Today we've got a story with real dramatic flair. Its protagonist is Groq — once one of the brightest unicorns in AI chips, loudly proclaiming it would dethrone Nvidia's GPUs with its homegrown LPU silicon. Today? It just raised a $350 million Series A at a $3.5 billion valuation. Sounds like a lot, right? Except last September it was valued at $6.9 billion. Roughly cut in half.

And the kicker? The round is led by Disruptive, and Nvidia is planning to participate.

Yes, you read that correctly. The company that once wanted to destroy Nvidia is now Nvidia's partner, and Nvidia is even investing in it. The plot twists harder than a prestige drama.

Rise and Fall of the Dragon-Slayer

Let's rewind a few years. Groq was founded in 2016 by Jonathan Ross, a former Google TPU engineer, with a bold thesis: build its own LPU (Language Processing Unit) chips that deliver lower latency and higher throughput than GPUs for AI inference. Back then, Groq was the star startup — from seed through Series D, it raised over $1.75 billion from a who's who of investors: Social Capital, Tiger Global, D1 Capital, BlackRock, Cisco, Samsung… the list goes on.

Groq's narrative back then? "We're going to be the Nvidia killer of the inference era."

Then December 2025 happened. The plot took a sharp turn. Nvidia and Groq announced a technology licensing deal worth approximately $20 billion. The details? Nvidia got access to Groq's technology, and — here's the kicker — founder and CEO Jonathan Ross, president Sunny Madra, and multiple core employees all joined Nvidia.

Wait, this wasn't an acquisition. Groq remained an independent company. But the core team left for Nvidia, and the technology was licensed to Nvidia. So what was left of Groq?

An empty shell, and a business that needed a complete reinvention.

From Chip Designer to "Nvidia Compute Reseller"

Team gone, technology licensed out — what does Groq do? Pivot, obviously.

From a homegrown LPU chip company to an AI inference cloud operator running on Nvidia systems. Translation: we don't make chips anymore. We buy Nvidia chips, build clusters, and sell inference compute to customers.

Sound familiar? Yeah — that's basically what cloud providers do. The once-mighty "Nvidia killer" is now an Nvidia customer, helping Nvidia sell compute.

And the reason this round is called "Series A" is because it's effectively the first round of a new phase — new team, new business, new capital structure. The old valuation is gone. $6.9B → $3.5B, nearly halved. Groq says this isn't a "down round" — it's a "re-valuation" after the Nvidia technology licensing deal.

Sure. Whatever you say.

What's even more interesting is that Groq already raised $650 million in bridge financing back in June. Add this $350 million, and that's $1 billion in two months. Translation: the burn rate is high. Pivoting to inference cloud is a capital-intensive business — buying chips, building data centers, operations and maintenance — everything costs money.

How's Groq Doing, Really?

By Groq's own account, things are going pretty well. Thirteen data centers in operation, serving over 6 million developers, processing trillions of tokens per week. Plans to expand compute capacity from 54 MW to over 200 MW next year — a 270% increase.

But — what's the revenue? What's the customer pay rate? What's the data center utilization? Is it profitable? What's the split between Nvidia GPUs and homegrown LPUs in the new capacity? On these key questions, Groq has… nothing to say.

Draw your own conclusions.

And Nvidia's involvement is also a bit cryptic — "plans to invest," "pending closing." Which means the money isn't actually in yet. Is this a real investment, or just站台 (standing on stage) to help Groq keep the story going?

If Nvidia actually does invest, it would be the blackest comedy moment in AI history: the rival that once vowed to overthrow you ends up becoming your distributor, and you throw it a symbolic investment to show the world — look, even my enemies eventually become my partners.

A Few Hard Truths

Story time's over. A few honest takeaways.

First, if you're building hard tech, watch out for dimensional reduction strikes. You spend years building technology, and then a giant shows up and acquires your team and your IP with one deal and one check. You think you're competing with them — but they don't even see you as a competitor. You're just a talent pool and a technology reserve.

Second, "destroying the giant" is a narrative, not a strategy. Most startups that shout about disrupting the industry end up with their best outcome being acquired or absorbed by the very giant they targeted. Groq is actually lucky — at least the team got a high-priced "recruitment," and investors got paid back through the licensing deal. Most chip startups don't even make it that far.

Third, in the AI compute business, what matters in the end isn't technology — it's capital and scale. No matter how impressive your tech is, without billions to throw at data centers, you can't do business. Groq's pivot to inference cloud essentially transforms it from a technology company into a capital-heavy operator. And in heavy assets, the giants have crushing advantages.

So yeah — dragon-slayer stories are fun to listen to. In reality, most dragon-slayers either get recruited by the dragon… or they just don't make it.

That's it for today's roast. More tomorrow.