What the data shows
Constitutional AI alignment research sits 2.1 times above the $165K baseline — $250K–$380K base, $400K–$620K total compensation. That is the widest skills-level pay spread ENTRA has measured in a single index across the 2026 AI labor market. The second-ranked skill, RLHF, comes in at 1.9×. By rank 10, Prompt Engineering is at 1.15×. By rank 20, classic NLP is at 1.1×. The gap between the top of this ranking and the bottom is not ten salary points — it is a structural divide between skills that only a handful of employers in the world can absorb and skills that any enterprise software company can hire for with a two-week job posting.
The core finding of this ranking is that the pay premium hierarchy mirrors the frontier-to-commodity spectrum with near-perfect fidelity. Lab-specific capabilities — alignment research, RLHF, GPU kernel engineering, multimodal architecture — occupy the AAA tier with premiums above 1.75×. The key variable is employer concentration: these skills have 5–15 employers globally who can meaningfully deploy them at scale, which means every hire is a direct negotiation without a market comparator. Infrastructure skills — MLOps, vector databases, prompt engineering — cluster around 1.0×–1.2× because the employer base runs into the thousands and market-rate discovery is efficient.
The transparency paradox runs in the exact opposite direction. The highest-premium skills have the lowest salary disclosure rates. Constitutional AI: 19% of postings include a band. Foundation Model Pre-training: 18% — the lowest in this index. Classic NLP: 67% — the highest. This is not accidental. Frontier labs treat their compensation floors as competitive intelligence. A disclosed band for a Constitutional AI researcher communicates both the comp floor (which competitors will immediately use to calibrate counter-offers) and the headcount signal (how many roles exist at this level). The opacity is structural and deliberate. The practical implication for candidates: the higher the premium on a skill, the less market information is publicly available, and the more a candidate must rely on direct network intelligence and recruiter relationships to understand the actual comp range.
Three tiers of skill premium
Tier 1 — AAA (2.0× and above): Lab-monopolized skills. Constitutional AI, RLHF, GPU Kernel Engineering, and Multimodal Model Architecture constitute the AAA tier in this index. The defining characteristic is employer concentration: only 5–8 organizations globally can absorb these skills at a meaningful scale. Anthropic and OpenAI are the primary consumers of Constitutional AI and RLHF talent. NVIDIA and AMD anchor the CUDA/ROCm engineering market, alongside the AI-chip startup cohort (Cerebras, Groq, Tenstorrent). Google DeepMind, Anthropic, OpenAI, and Meta FAIR employ the vast majority of multimodal architecture talent. When the employer base is this narrow, compensation is set through direct negotiation rather than market discovery, and the result is a structural 2.0× floor that resists compression even as the broader AI job market normalizes.
Tier 2 — AA (1.5×–2.0×): Lab-plus-applied skills. AI Safety / Red-Teaming, Foundation Model Pre-training, and LLM Fine-Tuning constitute the AA tier. The employer base here is larger — 50–200 organizations — but still concentrated enough to sustain a material premium. LLM Fine-Tuning sits at the boundary between Tier 1 and Tier 2 in transparency terms: its 52% disclosure rate is the highest of any skill with an AA rating or above, reflecting the fact that fine-tuning roles are now appearing in Fortune 1000 hiring pipelines alongside the frontier labs, and Fortune 1000 postings are subject to state pay transparency mandates. The premium here (1.4×–1.7×) is real and durable in the near term, but fine-tuning specifically is at risk of Tier 3 migration as instruction-tuning toolchains commoditize.
Tier 3 — A through BB (1.0×–1.5×): Enterprise-broad skills. RAG, MLOps, Prompt Engineering, Vector Database Engineering, AI Agent Architecture, and the remaining skills in this ranking all occupy a world where thousands of employers can hire for the role. The premium is real — 1.1×–1.3× is still material on a $165K baseline, representing $18K–$50K in additional compensation — but it is compressing. Supply is scaling faster than demand across the entire Tier 3 cohort. The clearest case is classic NLP: from 1.4× in 2024 to 1.1× in 2026, a 0.3-point compression in two years driven by LLM displacement of classical pipelines. RAG and Vector Database Engineering are in earlier stages of the same trajectory.
The transparency paradox
The inverse relationship between pay premium and disclosure rate is the sharpest finding in this index. Skills with the highest premiums have the lowest transparency; skills with the lowest premiums have the highest transparency. This pattern holds across all 20 skills without exception.
