The US AI labor market is bifurcating on geography in a way that matters for every engineer choosing an offer. At one end: Anthropic, OpenAI, and xAI running office-first concentration plays, with fewer than 8 percent of open roles carrying any remote designation. At the other: Hugging Face with 89 percent of its open US postings listed remote, Cohere distributing 252 of roughly 1,100 employees outside any named city, and Scale AI operating a global contractor workforce that dwarfs its San Francisco headcount by an order of magnitude. The July 2026 map of US remote AI hiring is not a story about the death of the office. It is a story about two fundamentally different theories of what produces frontier AI work — and a compensation spread that is compressing as remote-first labs get serious about competing for the same senior candidates.
What's Happening
Only 3.55 percent of US job postings offered fully remote work in early 2026, according to cross-industry data tracked by Decanaria. But inside the Technology, Information and Internet sector, that figure is closer to one in four postings — and within AI specifically, the distribution is more polarized than any aggregate captures.
The office-first bloc. Anthropic's job board carried 392 open roles as of early June 2026. Thirty-five of those — roughly 8 percent — listed a remote option, concentrated in AI Safety and Policy, Security Engineering, and select infrastructure roles. The company's posture is deliberate: it occupies more than one million square feet of office space along Howard Street in San Francisco and expects most employees in-office at least one week per month. The lab's internal argument is that interpretability and safety research benefit from physical density in a way that cannot be replicated asynchronously. OpenAI runs a near-identical policy — of 681 analyzed roles, just 28 (4 percent) are remote-eligible, concentrated in security, infrastructure, and select deployment engineering. Sam Altman has publicly called remote work "a mistake" for frontier AI development. xAI, now operating as the AI division of SpaceX following Musk's May 2026 restructuring, is the most location-concentrated of the three: its Palo Alto headcount and the Colossus supercomputer campus in Memphis define its physical footprint, and remote engineering roles are not a documented part of its hiring playbook.
The distributed bloc. Hugging Face is the clearest counterexample in the US market. Of the company's roughly 730 employees globally, 89 percent of open US-facing postings list the role as remote. The company's 56 percent of postings focused on US locations but hiring from anywhere reflects a genuine work-from-anywhere policy supported by stipend infrastructure — not the "remote-friendly with SF travel" formulation that Anthropic uses. Hugging Face co-founder and CEO Clément Delangue has described the architecture as a structural advantage: the open-source model requires contributors and employees to function asynchronously across time zones by default, so the tooling and culture required for distributed work was built into the company's operating system before it had 100 employees.
Cohere sits in a middle tier that has become more interesting since its April 24, 2026 merger announcement with Germany's Aleph Alpha. Pre-merger, Cohere operated 122 open roles with 72 in Toronto and 57 in New York — but 252 employees were classified in non-named city locations across the combined workforce of roughly 1,100. The Aleph Alpha transaction, anchored by a $600 million commitment from Schwarz Group and valuing the combined entity at approximately $20 billion, will by design need to sustain a transatlantic distributed workforce. Aidan Gomez, Cohere's co-founder and CEO, described the deal as building "the world's first truly transatlantic AI company" — a framing that implies distributed team architecture as a feature, not a constraint.
Scale AI presents a different model again. The San Francisco-headquartered company employs approximately 1,000 full-time staff, but its core product is a globally distributed workforce of contractors and specialists. The Outlier and Remotasks platforms route RLHF and data-labeling work to credentialed specialists worldwide. RLHF contractor rates from the Scale pipeline range from $18 to $35 per hour for general tasks, rising to $35 to $60 per hour for credentialed specialists in chemistry, law, medicine, and STEM domains. That distributed architecture is not a hiring preference — it is the product. The Mercor marketplace operates adjacent to this model: the San Francisco-based platform placed over $50 million (est., per ENTRA tracking) of contractor compensation in 2025 and functions as an informal employer of AI domain experts distributed across the US and globally.
Why It Matters
The office-first versus distributed split maps almost exactly onto two different theories of where frontier AI value gets produced.
Anthropic's and OpenAI's arguments for physical density are research-specific and credible on their own terms. Mechanistic interpretability — reverse-engineering how trained models map parameters to meaningful algorithms — has a global credentialed candidate pool that numbers in the hundreds. Concentrating those people in one building in San Francisco is not an arbitrary preference. It is a bet that the marginal gain from hallway proximity is real and measurable against the loss of candidates who will not relocate. The data, so far, supports the bet: Anthropic has been able to compress its median time-to-offer for all open roles to 19 days against an industry benchmark of 47 to 64 days, suggesting it is not losing the offer competition even with a location requirement attached.
