2025-12-25

NVIDIA acquires Groq for $20 billion — the LPU era ends

NVIDIA acquired Groq for $20 billion on Christmas Eve 2025, absorbing the LPU team in its largest-ever deal. Groq founder and 90% of engineers joined NVIDIA.

On December 24, 2025, NVIDIA announced the acquisition of Groq for $20 billion in a deal that closed before the new year. It is the largest acquisition in NVIDIA’s history, eclipsing the failed Arm deal ($40B proposed but blocked in 2022) in strategic significance if not price. Groq no longer exists as an independent company.

What Groq was

Groq built the LPU (Language Processing Unit) — a deterministic, streaming chip architecture purpose-built for transformer inference. Unlike GPUs, which batch work and schedule dynamically, the LPU executes a fixed dataflow graph at wire speed. The result: 300–800 tokens/second per chip for LLaMA-3-70B, with sub-100ms TTFT that made it the fastest public inference endpoint for much of 2024.

Groq’s public cloud API became the benchmark that other inference providers chased. At its peak in Q3 2024, GroqCloud served over 1 billion tokens per day.

Why NVIDIA paid $20 billion

NVIDIA’s H100/H200 dominance in training is uncontested. Inference is the battleground. The LPU’s deterministic execution model offered a fundamentally different architecture — one that NVIDIA couldn’t easily replicate in CUDA. By acquiring Groq, NVIDIA:

Removed the most credible public NVIDIA alternative in inference performance
Absorbed ~400 chip architects and compiler engineers, including founder Jonathan Ross and the team that built Google’s TPU v1
Gained LPU patents covering statically scheduled dataflow execution for transformers

The deal was structured as a technology and talent acquisition — existing GroqCloud customer contracts were wound down or migrated to NVIDIA cloud.

What happened to the LPU

Jonathan Ross and 90% of Groq engineers joined NVIDIA. The LPU architecture is reportedly being integrated into next-generation NVIDIA inference hardware, potentially surfacing in a 2026 or 2027 product. The TSQ-1 (Groq’s next-gen LPU that was in development) was folded into NVIDIA’s roadmap under a different name.

Groq’s rack-scale systems (LanguageModel System 1, or LMS-1) — deployed at partners including Saudi Arabia’s SDAIA and Dell — remain operational but are no longer supported for new deployments.

Market reaction

The acquisition confirmed that NVIDIA views inference-specialized hardware as an existential risk, not an incremental competitor. AMD’s MI300X, Intel Gaudi 3, and now the departed Groq were all chipping at inference — but the $20B price tag signals NVIDIA’s willingness to pay to maintain architectural lock-in.

For the remaining AI chip startups — Cerebras, SambaNova, Tenstorrent, d-Matrix — the Groq acquisition cut both ways: validation that the space has real value, but a reminder that NVIDIA is willing to acquire rather than compete.

Practitioner note

If you were using GroqCloud, your API access ended Q1 2026. Migration options: Cerebras Inference (similar latency profile), Together AI, or Fireworks. For the LPU’s core use case — real-time voice and interactive coding at sub-100ms TTFT — Cerebras is the closest public alternative at scale. The Groq acquisition is also a reminder that betting infrastructure on a single-vendor inference startup carries real discontinuity risk; multi-vendor API abstraction layers (LiteLLM, LangChain) are not just performance tools, they’re business continuity tools.