Six trillion dollars of compute for a labor shock that hasn't arrived yet.
The world is building the largest capital project in the history of technology to automate work. So far, the measurable hit to jobs is counted in the tens of thousands — not the tens of millions. The distance between what's projected and what's realized is the most important number in the economy. Here is what the data actually says — and how we read it as investors in the green economy.
Between 2026 and 2028, artificial intelligence presents a paradox that most headlines get exactly backwards. On the supply side, capital is being deployed at a scale with no precedent: an estimated $6.7 trillion of data-center investment by 2030, with AI-specific compute capacity growing roughly 3.5× to make up about 70% of all data-center capacity.1 On the demand side — the actual labor market — the realized impact through early 2026 is small enough to be a rounding error: Goldman Sachs estimates AI has trimmed U.S. monthly payroll growth by about 16,000 jobs and lifted unemployment by roughly 0.1 percentage points.3
This is not a contradiction. It is a timing problem. The exposure is genuinely vast — the IMF puts ~40% of global employment in AI's blast radius,4 Goldman the equivalent of 300 million full-time jobs2 — but exposure is not displacement. Roughly half of exposed work is expected to be augmented, not replaced, and even Goldman's aggressive base case sees only 6–7% of workers displaced over a full decade of adoption.3 The next three years are the hinge: capital has been committed, capability is arriving, but the labor and productivity effects are still climbing the early, flat part of the J-curve.
This briefing maps that hinge in seven moves — the capital buildout, the exposure map, the gap between projection and reality, the industries pulling ahead, the specific jobs that win and lose, and the skills being repriced in real time. Every figure is sourced to a named primary institution and dated. Where the number is a model, not a measurement, we label it as one.
The most expensive bet in the history of technology
Start with the money, because the money is the part that is already real. In its 2025 analysis The cost of compute, McKinsey projects that meeting expected demand will require the world's data centers to absorb roughly $6.7 trillion in capital expenditure by 2030. About $5.2 trillion of that is for AI-capable infrastructure specifically; the remaining $1.5 trillion is conventional IT.1 Total data-center power capacity is projected to grow from about 82 gigawatts in 2025 to 219 GW by 2030 — a near-tripling — with the AI-specific slice rising from 44 GW to 156 GW, a 3.5× jump.
Capital at this scale only makes sense if demand follows, and adoption data suggests it is. NVIDIA's State of AI 2026 survey (fielded late 2025 across 3,200+ organizations) found 64% of organizations reporting active AI usage, another 28% in an assessment phase, and only 8% with no plans at all. Among large enterprises with 1,000+ employees, active usage rises to 76%.7 The frontier is shifting from assistive chatbots toward agentic AI — systems that execute multi-step tasks autonomously — which analysts expect to become a mainstream enterprise pattern by 2028.9
The capital has been committed. Capability is arriving on schedule. The only variable left is how fast that capacity converts into changed work — and that conversion is where every forecast diverges.
What that $6.7 trillion buys, if the optimistic case holds, is enormous. Goldman Sachs' widely-cited 2023 estimate is that widespread generative-AI adoption could eventually raise annual global GDP by about 7% — roughly $7 trillion — and lift labor-productivity growth by around 1.5 percentage points per year over a decade.2 That is the prize. The rest of this briefing is about who captures it, who pays for it, and how quickly.
Forty percent of the world's jobs are in the blast radius
The single most important — and most misread — statistic in this debate comes from the IMF's January 2024 staff analysis: almost 40% of global employment is exposed to AI. That exposure is wildly uneven by development level. In advanced economies it reaches about 60% of jobs; in emerging markets roughly 40%; in low-income countries about 26%.4 Richer economies have more to gain and more to lose, because their jobs are concentrated in exactly the cognitive, task-based work that language models touch.
