The AI/ML band split in two

“AI/ML” has become a band that describes two jobs pretending to be one. At one end sits established machine-learning engineering, feature pipelines, model training, MLOps, a deep, well-supplied talent pool whose price has behaved much as it did a year ago. At the other end sits applied generative work: people who've actually shipped retrieval systems, evaluation harnesses and agentic workflows into production, not just prototyped them. That pool is shallow, and the market has begun to price it accordingly.

The result is a band that split. The midpoint you quote now sits in a valley between two populations, describing neither. That's why the number feels simultaneously too high for the candidates you can find and too low for the ones you want.

The distinction matters because the market it sits inside isn't soft. India's Net Employment Outlook was +68% for Q2 2026, the strongest reading since 2008, yet 82% of employers reported difficulty finding the skills they need, according to ManpowerGroup's Employment Outlook Survey. High intent colliding with scarce supply is precisely the condition under which a single band fractures: demand pushes hardest against the thinnest part of the pool, and the price there detaches from the average.

The traditional-ML end has been supplied for a decade. India built that bench through the analytics and data-science boom, and the pipeline that feeds it, degree programmes, bootcamps, on-the-job MLOps, is mature. Depth keeps price honest: when many capable people can do the work, no one candidate can hold a premium for long. That's why this end has stayed close to where it was.

The applied-GenAI end is younger than the pipelines that would supply it. The capability the market now wants, production retrieval, evaluation, agentic systems, barely existed as a job 2 years ago, so almost nobody has 3 years of it. Supply is a lagging function of demand here, and the lag is the whole story. The aggregate signals confirm the direction: the World Economic Forum's Future of Jobs Report 2025 places AI and big data at the very top of the fastest-growing skills to 2030, and LinkedIn's Economic Graph reports the global economy added roughly 1.3 million AI-related jobs in two years, with India (+40%) among the fastest-growing AI-hiring markets. The constraint is proven supply at the production end.

Two adjacent skills, two different clocks. One end is priced by a deep, settled pool; the other by a scramble for the few people who've actually shipped. A single band assumes one clock. There are two.

The pull at the top has a cost in the middle. As the proven-GenAI end runs ahead, the mid-tier of the band gets compressed against a ceiling that's climbing faster than their pay. A strong mid-level engineer sees the offers going to the production-GenAI hires and reads the gap on their own req before you do. That's how a splitting band quietly hollows out the people you were relying on to grow into it.

This isn't a story about a few hot job titles. The gap sits under the whole AI labour market. Deloitte and Nasscom estimate India's AI-talent demand will exceed 1.25 million by 2027, up from roughly 600,000–650,000 in 2022, against a shortfall of about 50% in 2024, with only around 16% of IT professionals AI-skilled. When barely 1 in 6 practitioners can do any AI work, the subset who've run generative systems in production, at scale, is a fraction of a fraction.

Hiring volume is tracking the same asymmetry. Naukri's JobSpeak Index recorded AI/ML roles growing around 45% over the year, far ahead of the roughly 8% for white-collar hiring overall, while the senior 20-plus-LPA band rose about 16%. Postings surge fastest exactly where the proven pool is thinnest, which is the mechanical definition of a market re-pricing at one end. And on the supply side, India ranks first globally in AI skill penetration, per Nasscom's State of Data Science and AI Skills report, a strength that's real but sits mostly at the practitioner layer, not at the production-GenAI edge that scarcity is concentrated in.

India has more AI talent, and more concentrated, than almost anywhere. The tension is that proven production-GenAI talent is scarce inside an otherwise deep market, which is precisely how one band comes to contain two prices.

Global capability centres are where this concentrates. F500 India GCCs already hold more than 126,600 AI professionals, roughly 22.5% of the country's AI talent pool, according to ANSR's enterprise-AI workforce release. And LinkedIn's Economic Graph identifies Bengaluru specifically as a hub concentrating generative-AI roles. GCCs aren't dabbling at the edges of this market; they sit at its centre, which means they feel the fracture before anyone else does.

For a GCC building an AI function, the practical consequence is that the same requisition covers two hires the market prices differently. Staff the pipeline and the platform work from the deep pool at a fair rate. Staff the person who owns the generative system in production from a pool that barely exists, and be honest that this is a different negotiation. Run both through one band and the offer letter is wrong twice: generous to the hire you could have sourced cheaply, and short to the one you actually needed to win.

The parent organisation rarely sees this. It reads a single AI/ML budget line and a single market average and assumes they describe each other. They no longer do. The averages the market publishes are true and useless in the same breath: true about the pool, useless about the position.

The diagnosis lives in the proof behind the work. Ask three questions before you set a number. Has this person shipped a generative system into production, or prototyped one? Did they own the failure modes, the evaluation, the drift, the cost of a bad answer at scale, or hand them off? And is the scarce part of the role the modelling, which the deep pool supplies, or the production judgment, which it doesn't?

Where the answers point to pipelines, training and platform work, you're in the traditional-ML market: price it at a fair market rate, because depth of supply will hold that rate for you. Where they point to a live generative system with real users and real consequences, you're in the applied-GenAI market: the premium tracks demonstrated production experience. Two roles that read identically on a job board can belong to different labour markets, and the interview is where you find out which.

The centres closing GenAI hires cleanly have quietly abandoned the single band. They ask the sharper question, has this person shipped generative systems into production, at scale, with the failure modes that teaches, and they price the answer. The premium tracks demonstrated production experience.

Do that and two things happen. You stop over-paying for traditional ML capability you can source at a fair market rate, and you free up the headroom to actually close the scarce end. Splitting the band is how you spend the same budget more accurately, and in a market where intent is at a post-2008 high but 4 in 5 employers can't find the skills, per ManpowerGroup, accuracy at the scarce end is the difference between an open req and a closed one.