HN August 2026 "Who Is Hiring": SF-Onsite Is the New AI Signature

If you are hiring AI engineers from outside the Bay Area this quarter, the August 2026 "Ask HN: Who is hiring?" thread is bad news you need to read carefully. The top of the thread is dominated by AI-native seed to Series B companies posting "SF, onsite, no visa sponsorship," and that combination is the most useful hiring intel you will get this quarter.

This is not the "remote is dead" story. Aggregate remote demand is still growing. What is happening is narrower and more actionable: a specific slice of well-funded AI-native companies has decided physical co-location is a velocity bet, and that decision is redrawing where the best AI ICs are willing to sit.

What the August 2026 HN thread actually shows

Among the seed to Series B AI-native cohort in the August 19, 2026 thread (item 49156683), "SF, onsite" is now the default work-mode tag and remote is the exception. HN's own posting convention requires marking mode ("REMOTE" or "ONSITE"), so month-over-month comparisons are apples to apples.

A few postings anchor the pattern:

• Bucket Robotics (YC S24, SF, 5 employees) posted three engineering roles, all SF-onsite. Their Senior Frontend listing is $120K to $160K base, 0.50% to 1.50% equity, US citizen or visa holder only. Founder Matt Puchalski is ex-Argo AI and Stack AV. Canonical seed-stage AI-native shape.
• Tasklet, from ex-Firebase founder Andrew Lee and Jonny Dimond, raised $42M from USV, Lightspeed and YC. Team of 12 in SF. The listing reads: "All roles are in person at our office in SF, no exceptions, and we're not sponsoring visas at this time." Their Head of Design Engineering IC role is $300K to $450K base, 0.35% to 0.70% equity.
• Proxima Fusion (Munich, Max Planck spin-out, €650M+ raised) posted a Senior SWE role marked ONSITE in Munich. Not SF, still onsite. The pattern travels.

Compare that to the historical HN baseline: at the remote peak, hntrends.com clocked 76% of postings tagged remote, drifting down to 64% by the March 2023 snapshot. Filter the August 2026 thread to AI-native postings from companies under 50 people and onsite dominates.

Why concentrated-capital AI companies are choosing onsite

Concentrated-capital AI startups are choosing onsite because they are optimizing for iteration speed on hard, undefined problems, not for HR policy or real estate. When your competitive edge is "we ship the next eval run faster than DeepMind's team down the street," a whiteboard beats a Zoom.

Three mechanisms are doing the work:

1. Iteration velocity on undefined problems. Bucket Robotics is doing CAD-to-vision defect detection. Tasklet is building agent tooling. Proxima is building a stellarator. None of these have a spec; the spec emerges from arguing at a whiteboard. Remote-first is great for shipping known scope. It is slower for defining scope.
2. Recruiting compression. A 12-person team hiring their 13th engineer cannot afford a 40-day loop. Onsite mandates cut the funnel to people who already live in a coffee-walk radius, which is why "US citizen or visa holder, San Francisco" shows up as a hard filter, not a preference.
3. Comp band inflation. Tasklet is posting a $300K to $450K IC base for Design Engineering. That number does not exist for a remote hire at a 12-person company, because it is a compression play: pay top of market for someone who will be in the room every day and delete two org layers.

The "no visa sponsorship" language is the tell. Tasklet's "no exceptions" and Bucket's "US citizen or visa only" repeat across the thread, and combined with onsite they shrink the addressable pool to a few thousand people who are already in the Bay Area with work authorization sorted. That is a much smaller and much more sourceable universe than the aggregate US AI engineer count.

The addressable pool, in real numbers

There are 21,022 US-based professionals holding AI Engineer, ML Engineer or Research Engineer titles in Refolk's index as of the August 2026 pull. That is the top of the funnel. The SF-onsite carve-out is a small slice of that number, and it is the slice everyone in this thread is fighting over.

21,022

The more revealing cut is how few of these people self-identify as remote. Only 10 profiles in that 21,022 pool have "remote" in their headline. That is roughly 0.05%. Whatever remote AI engineers are doing, they are not branding themselves that way, and they are not the ones AI-native founders are chasing.

The top current employers of that 21,022 pool, per the same index, are Google DeepMind, Meta, ElevenLabs, Distyl, Notion and Harmonic. All six have an SF Bay Area presence. If you are a non-SF founder trying to hire off this list, you are not competing with a job board. You are competing with a coffee-walk radius.

How to source against SF-onsite gravity if you are not in the Bay Area

If you are outside SF, stop trying to out-post the thread. Source outbound, target the specific micro-segments that do not want to relocate, and lead with the two things onsite cannot offer: async depth and equity in a smaller pond.

Four segments that are sourceable right now:

• The "quietly onsite elsewhere" cohort. Proxima Fusion in Munich is the proof that onsite AI hiring is not SF-exclusive. There are equivalent pockets in Zurich, London, NYC, Toronto, Boston, Seattle. Source engineers already living in your city who would rather not move.
• Post-IPO engineers with mortgages. People at Google DeepMind, Meta and (increasingly) ElevenLabs who are five years in, have kids in school, and would take a 20% pay cut to skip the daily commute. These do not show up in HN threads. They show up when you search for them.
• Ex-YC founders returning to IC. The 2022 to 2023 YC batches produced a lot of technical founders whose companies wound down or acqui-hired. Many now want a senior IC role, not another founder grind. They are in Austin, Denver, Portland, London.
• Distyl and Harmonic alumni. These two show up in the top-10 current employers of the AI-title pool, and both have generated a diaspora of engineers who have now seen how AI-native companies actually operate and can bring that operating model to a remote team.

