River AI's $1.1B Hires From a Pool of 43. Here's the Named List.

On August 11, 2026, River AI announced $1.1B in seed and Series A at a reported $5B valuation, led by General Catalyst and AMP PBC with Nvidia, AMD Ventures, Y Combinator and Temasek. Igor Babuschkin, the ex-xAI co-founder running it, is buying compute, but he is also buying access to a talent pool that fits inside a single conference room. If you are recruiting against him, or trying to sell into him, the pool you actually care about is 43 people, not thousands.

The pool River AI is actually hiring from is 43 people

The intersection of "reinforcement learning post-training" and "LoRA fine-tuning on frontier open weights" does not exist as a labeled category. In Refolk's index of professional profiles, exactly 43 people globally list "reinforcement learning post-training" as a distinct headline concept in their current role, and a keyword filter for RLHF plus LoRA plus open-weight post-training returns zero profiles. That zero is the story. River AI's product scope, per its own release, is "state-of-the-art LoRA fine-tuning and reinforcement learning for frontier open weight models," and no one on the open market has bothered to label themselves that way yet.

That means two things for anyone competing for River AI hiring:

• Keyword sourcing on LinkedIn will miss the pool entirely. The label does not exist.
• The talent has to be assembled from adjacent cohorts (DeepMind RL, OpenAI post-training, xAI, Meta Superintelligence Labs) and filtered by whoever has shipped LoRA adapters at scale.

43

The top employer of those 43 is Meta with 3, then Amazon with 2, with singletons at Google, NVIDIA, Apple and AWS. Roughly 3 sit in the Bay Area. Another 5 or so are spread across Seattle, Bellevue and Sunnyvale. That is the entire visible US supply for a category River AI has to staff to double digits.

Why the $1.1B is a hiring instrument, not a compute instrument

Read the round size as a payroll commitment, not a GPU order. Meta has reportedly offered packages worth up to $300M over four years to retain top AI researchers. A single competitive counter from Zuckerberg eats about 1% of River AI's round. With 20 to 40 hires needed against a labeled global pool of 43 RL post-training specialists, Babuschkin has to convert roughly 50% to 90% of the visible market. That is not a compute problem. That is a war chest.

Three mechanics explain why the price per hire has moved to eight figures:

1. Attrition is concentrated. 50+ researchers and engineers have departed xAI since February 2026, and 11 of 12 co-founders are gone. That is one supply-side event, not a steady drip.
2. The counter-bidders are named and rich. Meta Superintelligence Labs has absorbed at least 11 xAI leavers. Thinking Machines Lab, run by Mira Murati, has absorbed at least 7. Anthropic and Google DeepMind pay similar bands.
3. Nvidia and AMD both invested. That is unusual and it narrows the target: engineers who know both CUDA and ROCm. The subset of the 43 that has shipped serious ROCm work is smaller still.

The round is a poaching-war war chest. The compute is almost incidental.

LoRA supply is offshore. RL supply is Palo Alto. That mismatch is the entire hiring plan.

The LoRA labor market and the RL post-training labor market barely overlap geographically. A LoRA-specific keyword search in Refolk's index returns 177 people, but only about 2 sit in the Bay Area or NYC. The modal LoRA engineer is in Bengaluru: 7 of the top 25 profiles. Meanwhile RL post-training clusters in the Bay Area and greater Seattle. River AI's Palo Alto HQ implicitly picks the RL side of that split and outsources or acquires the LoRA-scale-engineering side, which is exactly what the Nvidia and AMD strategic checks are for.

If you are sourcing open-weight LLM engineers against this posting, the shape of the market looks like this:

The practical read: you cannot hire the profile River AI describes off a shortlist that already exists. You have to assemble it, which is the exact gap Refolk closes. Describe the person in plain English (a DeepMind or OpenAI RL alum who has shipped LoRA adapters at 70B or above, currently in the Bay Area or willing to relocate, comfortable with both CUDA and ROCm) and get a ranked shortlist back, without pretending the label exists on LinkedIn.

The named sources you should actually work from

Start with four adjacent cohorts, in this order. This is the sourcing map every recruiter chasing reinforcement learning post-training talent should have taped to their monitor.

1. DeepMind alumni with RL from the Gemini post-training lineage. Babuschkin's own pedigree runs through DeepMind, where he worked on generative modelling and reinforcement learning before OpenAI and xAI. The RLHF playbook that transfers to open-weight models has deep roots there.
2. OpenAI RLHF leavers. Babuschkin led large-scale training at OpenAI before co-founding xAI, and the post-training and alignment groups are where the "reward model plus PPO on a frontier base" muscle actually lives.
3. xAI's post-training cohort. With 50+ departures since February 2026 and 11 of 12 co-founders gone, this is the single largest supply event of the year. Babuschkin is the natural aggregator, but the pool is contested by Meta and Thinking Machines. Names to watch include Manuel Kroiss, the last co-founder to leave xAI alongside Ross Nordeen.
4. Together AI, Fireworks AI and Predibase engineers. These three commercial open-weight fine-tuning platforms are the most poachable non-frontier-lab source of LoRA-at-scale expertise. Their infra engineers routinely serve LoRA adapters against large open-weight base models in production, which is the exact operational skill River AI's release language ("token-metered billing and instant deployment") requires.

