Baseten's $1.5B Round Wants to 3x Headcount. The Global RL Pool Is 107.
Baseten's $1.5B Series F plans to triple headcount, but the global post-training engineer pool is 107. Here's where to actually find them.
Baseten closed a $1.5B Series F on June 22, 2026 at up to a $13B valuation, and its press release commits to tripling headcount across engineering, research, operations, and GTM. The problem for anyone recruiting into that plan: the specific talent that justifies the multiple, engineers who have shipped production RLHF/RLAIF and continual-learning loops at inference scale, is one of the smallest addressable pools in the AI stack. In Refolk's index, the entire global set of named post-training engineers is 107 people, and Baseten already employs two of them.
The math behind Baseten's hiring wall
Baseten needs to hire into a global exact-title pool of 107 post-training engineers, of which only 15 sit in the United States. That is the addressable universe before any filter for seniority, visa status, or whether the person has actually shipped RL loops at billion-call/day scale.
The round itself is real and large. Baseten raised $1.5B led by Altimeter Capital, Conviction, and Spark Capital, with Sands Capital and Wellington co-leading, across two tranches at $13B and $11B. Business Wire pegs it as a 160% valuation increase in under five months, on top of the $300M Series E in January 2026 (IVP, CapitalG, Nvidia) at $5B. Sacra reports $600M ARR in March 2026, up roughly 1,900% year over year from $200M in December 2025. The platform now serves more than 1 billion inference calls per day across 87 clusters and 18 clouds.
That is the growth story the $13B tag is pricing. The hiring story is different.
Where the 107 actually work
Baseten sits inside the pool, not outside it. Refolk's index shows the top employers of the 107 as AECOM, Baseten (2), Cruise, OpenAI-Microsoft AI, and NVIDIA. Baseten's existing share is roughly 1.9% of the entire named global pool, which is significant leverage for a Series F company but also a warning: to double that share, Baseten has to poach from OpenAI, NVIDIA, or a direct inference competitor. There is no fourth option.
Why the "post-training engineer" title barely exists
The role is new. RLHF as a productionized pattern only took off in 2023, so the seniority curve has not had time to fill, and most people doing this work either carry a generic "Research Engineer" title or a lab-specific one that does not show up in a Boolean search.
That is the mechanism behind the small number. It is not that the work is rare; it is that the resume language is still forming. When you search LinkedIn for "post-training engineer" you miss the DeepMind researcher who spent two years on Gemini's alignment stack under the title "Senior Research Engineer, LM Post-Training." You miss the Anthropic staffer whose LinkedIn just says "Member of Technical Staff." Exact-title matching is a floor, not a ceiling, but the ceiling is not much higher: filter for people who have actually shipped RLHF/RLAIF loops on billion-call/day systems (essentially OpenAI, Anthropic, Meta, Google DeepMind, xAI, and now Baseten) and the addressable set drops well below 50.
This is exactly the gap Refolk closes. Instead of hoping "post-training" appears in a title, you describe the person in plain English (shipped an RL loop against a production inference cluster, comfortable with vLLM internals, published or contributed to a public RLHF codebase) and get a ranked shortlist across GitHub, LinkedIn, and the open web.
Refolk's index: the numbers Baseten's recruiters need to see
Every RL/post-training segment in Refolk's index is countable in three digits or fewer, and most are countable in two. This is what a $13B company is actually scaling into.
| Segment | Count | Notable employers |
|---|---|---|
| Global named post-training / RLHF / RL / Alignment engineers | 107 | AECOM, Baseten (2), OpenAI-MSFT, NVIDIA |
| US exact-title "Reinforcement Learning / Post-Training Engineer" | 15 | APQX (5), NVIDIA (2) |
| UK RLHF / post-training keyword matches | 2 | Causaly, Mercor Intelligence |
| Baseten's current share of named global pool | ~1.9% (2 / 107) | - |
| Implied post-training hires if 5% of net-new headcount is RL | ~20 | - |
| Baseten YoY revenue growth | ~20x ($200M to $600M ARR) | - |
The 20-hire figure at the bottom is the honest read of the press language. Baseten's release says "engineering, research, operations, and GTM," and at $600M ARR with an enterprise motion, the majority of net-new hires will be AEs, SEs, and platform engineers. The RL bench stays tiny by design, because the pool forces it. Founders and recruiters reading Baseten's headline should not assume 100+ open RL roles.
The London insight nobody is pricing
London, not the Bay Area, is the highest-density post-training market per capita, and it is where Baseten's second office belongs. Refolk's index shows 5 of the top 10 global regions for these titles clustered in the UK, driven by the Oxford and Cambridge alignment pipelines that produced Parsed itself.
Parsed, the RL post-training startup Baseten acqui-hired on December 10, 2025, was founded at Oxford by Chief Scientist Charles O'Neill and Rhodes Scholars Mudith Jayasekara and Max Kirkby. The team came out of ML Alignment & Theory Scholars (MATS), Cambridge, Stanford, NASA, and Johns Hopkins. Five months later, in May 2026, the ex-Parsed team shipped Baseten Loops, an RL training SDK for frontier AI systems. That product is the tip of Baseten's post-training org chart, and it is the artifact recruiters should reference when they cold-message candidates.
The playbook writes itself:
- Search MATS cohorts 2022, 2023, and 2024.
