Two senior ML engineers. Same seniority band. Same skill set: production-grade LLM fine-tuning, RLHF pipeline experience, distributed training at scale. One lives in San Francisco. One lives in Warsaw. Under a location-indexed Big Tech AI pay policy, the San Francisco engineer earns $280,000 in base salary. The Warsaw engineer earns approximately $154,000 — 55 percent of the SF rate, reflecting the local cost-of-living adjustment that Google, Meta, and Microsoft apply to employees outside their tier-one markets. The employer saves $126,000 per year on that single Warsaw hire. But a competing employer — one running a market-rate remote policy — offers the same Warsaw engineer $280,000, identical to the San Francisco rate. The Warsaw engineer leaves.
That scenario, repeated across the 16 cities ENTRA has covered in The Remote Issue this month, defines the central tension in AI talent architecture entering H2 2026. The remote AI compensation market is not a single market. It is three distinct markets operating simultaneously, governed by three incompatible philosophies about what remote pay is for — and the employer that picks the wrong model for its talent-acquisition context is either leaving $126,000 per engineer per year in savings unrealized, or bleeding its best distributed hires to competitors who price globally.
This report synthesizes what 15 weeks of city-level coverage across four continents has revealed about how AI companies decide what to pay engineers who live in Warsaw versus San Francisco versus Nairobi — and maps where the market is heading in the six months that remain in 2026.
Section 1: The Three Models
The remote AI compensation market in H2 2026 has fragmented into three structurally distinct approaches. Calling them "models" is not imprecise — each reflects a coherent internal logic about what compensation is doing in the organization.
Model A: Location-Indexed Pay
The incumbent model at Google, Meta, and Microsoft AI deploys a cost-of-living index to adjust nominal salaries by geography. A senior ML engineer at Google DeepMind at L5 earns a San Francisco base of approximately $275,000–$305,000 (per Levels.fyi Q2 2026). The same engineer, working remotely from Warsaw, earns approximately $151,000–$168,000 — 55 percent of the SF rate on gross nominal terms. From the employer's perspective, the logic is purchasing-power equivalence: the Warsaw engineer's $155,000 buys the same lifestyle, and then some, as the San Francisco engineer's $280,000. The employer captures the geographic arbitrage. The engineer, in PPP-adjusted terms, is roughly whole.
Location-indexed pay made sense as a permanent model when AI engineering talent was abundant enough that engineers had limited outside options. That condition no longer holds at the senior level. Anthropic briefly explored a location-indexed framework for distributed roles in early 2025, per ENTRA recruiter network sources, before concluding that the model was structurally incompatible with the senior-research talent competition it was engaged in. The company rolled back geographic adjustment for roles above L4-equivalent in Q3 2025. The lesson was not that location-indexed pay is wrong in principle; it is that it is wrong for the specific talent segment frontier labs compete for.
Model B: Market-Rate Remote (Location-Independent)
The insurgent model — adopted by Anthropic for research roles above L4, extended by OpenAI to its "distributed research" track in Q4 2025 — eliminates geographic adjustment entirely. Total compensation is tied to the role, the level, and the employer's global pay philosophy. A Senior Research Scientist at Anthropic earns $480,000–$740,000 in total compensation regardless of whether they live in San Francisco, London, or Warsaw (per ENTRA recruiter network and Levels.fyi Q2 2026). A Senior ML Engineer on the applied track at the same firm earns $360,000–$540,000 on the same basis.
Model B is not a cost-reduction strategy. It is a talent-acquisition investment. The employer pays the same total compensation per head whether that head is in a tier-one city or a sub-Saharan capital. What the employer gains is access to a global talent pool unrestricted by geographic willingness to relocate — including the 60-plus percent of qualified senior researchers who, per ENTRA's Q2 2026 Remote AI Work Survey (n=890), would not relocate internationally for a role they could access remotely. The cost structure of Model B is identical to running a San Francisco team; the talent-supply structure is categorically larger.
