The most distributed frontier AI research organization in the world today is not headquartered on a campus. It does not require its researchers to live within commuting distance of a central facility, and its ability to produce the research that defines the state of the art in alignment, interpretability, and model evaluation does not depend on any single geography. Google DeepMind employs researchers across 25 or more countries and every inhabited continent. Meta FAIR operates eight formal research nodes from Menlo Park to Tel Aviv to Montreal, with an additional 200-plus researchers on fully remote arrangements. Anthropic's research organization — fewer than five years old — spans 20 or more countries, with researchers concentrated in six time-zone clusters rather than physical offices. OpenAI, the frontier lab most associated with a physical campus identity, employs its largest cohort of researchers outside San Francisco in London, where its approximately 400-person presence following a major 2025 expansion represents a distinct research organization rather than a satellite function. The campus model — the organizing structure of every significant technology research institution from Bell Labs to Xerox PARC to Google Brain circa 2017 — is no longer the default mode of frontier AI research in 2026. What has replaced it, and what organizational infrastructure that replacement requires, is the subject of this report.
The shift from campus-centric to distributed research did not happen because remote work became normalized during the pandemic years. The pandemic accelerated hiring across geographies and lowered institutional resistance to asynchronous work, but the structural driver was different and more durable: the global distribution of the talent pool that frontier AI labs must compete to hire. ENTRA's Q2 2026 Talent Index estimates the global pool of qualified senior AI researchers — those with frontier-applicable skills in pretraining, RLHF, interpretability, or advanced evaluation — at approximately 38,000 individuals worldwide, a number that has not grown proportionally to frontier-lab demand. Against more than 12,000 active senior-research postings in Q2 2026 alone, the utilization rate of available specialized talent exceeds 85 percent. That constraint cannot be solved by building a better campus in San Francisco. It can only be solved by removing geography as a hiring filter.
The question this report addresses is not whether distributed research works at the frontier lab tier. The output record of Google DeepMind's distributed AlphaFold teams, Anthropic's distributed interpretability group, and Meta FAIR's distributed language-research pod in Paris is answer enough. The question is what organizational architecture makes it work. What specific structures, tools, rituals, and talent decisions separate the frontier labs that have built genuinely distributed research organizations from the enterprise AI teams that have attempted the same and failed? And what does the emergence of distributed frontier AI research at scale mean for the global AI talent map — for which cities become research nodes, which roles can be distributed and which cannot, and how the 23 percent remote premium that ENTRA's July 3 Global Remote AI Labor Report documented at the senior-IC tier is structurally produced and sustained by the distributed model?
The answer requires a lab-by-lab structural analysis, because there is no single model of distributed frontier AI research. Anthropic, OpenAI, Google DeepMind, Meta FAIR, Mistral, and xAI have each arrived at different organizational architectures reflecting their research priorities, founding cultures, compliance requirements, and hiring strategies. What they share is not a common structure but a common problem: how to coordinate research across time zones, sustain the synchronous-collision density that produces research insights, and maintain organizational coherence at 1,000-plus engineers when no single campus can serve as the institutional spine. The answers they have found are the subject of the lab profiles that follow.
The Campus That Wasn't Built
The campus model has a specific meaning in AI research history. When OpenAI launched in December 2015 as a nonprofit in San Francisco's Mission District, the physical co-location of its founding team was not incidental to the research program. It was a research philosophy. The belief, inherited from Bell Labs and reinforced by the Google Brain experience, was that frontier research required the accidental collision that only physical proximity could reliably produce: the hallway conversation that reshapes a research direction, the whiteboard session that crystallizes a month of confused thinking, the informal lunch that becomes the nucleus of a new paper. DeepMind, founded in London in 2010, was built with the same philosophy — a specific neighborhood in King's Cross, eventually Pancras Square, designed as a physical container for cognitive density. Anthropic, founded in 2021 by departing OpenAI researchers, initially followed the same logic: Hayes Valley in San Francisco, walkable to the major employer clusters and convenient for the founding team.
The campus model's disruption was not the pandemic. The structural break came from the intersection of two forces that became load-bearing simultaneously in 2022 and 2023. The first was the exhaustion of the addressable San Francisco Bay Area senior research talent pool. When the 2022 hiring acceleration pushed the top five frontier labs past a combined 3,000 senior researchers and engineers, the supply of qualified candidates within commuting distance of any specific San Francisco address had already been substantially consumed. Labs were either hiring remotely or they were not growing. The choice was empirical, not philosophical. The second structural driver was the emergence of deep AI research talent pools outside the US — in the UK, France, Germany, Israel, Canada, and the Gulf — that could not be accessed without offering either relocation packages large enough to compete with high-quality-of-life alternatives abroad, or remote-eligible employment that removed the relocation requirement entirely. Anthropic's decision to expand into London, Mistral's decision to structure itself as a European-distributed organization from founding, and Google DeepMind's decision to treat its Zurich and Paris pods as research nodes rather than satellite offices all reflect the same structural logic.