Job-listing tracking across OpenAI, Anthropic, Google DeepMind and Meta's Superintelligence Labs ranks reinforcement learning specialists, evaluation and red-team engineers, post-training researchers, and inference optimization experts at the top of demand. Every one of those roles is a River AI role too. Every offer River makes will be counter-bid by Meta, Anthropic, Google DeepMind, or Thinking Machines. Plan the outbound sequence accordingly: assume two competing offers on the table by the time your candidate says yes to a first call.

Why Babuschkin can plausibly close this hire plan

River AI is not a cold start. It is a diaspora aggregator, and that is the only reason a four-month-old company (Nevada-incorporated on April 20, 2026) can credibly promise to hire 30 people who each cost eight figures. Read the round through that lens.

Three signals in the round itself support the read:

• AMP PBC leading alongside General Catalyst. AMP was founded in 2026 by Anjney Midha, previously a general partner at Andreessen Horowitz and a backer of Black Forest Labs, Mistral AI, LMArena and OpenRouter. That is the open-weight ecosystem's connective tissue. Warm intros to every relevant candidate flow through that network.
• Babuschkin's $100M personal check. He is reportedly committing up to $100M of his own money. Candidates read that as skin-in-the-game and as headroom for equity refreshers.
• Nvidia and AMD on the same cap table. Strategic silicon vendors do not co-invest by accident. Both want a neutral open-weight training substrate, which gives River a plausible story about not being locked to one accelerator, which matters to the systems engineers on the target list.

The "xAI-in-exile" framing is the real pitch. If you are Meta or Thinking Machines Lab, your counter-offer is bigger cash and a bigger existing team. If you are River AI, your counter is "you already know half the people here, and you get to rebuild what you built at xAI without the parts that made you leave." Against 50+ available leavers, that is a workable pitch for maybe 20 hires. The other 10 to 20 come from the DeepMind and OpenAI cohorts, and those close on comp and mission, not on nostalgia.

What this means for everyone else sourcing the same profile

If you are hiring LoRA fine-tuning engineers or reinforcement learning post-training talent in Q3 and Q4 2026, price in River AI as the new top of the market for this specific intersection. Three concrete moves:

• Stop keyword-searching. The label "RL post-training + LoRA" does not exist in the wild, and any tool that only matches strings will return a near-empty set. Search by trajectory instead: who worked on which post-training project, who committed to which open-weight fine-tuning repo, who spoke at which workshop.
• Widen to Together, Fireworks and Predibase before your competitors do. Those engineers are cheaper than frontier-lab leavers, they know LoRA at production scale, and they are one training run away from being credible RL post-training hires.
• Assume every offer is a bidding war. Bring your best comp forward on the first written offer. The candidates on this list have three other offers in their inbox.

Refolk was built for exactly this kind of assembly problem: describe the Igor Babuschkin team member you want in one sentence, across GitHub, LinkedIn and the open web, and Refolk returns the ranked list including people who never labeled themselves the way your JD does. If your sourcing pipeline still starts with a keyword filter, River AI's next 30 hires will come out of your funnel before yours do.

FAQ

How many people can actually do LoRA plus RL on frontier open weights?

The labeled global pool is effectively zero. Refolk's index shows 43 people worldwide with "reinforcement learning post-training" in their current role and 177 with "LoRA fine-tuning," but the intersection returns no profiles. The real working pool has to be assembled from adjacent cohorts at DeepMind, OpenAI, xAI, Meta Superintelligence Labs, Together, Fireworks and Predibase, and it is realistically in the low hundreds globally.

Who is River AI competing against for the same hires?

Primarily Meta Superintelligence Labs, Thinking Machines Lab, Anthropic and Google DeepMind. Meta has absorbed at least 11 xAI leavers since February 2026 and Thinking Machines at least 7. Meta has reportedly written retention packages worth up to $300M over four years for top AI researchers, which sets the ceiling River has to match or beat.

Why did both Nvidia and AMD invest in the same round?

Both silicon vendors want a neutral open-weight training and serving substrate that does not lock customers to a single accelerator. Their co-investment also signals that River AI plans to run on both CUDA and ROCm, which narrows the hire target further to systems engineers fluent in both stacks. That subset of the 43 is small, which is why the round is so large.

Where should recruiters focused on open-weight LLM engineers start today?

Start with four sources in order: DeepMind RL alumni, OpenAI post-training and RLHF leavers, the xAI diaspora (50+ people since February 2026), and infra engineers at Together AI, Fireworks AI and Predibase. Do not filter by keyword. Filter by shipped work: who has committed to LoRA training or serving code on large open-weight models, and who has published or presented on RL post-training recently.