Nvidia Took Ross. Groq Kept the Cloud. Source Both Sides of the $20B Split.

On December 24, 2025, Nvidia paid roughly $20 billion in cash to license Groq's inference technology and hire Jonathan Ross, Sunny Madra, and a chunk of the senior engineering bench. On June 22, 2026, the Groq that was left behind closed $650M led by Disruptive and Infinitum to scale its inference cloud. Two talent pools now exist where there used to be one, and they are not interchangeable.

If you are sourcing for AI infrastructure in the back half of 2026, this is the cleanest acquihire split on the board. It is also the one most recruiters are getting wrong, because they are still searching for "LPU engineers" as if that were a single bucket.

The deal, in the shape a sourcer actually needs

The Nvidia transaction was not a full acquisition. It was a non-exclusive technology licensing agreement bundled with an acquihire of Ross (founder/CEO, ex-Google TPU co-creator), Sunny Madra (president, who had arrived via the 2024 Definitive Intelligence deal and was running GroqCloud), and other senior leaders. Jensen Huang told Nvidia staff the plan was to "integrate Groq's low-latency processors into the Nvidia AI factory architecture." Three months later, at GTC March 2026, Nvidia unveiled the LPX inference platform, branded "Nvidia Groq 3 LPX," built directly on the licensed IP.

The $20B number matters for context. It is Nvidia's largest transaction on record, dwarfing the ~$7B Mellanox deal in 2019. That is the size of check a company writes when it wants an architecture and the people who can defend it, not when it wants to shut a competitor down.

$20B

Now the other side. Groq had already raised $750M at a $6.9B valuation in September 2025, right before the Nvidia deal. Existing investors got paid out at the December close and then, six months later, backed a $650M re-raise led by Disruptive and Infinitum. Per Axios reporting, that dual outcome (cash out, then reload) is unusual. Most acquihire "leftovers" cannot re-raise. This one did, in size.

That is the first sourcing signal. The residual Groq team is not the B-team.

What Nvidia took vs. what Groq kept

Nvidia took the silicon origin story. Ross co-created the TPU at Google. He built the LPU. The IP that became LPX is his and his direct reports'. That pool skews compiler, custom-ASIC verification, deterministic-latency systems, and low-level runtime. Small, expensive, mostly PhDs.

Groq kept the cloud. The company runs 13 data centers across North America, Europe, MEA, and APAC, serves more than 5 million developers, processes trillions of tokens per week, and is targeting 200 MW of capacity by end of 2027. You do not run that operation with chip designers. You run it with distributed-systems engineers, capacity planners, network architects, and DC-ops leaders.

The new executive bench makes the pivot explicit. Alan Rice joined as COO from xAI, before that Meta Datacenters, and before that US Navy nuclear submarines. Sinclair Schuller comes in as CTO from Apprenda and Nuvalence, enterprise software pedigree. Rakesh Malhotra is CPO after roughly a decade at Microsoft. None of those hires would make sense if the plan was to design another chip. All three make sense if the plan is to run the largest independent inference cloud on someone else's licensed silicon.

13

The two skill signatures, side by side

The Nvidia pool (ex-Groq, now inside Nvidia's LPX org):

• TPU/XLA compiler experience
• Custom-ASIC verification, RTL, timing closure
• Deterministic-latency systems, streaming architectures
• Ex-Google Brain/TPU alumni
• Compiler-design PhDs from Waterloo, Toronto, Stanford, CMU

The Groq pool (still at Groq post-raise):

• Multi-region inference cloud operations
• Kubernetes at DC scale, capacity forecasting
• Network engineering for low-latency serving
• Developer platform, API, billing, quotas
• Enterprise sales engineering and solutions architecture

If you are running a single Boolean for "LPU engineer" you are getting maybe 10 people worldwide and none of the roles you actually need to fill.

The Toronto compiler cluster nobody is talking about

Here is a data point from our own index that we have not seen anyone else write up. When you search for "Compiler Engineer" with "Groq" as a keyword, you get a very small, very concentrated pool. Of the top five profiles, four are currently at Groq and one is at Nvidia. Geographically, they cluster in Toronto and Mountain View, with a clear Toronto center of gravity for the compiler team.

That matters for two reasons. First, LPU-adjacent compiler talent is globally scarce, and any Toronto-based compiler engineer with Groq on their resume is a top-of-funnel target for every inference startup on the planet, whether they know it or not. Second, if Nvidia pulled only the stateside subset of Ross's team into the LPX org, the Canadian compiler bench may still be at Groq (working on runtime and toolchain for whatever comes next) or already free-agent. Either state is a sourcing opportunity.

This is exactly the kind of query that dies in a LinkedIn Recruiter session, because the title "LPU engineer" does not exist externally and "Compiler Engineer, Groq, Toronto" is a five-second filter that returns a list too small to justify a seat license. It is a better fit for Refolk, where you describe the person in plain English (compiler engineers who worked on Groq's LPU stack, currently in Toronto or open to remote) and get a ranked shortlist across GitHub, LinkedIn, and the open web, without pretending the role has a canonical title.

Named departures, named stayers

Sourcing an acquihire split starts with a roster. Here is what the reporting has confirmed.

Went to Nvidia:

• Jonathan Ross, founder and CEO. Ex-Google TPU co-creator. The archetypal "went with the deal" persona. If you are trying to find engineers who followed him, look at anyone who has worked directly under Ross since 2016.
• Sunny Madra, president. Arrived at Groq via the 2024 Definitive Intelligence acquisition and had been running GroqCloud. Notable because he was running the cloud side and still left. Read into that.
• Other senior leaders and "key engineering talent," unnamed in the public reporting but named in internal Nvidia comms.