The campus also encountered a scientific constraint specific to AI research. Unlike biology or chemistry, where laboratory equipment creates a physical anchor, AI research requires compute and collaboration — both of which can be provisioned and sustained remotely. A researcher running pretraining experiments needs access to GPU clusters, not a desk in a specific building. A researcher working on interpretability needs access to model weights, inference infrastructure, and co-authors — none of which require physical proximity. The scientific substrate of AI research is more compatible with distributed execution than almost any other research-intensive discipline. That compatibility does not make distributed AI research costless; it makes the costs organizational rather than scientific, and organizations can be redesigned in ways that physics cannot.
How Anthropic Built a Distributed Research Org
Anthropic's distributed architecture is the most deliberately designed of the frontier labs — a product of an explicit founding-team decision, formalized in organizational policy in 2023, that building a campus-centric research organization would structurally limit the quality and resilience of the safety-research mission. Dario Amodei has articulated the underlying reasoning across multiple public forums: concentrating all frontier safety research in a single geographic location creates organizational brittleness that is specifically bad for safety. A lab whose entire interpretability team lives within five miles of each other has single-point-of-failure exposure to everything from natural disasters to local regulatory action to the cultural homogeneity that limits the intellectual diversity of research output.
The organizational structure Anthropic built reflects this reasoning. Rather than offices, Anthropic uses what its internal documentation describes as "research clusters" — groupings of researchers organized by time-zone overlap and research proximity rather than physical location. The Constitutional AI research cluster, which does foundational work on alignment techniques and model values, is distributed across the North American West Coast (primary), a European morning-overlap cohort (secondary), and a small remote-Pacific group. The Interpretability cluster — which ENTRA estimates at 120 to 140 researchers based on Q2 2026 job-posting analysis and recruiter network triangulation — is the most geographically dispersed of Anthropic's core research functions, with significant nodes in the UK and Canada. The Evals and Red Team cluster operates on the opposite model: deliberately concentrated in San Francisco, with an in-person cadence driven by the adversarial-testing methodology that requires rapid synchronous iteration.
Anthropic's compensation structure reinforces the distributed model at the talent level. The lab moved to location-agnostic compensation bands for all roles above L4-equivalent in Q3 2025 — a policy that ENTRA's July 3 Global Remote AI Labor Report identified as a key structural driver of the 23 percent remote premium across the frontier-lab tier. For research roles at L6-equivalent (Senior Research Scientist) and above, Anthropic's bands run $480K to $740K in total compensation regardless of where the researcher is located. For applied engineering at the same seniority, the band runs $360K to $540K. The bifurcation between research and applied tracks is standard across frontier labs; the absence of geographic adjustment within each track is distinctively Anthropic, and it is now effectively industry-standard for safety-oriented lab research roles following OpenAI's parallel policy move in Q4 2025.
ENTRA estimates that approximately 40 percent of Anthropic's research and engineering headcount — which ENTRA's H1 2026 AI Hiring Monitor places at 3,800 to 4,400 total employees — is located outside San Francisco. The UK is the largest non-US node by ENTRA's estimate, at approximately 380 to 420 researchers and engineers. Canada (Toronto and Montreal) constitutes a secondary cluster of an estimated 180 to 220. A third cluster spanning Berlin, Amsterdam, and Edinburgh accounts for approximately 140 to 160. Across ENTRA's analysis, Anthropic has documented researcher presence in at least 20 countries, with the remaining international cohort distributed primarily in Australia, Japan, Singapore, Israel, and across Western Europe via fully remote arrangements.
The internal tooling that sustains this distribution is centered on documentation culture. Anthropic's internal wiki, by researcher accounts corroborated through ENTRA's recruiter network, functions as the primary knowledge substrate — not a secondary artifact of meetings, but the first artifact of any research direction. The formalized async review process requires that research proposals above a defined scope threshold go through a written review cycle, with a 24-to-48-hour comment window, before any synchronous discussion is scheduled. The in-person synchronization mechanism is the quarterly "Research Intensive Week" (internally abbreviated as RIW): a five-day structured research sprint, held in person, in which the agenda is pre-published two weeks in advance, cross-team research reviews are conducted synchronously, and the whiteboard-collision density that a campus normally provides is compressed into a defined window. The model is costly — global travel for 1,500-plus researchers, four times a year, is a significant operating expense — but ENTRA's recruiter network sources consistently describe the RIW as the primary mechanism by which Anthropic maintains research coherence that pure-async operation cannot sustain.