Model C: Zone-Based Bands
The emerging middle ground structures compensation into three to five geographic tiers, each with its own pay band. Tier 1 covers San Francisco, New York, London, and Zurich — the markets where AI employers face direct in-person competition from co-located firms. Tier 2 covers Berlin, Warsaw, Tel Aviv, Amsterdam, and Toronto — markets with deep AI talent density but a 30–50 percent nominal cost advantage versus tier-one cities. Tier 3 covers Nairobi, Medellín, Tbilisi, and Casablanca — markets where top-tier talent is available at 60–80 percent below tier-one rates on a nominal basis.
Hugging Face and Weights & Biases have both operated on zone-based architectures since 2024, per ENTRA's employer-network intelligence (n=12 companies, anonymized). The model is gaining traction because it threads the needle: it preserves meaningful cost arbitrage versus full market-rate remote (Model B) while avoiding the talent-defection risk of pure location-indexing (Model A). The internal equity argument for Zone-Based pay is stronger than for strict location-indexing: two engineers in the same zone receive the same treatment, rather than every individual city or postcode generating a different modifier. The zone becomes the unit of pay equity, not the address.
The friction in Model C is zone definition and zone maintenance. Berlin and Warsaw are both Tier 2 under most frameworks, but they are not economically equivalent: a Warsaw B2B contractor with the IP Box tax structure nets significantly more on the same gross than a Berlin engineer facing marginal German income tax. Employers that adopt zone-based pay often find that the zones require annual recalibration as local market rates move, adding HR infrastructure overhead that neither Model A (index-linked) nor Model B (flat rate) requires.
Section 2: The Data from 15 Cities
Fifteen weeks of city coverage across The Remote Issue has produced the most granular cross-market AI compensation dataset ENTRA has assembled. What follows is the picture that dataset paints — market by market, model by model.
San Francisco (Tier 1 anchor)
The SF senior ML engineer remains the global compensation benchmark. At frontier labs — Anthropic, OpenAI, xAI — total compensation for L5/L6 equivalent Senior ML Engineers runs $340,000–$560,000 (per Levels.fyi Q2 2026). The range reflects a bifurcation within the frontier tier: research-track roles (Research Scientist, Research Engineer) sit at the top of the band; applied-track roles (ML Engineer, Applied Scientist) cluster toward the lower end. At Big Tech AI divisions — Google DeepMind's US operations, Microsoft AI, Meta GenAI — the equivalent band runs $260,000–$380,000 (per Levels.fyi Q2 2026). The gap between frontier labs and Big Tech at the senior IC level is now $80,000–$180,000 in annualized total compensation — wide enough to constitute a structural market segmentation, not a band overlap.
Warsaw (Tier 2 — Central and Eastern Europe)
Warsaw is the report's pivotal data point because it sits at the intersection of all three compensation models simultaneously. Under Model A (location-indexed), a Warsaw-based Senior ML Engineer at a Big Tech AI division earns approximately $143,000–$209,000 (approximately 55 percent of the equivalent SF rate, per ENTRA estimate based on published location modifier guidance and Levels.fyi Q2 2026 submissions). Under Model C (zone-based, Tier 2), the same engineer earns from roughly $180,000–$260,000 at employers whose Tier 2 bands track Western European market rates. Under Model B (market-rate remote), the engineer earns $340,000–$560,000 from a frontier lab — identical to the San Francisco rate.
Warsaw's local B2B contractor market adds a fourth reference point: senior AI/ML engineers on B2B contracts with European and US employers earned USD 70,000–130,000 annually in H1 2026 (per ENTRA Warsaw briefing, July 7, 2026, sourcing ITMAGINATION and Freenance Tech Salaries Poland 2026), reflecting rates set by US and UK employers who are applying an implicit location discount below even the formal Model A bands. This is the informal gray zone beneath Model A — employers who are not running a formal geographic adjustment framework but are simply paying below their domestic rates because Warsaw engineers accept it in exchange for the premium over local Polish enterprise alternatives (approximately PLN 336,000–541,000 annually at Allegro, DocPlanner, and equivalents; converted from PLN at 1 PLN = €0.234, Q2 2026, per ENTRA Warsaw briefing).
