Nvidia dropped a surprise announcement on Christmas Eve: a $20 billion deal to license AI chip startup Groq’s technology and bring over most of its team, including cofounder and CEO Jonathan Ross. It was a move that hinted Nvidia is no longer assuming its GPUs will be the only chips useful for the next big phase of AI deployment: running already trained AI models to do everything from answering queries and generating code to analyzing an image—a process known as inference—and doing so at a huge scale.
The Groq deal bolsters the standing of other startups building their own AI chips, including Cerebras, D-Matrix, and SambaNova—which Intel has reportedly signed a term sheet to acquire—as well as newer players like U.K.-based chip startup Fractile. It also lifts AI inference software platform startups like Etched, Fireworks, and Baseten, strengthening their valuations and making them more attractive acquisition targets in 2026, according to analysts, founders, and investors.
Karl Freund, founder and principal analyst at Cambrian-AI Research, pointed to the Microsoft-backed D-Matrix, which raised $275 million last month at a $2 billion valuation. Like Groq, D-Matrix is focused on trading some of the flexibility of Nvidia’s GPUs for greater speed and efficiency when running AI models. “I’m sure D-Matrix is a pretty happy startup right now,” Freund said. “I suspect their next round will be at a much higher valuation.”
Cerebras, another inference-focused chip company, also appears well positioned. Known for its dinner-plate-size “wafer-scale” chip designed to run extremely large models on a single piece of silicon, Cerebras has filed for an IPO after a previous delay. Freund said the company has increasingly been viewed as a potential acquisition target as well. “You don’t want to wait until after the IPO, when it’s more expensive,” he said. “From that perspective, Cerebras is sitting pretty right now.”
Nvidia-Groq deal has clarified market’s direction
Executives at these companies say Nvidia’s move has helped clarify the market’s direction. “When [the Nvidia-Groq deal] happened, we said, ‘Finally, the market recognizes it,’” Sid Sheth, CEO of D-Matrix, told Fortune. “I think what Nvidia has really done is they said, Okay, this approach is a winning approach.”
And Cerebras CEO Andrew Feldman posted on X that, in the past, the perception that Nvidia GPUs were all you needed for AI acted as a moat, keeping AI chip startups from nibbling away at Nvidia’s market share. But that moat is now gone with the Groq deal, Feldman wrote. “It reflects a growing industry reality—the inference market is fragmenting, and a new category has emerged where speed isn’t a feature—it’s the entire value proposition. A value prop that can only be achieved by a different chip architecture than the GPU.”
Still, not everyone is convinced that every inference chip startup will benefit equally. Matt Murphy, a partner at Menlo Ventures, said the chip sector remains a difficult one for venture investors, given the high capital requirements and long timelines. “A lot of VCs stopped investing in chips 10 or 15 years ago,” Murphy said. “It’s capital-intensive; it takes years to get a product out; and the outcomes are hard to predict.”
That said, he pointed to Fireworks, an AI inference platform that raised $250 million at a $4 billion valuation in October, as a startup with a technical advantage, thanks to a founding team filled with engineers who built PyTorch. But he added that it remains unclear how much of the current enthusiasm reflects genuine technical differentiation. “It’s hard to tell who’s really got something significant versus the tide is [raising] all boats, which is what seems to be going on,” he said, adding that consolidation across the sector now appears increasingly likely.
New entrant seeks true disruption
But at least one veteran of the AI hardware world argues that even today’s inference-focused startups are not truly disruptive.
Naveen Rao, former SVP of AI at Databricks and founder of MosaicML, recently left Databricks to start Unconventional AI, which last month confirmed a massive $475 million seed round led by Andreessen Horowitz and Lightspeed Ventures. His critique: Companies like Groq, D-Matrix, and Cerebras may be well positioned in today’s market, but they are still optimizing within the same digital computing paradigm.
After Nvidia’s Groq deal validated demand for faster, more efficient inference, startups that fit neatly into today’s AI stack suddenly look far more valuable—not because they reinvented computing, Rao argues, but because they work within it. Unconventional AI is pursuing a more radical path: building new hardware that exploits the physical behavior of silicon itself, and redesigning neural networks to match it.
“We’ve been building the same fundamental machine for 80 years, a numeric digital machine,” he said. “But there was never one workload that dominated even more than 2% of all compute cycles.” That is changing, he explained: In a few years, 95% of all compute will be used for AI.
From that standpoint, it’s important to construct an entirely different machine than what is built today, he said. However, Rao says the effort could take five years or more to bear fruit—and is not intended to capitalize on the current inference boom.
