Groq

What they built

Groq designs and deploys the LPU — a tensor-streaming processor architecture purpose-built for language-model inference. The product stack spans the physical LPU silicon, GroqCloud (the public inference API and SaaS product), and enterprise deployments for customers who need latency-critical or on-premise inference. The architectural bet — that inference workloads look different enough from training that a dedicated accelerator would outperform GPUs — has been validated both technically (measured latency and throughput numbers) and commercially (NVIDIA’s $20B move to control the technology).

How they got here

Jonathan Ross co-invented Google’s first TPU before leaving in 2016 to found Groq on the thesis that inference would eventually be bigger than training, and the accelerator economics would diverge. The first six years were quiet — hardware startups move slowly, enterprise sales cycles are long, and the AI category was still dominated by training spend. The inflection came in 2023–2024, when LLM inference workloads went from research to commercial production and the latency properties of Groq’s LPU became a material competitive advantage for real-time applications.

The September 2025 $750M round at $6.9B was the last standalone valuation. In February 2026, NVIDIA announced a $20B asset-acquisition structure — a deal shape similar to the Microsoft-Inflection transaction that distributed $7.6B to shareholders (roughly $64 per share, about 75% of stock ownership) while the company continues to operate independently under new CEO Simon Edwards. Jonathan Ross and President Sunny Madra joined NVIDIA to support scaling the licensed technology. The FTC announced in early 2026 that it would examine the pattern of such “merger-in-disguise” arrangements across AI, which is now a regulatory factor.

What’s ahead

Three things matter over the next 12 months. First, GroqCloud growth: the standalone business needs to continue growing inference API and enterprise revenue to justify the post-deal capitalization. Second, regulatory: the FTC’s examination of the NVIDIA-Groq structure could reshape how similar deals are executed going forward. Third, technology independence: how tightly the Groq-NVIDIA technical roadmap integrates will determine whether Groq remains a meaningful competitive alternative or becomes effectively an NVIDIA subsidiary.

Why it matters

Groq is the reference case for what happens when a genuinely differentiated AI accelerator reaches commercial scale — and what the incumbent ecosystem does about it. For founders in AI hardware, the Groq story is the playbook: patient capital, real technical differentiation, and enough commercial traction to force the dominant player’s hand. For investors, understanding the NVIDIA-Groq deal structure is essential for pricing other AI accelerator companies (Cerebras, SambaNova, Tenstorrent) — because it effectively set the ceiling and the exit pattern for the category.