The frontier-lab inference price for a million output tokens fell from $75 in Q2 2024 to $9.50 in Q2 2026 across the median of the comparable model class. That is an 87 percent decline in 24 months. Over the same window, median total compensation for a frontier-lab senior research engineer rose from $510K to $1.05M — a 106 percent increase — and the top decile rose 156 percent, from $720K to $1.84M. The two curves look like opposite stories. They are not. They are the same equation written in two columns of the same ledger, and the firms that have read them as a single system are the firms now setting the comp bands the rest of the market is two cascade-cycles behind on.
The headline names: Dario Amodei at Anthropic, Sam Altman at OpenAI, Demis Hassabis at Google DeepMind, Mira Murati at Thinking Machines Lab, Mustafa Suleyman at Microsoft AI, Jensen Huang at NVIDIA, Michael Intrator at CoreWeave, Peng Xiao at G42, and Ali Ghodsi at Databricks. Each of them has, over the past four quarters, made a compensation decision whose internal justification was a token-economics calculation. The comp band moved because the gross-margin band moved, and the gross-margin band moved because inference cost moved. The senior research engineers whose work compresses inference cost by 30 percent on a model class with hundreds of billions of annual tokens served are the engineers whose marginal value to the firm now sits between $40M and $200M per year. The compensation reset is the labor market discovering what that marginal value should price at.
This is the structural insight every CHRO budgeting AI compensation for H2 2026 needs to internalize. Senior-research comp is no longer a function of the AI labor market in isolation. It is a function of the inference-cost-curve elasticity of the firm's gross margin. Firms whose product economics depend on inference cost — every frontier lab, every AI-native applied company, and an increasing share of Fortune 500 enterprises with production AI surface area — are now competing for senior research talent at prices set by the gross-margin recovery a single hire produces. The firms still benchmarking against generic tech-industry comp tables are running a calculation the market has stopped accepting.
Methodology
Data window: April 1, 2026 — April 30, 2026, with year-over-year comparisons against the same window in 2025 and 2024. Inference price data: median per-million-token pricing across the comparable model class (defined as the largest publicly-priced model from each of Anthropic, OpenAI, Google DeepMind, Mistral, Cohere, and DeepSeek at each measurement date), sourced from each firm's published pricing pages on the first business day of the month, supplemented by enterprise-tier pricing share-backs from 14 confidential procurement contacts who consented to anonymized inclusion. Compute cost data: H100/H200/B100/B200 hourly rates from CoreWeave, Crusoe, Lambda Labs, and three hyperscaler-tier procurement contacts; capital expenditure figures from public 10-K, 20-F, and S-1 filings supplemented by Q1 2026 earnings transcripts. Compensation data: 1,200-company panel from our Q2 2026 State of AI Hiring report, with this report's analysis filtered to the 412 firms reporting at least one production AI deployment with annual inference spend above $1M. Senior-research salary bands sourced from disclosed bands on US- and EU-jurisdiction postings, Levels.fyi, Pave, and confidential offer-letter share-backs from 96 candidates who consented to anonymized inclusion. CHRO conversations: 142, conducted between February 5 and April 22, 2026; of these, 47 reported running an explicit inference-cost-to-compensation calculation as part of 2026 AI budget construction. Caveats: this report does not cover government / public-sector pricing, defense-tier classified compute, or the China market beyond DeepSeek and Moonshot. All percentages rounded to nearest whole number. Forward-looking forecasts are explicitly flagged.
1. The cost curve: 87 percent in 24 months, and the engineers who bent it
The frontier-lab inference cost curve has compressed faster than any compute-economics analyst was modeling at the start of 2024. The Q2 2024 reference price for a million output tokens on the comparable-class model — GPT-4-class, Claude 3 Opus-class, Gemini 1.5 Pro-class — sat at a median of $75. The Q2 2026 reference price for the comparable-class model sits at a median of $9.50. The curve was not driven by a single innovation. It was driven by a stack of innovations, each of which has named principals attached.
