation AI accelerator chip. Juniper is also an investor in Recogni, having come in on the company’s recent $102 million Series C round.

The companies will work together to develop a rack-scale generative AI inference system for the data center, Recogni CEO Marc Bolitho told EE Times.

“We see that we need to solve the generative AI inference problem at a system level to be successful,” he said. “You have to take many different considerations and variables [into account]. You need to have a very high compute with low energy and low total cost of ownership, you have to have the requisite memory for these very large models, and you also have to have a network to interconnect the chip.”

Recogni’s series C was led by existing investors who backed the company when it was making an automotive vision chip.

“I think that what everyone could see very quickly is that the architecture we’ve developed for doing AI applied very well in the generative AI inference space,” Bolitho said. “Looking at that opportunity and the differentiation that we have, it was not that difficult to get everyone focused on that generative AI data center direction.”

The company has raised $175 million total to date. Startup competitors in this space are two to five years ahead, with many hundreds of millions in funding. Can Recogni compete?

“It’s about having the right people—the right people on the math, the right people on chip development, and on hardware-software codesign to develop this innovative architecture,” he said. “As far we what we have today, and what we have planned for future funding, absolutely, we’ll have the requisite capital we need in order to bring the system forward.”

Recogni’s Pareto math system sets it apart in terms of compute efficiency for inference, Bolitho added.

“Nvidia is clearly the leader in this space, but in the end, total cost of ownership is going to matter for inference,” he said. “Inference is a profit center, and it’s going to be about who has the best purpose-built solution and who can deliver lower total cost of ownership and lower energy.”

Target customers will be hyperscalers, cloud providers and enterprises who need a full-system solution. Bolitho is confident that despite a distinct lack of success for startup chips in the hyperscaler market so far, this remains a viable market.

“I think as we go through this inflection point and the growth of inference and inference compute, everyone is looking for the best solution, the one that can provide the lowest TCO,” he said. “I think even the hyperscalers will look for the best solution, the solution that provides them the most benefit, whether it’s internal or external.”

Recogni’s system is due to go into production in 2026, and the company can build on the software stack it already built for its first-generation chip.

“We’ve developed a full software stack for that device which includes an SDK and compiler that allows us to quantize, compile and deploy onto that chip,” Bolitho said. “This will be our second generation of software stack, so we’re very confident in being able to have frictionless deployment and being able to take any trained models and being able to use our SDK for easy deployment by developers.”

All of Recogni’s development effort will go towards its new generative AI chip and system over the next 18 months or so, but Bolitho would not rule out a return to the edge space in the future. “As we look forward, we believe our architecture is well suited to other areas, other AI spaces, like robotics, which could include automated vehicles as well, and we see there’s a market and a demand for that.”

From EETimes