OpenAI's Hybrid Architecture
OpenAI's distributed model is structurally different from Anthropic's and more internally contested. The lab has maintained a San Francisco-centric gravity throughout its growth — the Mission District original, the current 575 Market Street headquarters, and the research-intensive culture built around in-person collaboration remain defining features of OpenAI's institutional identity. But OpenAI now employs approximately 400 researchers and engineers in London, following a major expansion that began in 2024 and accelerated through 2025, plus documented presence in Dublin, Singapore, Tokyo, and Tel Aviv. The non-SF headcount, ENTRA estimates, constitutes approximately 30 to 35 percent of OpenAI's total research and engineering organization of roughly 2,800 to 3,200 positions. Across all geographies, OpenAI has documented researcher and engineering presence in at least 35 countries — the widest geographic footprint of any frontier lab in this analysis.
The model OpenAI has converged on is what ENTRA's recruiter network characterizes as the "presence anchor" model: a structure in which distributed researchers have full remote-working capacity for day-to-day execution but are expected to be in a physical OpenAI location — not necessarily San Francisco — for a minimum of four weeks per quarter. The presence anchor differs from a standard hybrid mandate in that it does not require a specific daily cadence of office attendance; it requires periodic physical immersion at defined intervals. The London office, at approximately 400 researchers and engineers, functions as the presence anchor for OpenAI's European distributed cohort. The Tokyo office, opened in early 2025, functions similarly for its Asia-Pacific cohort. The underlying logic is that four weeks of quarterly in-person time — roughly one week per month averaged — is sufficient to sustain the cross-team relationship density that drives both research collaboration and the informal sponsorship networks that influence promotion decisions at the senior-IC tier.
The internal tension in OpenAI's distributed model is a bifurcation between safety research and systems research. Safety research — the Preparedness team, alignment evaluation, the interpretability work that runs in parallel with Anthropic's — is substantially compatible with distributed execution. The work is primarily cognitive and document-producing; collaboration dependencies are manageable across time zones with well-designed async review processes. Systems research — hardware-software co-design, distributed training optimization, inference latency reduction — is materially more location-sensitive. The engineers who optimize OpenAI's training infrastructure need rapid-fire synchronous collaboration with the compute operations team, the monitoring function, and the researchers simultaneously running experiments on shared infrastructure. Time-zone fragmentation in that collaborative environment is not a minor friction; it is a capability constraint that compounds with each additional hour of timezone spread.
OpenAI has responded to this bifurcation by maintaining a significantly higher in-person expectation for systems research roles — an effective three-day-per-week presence requirement for SF-based systems engineers — while extending greater flexibility to safety research and applied product engineering. ENTRA's Q2 2026 analysis of OpenAI job postings shows the split explicitly: systems and infrastructure roles listed "hybrid, San Francisco" in 73 percent of cases; safety research and alignment roles listed "remote-eligible" in 61 percent of cases. The taxonomy is not explicit in OpenAI's public communications, but it is legible in the posting data and consistent with what ENTRA's recruiter network sources describe as OpenAI's internal role-categorization framework.
Google DeepMind's Multi-Campus Model
Google DeepMind's distributed architecture predates the current wave of frontier-lab expansion by nearly a decade. DeepMind opened its Montreal office in 2017, its Edmonton and Paris offices in 2018, and its New York office in 2019 — all before the pandemic, all as deliberate research-diversification strategies rather than responses to remote-work norms. What has changed in the 2024-to-2026 period is the scale and organizational weight of the non-London nodes relative to the King's Cross anchor. The Zurich research pod, which ENTRA estimates at approximately 250 researchers as of Q2 2026, has grown into the lab's primary hub for hardware-adjacent AI research and inference optimization work. The Paris pod, which ENTRA estimates at approximately 160 to 180 researchers, functions as the anchor for DeepMind's Gemini EU research contributions and European regulatory engagement. The Singapore pod, opened in 2024, serves the Asia-Pacific research and safety function. Across all locations, ENTRA estimates DeepMind has documented researcher presence in at least 25 countries.
DeepMind's organizational model for managing this distribution is the "lead researcher" framework. Research programs at DeepMind are owned by a named lead researcher rather than by a team in a specific location. The AlphaFold program was structurally a London program in its early years, but its expansion to cover protein interactions required researchers across multiple time zones; the lead-researcher frame — Demis Hassabis's long-standing organizational principle that research accountability should follow individuals rather than locations — allowed that expansion without requiring geographic consolidation. The same model has been applied to the Gemini research program: Gemini has research contributors in London, Zurich, Mountain View, New York, and Paris, coordinated through a lead-researcher structure that assigns ownership of each research direction to a specific individual who manages cross-timezone collaboration. The AlphaFold model is now the institutional template: programs originate with a lead researcher's vision, execute across distributed contributor nodes, and are integrated through the lead researcher's coordination authority rather than through geographic co-location.