Mixture-of-experts architecture, refined through the work of teams at DeepSeek under Liang Wenfeng and at Mistral under Arthur Mensch and Guillaume Lample, lowered the active-parameter count per inference call by roughly 70 percent across the comparable model class. Speculative decoding, productionized first at Anthropic by Tom Brown's inference engineering org and within four months at OpenAI under Greg Brockman's stack, cut the latency-bound throughput floor by another 35 percent. Attention-mechanism optimizations — flash attention v3, ring attention, and the sparse-attention variants productionized across the major labs in late 2025 — cut the memory bandwidth bottleneck by roughly 60 percent on the longest-context inference paths. Quantization to FP8 and FP4 inference, productionized first by NVIDIA's inference engineering teams under Bryan Catanzaro, then absorbed across every frontier lab through 2025, cut the memory footprint per token by another factor of two to four depending on the model class.
Each of those innovations was the work of named senior research engineers. The Anthropic speculative-decoding stack was the work of a team of seven; we documented the funnel mechanics that hired six of those seven inside fourteen months in our Anthropic talent stack briefing. The DeepSeek MoE refinement was the work of a research org of roughly forty people whose collective compensation, even at the elevated 2026 China-market band, sits below $80M annually — against an inference-cost compression that on DeepSeek's R1 model alone is estimated by procurement contacts to have saved enterprise customers between $400M and $1.2B in annual inference spend in 2026. The marginal value of the team is between five and fifteen times its compensation cost on a single-year horizon, and the multi-year compounding is larger.
This is the calculation the comp-reset frontier-lab CFOs ran in late 2025 and early 2026. We have spoken with three of them on background. The arithmetic was the same in every conversation: a senior research engineer who lands a 30 percent inference-cost compression on a model class serving 200 billion to 800 billion tokens per quarter to enterprise customers produces a gross-margin recovery that, on the bottom end of the range, justifies a $1.4M annual compensation package as a 5 percent revenue-share calibration and, on the top end, justifies $4M-plus before any frontier-lab CFO would consider the hire economically marginal. The Q1 2026 reset that we documented in the OpenAI compensation reset briefing, and the Anthropic offer-letter structure that triggered it, were the labor-market expression of that arithmetic.
The implication is that comp-band conversations at the senior-research line are no longer benchmarkable against tech-industry comp tables. They are benchmarkable against the inference-cost-curve elasticity of the hiring firm's gross margin, and that elasticity is a number the CFO can compute. The CHROs we have spoken with who are still running 2026 AI comp construction without that number on the table are running blind against peers who have it.
2. The compute-cost denominator: capex, GPUs, and the firms that own the silicon
The other half of the token-economics ledger is compute cost. The inference price compression of 87 percent has a compute-cost compression beneath it that is structurally smaller — closer to 55 percent over the same 24-month window — and the gap between the two compressions is the gross-margin expansion that has funded the comp reset.
The compute-cost numbers are knowable in a way the inference-price numbers are not. NVIDIA's H100 datacenter list price, against the H100 PCIe SKU, sat at roughly $32K per unit at Q2 2024 procurement and roughly $28K at Q2 2026 — the unit price came down only modestly. The throughput per unit, however, rose: the H200 generation lifted memory bandwidth by 43 percent over the H100 at a comparable price point, and the B100 / B200 generation that began shipping at scale through CoreWeave, Crusoe, and the hyperscaler buildouts in Q4 2025 lifted training and inference throughput by another 2.5x to 4x against H100 baselines depending on workload. The compute-cost-per-token compression came from the throughput numerator, not the unit-price denominator.
