Cursor's 250 New Hires: The Real SF Pool Is Under 3,000 People

Anysphere went from roughly 50 employees at $2B ARR in early 2026 to about 300 by mid-2026, with a Series E in advanced talks at a ~$50B pre-money valuation and no traditional recruiting function to speak of. Every AI founder now wants to copy the pipeline. Most are copying the wrong half of it.

The story that gets repeated ("Cursor has no recruiters, just founders and referrals") is half true and completely useless as a playbook. The half that matters is that the addressable pool is tiny, geographically concentrated, and already saturated by four employers. If you understand that, you can actually source against it.

The "no recruiters" story is a meme, not a playbook

Cursor does have a recruiting function. It hired Adam Ward, previously at Meta and Pinterest, to run recruiting-as-a-team-sport, and publishes a founder-owned sourcing playbook through Sierra Ventures. What Cursor rejects is the model where a recruiter owns a req and pushes candidates into a funnel. What it runs instead:

• Names, not roles, are the atomic unit. Candidates get posted in a #hiring-ideas Slack channel and swarmed by the team.
• Final round is a two-day in-office paired build with the actual team, meals included, ending in a demo.
• If mutual, the candidate starts Monday.
• First-round technical interviews ban AI beyond autocomplete. Founder Michael Truell calls unaided programming "a really great time-boxed test for skill and intelligence."

So when you read "no recruiters," read "founders own sourcing, and a small recruiting team runs logistics." That distinction is the whole article. You can copy the logistics. You cannot copy the pool without knowing what it looks like.

The real Cursor source pool, sized

The addressable pool for a Cursor-grade product engineer in the US is roughly 50,000 people at the outer edge and under 3,000 at the tight edge. Everything else in the pipeline (acqui-hires, in-person SF requirement, the AI ban in interviews) is a rational response to that number.

Here is the pool, sized from Refolk's index of US professional profiles:

2,700

The MoTS pool is where it gets uncomfortable. In Refolk's index, of the top 25 profiles carrying "Member of Technical Staff" or "Founding Engineer" as their current title, OpenAI alone accounts for 11. Anthropic, Databricks, and SpaceX/xAI take another 8 between them. That's 76% of the visible frontier-lab pool sitting inside four companies, and OpenAI is the largest single donor.

This is why Cursor's actual hiring bar looks the way it does. When your bullseye pool is a few thousand people and 44% of the frontier-lab tier is inside one competitor, you don't run a funnel. You run a name list.

What the founders actually sourced from

Cursor's founders sourced their first 20 engineers from five explicit pools, none of which are LinkedIn Recruiter searches. Every one is still workable in 2026:

1. Authors on LLM, compilers, and code-generation papers (NeurIPS, ICLR, PLDI).
2. GitHub commit histories on relevant OSS (LLVM, tree-sitter, Neovim ecosystems, model runtimes).
3. Ex-OpenAI, DeepMind, and Anthropic engineers, particularly Members of Technical Staff.
4. Conference speakers in the same subfields.
5. Referrals from early employees, weighted heavily to the MIT PDOS and Recurse Center lineages that produced the founders themselves.

The paper-authors channel is worth naming, because it explains why fresh grads with only AI-tool experience get filtered out. An arXiv author on a 2022 code-generation paper had to write the compiler-adjacent code themselves. That's the same signal Cursor's AI-ban interview extracts. The interview policy and the sourcing pool are the same filter, applied twice.

For a founder or recruiter trying to work these channels without a two-year head start, the friction is real: you need to cross-reference paper authors against GitHub handles against current employer against SF-area residency, and do it fast enough that the person is still gettable. That cross-reference is the exact gap Refolk closes. You describe the person in plain English ("US-based engineers who authored a code-generation paper at NeurIPS 2022 to 2024 and currently ship PyTorch code, ideally in SF") and get a ranked shortlist across GitHub, LinkedIn, and the open web in one query.

Acqui-hires are the recruiting channel, not the M&A strategy

Roughly half of Cursor's hires arrive through referrals, per third-party estimates. The other half of the last 250 hires (roughly 125 people) came from founder cold outreach and, structurally, from acqui-hires. At 300 headcount you need on the order of ten acqui-hire-scale events to fill that gap, and Cursor has run them on schedule:

• Supermaven, November 2024. Brought in Jacob Jackson (ex-Tabnine co-founder) and the Supermaven engineering team.
• Koala, July 2025. Staffed a new enterprise-readiness team. Koala's CRM product was not adopted and was later shut down. The point was the team.
• Graphite, December 2025. Cash plus equity above Graphite's $290M last valuation, bringing in co-founders Merrill Lutsky, Tomas Reimers, and Greg Foster, plus a team that had served Shopify, Snowflake, Figma, and Perplexity.
• Travis McPeak, individually. Co-founder and CEO of Resourcely, hired to run Cursor's security team. The pattern: hire the founder, not the IC.

Cursor's acqui-hires are not opportunistic. They are the arithmetic answer to a pool that runs out at ~3,000 people.

If you're a founder trying to copy this, the honest read is that referrals cap around 50% of hires, and past ~150 headcount you need either a very large operator recruiting team or a corp-dev checkbook. Cursor chose the checkbook. Most Series A companies cannot, which means they have to work the long tail of the pool harder than Cursor does.