The practical limitation of DeepMind's multi-campus model is its persistent London-centricity at the senior-research level. ENTRA estimates that approximately 70 percent of DeepMind's Distinguished Researcher and Research Fellow-level headcount — its most senior research tier — remains located in London. The lead-researcher model works well for managing cross-timezone execution of defined research programs. It works less well for the informal senior-research conversations that shape research direction before formal programs exist. The probability that two senior DeepMind researchers encounter each other at the Pancras Square espresso bar and develop a research direction that would not have emerged through formal channels remains higher in London than between any two non-London nodes. The organizational challenge DeepMind faces in H2 2026 is whether its non-London pods have reached the critical mass at which they generate their own informal-collision density, or whether they remain execution hubs for programs whose intellectual genesis happens in King's Cross.
Meta FAIR: The Research Lab That Never Required a City
Meta FAIR's distributed architecture is the most instructive in the frontier-lab set because it predates every other lab's distributed experiment by years. When Yann LeCun established FAIR in 2013, he structured it from the first day with simultaneous nodes in Menlo Park, New York, and Paris — reflecting both his own split residence between New York and Paris and his foundational research philosophy that fundamental AI research does not require geographic concentration. LeCun's position, stated consistently over more than a decade of interviews and public communications, is that research quality depends on the quality of the researchers and the quality of the research environment — not on the proximity of the researchers to each other. The corollary — that the best researchers should be hired where they are rather than required to relocate — shaped FAIR's entire talent strategy from the start.
That founding model scaled into FAIR's current structure: eight-plus offices globally, with formal research nodes in Menlo Park (systems and foundation model research), Paris (language AI and multilingual research; ENTRA estimates approximately 180 researchers), New York (computer vision and robotics; ENTRA estimates approximately 210 researchers), London (AI robustness and safety research), Tel Aviv (video understanding and generative AI), Montreal (foundational ML research in collaboration with MILA), and Seattle and Pittsburgh as smaller applied-research pods. The full FAIR research population ENTRA estimates at 1,100 to 1,300 researchers — making it the largest single distributed AI research organization in this analysis by headcount, and the one with the longest operational history at scale.
The research-node model is FAIR's most portable contribution to distributed AI research design. The principle — that each geographic node should have a defined research specialty it owns globally rather than trying to replicate the full lab in miniature — is what differentiates FAIR's model from the "offices everywhere" approach that produces organizational fragmentation without research coherence. FAIR Paris is not a generalist FAIR outpost that executes whatever Menlo Park prioritizes. It is the global center of FAIR's multilingual and language-model research, with a specialty deep enough that the node's researchers are world-recognized contributors in that domain and maintain a research identity that does not depend on what the other nodes are doing in real time. FAIR New York's computer vision identity is equally defined. That specialization is what makes the distributed model research-coherent rather than simply geographically convenient.
The LeCun lab model — the principle that a lead scientist of sufficient prestige and intellectual influence can sustain a distributed research organization through the force of research direction rather than physical co-location — has been both FAIR's greatest organizational strength and, since LeCun's departure in November 2025 following the restructuring triggered by Meta's acquisition of a 49 percent stake in Scale AI, its greatest organizational challenge. LeCun's ability to hold together a globally distributed research organization depended in part on his intellectual authority as a field founder — an authority that does not transfer mechanically to successors and cannot be replicated through organizational charts. FAIR's current challenge under its post-LeCun leadership structure is reconstructing the institutional glue that LeCun provided personally, without the benefit of a physical campus to serve as a substitute.
Mistral's "European Distributed" Model
Mistral AI is five nodes and no campus. The Paris headquarters — where Arthur Mensch, Guillaume Lample, and Timothée Lacroix founded the company in April 2023 — is a legal address and an organizational anchor, but it is not the geographic attractor that Paris-based technology companies have historically relied on to build teams. Mensch has been explicit, in interviews with Le Monde, MIT Technology Review, and ENTRA's Q1 2026 briefing request, about his decision to build Mistral as a non-Paris-locked organization: "We are a Paris company. We are not a Paris-only company. The researchers we want are in Edinburgh, in Berlin, in Amsterdam, in London. We go to them."