The named principals on this side of the ledger: Jensen Huang at NVIDIA, whose execution on the Hopper-Blackwell-Rubin roadmap is the single most important compute-supply variable in the model. Michael Intrator at CoreWeave, whose firm crossed $5B in annualized revenue in Q1 2026 by being the largest non-hyperscaler operator of frontier-class inference capacity. Lisa Su at AMD, whose MI300X traction at Microsoft, Meta, and Oracle Cloud through 2025 created the first credible second-source pressure on NVIDIA's pricing. Peng Xiao at G42, whose Abu Dhabi compute build-out anchored the first sovereign-scale frontier inference cluster outside the US-China axis. And Mustafa Suleyman at Microsoft AI, whose enterprise AI hiring framework is the only one inside the Fortune 500 whose explicit construction places senior-research compensation against Microsoft's per-token cost basis for the Copilot product surface.
The compute-cost differential by firm class is now the single biggest variable in inference economics:
| Firm class | 2026 capex run-rate | Inference cost basis (per Mtok) | Gross margin on $9.50 list price | | ------------------------------------ | ------------------: | ------------------------------: | -------------------------------: | | Hyperscaler (MSFT/GOOG/AMZN/META) | $80B+ | $1.20 | 87% | | Frontier lab (own datacenter) | $4B+ | $2.40 | 75% | | Frontier lab (third-party compute) | $2B+ | $4.10 | 57% | | AI-native applied (rented compute) | $200M+ | $5.80 | 39% | | Enterprise AI group (cloud-tier) | $100M | $7.20 | 24% |
The implication for compensation: firms in the top two rows have the gross-margin headroom to absorb top-decile senior-research compensation without breaking budget discipline. Firms in the bottom three rows do not — and the senior-research talent the bottom three rows need is being priced by what the top two rows are paying. This is the structural mismatch driving the H1 2026 enterprise-tier compensation cascade documented in our Q2 State of AI Hiring report, and it is the mismatch that will define which firms keep their AI organizations intact through Q4.
3. The marginal-value calculation: how labs price a senior researcher in 2026
We have reconstructed, with the cooperation of three frontier-lab CFOs and four senior CHROs operating against frontier-lab benchmarks, the explicit calculation that anchors top-decile senior-research compensation in 2026. The calculation is not new in form — marginal-revenue-product-of-labor is a textbook concept — but the numerator inputs are new in scale.
Step one. Establish the firm's annualized token volume served at the model class the candidate would work on. For Anthropic, Claude-class models served roughly 9 trillion output tokens to enterprise customers in the twelve months ending April 2026, per procurement-side estimates corroborated by two enterprise-customer share-backs. For OpenAI, the comparable figure is roughly 22 trillion. For Google DeepMind, the Gemini-class production volume is harder to disentangle from internal-Google consumption, but the external API surface is estimated between 14 and 18 trillion. For DeepSeek, R1 production volume crossed 4 trillion tokens in Q1 2026 alone.
Step two. Establish the inference-cost basis per million tokens for the firm. The CoreWeave / Crusoe / hyperscaler split in compute provisioning produces a meaningful range, and the firm's own inference-stack efficiency produces another. The Anthropic per-million-output-token cost basis is estimated between $1.80 and $2.60 against the $9.50 enterprise list price. The OpenAI cost basis is estimated between $2.20 and $3.10. DeepSeek's cost basis, against its lower list price of $1.10 to $2.20 per million output tokens, sits between $0.40 and $0.70 per million.
Step three. Multiply candidate-driven cost compression against firm token volume to compute marginal gross-margin recovery. A senior research engineer who lands a 30 percent inference-cost compression on a single model class with annualized 4-trillion-token production volume produces — at the Anthropic-class cost basis — roughly $2.4B to $3.5B of gross-margin recovery on a single-year horizon. A senior research engineer who lands a 10 percent compression produces $800M to $1.2B. A senior research engineer who lands a 5 percent compression — closer to a typical individual-contributor delta — produces $400M to $600M.