The mechanism is straightforward: frontier labs compete for a fixed, tiny pool of practitioners whose skills are non-substitutable. Posting a salary band publicly communicates the comp floor, the headcount strategy, and the relative prioritization of the team. For a Constitutional AI team of eight researchers, none of that information should be public. For a Prompt Engineering team of forty, none of that information is sensitive. The employer type determines the disclosure incentive, and the employer type correlates directly with the skill tier.
State pay transparency laws (CO, CA, NY, WA, NJ, IL) are beginning to force partial disclosure even in Tier 1 and Tier 2 roles, but frontier labs have responded with qualification: they post ranges so wide as to be uninformative ($250K–$650K for a "Research Scientist" covers four functional levels and two career tracks), or they post roles in jurisdictions with no mandate and geo-restrict the role after the fact. The transparency laws are working at the margin — LLM Fine-Tuning's disclosure rate jump from 31% to 52% in two quarters is materially driven by transparency law compliance among non-frontier enterprise employers — but they have not yet penetrated the frontier lab compensation architecture.
Three standout data points
LLM Fine-Tuning has the fastest transparency growth in this index. The disclosure rate for roles explicitly requiring fine-tuning or instruction-tuning experience rose from 31% in Q4 2025 to 52% in Q2 2026 — a 21-percentage-point jump in two quarters. The driver is structural: fine-tuning has migrated from frontier labs to the Fortune 1000, and Fortune 1000 postings in CO, CA, NY, and WA are required to include salary ranges. As the skill democratizes, it enters jurisdictional compliance mandates that apply to general enterprise software hiring. By Q4 2026, ENTRA projects the disclosure rate for LLM Fine-Tuning will reach 60%+, converging with Tier 3 skills despite retaining a Tier 2 pay premium.
GPU Kernel Engineering has the widest total compensation range in this index. The base band ($210K–$320K) is relatively narrow — $110K from floor to ceiling — but equity components swing from $110K to $230K annually, producing a TC window of $320K–$550K, a $230K range. The asymmetry reflects NVIDIA's stock performance through 2024–25: engineers hired at the base-band floor on a standard NVDA RSU package realized total compensation materially above published Levels.fyi numbers as the stock appreciated. The skill's 1.8× median pay premium understates the p75 and p90 outcomes for practitioners inside NVIDIA's equity cycle. For candidates comparing cash-heavy frontier lab offers against NVIDIA equity packages, this asymmetry is the single most important data point in the negotiation.
Classic NLP's premium has compressed from 1.4× in 2024 to 1.1× in 2026. This is the LLM displacement thesis confirmed in compensation data. Classical NLP pipelines — tokenization, NER, relation extraction, parsing — have been systematically replaced by LLM-based approaches across enterprise applications, and employers have responded by reclassifying NLP roles, lowering comp bands, and redirecting hiring toward fine-tuning and prompt engineering. The practitioners most affected are those with pure classical NLP backgrounds who have not transitioned into the LLM ecosystem. The 67% disclosure rate — highest in this index — is itself a symptom: when a skill commoditizes, the employer base broadens into transparency-law jurisdictions and market-rate discovery becomes efficient. The comp compression is the price of commoditization, and for classic NLP, that compression is already underway.
METHODOLOGY
Data period: Q2 2026 (April–June 2026)
Sources: ENTRA Salary Survey Q2 2026 (n=2,140 AI practitioners); ENTRA Job Signal Index Q2 2026 (n=47,000+ AI postings monitored); Levels.fyi Q1-Q2 2026 offer corpus (n=3,800+ verified offers); Glassdoor Q2 2026
Scoring: 5 equally-weighted dimensions (20 pts each):
- Pay Premium Score — base salary premium above $165K US senior AI engineer median
- TC Premium Score — total compensation premium (base + equity + bonus)
- Demand Heat Score — YoY growth in postings requiring this skill
- Supply Scarcity Score — demand-to-supply ratio for verified practitioners
- Transparency Score — percentage of postings including salary bands
Rating bands: 90–100 AAA · 80–89 AA · 70–79 A · 60–69 BBB · below 60 BB
All data sourced from ENTRA proprietary monitoring; not based on self-reported surveys alone. First edition — all entries marked yoyDelta=NEW. Methodology inquiries: methodology@entracareers.com