The organizational structure is a five-node architecture: Paris (headquarters; foundation model research and pretraining), London (safety research and applied engineering; ENTRA estimates approximately 140 to 160 researchers), Berlin (infrastructure engineering and European enterprise deployment), Edinburgh (academic research collaboration, serving as the primary channel to the University of Edinburgh's language AI research program, one of Europe's deepest), and Amsterdam (EU regulatory compliance and European enterprise go-to-market). Mistral's fully remote cohort — researchers who work for the company but are not affiliated with a specific node — ENTRA estimates at 80 to 120 globally. Across nodes and remote, Mistral has documented AI research and engineering presence in at least 12 countries, concentrated in Europe and the UK.
Mistral's compensation structure is the clearest example in this report of a non-US frontier lab offering competitive compensation without US-equivalent absolute numbers. Senior researchers at the L5-to-L6 equivalent seniority — five or more years of experience, primary ownership of a research direction or engineering system — earn €200K to €310K in total compensation, per ENTRA's EU Bureau sourcing and Levels.fyi submissions. After French income tax and social charges, take-home is materially lower than US frontier-lab equivalents. Mistral's talent retention proposition rests on three non-compensation elements: the European AI sovereignty thesis (the argument that building European-sovereign AI capability is a mission distinct from and separately valuable relative to the US frontier labs), the quality-of-life proposition of Paris, Edinburgh, and Berlin relative to San Francisco for researchers with European roots or family ties, and the equity structure of a company at an approximately €11.7B valuation (per the September 2025 Series C; Bloomberg reported discussions at a €20B valuation in June 2026) with credible liquidity optionality. That proposition is not universal — it does not compete for the subset of senior researchers who are primarily motivated by maximizing US-equivalent take-home. But it is credible for a specific and consequential segment of European AI research talent that it competes for effectively.
The Culture Infrastructure Required
Distributed frontier AI research does not work by accident. Every lab in this analysis that has built a functional distributed research organization has also built a specific set of organizational systems that do not emerge spontaneously from putting talented researchers in different time zones and giving them Slack access. The systems cluster into four categories.
Documentation culture. All knowledge must be in writing before it enters a room. This principle sounds obvious and is widely violated. At Anthropic, the internal wiki is the primary knowledge substrate — not a secondary artifact of meetings, but the first artifact of any research direction. Research proposals go through a written review process before any synchronous discussion is scheduled: the lead researcher writes a structured proposal, relevant reviewers have a defined comment window (Anthropic's standard is 24 to 48 hours, depending on scope), and the synchronous discussion — if one is called — happens after, not instead of, the written review. This is a significant cultural inversion from the campus model, where the meeting is the knowledge-creation event and the document is the record of the meeting. In a distributed lab, the document is the knowledge-creation event and the meeting is the document's synthesis. The inversion is difficult to execute because it requires researchers who are accustomed to thinking in conversation to learn to think in writing first — a cognitive-mode shift that compounds the organizational challenge of distribution with a personal-practice challenge for individual researchers trained in campus environments.
Async review architecture. Research review — the process by which a research direction is evaluated, critiqued, and either approved or redirected — is the most synchronization-intensive function in a research organization. Frontier labs have developed specific mechanisms for conducting it asynchronously. The "reviewer token" model, which sources inside Anthropic's research organization describe and which appears in variant forms at OpenAI and DeepMind, assigns explicit review responsibility to named individuals with a defined response deadline. The token creates accountability without requiring a scheduled synchronous meeting. Decisions accumulate asynchronously through the reviewer token chain; a synchronous meeting is called only when written review has reached an impasse that cannot be resolved through the written medium. The model requires that reviewers treat the written review window as a hard commitment rather than a soft suggestion — a cultural norm that labs enforce through social expectation rather than technical enforcement, and that breaks down when meeting culture reasserts itself under deadline pressure.
In-person synchronization cadence. Fully async research is not sustainable at the frontier. Every lab in this analysis runs a structured in-person synchronization cadence: Anthropic's quarterly Research Intensive Weeks, DeepMind's semi-annual "Frontier Weeks" at the London campus with video-linked international participation, Meta FAIR's node-level monthly in-person days, and OpenAI's presence-anchor requirement of four weeks per quarter. The common design principle across all of these is that in-person time is structured research time with a specific agenda, not social time that happens to include researchers in the same room. The labs that run in-person gatherings as offsites — with team-building exercises and social dinners as the primary content — report lower research-coordination benefit than those that treat the in-person window as a compressed research sprint. Research Intensive Weeks at Anthropic are pre-programmed with research reviews, cross-team design sessions, and structured whiteboard time. The social element is real but secondary; the primary output is research decisions that the async channel was not resolving.