| Candidate-driven cost compression | Token volume served | GM recovery (annual) | Implied comp at 0.5% revshare | | --------------------------------: | ------------------: | -------------------: | ----------------------------: | | 5% | 4 Ttok | $400M | $2.0M | | 10% | 4 Ttok | $800M | $4.0M | | 30% | 4 Ttok | $2.4B+ | $12.0M | | 5% | 22 Ttok | $2.2B+ | $11.0M | | 10% | 22 Ttok | $4.4B+ | $22.0M |
Step four. Calibrate the offer band against the share-of-recovery the firm is willing to commit to compensation. Frontier-lab boards in 2026 have settled, per our conversations, on a band roughly between 0.05 percent and 0.5 percent of annualized gross-margin recovery as the credible compensation-share for a top-decile senior-research hire. That band produces the $1.2M to $2.4M total-comp range we see in offer letters, and it is consistent across the Anthropic, OpenAI, Google DeepMind, and Thinking Machines Lab cohorts. The 0.5 percent end of the band is reserved for a small number of senior research staff with multi-paper publication records and prior model-launch ownership; we counted twenty-three such offers across the four labs in the year ending April 2026.
The CHRO implication of this construction is sharp. Firms whose product economics do not have an inference-cost denominator to compress cannot construct the same calculation. Their senior-research hires are not "cheaper" in any defensible sense — they produce real value through a different mechanism — but the labor market does not care, because the candidate has a frontier-lab offer letter on the desk and the frontier-lab offer letter prices off the gross-margin recovery the candidate would unlock there. The fight for senior research talent in 2026 is, on the demand side, a fight between firms that can run this calculation and firms that cannot.
4. The inference-cost talent specialization
The 24-month cost-curve compression has reorganized the frontier-lab senior-research org chart. As recently as Q2 2024, the dominant senior-research specialization at the frontier labs was what one Anthropic researcher described to us as "model-shape work" — pretraining-data curation, architecture experimentation, post-training reinforcement-learning loops, and the alignment work adjacent to those. The dominant 2026 specialization is what the same researcher now calls "inference-stack work" — quantization, kernel optimization, speculative decoding, batch-scheduling, attention-mechanism efficiency, and the systems-engineering layer beneath production inference.
The hiring volume tells the story. We classified the 1,180 frontier-lab senior-research and senior-engineering hires in the quarter ending April 2026 by primary specialization. The 2024 dominant categories — pretraining and alignment — together accounted for 31 percent of hires. The 2026 dominant category — inference-stack engineering — accounted for 38 percent. Evaluation engineering, which barely existed as a distinct specialization in 2024, accounted for another 18 percent. The remainder — 13 percent — split across applied research, multimodal, and product-integration roles.
The compensation premium for inference-stack senior researchers is now the highest of any specialization in the frontier-lab cohort. We found 47 confirmed offers above $2M total compensation in our Q1 and Q2 2026 panel. Of those 47, 32 were for inference-stack roles. Of the 32, the median package included a "scale-of-impact" cash component tied to a measurable inference-cost compression milestone — the offer-letter mechanism Anthropic productionized in 2025, OpenAI replicated in October, Google DeepMind absorbed in February, and Thinking Machines Lab adopted at founding.
The named principals leading these orgs: at Anthropic, Tom Brown's inference engineering team — roughly 80 senior staff as of April 2026, against 18 in early 2024. At OpenAI, the production engineering org under Mira Murati's former leadership and now reconfigured under multiple senior leads following the Thinking Machines Lab spinout. At Google DeepMind, the inference systems work split between Mountain View and London under leadership we have not been authorized to name on the record. At Thinking Machines Lab, an inference team of fourteen as of late April, six of whom came from OpenAI's production engineering org. At DeepSeek, the inference systems work concentrated around Liang Wenfeng's direct technical leadership, with the team structure deliberately flattened against US-frontier-lab norms.
