Cognition's SWE-1.7 Named 4 Datacenters. That's Your RL Infra Shortlist.
Cognition's July 8 SWE-1.7 launch disclosed a globally distributed RL pipeline. Here's how to reverse-engineer the sourcing map before December vesting cliffs close it.
On July 8, 2026, Cognition shipped SWE-1.7 and, in the same breath, published something rarer than a model: an org-chart-shaped disclosure. Four datacenters, three continents, one US trainer cluster streaming compressed weight deltas to inference rollouts on rented Fireworks and Cerebras capacity. If you are hiring for frontier RL infrastructure, that blog post is a targeting document, and it has a shelf life.
The window is defined by a specific piece of paper. When Cognition closed the Windsurf acquisition in December 2025, it granted 100% accelerated vesting to every Windsurf employee for work to date. Roll that forward twelve months and you get a Q4 2026 / Q1 2027 cohort of engineers whose golden handcuffs are already off. Combine that with a $492M ARR run rate, 50% month-over-month enterprise growth, and a public admission of exactly which compute partners were in the loop, and you have the cleanest RL infrastructure engineer sourcing map of the year.
What SWE-1.7 actually disclosed
Read the technical blog carefully. Cognition trained SWE-1.7 from a Kimi K2.7 base using a globally distributed RL pipeline. A single US trainer cluster produces weight updates. Those updates are compressed into deltas, shipped through cloud object storage, and pulled by inference rollout clusters on three continents. The compression shrinks each transfer by more than 99%. Cross-continental weight updates for a trillion-parameter model complete in one to two minutes. Inference pauses for three to four seconds per sync.
That is not a description of a product. It is a description of a team.
To ship this you need people who have written or contributed to PipelineRL-style systems (arXiv:2509.19128 is the closest public reference), engineers fluent in top-p sampling with sampling distribution replay, off-policy RL correction, entropy collapse mitigation, and weight-delta streaming through object storage. Each of those primitives is name-searchable in commit history and paper co-authorship. That is your first-pass query, not "reinforcement learning engineer."
The named-byline roster
The SWE-1.7 blog and the FrontierCode 1.1 authorship list surface a public roster: E. Lu, B. Pan, D. Birlikci, S. Lee, R. Wang, R. Choudhury, F. Ma, TC Qin, C. Baronio, S. Alberti. Add the founders (Scott Wu, Steven Hao, Walden Yan) and the early hires like Gennady Korotkevich, and you have the closest thing to a directory of Cognition's frontier research group that will ever be published. This is not the list to poach from first. It is the list whose former reports and collaborators you poach from first.
The intersection nobody is searching
Most recruiters running Cognition Devin hiring searches will pull up the ~305 current employees on LinkedIn, filter for "reinforcement learning," and stop. That is the wrong query. The people who actually shipped July 8 sit at the intersection of three org charts, not one:
- Cognition's internal RL team (trainer side)
- Fireworks AI's distributed inference engineering (rented rollout capacity)
- Cerebras Systems' kernel and serving engineers (1,000 TPS inference)
Anyone who touched the SWE-1.7 pipeline is fluent in at least two of those three stacks. Add Moonshot AI's post-training team in Beijing (they authored the K2.7 Code paper that seeded the base model) and you have the invisible fourth lab. Fewer than 100 people on earth qualify against that intersection.
The LinkedIn filter set that produces this list does not exist. You cannot express "worked on distributed inference rollouts that plausibly touched a customer running trillion-parameter deltas through S3" in a Boolean string. This is exactly the kind of query Refolk was built for: describe the person in plain English, get a ranked shortlist across GitHub, LinkedIn, and the open web, with the paper co-authorships and commit histories already pulled in.
Windsurf engineers post-acquisition: the expiring arbitrage
Here is the part with a date on it.
Cognition acquired Windsurf for approximately $250M in December 2025. Windsurf brought $82M ARR, 350+ enterprise customers, and 210 employees into the deal. Roughly six months earlier, Google had already skimmed the top: on July 11, Varun Mohan, Douglas Chen, and other senior R&D leaders left for a $2.4B compensation-and-licensing package at DeepMind. The people who stayed for the Cognition deal are the ones you can actually reach.
Then Cognition did something recruiters should tattoo on their forearms. It granted 100% accelerated vesting to every remaining Windsurf employee. No four-year cliff. No two-year holdback. Fully accelerated for work to date.
The retention clock at Cognition is not private. It is stamped on a blog post, and it starts running out in December 2026.
That means the ordinary golden-handcuff timeline that keeps ex-acquisition talent stapled to a chair for 24 to 48 months is compressed to roughly twelve. A meaningful cohort of ex-Windsurf engineers hits a natural leave-of-absence window in Q4 2026 through Q1 2027. If you are building a distributed reinforcement learning talent bench, that is the window. Not next year. Not "when the market settles." Now, with outreach warm by October.
Jeff Wang (ex-Windsurf head of business, now inside Cognition) is the reference name to search around. His peers on the R&D side who did not leave with Mohan are the ones whose LinkedIn is quiet and whose inbox is not.
The nine-month severance filter cuts both ways
Cognition's CEO publicly offered nine months of severance to anyone unwilling to commit to the company's grind. Read the culture blog and it reads like a filter. It is. But filters have exhaust streams.
The engineers still there are the hardest people to poach on comp alone. They opted in with money on the table. Skip them for your first three quarters of outreach.