What "sourcing without warm intros" actually looks like in this pool

If you don't have Cursor's referral graph, you have to reconstruct it from public signal. The signal is more legible than most recruiters assume, because the pool is small and the artifacts are public.

Concrete channels that map to Cursor's own five-pool list:

• GitHub contribution histories on LLVM, tree-sitter, Neovim, PyTorch, vLLM, and llama.cpp, filtered to committers with 2019-2023 activity (the pre-hype filter).
• arXiv and OpenReview authorships in code-generation, retrieval, and inference-optimization tracks.
• Ex-MoTS from OpenAI, Anthropic, DeepMind, Databricks, and xAI, weighted by tenure. In Refolk's index of 16,946 US MoTS/Founding Engineer profiles, these five employers dominate.
• Conference talk lists from NeurIPS, ICLR, PLDI, MLSys, and the smaller compiler workshops.
• Recurse Center batch alumni and MIT PDOS graduates, per the founders' own lineage.

The tactical mistake most non-Cursor founders make is treating this as a LinkedIn problem. It isn't. LinkedIn shows you the current title. GitHub shows you the pre-hype filter. Papers show you the intellectual lineage. You need all three joined on the same person, and you need the join to be fast, because at ~2,700 people the pool is smaller than a single Meta org.

Why Cursor's revenue-per-employee changes the math

Cursor's revenue-per-employee is roughly $10M at $4B ARR. Salesforce sits around $530K. Atlassian around $700K. That's a ~19x multiple over category leaders, and it's the reason the bar doesn't move.

19x

When each hire has to clear $10M in expected revenue contribution, "close enough" hires are a tax you cannot afford. That's why the two-day paired build exists, why the AI-in-interviews ban exists, and why the acqui-hire premium over Graphite's $290M valuation is rational rather than reckless. A $50B pre-money valuation on a 300-person company implies each seat is worth roughly $167M in enterprise value. Buying a team at a premium to their last round is cheap by that math.

For founders trying to copy the model at 30 or 50 headcount, the takeaway is not "hire like Cursor." It's "your pool is the same pool, minus the OpenAI-alumni gravity Cursor gets for free." You are fishing in the same 2,700-person pond, against a competitor whose interview loop starts on Friday and ends with a Monday start date. Speed is the only variable you actually control, and the sourcing step is where most teams lose a week they didn't need to lose. That is the specific friction Refolk removes: a plain-English query that would take a sourcer two days of Boolean and CSV work returns in a minute, so the founder's calendar becomes the bottleneck instead of the pipeline.

The playbook a Series A founder can actually run

The transferable pieces of Cursor's model, in order of ROI for a company that isn't yet Cursor:

1. Own sourcing at the founder level. Not "review the pipeline." Actually read GitHub profiles and paper PDFs yourself for the first 30 hires.
2. Publish a name list, not a job spec. A shared doc of 200 named engineers, updated weekly, outperforms any careers page in this pool.
3. Compress the loop. Two-day paired build, Monday start. Every day of latency loses candidates to the four employers that already dominate the pool.
4. Budget for one acqui-hire per 50 hires past headcount 100. The referral channel caps around 50%. Corp-dev fills the rest, or you slow down.
5. Treat SF residency as a filter, not a preference. Around 60% of the MoTS sample lives there. Remote-first sourcing shrinks your workable pool by more than half.
6. Filter for pre-2023 GitHub activity. Same filter as Cursor's AI-in-interviews ban, applied at sourcing time instead of assessment time.

None of this requires a $50B valuation. It requires knowing that the pool is 2,700 people, not 2.7 million, and behaving accordingly.

FAQ

How many engineers actually fit Cursor's hiring bar in the US?

The outer generalist pool (US Software/Sr/Staff engineers with TypeScript + LLM on their profile) is around 49,950 in Refolk's index. The realistic bullseye (Software/Founding/Staff engineers with PyTorch + CUDA) is 9,686, and the SF-Bay slice of that is roughly 2,700. The frontier-lab feeder pool (US Member of Technical Staff and Founding Engineer titles) is 16,946, but 76% of the top sample sits inside OpenAI, Anthropic, Databricks, and xAI.

Does Cursor really have no recruiters?

No. Cursor hired Adam Ward (ex-Meta, ex-Pinterest) to lead recruiting and runs a founder-owned sourcing model with a small recruiting team behind it. The meme conflates "no traditional req-owning recruiter funnel" with "no recruiting function." The distinction matters because the founder-sourcing part is copyable and the "we don't need a recruiter" part is not.

How do I source ex-OpenAI or ex-Anthropic engineers without warm intros?

Cross-reference public artifacts: GitHub commit histories on model runtimes and inference stacks, arXiv or OpenReview authorships in the relevant subfields, conference speaker lists, and current LinkedIn titles. The join across those sources is where most sourcing tools fail, because they only see one layer. Refolk is built for exactly that join: describe the person in plain English and get a ranked shortlist across GitHub, LinkedIn, and the open web.

Why do acqui-hires matter so much for Cursor's headcount growth?

Because referrals cap around 50% of hires, and the non-referral half has to come from somewhere. At 300 headcount with a ~2,700-person addressable SF pool, buying teams (Supermaven in Nov 2024, Koala in Jul 2025, Graphite in Dec 2025) is the fastest way to add 10+ senior engineers in a quarter without spending six months per hire on cold outreach. It is a sourcing channel dressed up as corp-dev.