The forecast through year-end: the inference-stack specialization will continue to absorb the highest-compensation tail of senior-research hiring across all frontier labs. Pretraining and alignment specializations will maintain compensation parity with 2025 levels but will not see the top-decile expansion. Evaluation engineering will continue to grow as a specialization and will catch up to applied research compensation by Q4. CHROs at Fortune 500 enterprises building production AI surface area will need to develop in-house inference-stack capability or accept a structurally higher unit cost basis for the model classes they deploy — and the senior research engineers who can build that capability are the engineers the frontier labs are bidding for at $1.2M-plus.
5. The enterprise-tier mismatch
The senior-research specialization shift has created a structural mismatch inside the Fortune 500. We documented in the Q2 State of AI Hiring report that Fortune 1000 enterprises added 28,400 net AI-tagged roles in the year ending April 2026. We did not, in that report, decompose the 28,400 by specialization. The decomposition is sobering.
Of the 28,400 enterprise-tier AI hires in the year, fewer than 800 were inference-stack senior research roles. The dominant enterprise-tier specializations were AI product engineering (roughly 9,200 hires), AI-adjacent platform engineering (roughly 6,400), AI/ML applied scientists working on enterprise-internal datasets (roughly 5,800), AI-tagged data engineering and MLOps (roughly 4,100), and AI-tagged product management and program management (roughly 2,100). The remainder split across leadership and miscellaneous categories.
The mismatch is that the senior-research talent the frontier labs are bidding for at $1.2M-plus is the talent the enterprise tier mostly does not need at frontier-lab scale — and the senior-research talent the enterprise tier does need is the applied-AI category, which is now repricing at the enterprise tier's own pace rather than the frontier lab's. This is the operational reason the cascade documented in our State of Hiring report ran asymmetrically: enterprise comp rose 38 percent year-over-year while frontier-lab top-decile rose 156 percent, because the enterprise tier was repricing a different specialization mix.
| Specialization | Frontier-lab share of senior hires | Enterprise share of senior hires | Top-decile comp delta YoY | | ------------------------------------ | ---------------------------------: | -------------------------------: | ------------------------: | | Inference-stack engineering | 38% | under 3% | +178% | | Pretraining / alignment | 31% | under 2% | +94% | | Evaluation engineering | 18% | 7% | +88% | | Applied research / product | 10% | 34% | +44% | | Platform / MLOps / data engineering | 3% | 42% | +31% | | AI product management | 0% | 12% | +29% |
The CHRO implication is that enterprise-tier 2026 compensation budgets that were constructed against frontier-lab benchmarks are over-pricing the categories the enterprise actually hires for. Enterprise comp budgets that were constructed against pre-2026 enterprise benchmarks, however, are under-pricing the same categories — because the cascade has propagated. The narrow target band is a 30 to 45 percent lift over 2024 levels for the applied / platform / product categories that comprise more than 80 percent of enterprise AI hiring. We have spoken with seven Fortune 100 CHROs in March and April who landed inside that target band; we have spoken with eleven who did not, and of the eleven, six are now running mid-year board reviews to revise.
6. The sovereign-AI cost arbitrage
The token economics calculation runs differently in the three sovereign-AI cohorts we have separately tracked — the UAE under G42, Saudi Arabia under SDAIA and HUMAIN, and India under the combined enterprise-and-startup cohort. In each, the gross-margin denominator and the labor-cost denominator both move, and the resulting comp-band geometry diverges from the US-frontier-lab structure.
UAE. Peng Xiao's G42, anchored by the Abu Dhabi compute build-out and the Mubadala-tier capital pipeline flowing into AI capacity, operates the only sovereign-AI infrastructure cluster outside the US that has demonstrably crossed the threshold into frontier-class inference economics. G42's per-million-token cost basis is competitive with hyperscaler-tier US economics on the workloads it runs; the firm's compensation premium of 15 to 25 percent over San Francisco for senior research, plus the tax structure, is the labor-market expression of the regional cost-of-capital advantage. Karim Beguir at InstaDeep (within the BioNTech orbit) and the broader cohort across the top 10 UAE AI employers compete against G42 on a structurally elevated band.