Here is where the reading usually goes wrong. "Exposed" is not "eliminated." The IMF splits exposure into two halves that point in opposite directions. Of the ~60% of exposed jobs in advanced economies, roughly half are candidates for complementarity — AI makes the worker more productive, raising demand for their labor. The other half face substitution — AI does enough of the task to lower labor demand, wages, and hiring.4 The same technology is a tailwind for one worker and a headwind for the next, and which one you get depends almost entirely on whether the AI amplifies your judgment or replaces your keystrokes.
There is a second, more unsettling wrinkle. Every prior wave of automation hit middle-skill routine work hardest — factory lines, back-office processing. Generative AI is different: its reach extends up the wage ladder into higher-paid cognitive roles. The IMF finds that where AI strongly complements high earners, those workers capture a more-than-proportional increase in labor income — which means AI is positioned to widen income inequality, rewarding those best placed to direct it while squeezing those whose tasks it simply absorbs.4
Every giant headline number — 40% of jobs, 300 million roles, two-thirds of occupations — measures exposure to task automation, not jobs eliminated. Conflating the two is the single most common error in AI-and-work commentary. About half of exposed work is expected to be augmented.
The model says millions. The payroll says thousands.
Put the biggest forecast next to the latest measurement and the gap is the whole story of 2026–2028. Goldman Sachs' modeled base case is that full AI adoption unfolds over roughly a decade, during which 6–7% of workers are displaced (a wide uncertainty band of 3–14%) and unemployment rises by about 0.6 percentage points before productivity gains re-absorb the slack.3 Now the measurement: as of April 2026, the realized drag Goldman can attribute to AI is a net reduction of about 16,000 jobs per month in U.S. payrolls (≈25k substitution losses offset by ≈9k augmentation gains) and roughly 0.1 points of unemployment.3
Why the gap? Because deployment is not adoption, and adoption is not reorganization. A model can be bought in a quarter; the rewiring of workflows, the retraining of teams, the rebuilding of trust and compliance around autonomous systems takes years. This is the classic productivity J-curve: measured output stalls while organizations absorb a general-purpose technology, then rises steeply once the complementary changes are in place. Economists disagree sharply on the curve's ultimate height — Goldman models a ~15% long-run U.S. productivity lift, while MIT's Daron Acemoglu estimates something closer to 1% of cumulative total-factor productivity over a decade.3 That is more than an order of magnitude of disagreement, and it is unresolved.
For the next three years, the honest base case is not mass unemployment. It is a slow, uneven redistribution — hiring freezes at the exposed margin, not layoffs at the core — while the big productivity payoff stays mostly on the forecast page.
The near-term risk is therefore not a cliff but a ratchet: roles quietly not backfilled, entry-level rungs thinning, "we'll wait and see if AI can do it" becoming the default answer to a hiring request. That shows up in the exposed occupations long before it shows up in the unemployment rate — which is exactly what the industry-level data below already reveals.
The gap between exposed and unexposed industries is already a chasm
If you want to see AI's economic effect before it reaches the unemployment statistics, look at productivity by industry — and the divergence is already stark. PwC's 2025 Global AI Jobs Barometer, built on analysis of nearly one billion job advertisements, found that industries most exposed to AI (financial services, software publishing, information services) saw revenue-per-employee growth nearly quadruple — from 7% over 2018–2022 to 27% over 2018–2024. The least-exposed industries (mining, hospitality, construction) barely moved, slipping from 10% to 9%.6
Crucially — and this cuts against the "robots take the jobs" narrative — AI exposure at the industry level has so far correlated with more employment, not less. A 2026 WEF analysis found that industries with AI exposure one standard deviation above their peers saw, in a single year, about 10% higher productivity, 3.9% higher job growth, and 4.8% higher wage growth.10 In the near term, the winning industries are hiring and paying more, because they are growing faster than automation is shrinking their task load.
Where the wins and losses land
The pattern is consistent across sources. The industries positioned to win are those where AI amplifies an expensive expert: financial services, enterprise software and IT, professional and business services, biotech and drug discovery, and the entire compute-and-energy supply chain feeding the buildout. The industries under structural pressure are those built on standardized information handling: business-process outsourcing and call centers, routine back-office and clerical operations, and entry-level content and data processing — the work most cleanly described in a prompt.