This is the exact gap Refolk closes: you describe the person in plain English ("senior ML engineer, ex-DeepMind or ex-Meta AI, currently in Seattle or Toronto, not actively looking, has open-source work on eval infrastructure") and get a ranked shortlist across GitHub, LinkedIn and the open web. You do not need to guess titles or Boolean-string your way to a dead end.

The inbound trap the HN thread hides

Inbound applications are the wrong metric for SF-onsite AI roles, because the applicant math is inverted. Remote roles are roughly 8% of LinkedIn postings but receive about 35% of applications. That is a 4x to 5x demand overhang. SF-onsite gets the reverse: fewer applications per posting, but the applications are lower quality because self-selected relocators dominate the pile.

Onsite postings look uncrowded from the outside. From the inside, the pile is smaller and worse.

The implication for a founder reading the HN thread: if you post an SF-onsite role and rely on inbound, you will get a thin pile of self-selectors, and the best candidates in the 21,022 pool will never see it. Outbound is not optional in this slice. It is the only mode that works.

This is where a sourcing tool that indexes across GitHub, LinkedIn and the open web earns its keep. The best candidates are working, not applying, and their strongest signal is a repo or a conference talk from six months ago, not a resume update. Refolk indexes all three surfaces at once, which means "find me people with recent CUDA kernel commits who also list ex-DeepMind on LinkedIn and gave a talk at MLSys" is one query, not five tools.

What the SF-onsite shift means for comp

SF-onsite AI-native companies are posting the top of the comp band, and if you are not matching it, you need to compete on scope, not cash. Tasklet's $300K to $450K IC base for Design Engineering is not the ceiling; it is the floor for a 12-person Series A. Bucket's $120K to $160K for a Senior Frontend at a 5-person YC S24 is lower cash but 0.50% to 1.50% equity, which at a fusion or robotics outcome dominates the base.

If you are hiring remote for the same profile, you have two levers:

• Scope compression. Offer the person "you own the entire model serving stack" in a way a 12-person SF team cannot, because they already have someone for that.
• Equity in a smaller cap table. A remote Series B with 0.25% at a $200M valuation beats an SF Series A with 0.10% at a $500M valuation on expected value, if you are honest about dilution.

Cash-matching an SF-onsite offer from remote is a losing move. The candidate will take the SF offer and tell themselves the office is fine.

What to actually do this week

Read the August 2026 HN thread as a sourcing map, not a job board, and rebuild your outbound list around the specific companies you are now competing against. Concretely:

1. Pull the SF-onsite AI-native postings from the thread. hnhiring.com and hntrends.com both give you clean filters. That list is your competitive set for Q4.
2. Map each to their top 20 engineers. These are the people your candidates will be compared against, and the people you will be recruiting from in 18 months when the seed rounds do not extend.
3. Build a target list from the six top employers. Google DeepMind, Meta, ElevenLabs, Distyl, Notion, Harmonic. Filter to engineers who have been there 2+ years and live outside SF. That is your most sourceable segment.
4. Write outbound that acknowledges the SF-onsite pull. Do not pretend it does not exist. Lead with the specific thing your remote role offers that Tasklet's office cannot: full ownership, no relocation, real equity, or a specific technical problem SF is not working on.

The founders who will win Q4 hiring are not the ones who read the HN thread and moved to SF. They are the ones who read it and understood exactly which people they are now competing for, and started calling them.

FAQ

Is remote AI hiring actually dying?

No. Global demand for remote and hybrid tech professionals rose 19.8% YoY in Q4 2025, per Toptal. What is happening is a specific carve-out: AI-native seed to Series B companies with concentrated capital are choosing onsite as a velocity bet. Aggregate remote demand keeps growing. The HN thread just is not where you see it.

How do I read the HN "Who is hiring" thread as a sourcing signal, not a job board?

Ignore the roles and read the metadata. Which stage of company is posting, what work mode they are demanding, what comp bands they are willing to publish, and what visa or citizenship filters they are applying. In August 2026, the answer is "seed to Series B, SF-onsite, top of band, US-authorized only," and that tells you where the AI talent gravity actually sits going into Q4.

Where does Refolk fit in outbound sourcing against SF-onsite roles?

Refolk indexes GitHub, LinkedIn and the open web, so a single plain-English query returns candidates who match on repo activity, current employer, location and title signals at once. For SF-onsite competition specifically, that lets you find the people who fit the profile but have already opted out of SF (or never lived there), which is the exact segment inbound will never reach.

What is the smallest useful cut of the 21,022 US AI engineer pool?

Filter by current employer among Google DeepMind, Meta, ElevenLabs, Distyl, Notion or Harmonic, then by tenure of 2+ years, then by current city outside the SF Bay Area. That gets you to a few hundred people, all of whom have direct exposure to how AI-native companies operate and none of whom will show up in your inbound pile. That is your Q4 outbound list.