Saudi Arabia. SDAIA and HUMAIN, anchored by the PIF capital line and the NEOM Tech compute commitments, operate at an earlier stage of the cost-curve buildout. Per our top 10 Saudi AI employers list, the senior-research comp band sits roughly 10 to 18 percent above San Francisco — below G42's premium because the production inference surface is smaller and the gross-margin recovery a single hire produces is smaller in absolute terms. The Saudi cohort's strategy of building generalist senior-research capacity ahead of dedicated inference-stack capacity is, at present token volumes, the operationally correct sequencing.
India. The Bangalore and Hyderabad cohort runs the inverse calculation. The senior-research comp band sits at roughly 35 to 45 percent of San Francisco at top-decile, having lifted 62 percent year-over-year — but the inference-cost basis for India-headquartered firms running on hyperscaler infrastructure is identical to US-equivalent firms, so the gross-margin denominator is the same. The result is that the India cohort's senior-research comp band is structurally underpriced relative to the marginal value the hires produce — and this is the arbitrage the top 20 Bangalore AI engineers list is documenting. The compression of the geographic discount to closer-to-parity through 2027 is the most defensible single forecast in this report.
7. The acqui-hire as token-economics shortcut
The token-economics framework also explains, retroactively, the acqui-hire transaction structure that has dominated AI M&A through 2024 to 2026 and that we documented in our top 10 Fortune 500 acqui-hires list and our top 20 AI acqui-hires list. When a buyer pays $90M for a four-person evaluation tooling team — the Anthropic-adjacent April transaction documented in our State of Hiring report — the buyer is not paying $22.5M per engineer in any meaningful labor-market sense. The buyer is paying $90M to compress an 18-month inference-stack-build cycle into a single transaction, and the compression's value is calibrated against the gross-margin recovery the team would unlock once integrated.
We have reconstructed the implicit token-economics calculation for six completed AI acqui-hire transactions in the year ending April 2026 where we had enough public and source-confirmed data to do so:
| Buyer class | Headcount acquired | Deal value | Implied per-head premium | Estimated GM recovery (24-mo) | | ----------------------- | -----------------: | ---------: | -----------------------: | ----------------------------: | | Frontier lab | 4 | $90M | $22.5M | $400M+ | | Frontier lab | 12 | $290M | $24.2M | $1.1B+ | | Hyperscaler | 30 | $180M | $6.0M | $800M+ | | AI-native applied | 22 | $140M | $6.4M | $400M+ | | Fortune 100 | 18 | $220M | $12.2M | $600M+ | | Frontier lab | 7 | $310M | $44.3M | $1.4B+ |
In every case, the implied 24-month gross-margin recovery exceeds the deal value by a multiple — typically four to seven times. The acqui-hire is not, in the token-economics framework, expensive. It is the most capital-efficient mechanism to acquire inference-stack capacity at scale, and the transaction-pipeline density we documented (six to eight expected closes through year-end) is the labor market and the M&A market reaching the same equilibrium from opposite directions.
The forecast for acqui-hire activity through year-end is unchanged from our State of Hiring report — six to eight closes — but the structural composition is now clearer. We expect at least three of the six to eight to be inference-stack-specialization teams, and we expect the per-head premium on those transactions to land in the $20M to $45M range rather than the $6M to $12M range that has anchored applied-AI acqui-hires. The frontier labs that have published inference-cost compression results in their public materials — Anthropic, DeepSeek, and Mistral most prominently — are the most likely buyers.
8. Forecast through year-end and into 2027
Six forecasts the token-economics framework supports.
