Why Nvidia Bought Groq and Kept It Alive — A Signal From the Inference Chip Market

What Happened

The order of events is strange. Usually when a giant absorbs a small chip startup for $20B, the startup disappears. Groq didn’t.

Here’s the timeline. In September 2025, Groq raised $750M in a round led by Disruptive at a $6.9B post-money valuation, with BlackRock, Neuberger Berman, and DTCP participating. Three months later, in December 2025, Nvidia announced a deal worth roughly $20B. It wasn’t an acquisition. Nvidia licensed Groq’s LPU hardware technology and moved some senior Groq staff to Nvidia. CNBC called it Nvidia’s biggest deal on record. And yet Groq stayed independent, with former CFO Simon Edwards stepping in as CEO. The industry labeled it a “not-acqui-hire”: take the tech and the people, but don’t kill the company.

Then in May 2026, reports surfaced that Groq is raising again — $650M from existing investors, with Disruptive and Infinituum agreeing to backstop the round. The money points at the inference “neocloud” business Groq has built on top of its own AI chip.

So what is Groq’s LPU that Nvidia spent $20B and still left the company breathing? The LPU runs deterministic execution. It keeps weights resident in SRAM to cut the memory bottleneck, and it drives time-to-first-token (TTFT) lower than a typical GPU. It’s a part built to fight Nvidia GPUs head-on — not on training, but on inference, and specifically on the latency war.

What It Means for Founders

The deal structure is itself the message. Nvidia got Groq’s technology and people, yet chose not to bury the company. Why? Because it doesn’t see the inference chip market collapsing into a single GPU. Better to embrace the tech through a license and let it keep running independently than to treat it as a threat to be eliminated.

The core read: training and inference are splitting into separate chip markets. Training still belongs to Nvidia GPUs — massive matrix math, enormous bandwidth, an ecosystem nobody else has. Inference is a different game. Once the model is built, the contest is processing each request as cheaply and quickly as possible. There, deterministic execution and low TTFT can matter more than a GPU’s generality. Groq leaning toward serving inference through its own cloud, rather than only selling chips, fits the same logic: become a company that sells inference, not silicon.

What this means in practice — including for YC-stage teams and the FAANG infra crowd:

If you consume inference infrastructure, drop the assumption that vendors will consolidate to one. Even Nvidia isn’t forcing inference onto a single chip. The reason to procure your training cluster and your inference backend from the same vendor is eroding. Inference goes separately, wherever it’s cheapest.
If you’re building an infra startup, study Groq’s exit structure. “License the tech + move the people + keep the company independent” is hardening into a new big-tech pattern. For founders, that’s more cards on the table — a way to monetize core IP without handing over the company.
A timing signal. Nvidia licensed the tech, and Groq is now scaling the same tech with a fresh $650M. Two parties running the same technology in parallel is the market saying inference demand can’t be met from one supply line. If inference is a large slice of your cost of revenue, treat the next 6–12 months — as backend options widen — as a pricing-negotiation window.

Actions You Can Take Now

If you’re locked into a single-vendor inference contract, check the renewal date and switching cost now. In a market where options are widening, lock-in is just cost.
Benchmark whether Groq LPU’s TTFT and deterministic execution actually help your real workload. The more your UX hinges on latency, the more the GPU gap shows up as a number.
If you hold infra IP, add “license + stay independent” alongside “full sale” in your exit scenarios. Groq is the precedent that makes that negotiation credible.