Tallinn (Tier 2 — Northern Europe/Baltic)
Senior AI engineers in Tallinn earn €65,000–€95,000 in base salary (approximately $71,200–$104,100 at ECB Q2 2026 rates), per ENTRA Tallinn briefing, July 16, 2026. Under a zone-based system, Tallinn would sit in Tier 2 alongside Berlin and Warsaw, but the nominal band is 20–30 percent below Berlin on gross terms. After Estonia's flat 20 percent income tax plus approximately 2.8 percent employee-side contributions, the net-of-tax position is favorable relative to German or French equivalents at the same gross. For a UK or US employer running Model C, Tallinn Tier 2 engineers represent the most cost-efficient access point in the EU after the Baltic and Balkan markets: EU legal perimeter, English-language operational norms, and a 3-hour UTC offset that fits European and early US East Coast windows.
Tel Aviv (Tier 2 — High end)
Israel sits at the expensive end of the Tier 2 category and in some frameworks bleeds into Tier 1. Senior ML engineers at Israeli AI companies — run.ai, Deci (acquired by NVIDIA), and the significant AI functions inside Israeli defense-tech companies — earn NIS 680,000–NIS 1,040,000 in annual total compensation (approximately $184,000–$281,000 at the ENTRA standard rate of 1 NIS = 0.27 USD). The Unit 8200 alumni cohort commands the upper end of that range or prices above it entirely: these engineers apply at market-rate and do not accept location-indexed offers. For employers running Model A, Tel Aviv is the market where the model breaks most visibly — the talent competes globally and prices accordingly, irrespective of what any local cost-of-living index suggests.
Nairobi (Tier 3 — East Africa)
Nairobi senior ML engineers working on Gulf remote contracts earn $60,000–$90,000 annually (per ENTRA Nairobi briefing, July 14, 2026). Against the local KES domestic market benchmark of $18,000–$35,000 at equivalent seniority (ENTRA East Africa AI Salary Index H1 2026), the remote premium is 2.6x–5x. For employers, Nairobi represents Model C Tier 3: world-class ML capability — validated by Zindi competition performance and Andela placement quality — at 15–25 percent of frontier-lab SF cost. Under Model B (market-rate remote), Nairobi engineers would receive $340,000–$560,000 for frontier-lab work; no frontier lab is currently offering this to Nairobi-based engineers at scale. The gap between what the talent can deliver and what the market rate is paying them is the most significant unresolved tension in the global AI compensation map.
Lagos (Tier 3 — West Africa)
Lagos senior ML engineers on Gulf remote contracts earn $35,000–$55,000 annually (per ENTRA Lagos briefing, July 16, 2026), representing a 65–70 percent cost reduction versus an equivalent Abu Dhabi on-site hire (ENTRA GCC AI Salary Index H1 2026). The UTC+1 timezone provides a six-hour sync window with UAE business hours — the widest functional collaboration window available from any sub-Saharan market. For Gulf employers under Model A or Model C, Lagos represents the bottom of the global AI pay structure: capable talent, documented through the Andela pipeline, at a compensation level that no Western employer can replicate.
Medellín (Tier 3 — Latin America)
Senior AI engineers in Medellín earn $38,000–$60,000 annually, with a reported median of approximately $52,000 (per ENTRA Medellín briefing, July 14, 2026, sourcing Deel international hiring data). The UTC-5 timezone produces exact overlap with US East Coast hours and near-complete overlap with Pacific Standard Time — the structural advantage that puts Medellín ahead of Warsaw or Bangalore in US startup sourcing. Colombia's 55 percent international hiring growth in 2024 (Deel, 2024 State of Global Hiring Report; figure represents Deel-platform hires and reflects Deel's own serviced market, not total Colombia remote employment) reflects US employers discovering that the timezone-plus-cost combination is superior to the Warsaw arbitrage for West Coast companies that need synchronous collaboration with distributed teams.
Prague, Athens, Copenhagen (European cross-section)
Prague-based senior ML engineers earn a CZK-denominated equivalent of approximately €55,000–€80,000 annually, consistent with the broader CEE Tier 2 range (per ENTRA estimate, sourced from Glassdoor Czech Republic Q2 2026 and regional salary benchmarks). Athens sits in a similar band: €60,000–€85,000 gross for senior AI engineering roles, against a net-of-tax position that is more favorable than it appears on gross given Greece's 50 percent tax exemption for returning diaspora talent. Copenhagen anchors the opposite end of the European AI pay spectrum at DKK-equivalent of approximately €110,000–€145,000 for senior ML engineers (per ENTRA estimate and Glassdoor Denmark Q2 2026) — effectively Tier 1 pricing in a Tier 2 cost-of-living context, reflecting Denmark's deep talent scarcity in applied AI and the competition between Danish employers and remote-eligible US lab offers.