Stayed at Groq:

• Doug Wightman, co-founder, ex-Google. TechCrunch reported he stayed and became CEO, though other outlets have cited Adam Winter as CEO and Matt Eng as CFO, with Simon Edwards (former CFO) initially named CEO at deal-close. The leadership title has churned publicly enough that if you are cold-outreaching Groq execs, verify the title on the day you send the note.
• Alan Rice, COO, joined post-raise.
• Sinclair Schuller, incoming CTO starting July.
• Rakesh Malhotra, CPO.

For a sourcer, "unnamed but reportedly key engineering talent" is where the actual work lives. You are looking for compiler, kernel, and hardware verification engineers who joined Groq between 2017 and 2023, are now showing Nvidia as current employer with a start date in Q1 2026, and whose LinkedIn geography ping either matches Santa Clara or is still showing their old city because they have not updated.

Nvidia took the silicon origin story. Groq kept the cloud. Those are two different Boolean strings.
</pull>

## The 12 to 18 month leak window inside Nvidia

Here is the contrarian read. Nvidia now has an internal team, sitting inside the LPX org, whose architecture is distinct from Blackwell and whose compiler stack is distinct from CUDA. That team is going to spend the next year and a half being politically absorbed, reorganized, and pushed to align with the mothership roadmap. Some of them will love it. Some will not.

Historically, acquihired teams start leaking around month 12. That means the "Nvidia pool" of ex-Groq talent will begin to show up on the market roughly December 2026 through mid-2027. That is a specific, time-boxed sourcing window. If you are staffing an inference startup or a hyperscaler's custom-silicon team, put a saved search on that cohort now and set the alert for early Q1 2027.

The Meta/Scale AI comparable is instructive. Meta's $14.3B "not-acquihire" of Scale AI happened about a year before the Groq/Nvidia deal. The remaining Scale team is reportedly on track for $1B in revenue post-deal. The acquihired Scale leaders inside Meta have already produced a visible trickle of departures. Expect the same pattern here.

## The competitor bench you should be cross-sourcing against

Groq's $650M is not happening in a vacuum. The same week, Baseten closed $1.5B at a $13B valuation. Capital is flooding the inference layer specifically. Every hiring plan Groq builds this year will collide with the same short list.

Cross-source against: Baseten, Together AI, Fireworks, CoreWeave, Nebius, Cerebras, Tenstorrent, SambaNova, d-Matrix, and Etched. If you are hiring for a Groq-adjacent role, half of your target list is already talking to at least two of those companies. The differentiator is speed of first contact and the specificity of the pitch, not the volume of InMails sent.

This is where plain-English search matters more than another Boolean. "Distributed systems engineers who ran multi-region inference at Together, Fireworks, or Baseten, open to a COO under Alan Rice at Groq" is not a search you can build in a keyword filter. It is a query. Refolk handles the query shape, ranks candidates against it, and shows you their current employer and any public signal that they are open. That is 80% of the work of an acquihire split.

## What to do this week

1. Build one list of ex-Groq engineers whose current employer is Nvidia, filtered to start dates in Q1 2026. Sit on it. Alert refreshes in 90 days.
2. Build a second list of current Groq engineers in Toronto, Mountain View, and remote NA, with compiler, kernel, or DC-ops signal. This is your active-poach list against Baseten, Together, and Fireworks.
3. Build a third list of ex-Scale AI engineers who left Meta between January and June 2026. That is your control group for what the Nvidia leak will look like in twelve months.

The Groq/Nvidia split is a rare, mappable case study. It will not stay this legible. Get the lists built while the graph is still fresh.

## FAQ

### Are "LPU engineers" really a distinct skill set from GPU engineers?

Yes, and that is the whole point. LPU (Language Processing Unit) is Groq's deterministic-latency, streaming-architecture chip. The compiler stack, the memory model, and the runtime are architecturally different from GPU. Engineers who worked on it have TPU/XLA compiler experience, custom-ASIC verification, and deterministic systems in their background, not CUDA kernel optimization. Do not search on the title "LPU engineer." Search on the pedigree: ex-Google TPU, compiler PhDs from Waterloo or Toronto, Groq or SambaNova or Tenstorrent tenure.

### How do I tell which ex-Groq people went to Nvidia vs. stayed?

LinkedIn current employer is the first pass but it is dirty. A better signal is start date on the Nvidia role. Anyone who joined Nvidia between December 2025 and February 2026 in a role touching inference, compiler, or ASIC is very likely part of the Ross cohort. Cross-reference against public GitHub activity on Groq's open source repos: contributors who stopped committing in December 2025 mostly moved, contributors who kept committing mostly stayed.

### Is Groq actually a good place to poach into right now?

Depends on the role. If you are hiring silicon or compiler talent, Groq is now a defensive story (the IP is licensed to Nvidia, the founder is gone) and it will be a hard pitch. If you are hiring for distributed systems, DC ops, developer platform, or enterprise GTM, Groq is arguably one of the strongest stories on the market: $650M in fresh capital, a working 13-DC cloud, 5M+ developers, and a leadership team explicitly built for scaling operations rather than designing chips. Match the pitch to the role.

### What is the fastest way to build the Toronto compiler shortlist?

Do not start with LinkedIn. Start with GitHub contributors to Groq's compiler and runtime repos, filtered by profile location or timezone. Then cross-reference against University of Toronto and Waterloo compiler-lab alumni pages. Then hydrate with LinkedIn to check current employer. Or describe the person in one sentence to Refolk and skip the three-tool workflow. Either way, the total addressable pool is small enough that speed of first contact is the only thing that matters.