LPU AI Inference Chip · 500 Tokens/sec · $12.5B Valuation
Last updated: June 18, 2026 · Status:Pre-IPO — No S-1 Filed · S-1: Not filed
The short version
Groq builds the LPU (Language Processing Unit) — a purpose-built chip for AI inference that's roughly 10x faster than NVIDIA H100 on LLM tasks, delivering 500 tokens per second. Founded by Jonathan Ross (former Google TPU architect), valued at $12.5B. No S-1 filed — but with Cerebras now public and AI inference demand exploding, Groq is a prime IPO candidate for 2026–2027.
IPO Quick Facts
Field
Data
Company
Groq Inc.
CEO / Founder
Jonathan Ross
Headquarters
Mountain View, CA
Founded
2016
Product
LPU (Language Processing Unit) — AI inference chip
Valuation
$12.5B
Total Funding
$1.497B
Key Investors
General Catalyst, Tiger Global, D1 Capital, Samsung
IPO Status
Pre-IPO — No S-1 Filed
Expected IPO
2026–2027 (estimate — no confirmed date)
Ticker
TBD
Primary Exchange
TBD (Nasdaq likely)
IPO Readiness Score
58
/ C+
IPO Readiness: C+ Grade
Massive $12.5B valuation, proven hardware, and fast inference performance. Blocked by: no S-1 filed, limited commercial revenue transparency, NVIDIA dominance, and the capital-intensive nature of chip manufacturing. IPO candidate in 2026–2027 if commercial traction accelerates.
Why Groq is Worth Watching
Groq's bet is that AI inference — running already-trained models — deserves its own silicon. NVIDIA GPUs were designed for training and adapted for inference; Groq's LPU was designed from scratch for transformer inference. The performance numbers (500 tokens/sec) are legitimately impressive and represent a 10x advantage over H100 in throughput.
Jonathan Ross's background is critical here: he was the lead architect of Google's TPU (Tensor Processing Unit) before founding Groq. He knows exactly what it takes to build competitive AI silicon, and he made the deliberate choice to focus purely on inference rather than training.
The risk is NVIDIA's ecosystem moat. Even if Groq's chip is faster, NVIDIA's CUDA software stack, developer tools, and cloud partnerships are extremely hard to displace. Groq needs to demonstrate that its speed advantage translates to real customer demand — either via GroqCloud (its hosted inference service) or as a chip embedded in customer infrastructure.
Technology Performance
Metric
Groq LPU
NVIDIA H100
Notes
Tokens/Second
~500
~50
10x faster for LLM inference
Architecture
Purpose-built inference
General compute adapted
LPU optimized for transformers
Memory Bandwidth
High (on-chip SRAM)
Shared HBM
No memory bottleneck
Availability
GroqCloud + on-premise
Widely available
Groq ramping deployment
IPO Timeline
2016
Groq founded by Jonathan Ross (former Google TPU lead architect).
Groq has not filed an S-1 as of June 2026. With the AI inference market growing rapidly and Cerebras now public, Groq is a prime IPO candidate for 2026–2027. The company has raised substantial capital and needs to demonstrate commercial scale before filing.
Groq was last valued at $12.5 billion from its Series D and related funding rounds in 2024. This reflects strong investor excitement about its LPU chip's performance advantage over NVIDIA GPUs in AI inference workloads.
Groq's LPU delivers approximately 500 tokens per second on standard LLM inference tasks — roughly 10x faster than NVIDIA H100 in comparable configurations. The key architectural difference: LPU uses on-chip SRAM with no memory bandwidth bottleneck, while GPUs rely on HBM memory that becomes a constraint at high batch sizes.
Groq competes with NVIDIA (H100, H200, GB200), AMD (MI300X), Cerebras (WSE-3), and custom silicon from Google (TPU), Amazon (Trainium), and Microsoft. Groq's differentiation is pure inference focus and the determinism of its architecture — no GPU overhead, no shared memory bandwidth constraints.
Pre-IPO Groq shares may be available on secondary markets like Hiive for accredited investors. Given the $12.5B valuation and high investor interest, shares would likely trade at a premium. No S-1 has been filed.
The AI inference market is massive and growing faster than the training market. If Groq can prove that its 10x speed advantage translates to customer cost savings and better user experiences, it has a genuine shot at disrupting NVIDIA's inference dominance. Jonathan Ross's TPU experience is a significant credibility signal. The question is whether Groq can build the software ecosystem and customer relationships to match its hardware performance.