The cross-city picture reveals a global AI pay floor and ceiling that the models must navigate. The floor — what Tier 3 talent receives from international employers — sits around $35,000–$60,000 annually. The ceiling — what frontier-lab senior researchers earn under Model B — reaches $560,000–$740,000. The distance between those numbers is not a market dysfunction. It is the structural expression of a global talent hierarchy in which the scarcest technical skills (frontier model research, interpretability, alignment) command a global premium that no local cost index can contain, while applied ML engineering talent — more abundant globally — remains subject to geographic price compression under Models A and C.
Section 3: The Talent War Dynamic
The three models are not equally stable. When they compete for the same engineer, Model B wins. Every time.
The mechanism is not complicated. A Senior ML Engineer in Tallinn with RLHF experience and a portfolio of production fine-tuning work is employable by any frontier lab running Model B at $360,000–$540,000 in total compensation (ENTRA estimate, applied-track band). The same engineer is employable by a Big Tech AI division running Model A at approximately $154,000–$209,000, depending on the location modifier applied to the Nordic/Baltic region. The engineer's outside option is a 2x–3x compensation premium. Under those conditions, Model A retention requires either equity structures, mission alignment, or a work environment that the frontier lab cannot offer — and often cannot credibly articulate to a candidate already holding a higher offer.
Anthropic's distributed strategy, documented in ENTRA's July 12 briefing from the ME Bureau, is characteristically precise about this dynamic. The lab moved to market-rate remote for research-track roles above L4 specifically after tracking a pattern of offer rejections from senior researchers in the UK and Europe who cited compensation delta as the primary factor. The strategic language internally — "targeted, not generous" per ENTRA recruiter network sources — reflects a deliberate framing: Anthropic is not offering location-independent pay to every remote employee. It is offering it to the specific seniority and track categories where the talent competition justifies the cost. Below L4-equivalent, Anthropic maintains a zone-informed approach closer to Model C. Above L4 on research track, it runs Model B without qualification.
OpenAI's bifurcated approach illustrates the practical ceiling of running different models within one organization. For frontier research and safety research roles listed under its "distributed research" track designation, OpenAI has adopted location-agnostic compensation since Q4 2025, with senior research scientist total compensation reaching $520,000–$810,000 (per ENTRA recruiter network and Levels.fyi Q2 2026). For applied engineering and product-adjacent roles, OpenAI continues to apply a location modifier — a model-A-consistent approach that generates compensation levels 20–30 percent below the research track for employees outside San Francisco. The internal equity friction this creates is documented: ENTRA's Q2 2026 employer survey flagged that organizations running different compensation models across role categories reported higher internal pay-equity complaint rates than organizations on a single model. The bifurcation is strategically rational; it is organizationally messy.
Hugging Face's zone-based approach offers the clearest case study in Model C execution. The company operates a three-tier zone structure — Tier 1 (San Francisco and New York), Tier 2 (London, Berlin, Paris, Tel Aviv, Toronto), Tier 3 (all remaining markets) — with published band ranges on its careers page and consistent application across role categories. A Staff ML Research Engineer at Hugging Face earns approximately $280,000–$380,000 in Tier 1, $210,000–$295,000 in Tier 2, and $140,000–$210,000 in Tier 3, per ENTRA employer-network intelligence (anonymized, n=12 companies). The Tier 2 band at Hugging Face is meaningfully above what a Big Tech AI division's Model A would generate for the same engineer in Warsaw or Berlin. That positioning — above Model A, below Model B — is intentional: Hugging Face competes for talent that values the open-source AI mission strongly enough to accept a pay band that is not frontier-lab level but is demonstrably above the location-indexed alternative.
Weights & Biases runs a structurally similar zone architecture, calibrated to the MLOps tooling market rather than frontier research. Its Tier 2 bands track closely to what Berlin and Warsaw ML engineers would earn from a well-compensated enterprise AI employer, without approaching frontier-lab levels. The appeal is not compensation maximization; it is a combination of compensation adequacy, a compelling product mandate, and the operational flexibility of a genuinely distributed team that has been running asynchronous-first since 2019.