Industries gaining ground
- Financial services7→27%
- Software & IT servicesexposed ▲
- Compute, semis & energy$6.7T
- Biotech & drug discovery▲
- Professional / advisory▲
Industries under pressure
- Call centers & BPO▼
- Back-office & clerical ops▼
- Data entry & processing−26%
- Routine content production▼
- Least-exposed, low-growth sectors9%
Which roles grow, which shrink, and which quietly disappear
Zoom from industries to individual occupations and the picture sharpens. The World Economic Forum's Future of Jobs 2025 — built from surveys of 1,000+ employers covering 14 million workers across 55 economies — nets out to +78 million jobs worldwide by 2030: 170 million created against 92 million displaced, with total churn equal to 22% of all jobs.5 An important caveat travels with that number: it reflects all structural forces — demographics, the green transition, geoeconomics — not AI alone. But AI/information-processing is the single most-cited driver, named as transformative by 86% of employers.5
The composition matters more than the net. The fastest-growing roles by percentage are almost entirely technical: Big Data Specialists, FinTech Engineers, AI and Machine-Learning Specialists, Software Developers. The largest absolute gains, by contrast, come from frontline work AI does not touch — farmworkers (the single biggest absolute gain, driven by the green transition), delivery drivers, and care workers.5 Meanwhile the largest declines fall squarely on clerical and secretarial roles: cashiers, administrative assistants, and — declining fastest in percentage terms — data-entry clerks, bank tellers, and postal clerks.
The U.S. Bureau of Labor Statistics' 2024–34 projections put hard numbers on the extremes. The fastest-declining occupation in the country is projected to be word processors and typists, shrinking −36%; data-entry keyers follow at −26% (a loss of ~36,700 jobs). At the other end, the fastest-growing occupation is wind-turbine service technicians at +50%, with solar-panel installers close behind — green-economy roles, not AI roles.8 The through-line: what disappears is codified information handling; what grows is either deep technical judgment or physical presence.
Roles on the rise
- Wind-turbine service techs BLS · fastest-growing+50%
- AI & machine-learning specialists▲ %
- Big-data & FinTech engineers▲ %
- Software developers▲ %
- Care workers & delivery drivers▲ abs.
- Farmworkers largest absolute gain+34M
Roles in decline
- Word processors & typists BLS · fastest-declining−36%
- Data-entry keyers / clerks−26%
- Bank tellers▼ %
- Postal service clerks▼ %
- Cashiers largest absolute decline▼ abs.
- Administrative assistants▼ abs.
The jobs that vanish are not the hardest jobs. They are the most describable ones — the work you could fully specify in a memo.
The market is already repricing skills in real time
You do not have to wait for the labor market to reorganize to see AI's effect on the value of skills — it is visible in wages right now. PwC's 2026 Barometer reports that the average wage premium for workers with AI skills reached 62% in 2026, up from 57% the year before — meaning an AI-fluent candidate commands, on average, 62% higher advertised pay than an otherwise-comparable peer without those skills. The premium ranges from about 16% in government and public-sector roles to as high as 118% in consumer markets.6 Jobs requiring AI skills are growing roughly 8× faster than the overall market.6
And the ground is moving fast beneath every worker, AI-skilled or not. The WEF projects that 39% of workers' core skills will be transformed or outdated by 2030 — though, notably, that figure is down from 44% in the 2023 survey and 57% in 2020, suggesting the pace of disruption, while high, may be stabilizing as adoption matures.5 The fastest-growing skills the WEF identifies are, in order: AI and big data, then cybersecurity, then technological literacy — followed closely by distinctly human capabilities like creative thinking, resilience, and flexibility.5
Value is concentrating at two ends: deep technical AI fluency (building, directing, and governing the systems) and irreducibly human judgment (creativity, complex communication, care, physical dexterity). The squeeze is on the middle — the standardized cognitive work that sits between them.