Deep work vs. meeting culture. Distributed research organizations tend toward over-meeting as a compensation mechanism for the physical proximity they lack — the instinct is to schedule more calls in place of the hallway conversations that no longer happen. The labs in this analysis that have avoided that failure mode have done so through explicit meeting constraints. Anthropic's default is no standing meetings for researchers except weekly research reviews with their direct cluster. OpenAI's distributed research track runs a "no meeting Wednesday" norm maintained since 2024. DeepMind's distributed nodes operate on a "core hours" model that defines three to four hours of guaranteed synchronous availability per day and explicitly protects the remaining hours for uninterrupted deep work. The constraint is not aesthetic; it is a recognition that the multi-hour uninterrupted concentration required to do research work cannot be scheduled around meeting blocks, and that distributed researchers without a physical workspace enforcing those boundaries need institutional norms to enforce them in their place.
The Talent Architecture
Distributing frontier AI research does not treat all roles equivalently. The talent architecture of a distributed lab is defined by which roles can be distributed without research-capability loss and which carry a location dependency that meaningfully constrains output quality.
The distributed-compatible roles are the majority. Interpretability research — the analytical work of understanding what large language models are doing internally, generating the publication streams from Anthropic's interpretability team and DeepMind's mechanistic interpretability group — is structurally suited to distributed execution. The work is primarily analytical, the outputs are documents and experimental results, and the collaboration dependencies are manageable over well-designed async channels. Alignment research, theoretical ML research, and evaluation framework design share this profile. ENTRA's Q2 2026 job-posting analysis shows that 78 percent of posted interpretability and alignment roles at the five frontier labs in this analysis are explicitly remote-eligible or location-agnostic. The roles that produce the field's foundational safety-relevant research are, paradoxically, the roles most amenable to the organizational model that makes their researchers most globally accessible.
The location-sensitive roles are specific and predictable. Systems research — training pipeline optimization, hardware-software co-design, distributed training infrastructure engineering — requires the rapid-fire synchronous collaboration that time-zone fragmentation penalizes severely. ENTRA's analysis of OpenAI and Anthropic systems-research job postings in Q2 2026 shows that 84 percent explicitly require or strongly prefer San Francisco or a specified hub location. Hardware-model co-design — the emerging research function requiring tight coordination between model researchers and chip teams at NVIDIA, Cerebras, and the labs' own ASIC programs — is even more location-sensitive, because the hardware engineers are concentrated in specific facilities and the co-design loop requires co-location, or at minimum same-timezone presence, to sustain the iteration velocity that produces results. The physical experiment cannot be distributed the way the analytical experiment can.
The global 38,000-person constraint on specialized AI talent, documented in ENTRA's July 3 Global Remote AI Labor Report, applies differently across these two role categories. Interpretability and alignment talent is globally distributed and benefits maximally from remote-eligible policy: the researcher in Edinburgh who would not relocate to San Francisco but can work for Anthropic's interpretability cluster in UK morning overlap hours represents a genuine supply expansion. Systems research talent is more geographically concentrated — disproportionately in the Bay Area, Seattle, and London, in proximity to the compute infrastructure and hardware partners that systems research requires. Remote-eligible policy for systems research does not expand supply as meaningfully as it does for alignment research, because the constraint on systems-research talent is not geographic willingness but specific capability depth that correlates with prior proximity to frontier-scale compute.
The compensation implication of distributing the distributed-compatible roles while concentrating the location-sensitive ones is a geographic pay structure more complex than a simple location-agnostic model. Anthropic and OpenAI have both moved to location-agnostic bands for distributed-compatible research roles — the L6 research band at Anthropic ($480K to $740K total compensation) and the equivalent band at OpenAI ($520K to $810K for senior research scientists) do not vary by geography. But both labs maintain implicit and occasionally explicit geographic concentrations for systems research that create de facto hub premiums for SF-based systems engineers competing in a tighter local talent market.
The visa and employer-of-record infrastructure required to sustain distributed frontier AI research at scale is substantial and often underreported as an operational cost. Every lab in this analysis uses EOR arrangements for researchers in markets where the lab has no legal entity; UK Global Talent visa sponsorship for UK-based senior hires; EU Blue Card processes for continental European hires; and immigration counsel embedded in recruiting pipelines for senior roles requiring work authorization processing. Anthropic's immigration operations budget, per ENTRA's recruiter-network estimate, runs approximately $8M to $12M annually — a figure reflecting both direct visa and EOR infrastructure costs and the opportunity cost of recruiting timelines extended by immigration processing. The infrastructure cost is real. The alternative — restricting hiring to the US labor market — forecloses access to the majority of the 38,000-person global senior-researcher pool.
What Non-Frontier Companies Get Wrong
The frontier lab distributed model is widely emulated and rarely replicated. The failure modes are specific and recurring enough to constitute a documented pattern across the enterprise AI teams ENTRA has tracked through its H1 2026 recruiter network.