The talent-defection pattern ENTRA's recruiter network has documented most frequently through H1 2026 runs as follows: a senior ML engineer in a Tier 2 city — Warsaw, Tallinn, Berlin, Athens — initially joins or stays with a Big Tech AI division under Model A. The engineer receives a competitive package relative to local alternatives. Over twelve to eighteen months, the engineer builds a portfolio of frontier-applicable work — fine-tuning, RLHF, evaluation design — that makes them recruitable by frontier labs running Model B. A recruiter from Anthropic or OpenAI reaches out. The compensation offer is 2x–3x the current total comp. The engineer accepts.
This defection cycle is not a moral failure of the Model A employer. It is a structural feature of a talent market in which two models operate simultaneously for the same talent pool. The employers who have recognized the defection pattern earliest are now using it proactively: building Tier 2 pipelines explicitly as a two-to-three-year development funnel, acquiring engineers at Model A or Model C rates while their skills are developing, and accepting that senior engineers at full frontier-lab caliber will be competed away. The model shifts from talent retention to talent development-and-recycle — a structurally different HR strategy with its own cost-benefit calculus.
Section 4: The Employer ROI Math
The financial architecture of each model is clearest at the team level.
A ten-engineer team of Senior ML Engineers in San Francisco, all at L5/L6-equivalent seniority, costs the employer $3,400,000–$5,600,000 in annual total compensation at $340,000–$560,000 per engineer (per Levels.fyi Q2 2026, frontier-lab applied track). Employer-side costs — payroll taxes, benefits, equipment, office infrastructure prorated per employee — add approximately 20–25 percent, bringing the fully loaded annual cost of a ten-engineer SF team to approximately $4,000,000–$7,000,000.
The same ten-engineer team, sourced under Model A from Warsaw, Prague, and Tallinn, costs the employer $1,400,000–$2,200,000 in annual total compensation — at location-indexed rates of $140,000–$220,000 per engineer weighted across those markets. Employer-side costs for distributed employees run lower on average (no office footprint, reduced benefits overhead in lower-cost markets) — bringing the fully loaded cost to approximately $1,600,000–$2,800,000 annually. The gross saving versus the SF team: $2,000,000–$3,200,000 per year for a ten-person team. Per engineer, the annualized saving is $200,000–$320,000 — significantly higher than the $126,000 headline figure because the headline uses a specific Big Tech band rather than the frontier-lab ceiling.
The $126,000 figure is the employer's saving at the specific intersection of a Big Tech AI division SF band ($280,000 base) and a Warsaw location-indexed equivalent ($154,000 base) — a narrower comparison reflecting the more compressed bands at the Big Tech tier rather than frontier-lab total comp. It is the correct number for the question "what does Google save per engineer in Warsaw?" It understates the number for "what does Anthropic save if it ran location-indexed pay instead of Model B?"
The key structural insight is that employer savings accrue only under Model A or Model C. Under Model B, the employer pays the same total compensation per engineer whether that engineer is in San Francisco or Warsaw. The cost structure is identical to the SF team; the talent supply is larger. Model B is an investment in talent acquisition breadth, not a cost reduction play.
The ROI calculation for Model B requires quantifying talent quality, not cost savings. If a frontier lab that runs Model B can access 40 percent more qualified senior ML researchers in its hiring pool than one that runs Model A — a figure consistent with ENTRA's survey-based estimate of the share of candidates who exclude location-indexed employers from their consideration set — then the return on Model B is measured in research output per dollar, not headcount cost per engineer. For frontier labs whose commercial value is built on research velocity and model capability, that is the correct ROI frame. For enterprise AI teams whose value is built on reliable applied-engineering delivery, the cost frame may dominate the capability frame — making Model A or Model C the rational choice.