What the next three years most likely hold
Synthesizing the sources, the base case for 2026–2028 is neither the utopia of the capex decks nor the apocalypse of the viral threads. It is a reallocation — capital, tasks, and wage premia moving faster than headcount.
1 · Capital keeps flowing; the payoff lags
The $6.7-trillion buildout continues, because the option value of AI capability is too large to sit out.1 But the productivity dividend — Goldman's ~15% or Acemoglu's ~1%, the truth likely between — arrives on a J-curve, and 2026–2028 is mostly the flat part. Expect a widening gap between AI's valuation and its measured GDP contribution, and a live debate about whether the buildout is overbuilt.
2 · The labor effect is a slow ratchet, not a cliff
Realized displacement stays modest in aggregate — tenths of a point on unemployment, not whole points.3 The mechanism is hiring restraint at the exposed margin: roles not backfilled, entry-level pipelines thinning, "can AI do this first?" as the new default. Clerical, data-entry, and routine content roles keep eroding; the pain is concentrated and often invisible in national averages.
3 · The winners compound
AI-exposed industries and AI-skilled workers pull further ahead — higher productivity, higher wages, and (for now) more hiring, not less.610 Inequality widens along the axis of who directs AI versus who is directed by it. The 62% wage premium is an arbitrage that rewards early skill-building and punishes delay.
4 · Geography diverges
Advanced economies capture most of the upside and absorb most of the disruption (60% exposure); lower-income economies see less of both (26%), risking a new productivity gap even as they are partly insulated from near-term displacement.4
Why this shapes how we build and back companies
We read this data as a green-economy investor, and three implications sit at the center of our thesis. First, the durable value is at the intersection of AI and the physical world. The roles and industries that grow are precisely the ones AI amplifies rather than absorbs — energy and water, agriculture, health, and mobility — where a model makes a scarce expert more productive but cannot replace the field, the farm, or the patient. That is the ground our portfolio is built on.
Second, the buildout is an energy story as much as a compute story. $6.7 trillion of data centers is a step-change in power and water demand — which turns the green transition from a cost centre into critical infrastructure for the AI economy itself.
Third, applied beats abstract. The premium accrues to those who direct AI toward a real problem. Across Kinetic, My Peptides AI, and My Farm AI, our conviction is the same: pair frontier AI with a domain that has physical stakes, and you capture the upside of this decade while sitting on the resilient side of the exposure line.
For 2026–2028: the money is real, the exposure is real, the mass displacement is not — yet — and the smartest position is to be the worker or the firm that directs AI rather than the one whose tasks it quietly absorbs.
How to read these numbers
This briefing synthesizes twelve independently verified findings drawn from ten primary institutional sources, each fact-checked against the underlying document. The most important interpretive cautions:
Primary sources
- McKinsey & Company — "The cost of compute: A $7 trillion race to scale data centers" (2025). mckinsey.com
- Goldman Sachs — "Generative AI could raise global GDP by 7%" (Briggs & Kodnani, 2023). goldmansachs.com
- Goldman Sachs — "How will AI affect the US labor market?" (incl. Peng, US Daily, Apr 2026). goldmansachs.com
- IMF — "Gen-AI: Artificial Intelligence and the Future of Work," Staff Discussion Note SDN/2024/001 (Cazzaniga et al., Jan 2024). imf.org
- World Economic Forum — Future of Jobs Report 2025 (Jan 2025). weforum.org
- PwC — Global AI Jobs Barometer, 2025 & 2026 editions. pwc.com
- NVIDIA — State of AI Report 2026 (survey fielded Aug–Dec 2025). nvidia.com
- US Bureau of Labor Statistics — Employment Projections 2024–34 (2025). bls.gov
- Gartner — Agentic-AI adoption forecasts to 2028 (via industry reporting, secondary). industry reporting
- World Economic Forum — "How AI is improving wages and job quality" (Feb 2026). weforum.org
A GreenLeafSource Research briefing · Compiled July 2026 · Every figure sourced to a named primary institution and dated. Projections are labeled as models; measurements as measurements.