The "fully remote without async culture" failure. The most common failure mode is adopting the physical distribution of the frontier lab model without the documentation culture that makes it functional. An enterprise AI team that declares itself fully remote but continues to treat meetings as the primary knowledge-creation venue has not built a distributed research organization. It has built a meeting-intensive organization whose participants happen to be in different cities, with the accumulated overhead of time-zone coordination and none of the research-coherence benefits of the distributed model. ENTRA's Q2 2026 survey of 890 AI professionals included 140 respondents who had worked on enterprise AI teams that had attempted to adopt distributed research models; 68 percent of that cohort cited "meetings replacing documentation" as the primary failure mode. The organizational fix — inverting the default so that documents precede meetings rather than follow them — is simple in description and deeply difficult in cultural execution, because it requires researchers trained in campus environments to develop new cognitive habits rather than simply changing where they sit.
The "timezone spread too wide" failure. Research teams that span 12 or more hours of timezone coverage cannot maintain the async review cycle that distributed research requires. When reviewer A is in San Francisco and reviewer B is in Singapore, the minimum latency on a written exchange is 24 hours and the realistic latency, accounting for work patterns, is closer to 36 hours. For decisions that require three rounds of review — a common threshold for significant research direction changes — the minimum elapsed time before consensus is three full days. That is not a blocker for individual research execution, but it is a blocker for the rapid-iteration feedback loops that safety evaluation and red-teaming require. The frontier labs have responded by managing timezone distribution at the team level rather than the organization level: Anthropic's research clusters are structured specifically to maintain a maximum 8-hour timezone span within a working cluster. ENTRA's analysis of failed enterprise distributed AI team deployments, attributable by recruiter network sources to unnamed financial-services and healthcare enterprises, includes at least three documented cases where teams spanning Pacific to APAC time zones without this constraint encountered research-cadence breakdown within six months of going fully distributed.
The "distributed without node specialization" failure. Enterprise AI teams attempting to replicate FAIR's research-node model often build geographic nodes without the specialization that makes the FAIR model functional. A distributed enterprise AI team with offices in five cities, each doing a mix of the same generalist AI engineering work, is not a research-node organization. It is a fragmented organization with the coordination overhead of distribution and none of the research-identity benefits of specialization. The research-node model requires each node to own a specific research domain globally — FAIR Paris's multilingual research, FAIR New York's computer vision — with depth sufficient that the node's researchers are the world-recognized authorities in that specialty. That depth of specialization is achievable for frontier labs with 1,000-plus researchers; it is aspirational for enterprise AI teams with 50 to 200 engineers. The teams that have attempted node specialization without the headcount to sustain it have typically found that nodes converge back to generalist execution under resource and deadline pressure, leaving behind the geographical fragmentation without the research specialization that justified it.
Methodology
ENTRA Distributed Lab Analysis — Q2 2026
Primary data sources: ENTRA Q2 2026 Talent Index covering 320,000-plus AI-related job postings globally across 1,400-plus employers, analyzed Q1 through Q2 2026. Frontier lab headcount and geographic distribution estimates are derived from LinkedIn Talent Insights employer-location mismatch signals (comparing declared employer to declared work location for senior IC profiles), company career portal and job-posting geotag analysis, arXiv submission author-affiliation tracking for research-scientist headcount in specific geographies, and ENTRA recruiter network triangulation (38 senior AI recruiters active across Anthropic, OpenAI, Google DeepMind, Meta FAIR, and Mistral accounts, interviewed in structured sessions during Q1-Q2 2026). LinkedIn senior-IC movement data covers departures and arrivals at tracked labs from January 2025 through June 2026.
Coverage: Anthropic, OpenAI, Google DeepMind, Meta FAIR, Mistral, xAI, Hugging Face — distributed footprint data reviewed for all seven organizations. Company-specific headcount figures outside headquarters are ENTRA estimates unless noted as confirmed through public disclosure or company announcement. Headcount estimates carry an estimated margin of plus or minus 15 percent, reflecting the structural opacity of EOR-employed and fully remote headcount in standard data sources.
Methodology for headcount-outside-HQ estimates: LinkedIn employer-location mismatch signal; company announcements and press releases on office openings and hiring programs; ENTRA recruiter network triangulation; arXiv author-affiliation tracking for research headcount in specific geographies; and cross-reference against LinkedIn Talent Insights hiring velocity signals by location.
Compensation data: Levels.fyi submissions (n=14,200 AI professionals, full 2025 through Q2 2026 window); ENTRA H1 2026 recruiter survey (n=38 senior AI recruiters); direct employer disclosure where available. Total compensation defined as base salary plus annual cash bonus plus annualized equity on a four-year vest schedule, in USD at Q2 2026 exchange rates.