The PPP-adjusted productivity comparison matters here. ENTRA's employer survey (n=12 companies, anonymized) asked employers about relative productivity of their distributed Tier 2 and Tier 3 engineers versus equivalent-seniority SF-based counterparts. The median response: within 10–15 percent for senior applied engineering roles. The range was wide — some employers reported no detectable productivity gap; others reported a 20–25 percent output disadvantage primarily attributable to timezone coordination overhead and async tooling friction. The productivity-adjusted cost comparison is therefore approximately: Model A Warsaw team at $1,600,000 fully loaded per year, producing output equivalent to an SF team at $4,000,000 fully loaded, adjusted for an estimated 12 percent productivity drag. Productivity-adjusted Warsaw cost: approximately $1,800,000. The employer still saves $2,200,000 per ten-engineer team annually — roughly $220,000 per engineer per year.
The employer ROI math is unambiguous for applied engineering under Model A: the savings are real, large, and productivity-adjusted favorable. The math breaks down specifically and only at the tier of talent that Model A cannot retain — the top quintile of senior ML engineers for whom Model B employers represent a 2x–3x outside option. For that cohort, the choice is not Model A versus Model B on cost terms. It is Model B to acquire the talent, or Model A with the understanding that the talent will eventually defect.
Section 5: H2 2026 Forecast and What Comes Next
The compensation model choice is hardening. The trial period — in which companies experimented with model switches and hybrid approaches — is giving way to structural commitments that become progressively more expensive to reverse.
Frontier labs that launched Model B for research roles before 2025 are under increasing internal pressure from headcount scaling. At 500 globally distributed engineers all receiving SF-equivalent total compensation, the organizational argument for maintaining Model B is purely competitive: the talent quality requires it. At 5,000 engineers, the finance function begins asserting that some role categories are not sufficiently talent-scarce to justify location-agnostic pay. The slide toward Model C begins. ENTRA's recruiter network sources describe this progression as already visible at OpenAI, where the "distributed research" designation that unlocks Model B pay has become more narrowly defined through H1 2026 than it was at launch in Q4 2025. The funnel narrows as scale increases.
Three structural forces will shape the compensation map through H2 2026 and into 2027.
EU AI Act enforcement. Article 50 of Regulation (EU) 2024/1689 — the transparency and information obligation for AI systems that interact with natural persons — applies from August 2, 2026. Annex III, covering high-risk AI systems in consequential decision-making domains, triggers its full compliance obligation from December 2, 2027. The compliance engineering wave that these deadlines are generating is already bidding up European AI engineer rates in markets with concentrated compliance expertise. Warsaw engineers with EU AI Act Article 9 risk management experience at Google Poland or Microsoft Azure are pricing at the upper end of CEE Tier 2 bands. Athens engineers with Annex III documentation credentials from Greek insurtech or healthtech employers are commanding premiums of 15–25 percent above the €60,000–€85,000 base market. ENTRA estimates that EU AI Act compliance expertise will add a 15–25 percent wage premium to European AI engineers in the relevant specializations by the time the December 2027 Annex III deadline forces large-scale hiring from enterprises that have deferred compliance investment. The EU is building a compensation microclimate inside the broader remote AI pay structure — and it is an upward one.
The Unit 8200 pricing floor. Israel's military technology alumni are establishing a de facto market-rate floor for the global senior AI research population. Engineers and researchers who completed advanced signals intelligence, adversarial ML, and systems security work within Israeli Defense Forces technical units exit into the private market with frontier-applicable skills and a consistent pricing posture: they apply at market rate globally and do not accept location-indexed offers. ENTRA's recruiter network uniformly reports that Unit 8200 alumni who are accessible on a remote basis price at or above Tier 1 equivalent compensation regardless of their Israeli residence, and that attempts to apply location modifiers to their offers are declined. The alumni cohort is not large enough to set the market on its own — but it represents the leading edge of a pricing posture that is spreading across the global senior AI talent population as market information improves and location-independent employers proliferate.
The Warsaw-Bangalore corridor. ENTRA's city coverage identifies Warsaw and Bangalore as the two markets where the intersection of talent quality and Model A or Model C affordability is most favorable for enterprise AI employers through 2026 and 2027. Both markets offer senior ML engineering talent at 40–55 percent of SF cost under location-indexed frameworks. Both markets have university pipelines — the University of Warsaw, IIT Bombay, IISc Bangalore — that continue to produce ML talent at the seniority level applicable to frontier-adjacent work. Both markets have a critical mass of established employer presence — Google Poland, Microsoft Azure Noida and Hyderabad, Anthropic's distributed engineering cohort — that validates the quality of the local talent bench to risk-averse HR leaders. For enterprise AI teams building applied engineering capacity at cost, the Warsaw-Bangalore corridor under a well-executed Model C represents the most credible cost-performance outcome available in H2 2026.
