Groq Seeks $650M For Inference Push After Nvidia's $20B Deal

The News

Groq, the AI chip startup, is seeking to raise around $650 million in fresh funding, drawn largely from its existing backers, as it pivots away from selling hardware towards running an inference cloud business. The figure was reported by Axios and relayed by TechCrunch. Two investors, Disruptive and Infinitium, have reportedly agreed to cover any part of the round that other shareholders leave unsubscribed.

The timing is striking. The raise lands only months after one of the strangest deals of the present AI cycle. In December 2025, Nvidia struck a roughly $20 billion arrangement with Groq that deliberately stopped short of a full takeover. Senior Groq staff moved across to Nvidia, and Groq's chip technology was licensed to the larger company. Existing Groq investors, in turn, walked away with cash payouts.

Having effectively handed its hardware crown jewels to Nvidia, Groq is now rebuilding around inference: the compute-heavy work of running already-trained models to answer live prompts, rather than training the models from scratch. That is the part of the AI stack where demand is currently deepest.

Why It Matters

The structure of the original Nvidia deal is the real story here. A licence-plus-talent transfer that avoids a formal acquisition has become a defining manoeuvre of this AI cycle. It echoes the wave of reverse acqui-hires seen through 2024, when Microsoft absorbed most of Inflection AI's team, Amazon took on Adept's founders, and Google paid to license Character.AI's technology while leaving the shell company standing. Each let a giant scoop up scarce talent and intellectual property while sidestepping the antitrust scrutiny a clean buyout would invite.

That Groq can return to the market for $650 million so soon afterwards signals two things. First, investor appetite for inference infrastructure remains strong even after the founding team has been thinned out. Second, the centre of gravity in AI spending is shifting from training ever-larger models to serving them cheaply and fast at scale. Whoever controls low-cost inference controls the economics of every chatbot, copilot and agent built on top.

Indian Angle

For India, the inference question is not academic. It is a margin question. Indian model builders such as Sarvam AI and Ola-backed Krutrim are training homegrown language models, but the recurring cost that decides whether their products are viable is inference, billed largely in dollars. A weaker rupee makes every served token more expensive, squeezing startups that price their apps for Indian wallets.

Cheaper, faster inference hardware of the kind Groq champions could lower that running cost for the entire Indian developer base, from fintech firms embedding chatbots to government-facing services under the IndiaAI Mission. MeitY has already committed public money to subsidised GPU capacity precisely because compute scarcity threatens to lock Indian founders out. A healthier market for inference silicon, with more than just Nvidia setting prices, is squarely in India's interest.

There is also a talent signal. India supplies a large share of the engineers who staff firms like Groq and Nvidia, so deals that move whole teams between companies quietly reshape where that talent, and the resulting intellectual property, ends up sitting.

FAQ

How much is Groq trying to raise?

Around $650 million, mostly from existing investors, according to an Axios report relayed by TechCrunch. Backers Disruptive and Infinitium have reportedly offered to cover any unsubscribed portion of the round, suggesting the company wants certainty that the full amount is secured.

What was the Nvidia deal?

In December 2025 Nvidia agreed a roughly $20 billion arrangement that licensed Groq's hardware technology and moved senior staff to Nvidia, without amounting to a formal acquisition. Existing Groq investors received cash payouts as part of the transaction.

Why does inference matter more than training right now?

Training builds a model once; inference runs it every time a user sends a prompt. As adoption scales, inference becomes the dominant recurring cost, which is why infrastructure money is flowing towards serving models cheaply rather than only building bigger ones.

What does this mean for Indian AI startups?

More competition in inference hardware could lower the dollar cost of running models, directly improving margins for Indian firms building on tight rupee budgets and supporting government-backed compute initiatives.

This story was reported by TechCrunch. Read the full original coverage at TechCrunch.