The distributed-first labs are making a different bet: that the candidate pool for open-source AI infrastructure, enterprise AI deployment, and RLHF quality work is large enough and geographically dispersed enough that concentrating in one city is a self-imposed constraint rather than a research advantage. Hugging Face's retention score of 90 on the ENTRA Talent Index — 42 percent above the weighted average for comparable employers — is the cleanest evidence that the distributed model can hold senior talent without frontier-lab compensation. At Hugging Face, senior ML engineers in the US earn $200,000 to $360,000 total compensation, per the company's published bands; software engineers peak around $183,000 for US remote roles per Levels.fyi reported data. Those numbers are meaningfully below Anthropic's Research Scientist median of approximately $746,000 — but the gap is partially explained by role category and partially by the mission-attribution value that open-source work provides. A senior Hugging Face engineer shipping code used in production at Anthropic, Google DeepMind, and Mistral simultaneously carries public attribution that is permanent. That is a retention mechanism with no equivalent at a closed frontier lab.
The compensation story is compressing at the distributed end. ZipRecruiter data for remote Cohere AI roles shows bands of $109,000 to $162,000 for mid-tier positions. Senior ML engineers at Cohere earn $240,000 to $420,000 total compensation per the company's published ranges — competitive with mid-senior levels at Microsoft and Amazon AWS AI, where Principal Applied Scientists top out around $520,000 and $380,000 respectively. What Cohere and Hugging Face cannot yet offer is Anthropic's or OpenAI's equity magnitude: RSUs at a $965 billion implied valuation (Anthropic's June 2026 confidential IPO filing), or membership in the model team that ships the product the entire industry benchmarks against. That gap will not close quickly. But the distributed labs are no longer farming the equity-agnostic segment alone — they are increasingly fielding offers to senior candidates who have weighed the remote premium against the SF density premium and decided.
The broader market context reinforces the divergence. Job postings mentioning AI terms surged more than 130 percent even as overall hiring weakened in early 2026, per Indeed Hiring Lab's January 2026 labor market update. That volume concentration in AI-specific roles — with everything else flat to negative — is intensifying competition for the senior AI engineer population that both blocs are pursuing. The BLS projects computer and IT occupations to add approximately 317,700 annual openings through 2034. But the senior AI research and engineering layer — the people Anthropic, OpenAI, and Hugging Face are all competing for — is not growing at BLS-projected rates. It is a thin market where location requirements function as real filters on the available pool.
What's Next
Three variables will determine whether the distributed bloc gains ground in H2 2026.
The first is the Cohere–Aleph Alpha integration. Merging a Toronto-headquartered workforce with a Heidelberg-based one under a $20 billion valuation creates a distributed org by structural necessity. How Gomez and the combined leadership team set compensation bands across North American and European time zones — whether they maintain geo-tiered structures or compress toward a unified remote band — will be one of the more closely watched comp policy decisions in enterprise AI this year.
The second is whether Anthropic's IPO changes the location-requirement calculus. Once Anthropic equity is liquid on a public exchange, the company can offer something no distributed-first lab can replicate: public-company RSUs at a frontier-lab valuation. The result may be that Anthropic's office-first requirement becomes easier to enforce, not harder — because the equity package makes the SF relocation math more tractable for senior candidates who currently choose remote-first employers partly because private-lab equity is illiquid and uncertain.
The third is Scale AI's full-time headcount trajectory following the competitive disruption from Surge AI, which sought up to $1 billion in its first external fundraise at a $25 billion valuation after capitalizing on customer uncertainty post-Meta's $14.3 billion investment in Scale. If Scale's full-time engineering headcount — the 1,000-person permanent core that manages the distributed contractor layer — begins pulling remote-eligible senior engineers to compete with Mercor and Surge, the RLHF talent market will see another comp round in what is already a two-tier market between $18 per hour general tasks and $60 per hour credentialed specialist work.
The map of US remote AI hiring in July 2026 is not a map of who has embraced flexibility. It is a map of who has decided that the talent they need is located everywhere — and who has decided that the talent they need is located in one zip code in San Francisco.
Global Career PlatformFind AI talent. Find your next role.
Booking is hotels. · Airbnb is apartments. · ENTRA is global careers.