The Zone-Based model standardizes. ENTRA's forecast: Model C (Zone-Based Bands) becomes the industry standard compensation architecture for applied AI companies, enterprise AI teams, and growth-stage AI startups by end of 2027. The convergence is driven by three forces simultaneously: Model A (location-indexed) proves too retention-porous at senior levels to sustain; Model B (market-rate remote) proves too expensive to sustain at scale; and Model C offers a stable middle ground that HR infrastructure can manage, recruiters can quote, and finance functions can model. The remaining question is not whether Zone-Based becomes the norm, but how many zones, where the boundaries fall, and whether the Tier 2 / Tier 3 bands are set competitively enough to retain the talent they need — or just generously enough to hire it and then watch it defect to the frontier lab that still runs Model B for the best researchers in every city.
The compensation architecture is not a stable equilibrium. It is an active competition between three models, each with internal logic, each with structural failure modes, and each being run simultaneously by employers competing for the same senior engineers. The employer that understands which model it is running — and why — has a structural advantage over the one that has never examined the choice.
METHODOLOGY
ENTRA Remote Compensation Analysis, H1–H2 2026
Salary data sourced from: Levels.fyi AI Engineer and Research Scientist submissions (cross-referenced Q2 2026, n=14,200 AI professionals); ENTRA Job Signal Index (LinkedIn Talent Insights hiring and salary signal data); public job postings with disclosed compensation bands; and ENTRA employer network interviews (anonymized, n=12 companies, conducted May–July 2026).
City-specific compensation data drawn from ENTRA Remote Issue briefings: Warsaw (July 7, 2026), Tallinn (July 16, 2026), Nairobi (July 14, 2026), Lagos (July 16, 2026), and Medellín (July 14, 2026). Local market salary benchmarks for Warsaw sourced from ITMAGINATION AI Salaries Poland 2025 and Freenance Tech Salaries Poland 2026. Tallinn figures from Glassdoor Estonia Q2 2026, cross-referenced with ENTRA EU Bureau sourcing. Lagos figures from ENTRA GCC AI Salary Index H1 2026 and Glassdoor UAE ML Engineer Median Q2 2026. Nairobi figures from ENTRA East Africa AI Salary Index H1 2026 and Central Bank of Kenya exchange rate (June 2026). Medellín figures from Deel International Hiring 2024 published data and ENTRA US Bureau sourcing.
PPP-adjusted figures use World Bank PPP conversion factors (2025) as published in ENTRA Remote AI Salary PPP Index (July 11, 2026).
Company compensation models (Model A, B, C classification) based on public job descriptions, Levels.fyi self-reports, Glassdoor submissions, and ENTRA recruiter-network intelligence (38 senior AI recruiters active across tracked lab accounts). Where individual company practices are not confirmed through public disclosure, compensation ranges are labelled "(per ENTRA estimate)" or "(per ENTRA recruiter network)."
Employer productivity comparison data from ENTRA employer network survey (n=12 companies, anonymized). Responses reflect self-reported employer assessment of relative output between SF-based and distributed Tier 2/Tier 3 engineers at equivalent seniority. Results are directional estimates, not audited performance data.
EU AI Act Annex III references: Regulation (EU) 2024/1689. Article 50 transparency obligations apply from August 2, 2026. Annex III full compliance deadline: December 2, 2027 (as extended by European Council Digital Omnibus agreement, May 7, 2026). EU AI Act compliance wage premium estimates are ENTRA forward projections based on demand/supply analysis, not confirmed market data.
Currency conversions: 1 USD = 0.92 EUR (ECB Q2 2026); 1 GBP = 1.27 USD (Bank of England Q2 2026); 1 NIS = 0.27 USD; 1 KES = 0.0077 USD (Central Bank of Kenya, June 2026); 1 PLN = 0.234 USD (NBP Q2 2026).
Global Career PlatformFind AI talent. Find your next role.
Booking is hotels. · Airbnb is apartments. · ENTRA is global careers.