Organizational and culture data: Internal tool names and organizational structures described in this report — including "Research Intensive Weeks," "research clusters," "reviewer token model," "presence anchor model," and "core hours" frameworks — are derived from ENTRA's recruiter network and researcher source network. They have not been independently confirmed through official company communication and are represented as ENTRA intelligence rather than confirmed company policy. Where these descriptions conflict with any future official disclosure, official disclosure supersedes.
Limitations: Distributed headcount data for xAI is significantly less complete than for other labs, due to xAI's limited public disclosure and lower LinkedIn profile density. Mistral's fully remote headcount is estimated with lower confidence than its node headcount due to the opacity of EOR arrangements in the Levels.fyi and LinkedIn data. Google DeepMind's internal research-node specialization descriptions reflect ENTRA analysis of publication patterns and recruiter intelligence rather than official organizational designations.
What's Next
Three forward signals define the trajectory of distributed frontier AI research in H2 2026 and into 2027.
The founding-architecture signal. The next cohort of frontier labs — Inception AI (the European research venture that has attracted several former Mistral and Google DeepMind researchers), H Company (the Paris-headquartered foundation model lab co-founded by former Google Brain and FAIR researchers), and SandboxAQ (the Alphabet-spinout quantum and AI research company operating across multiple geographies) — are making distributed-versus-campus organizational decisions at founding, not as growth-phase adaptations. H Company's early organizational signals are instructive: the lab launched with a deliberate two-node structure — Paris for foundational model research, London for safety and alignment — with explicit node specialization articulated before the company had its second dozen employees. That founding decision reflects a thesis that the distributed model is not a concession to talent geography but a first-principle organizational choice with better research outcomes than campus concentration. If the pattern extends across the next five to ten labs that are founded in H2 2026 and 2027, the distributed-by-design lab will be the organizational default for a new generation of frontier research organizations, rather than the distributed-by-necessity adaptation it has been for the current generation.
The regulatory concentration signal. EU AI Act enforcement and US AI security frameworks carry opposing implications for distributed research architecture. The EU AI Act's compliance requirements for general-purpose AI systems with systemic risk designations create strong incentives for EU-headquartered labs to maintain legal presence and governance accountability within EU jurisdictions — which pushes toward European geographic concentration of the research and documentation functions that generate GPAI compliance artifacts. The extended compliance deadline established by the European Council's Digital Omnibus agreement of May 7, 2026 — pushing the Annex III deadline from August 2026 to December 2027 — lengthens the runway without changing the destination. The US security framework, particularly the enforcement pattern of EO 14110 provisions through 2025 and 2026, creates incentives for US labs to concentrate frontier safety research in US-regulated locations to maintain control over the research output and the researchers who produce it. If both trends intensify simultaneously, the result could be a bifurcation of distributed research architecture: applied engineering and deployed-model monitoring remaining globally distributed, while safety research and frontier capability evaluation re-concentrate in specific regulatory jurisdictions. The geographic AI research map would then reflect regulatory geography rather than talent geography — a significant structural shift from the talent-driven distribution patterns documented in this report, and one that would change the talent implications of distributed research in ways that are not yet legible in the Q2 2026 data.
The physical infrastructure question. The frontier labs have to date built distributed research on commercial real estate infrastructure — leased offices in established urban markets, co-working spaces for smaller nodes, and employees' home offices for the fully remote cohort. The question that no lab has yet answered publicly is whether distributed research at 5,000 or 10,000 engineers across 50-plus countries requires purpose-built physical infrastructure: compute-proximate research facilities, purpose-designed collaboration spaces engineered around the mixed in-person and remote participation dynamics of distributed research sessions, or physical nodes designed around the specific security requirements of frontier safety evaluation work. The economics of purpose-built distributed research infrastructure — which would require CapEx commitments that frontier labs have historically avoided through leasing strategies — are not clearly favorable at current scale. Commercial real estate in the cities where frontier labs run nodes (London, Zurich, Paris, Singapore) is available and flexible enough to approximate the requirements. But as the distributed model matures and its organizational requirements become more precisely understood, the gap between what frontier AI research needs from a physical environment and what commercial real estate provides is likely to widen. The lab that builds the first purpose-designed distributed research node — a facility engineered from the ground up around high-speed compute access, synchronous-collaboration infrastructure designed for hybrid room-and-remote participation, and physical isolation for sensitive safety-evaluation work — will have answered the physical infrastructure question in a form that the rest of the industry will eventually be required to address. That lab has not yet announced its construction program. The conditions for it to exist are accumulating.
Global Career PlatformFind AI talent. Find your next role.
Booking is hotels. · Airbnb is apartments. · ENTRA is global careers.