- Search Oxford and Cambridge ML PhD graduates from 2021 to 2025.
- Search Rhodes Scholar cohorts with a CS or stats concentration.
- Search anyone who has starred, forked, or contributed to trlx, TRL, OpenRLHF, or verl on GitHub in the past 18 months.
Why Baseten's customers are also its poaching targets
Baseten's hiring is structurally zero-sum with its own customer base, and this is the conflict nobody at the company wants to say out loud. Mercor, Cursor, and OpenEvidence sit on the customer roster. They also employ the exact talent Baseten needs.
Refolk's index surfaces Mercor Intelligence as one of only two UK employers with named "Alignment & Post-Training" or Staff ML titles (Causaly is the other). The other named Baseten customers (Abridge, Clay, Cursor, Lovable, OpenEvidence) are all in the market for post-training engineers of their own. Every hire Baseten makes from this list damages a sales relationship the GTM team just closed.
Poaching a post-training engineer from Mercor damages the customer relationship the sales team just closed. </p>
The competitive set makes it worse. Baseten competes with Together AI, Fireworks AI, Modal, hyperscaler deployment tooling, and indirectly with Groq. Each of those companies is hiring against the same 107-person global pool, and each just watched Baseten set a $13B comp bar. Expect base salaries for named post-training engineers to move 15 to 25% in the next two quarters.
Where to actually find the next 20 hires
The Parsed alumni network is roughly 10 people, so treat it as a seed, not a market, and expand through five well-defined communities. Here is the sourcing order that works:
- The Parsed team and their Oxford / MATS cohort. Sub-500 people, fully enumerable.
- APQX engineers. Refolk's index flags APQX as the largest single US employer of exact-title RL engineers (5). Very few recruiters have this company on their list.
- NVIDIA's post-training research org. Two named engineers in the US pool; NVIDIA is also a Baseten investor, which complicates outreach but does not block it.
- OpenRLHF, TRL, trlx, and verl contributors on GitHub. Filter for people who have shipped, not starred.
- Ex-Cruise perception and RL engineers. Cruise appears in the top-5 employer list; the Cruise wind-down freed people who did production RL on hard latency budgets.
For a recruiter running this from scratch, the friction is not the message, it is the list. Building the list across five overlapping platforms is where weeks disappear. Refolk collapses that to a single plain-English query across GitHub, LinkedIn, and the open web, which is the exact motion this pool demands.
What this means if you are also hiring RL engineers
If you are not Baseten, you are now hiring in a market where a $13B company just committed to tripling headcount against your same 107-person pool, and your compensation package needs to move before their recruiters call your team.
Three concrete moves:
- Rewrite the JD to name the artifact, not the title. "Ship an RL post-training loop against our inference stack" pulls candidates that a "Senior ML Engineer" post cannot reach.
- Widen to adjacent titles. Alignment Engineer, Research Engineer with a post-training publication, ML Infra Engineer with a TRL contribution. The 107-person exact-title count is the floor.
- Move the interview loop to two weeks. The Baseten counteroffer cycle at $13B will be fast, and any process longer than 14 days will lose finalists.
The Baseten round is not a warning that AI hiring is overheated. It is a warning that the specific bench required to justify a $13B valuation in AI inference is countable, named, and mostly already employed by five companies. Recruiters who treat the pool as a list instead of a market will win. The rest will spend Q3 running Boolean searches against a title that barely exists.
FAQ
How many post-training engineers exist globally?
Refolk's index of exact-title matches for "Post-Training / RLHF / Reinforcement Learning Engineer / Alignment Engineer" returns 107 profiles worldwide. Only 15 of those sit in the United States, and just 2 named UK profiles carry an "Alignment & Post-Training" or Staff ML title (at Causaly and Mercor Intelligence). Filter further for engineers who have shipped RLHF or RLAIF loops on billion-call-per-day inference systems and the addressable pool drops below 50.
Who are the Parsed cofounders now inside Baseten?
Parsed was founded at Oxford by Chief Scientist Charles O'Neill and Rhodes Scholars Mudith Jayasekara and Max Kirkby, all of whom joined Baseten through the December 10, 2025 acqui-hire. The team came out of ML Alignment & Theory Scholars (MATS), Cambridge, Stanford, NASA, and Johns Hopkins, and shipped Baseten Loops (an RL training SDK) in May 2026. That product is the concrete artifact showing where ex-Parsed talent is deployed inside the company.
Why is the UK a stronger hiring geography than the Bay Area for this role?
London and the broader UK hold 5 of the top 10 global regions in Refolk's index for named post-training titles, driven by Oxford and Cambridge alignment pipelines including MATS and the Rhodes Scholar network. That is a per-capita density the Bay Area does not match for this specific skill, and it is where the Parsed founders came from. A London office for Baseten's post-training team is a stronger structural bet than a second US hub for this function.
Which companies compete most directly with Baseten for the same 107 people?
Baseten competes with inference-cloud peers Together AI, Fireworks AI, and Modal, with hyperscaler deployment tooling from AWS, Azure, and GCP, and indirectly with custom-silicon players like Groq. Each of those is hiring against the same 107-person global pool. On top of that, Baseten customers Mercor, Cursor, and OpenEvidence employ the exact talent Baseten needs, which turns every successful poach into a sales-relationship risk.