Groq

Groq builds Language Processing Units (LPUs) — a deterministic, software-scheduled inference architecture with large on-chip SRAM that targets the lowest, most predictable LLM latency; its viral GroqChat demo showed 500+ tok/s. Founded by ex-Google TPU lead Jonathan Ross. Raised a $750M Series E in September 2025 at a $6.9B valuation. In December 2025 Nvidia struck a reported ~$20B 'not-an-acquisition' — a non-exclusive license of Groq's hardware IP plus the hire of Ross, president Sunny Madra, and several senior engineers, with Groq's investors paid out in cash. Groq remains an independent company: GroqCloud keeps operating under interim CEO Adam Winter and CFO Matt Eng, and the company is raising ~$650M from existing backers to scale its inference cloud.

Groq's LPU posts among the fastest publicly measured LLM inference speeds — validation that drew a reported ~$20B Nvidia deal to license the technology and hire its leadership, even as Groq pushes on independently with GroqCloud.

groq.com AI AcceleratorsUSAprivate

At a glance

Total raised

$1.7B

last round series e

Headcount

—

no public estimate on record

Headquarters

Mountain View

USA

Founded

2016

10 years active

Status

Private

AI Accelerators

Latest valuation

$6.9B

Series E · 2025

For subscribers

Subscribe to see the full profile

Including Funding rounds, Bull / Bear thesis, Stock + earnings, Roster changes, Patents, News, and Open roles.

Founding

50% off

/month

91 of 100 spots left

Full

$10

/month

Available once all 100 Founding Member spots are claimed.

Get full access

Already subscribed? Sign in →

Funding history

Total raised

$1.7B

Rounds tracked

Latest stage

Series E

Latest valuation

$6.9B

Date	Stage	Amount	Valuation	Lead investors
Sep 17, 2025	Series E	$750M	$6.9B	Disruptive
Aug 1, 2024	Series D	$640M	$2.8B	BlackRock
Apr 1, 2021	Series C	$300M	$1B	Tiger Global Management, D1 Capital Partners
Jan 1, 2017	Seed	$10M	—	Social Capital

How they make money

Consumption-based infrastructurePattern

Groq runs a consumption-based infrastructure model: customers pay per token of inference processed through its cloud API. The company sells access to its Language Processing Units (LPUs) as a cloud service, where developers and enterprises submit AI inference requests and are billed on throughput (tokens per second) or compute time consumed. GroqChat, the viral demo that showcased 500+ tok/s, serves as both a public proof point and a lead-generation funnel for paid API access. The economics are structurally similar to other AI-inference-as-a-service offerings: gross margins are driven by hardware utilization rates across the LPU fleet, with fixed chip costs amortized over variable inference volume. Groq does not disclose whether it also sells LPU hardware directly to enterprise customers or cloud providers, but the primary revenue mechanism appears to be metered cloud access. Since the company is private and pre-revenue or in very early revenue stages, no margin or pricing data is available. The model scales favorably as fleet utilization increases, though the capital expenditure for fabricating custom ASICs (the LPU) is front-loaded and significant.

Generated by deepseek-v4-flash · June 2, 2026Analytical summary · not investment advice

Founder backgrounds

Jonathan Ross

Co-founder (former CEO; joined Nvidia Dec 2025)

Studied mathematics and computer science at NYU's Courant Institute, becoming the first undergraduate to complete Ph.D.-restricted courses. Joined Google, where he originated the Tensor Processing Unit (TPU) as a 20% project — spending two years designing and deploying the first-generation TPU, which went on to power over 50% of Google's internal compute. Subsequently worked at Google X's Rapid Eval Team incubating new Alphabet ventures. Co-founded Groq in 2016 with former Google colleagues, inventing the Language Processing Unit (LPU), an architecture prioritizing sequential inference throughput over parallel training. Groq achieved widespread recognition for dramatically faster token generation speeds. He joined Nvidia in December 2025 as Chief Software Architect.

AI bull / bear

Bull case

Nvidia just paid $20 billion to license Groq's IP for its Vera Rubin inference chips. This is the single strongest signal in the industry that Groq's LPU architecture — not GPUs — is the right answer for AI inference. Nvidia, the dominant player, effectively validated that Groq's deterministic, ultra-low-latency design is the path forward. With a $640M Series D, a founding team that built Google's TPU, and publicly measured speeds of 500+ tok/s on GroqChat, the company has both the technical proof and the capital to scale. The Nvidia licensing deal provides a massive revenue stream and strategic endorsement that opens doors with every hyperscaler and OEM. Instead of competing head-on with Nvidia's distribution machine, Groq gets paid to be the engine inside it. That is the only realistic path to challenging the incumbent, and Groq just executed it.

Bear case

The $20B Nvidia license is a poison pill disguised as validation. Nvidia now owns Groq's core IP inside its own silicon roadmap, meaning Groq has effectively licensed away its moat to the one company with the resources to commoditize it. Nvidia's Vera Rubin chips will ship at hyperscale volumes Groq cannot dream of matching — Groq still has "limited production scale" and only $640M in funding against Nvidia's $100B+ annual compute budget. Once Nvidia bakes LPU-style inference into its mainstream lineup, why would any customer buy from Groq directly? The company goes from being a disruptive threat to a paid-up supplier of a single customer. If Nvidia changes direction or renegotiates, Groq has no diversified revenue and no manufacturing scale to fall back on. That is not a partnership; it is an acquisition in slow motion at the buyer's price.

Generated by deepseek-v4-flash · June 2, 2026Refreshes quarterly · see disclosures

In the news

Groq · last 30 days

No articles ingested yet for Groq. Once the hourly news pipeline is live, every article the classifier tags as mentioning this company appears here with its one-line AI summary and sentiment.

Watch

Leadership

Jonathan Ross

Co-founder (former CEO; joined Nvidia Dec 2025)

Adam Winter

Interim CEO (Groq veteran; succeeded Simon Edwards, ex-CFO, who led briefly post-deal)

—

Matt Eng

Interim CFO

—

Dinesh Maheshwari

CTO

—

Products

LPU (Language Processing Unit)
A purpose-built AI inference chip using a deterministic, software-scheduled architecture with large on-chip SRAM instead of external HBM, eliminating the memory bottlenecks that limit GPUs on token generation. Designed by ex-Google TPU lead Jonathan Ross, the LPU delivers extremely low, predictable latency for large language models. It is Groq's core silicon and the basis for the fastest publicly measured LLM inference speeds.
GroqCloud
A developer-facing inference cloud that serves popular open models (Llama, Mixtral, Whisper, and others) at very high tokens-per-second over an OpenAI-compatible API. GroqCloud lets developers tap LPU performance without owning hardware and is Groq's primary commercial channel, with usage-based pricing. It remains the company's flagship service and continues operating after the December 2025 Nvidia deal.
GroqRack / On-Prem Systems
Rack-scale systems that interconnect many LPUs into a deterministic compute fabric for customers who need on-premise or sovereign inference deployments. GroqRack scales the LPU architecture beyond a single chip for high-throughput, low-latency serving of large models in private data centers, extending Groq's offering from cloud API access to dedicated enterprise and government installations.

Most recent patents

18 patents on file, but none with both an extractable figure and an abstract on Google Patents yet.