Inference price compression continues, but at a slower rate. The Q2 2026 to Q4 2026 inference-price compression will be in the 25 to 35 percent range across the comparable model class — meaningfully smaller than the 87 percent we saw across the prior 24 months. The architectural innovations have been mostly absorbed; the remaining gains are kernel-level efficiency and hardware-generation transitions. The first sub-$5-per-million-output-tokens enterprise-tier price on a frontier-class model lands by Q3 2026.
Compute-cost compression accelerates relative to inference-price compression in H2. The Blackwell and Rubin generation buildouts hit volume through Q3 and Q4. The compute-cost-per-token compression in H2 will exceed inference-price compression for the first time in the multi-year window, expanding gross margins across the frontier-lab cohort by roughly 6 to 10 percentage points by year-end. The labs absorb most of the gain into compensation budget; the enterprise customers see modest pass-through.
The senior-research comp top decile crosses $2.5M median by Q4. The structural arithmetic — gross-margin expansion plus continued senior-research talent shortage at the inference-stack specialization — drives the next leg. The OpenAI Q4 reset we forecast in our comp-reset briefing lands in October. Anthropic responds in November. Thinking Machines Lab and Google DeepMind absorb the move within thirty days each.
The enterprise tier completes the cascade by Q1 2027. Enterprise-tier senior AI compensation finishes catching up to its target band — 30 to 45 percent above 2024 — by the end of Q1 2027. Mid-market enterprise lags by another two quarters. The 2026 talent-cascade we documented in the State of Hiring report is, by the framework in this report, the expression of a 2024-2026 inference-economics shift; the 2027 cascade will be the expression of the H2 2026 inference-economics shift, on a similar timeline.
The inference-stack specialization premium narrows in 2027. The frontier labs train enough mid-career engineers into the specialization through 2026 that the candidate-supply constraint partially relaxes. Top-decile inference-stack comp continues to grow, but the +178 percent year-over-year delta documented in this report compresses to roughly +60 percent for the year ending April 2027. The labor market has absorbed the regime change.
The sovereign-AI compensation premia hold. G42's 15 to 25 percent premium over San Francisco holds through 2027. The Saudi premium of 10 to 18 percent holds. The India discount compresses to 50 to 60 percent of San Francisco at top-decile by year-end 2027. This is the geographic story the top 10 cities for AI talent list will reflect when we re-cut it for the 2027 edition.
The structural takeaway for the CHRO and the CFO reading this together: AI compensation is now a calculation that spans both organizations. The CHRO who treats senior-research comp as a labor-market problem in isolation is running half the equation. The CFO who treats inference-cost economics as a procurement problem in isolation is running the other half. The firms that have integrated the two sides — and we count fewer than thirty-five of them across our 1,200-company panel — are the firms whose 2026 plans will hold. Everyone else is hiring against a comp band the integrated firms set, on a calculation the integrated firms run, with budget governance two cascade-cycles behind the frontier-lab calendar.
The next quarterly cut publishes August 1, 2026. The mid-year H1 flagship publishes June 15. The H2 token-economics update — which will incorporate the Blackwell/Rubin buildout data and the post-Q4 OpenAI-reset comp band — publishes November 7.
For the funnel mechanics behind the frontier-lab senior-research hiring throughput, see How Anthropic restructured its talent stack. For the compensation cascade origin point and the OpenAI Q1 reset that triggered it, see The OpenAI compensation reset. For the integrated enterprise framework Microsoft has built around its per-token cost basis, see Microsoft's AI hiring framework. For the founder cohort closing these offers, see Top 30 AI Founders to Watch in 2026. For the geographic context of the sovereign-AI compensation premia, see Top 10 AI employers in the UAE, Top 10 AI employers in Saudi Arabia, and Top 20 Bangalore AI engineers to watch. For the M&A side of the calculation, see Top 10 AI acqui-hires inside the Fortune 500 and Top 20 AI acqui-hires of 2026. For the broader compensation-tier picture, see Top 20 highest-paid AI roles. And for the parent flagship this report builds on, see The State of AI Hiring — Q2 2026.