The engineers who took the exit are pre-qualified frontier-lab talent, already screened by Cognition's technical bar, already vested, and already available. Most recruiters running Cognition Devin hiring searches never look at this cohort because they do not know it exists. It is not on any list. It is people who quietly stopped listing Cognition as current employer between January and June 2026, whose GitHub activity shifted to personal repos, whose LinkedIn "about" line got a little vaguer.
Cerebras inference engineers and the adjacent-vendor pull
Cerebras runs SWE-1.7 inference at 1,000 tokens per second on wafer-scale hardware. That is not a plug-and-play integration. Someone at Cerebras wrote the serving kernels that let a Cognition-trained model actually run in production at that throughput. Those Cerebras inference engineers are the second most under-searched pool in this whole map.
The public signal set is thinner than for Cognition itself but denser than you would expect. Search for authors on Cerebras' engineering blog through H1 2026, the sglang and vLLM contributors who filed Cerebras-specific backend PRs, and the small group of ex-Groq and ex-Habana engineers who landed at Cerebras in the last eighteen months. Do the same exercise for Fireworks AI on the distributed inference side. This is a plain-English query problem, not a keyword problem, which is another place a tool like Refolk pays for itself: "senior inference engineers who shipped production kernels for wafer-scale or MoE serving in 2025 or 2026" is a sentence, not a Boolean.
The SI integration exhaust
The most-overlooked adjacent pool is the systems integrators. Infosys announced a Devin collaboration in January 2026 through its Topaz Fabric AI stack. Cognizant followed days later with a Flowsource integration covering both Devin and Windsurf. Those integration teams have hands-on Devin production experience, no Cognition equity lock-up, and salaries indexed to consulting comp bands rather than frontier-lab comp bands.
If you cannot afford to compete for the byline authors on the FrontierCode paper, you can absolutely compete for the Infosys engineer who spent Q1 2026 wiring Devin into a Goldman Sachs pilot. Same production surface area. One-fifth the cost of hire.
Why the non-US map matters
Cognition's public brand is San Francisco. The four-datacenter architecture is not. Rollout clusters on three continents almost certainly means contractor, partner, or hire-of-record engineers in EU and APAC. The Moonshot AI dependency puts serious research surface area in Beijing. The Fireworks and Cerebras partnerships have engineering staff in Bangalore and Toronto respectively.
A recruiter who searches only Bay Area LinkedIn for RL infrastructure engineer sourcing on this launch misses at least half the map. The people who authored the K2.7 Code paper are not on US recruiter radar. The Fireworks engineers who provisioned Cognition's rented rollout capacity are not filtered "San Francisco." Widen the geography, narrow the technical primitives, and the list gets shorter and better at the same time.
What to actually do this quarter
Three moves, in order.
First, build the exit cohort list. Ex-Windsurf employees who did not follow Mohan to DeepMind and did not opt into the nine-month grind. Their vesting cliff is visible on the calendar. Refolk queries phrased as "engineers who list Windsurf on their LinkedIn but not Cognition, who worked on inference or RL systems" surface this cohort in one pass.
Second, build the two-stack intersection list. Anyone with public work at both a distributed inference shop (Fireworks, Together, Anyscale, Cerebras) and an RL-heavy research lab. Cross-reference PipelineRL paper co-authors, SkyPilot and Ray heavy users, and EleutherAI Discord's RL channel regulars. This is the ~100-person pool.
Third, build the SI integration list. Infosys Topaz Fabric and Cognizant Flowsource engineers who touched Devin between January and June 2026. LinkedIn shows some of them. GitHub commit histories on internal-facing repos leak more of them. Their comp expectations are half of the frontier-lab ask.
The five-to-ten me-too RL-on-RL attempts that will launch in the next six months are already staffing up against this same pool. Every week you wait, the intersection gets smaller and the cost per hire climbs. The July 8 blog post handed you a map. The December 2025 vesting terms handed you a clock. Move before the clock catches the map.
FAQ
How many RL infrastructure engineers actually exist at Cognition's scale?
Measured in dozens globally, not thousands. Querying professional-network data for senior engineers whose profiles credibly combine distributed systems and reinforcement learning at trillion-parameter scale returns a vanishingly small pool. The SWE-1.7 launch adds specificity: engineers who have shipped cross-continental weight-delta streaming with sub-four-second inference pauses is a set of fewer than 100 people, and probably closer to 40.
When exactly does the Windsurf vesting window open?
The Windsurf acquisition closed in December 2025 with 100% accelerated vesting for work to date. In practice, ex-Windsurf engineers have unusually short golden-handcuff timelines relative to a normal acquisition, meaning Q4 2026 through Q1 2027 is when the natural leave-of-absence window arrives. Warm outreach starting October 2026 catches the cohort before competitors do.
Are Cerebras inference engineers really a poach target here?
Yes, but for a specific role type. The engineers who shipped Cerebras' 1,000 TPS serving of SWE-1.7 wrote wafer-scale kernels and serving infrastructure that few other companies have equivalents for. If you are building your own inference stack against MoE or long-context models, they are exactly the profile. If you are building trainer-side RL infra, the Fireworks distributed inference engineers are a closer match.
What does Refolk do differently on a search like this?
The core query behind this article ("engineers at the intersection of distributed inference and RL, who shipped in 2025 or 2026, and who are within one hop of the Cognition or Windsurf org charts") is not expressible in a LinkedIn Boolean. Refolk takes that sentence as a prompt and returns ranked candidates across GitHub, LinkedIn, and the open web, with paper co-authorships, commit signals, and current-employer inference already resolved. It is the difference between filtering a database and describing a person.