The AI economy is accelerating. Who gets to build it?

Buying access is not transformation. High performers redesign workflows; most organizations still bolt AI onto old ways of working.

Black-equity relevance: Not explicit in the survey; the gap matters for smaller Black-led institutions that cannot afford failed adoption.

Forecasts differ, but the preparation agenda does not: build better indicators, train people, and make public systems responsive.

Black-equity relevance: Not explicit; the scenarios help leaders prepare institutions before unequal shocks arrive.

The gap is not just access to software. About half of workers say training and a clearer understanding of usefulness would encourage adoption.

Black-equity relevance: Not explicit; the age, education, and training gaps identify where equitable access work can begin.

This turns “longer tasks” into “whole systems.” Epoch also warns that long autonomous runs change evaluation budgets and discloses possible memorization risk.

Black-equity relevance: Not explicit; it demonstrates the scale of technical capacity institutions can now access.

Today’s frontier is perishable. This project keeps the evidence linked, ranked, and updateable rather than freezing progress into one score.

Black-equity relevance: Not explicit; transparent evidence infrastructure helps institutions make independent choices.

It shows why institutions need a continuing evidence practice rather than a single frozen score or annual handout.

Black-equity relevance: Not explicit in the paper; useful as infrastructure for transparent, continuously updated decision-making.

Capability is becoming affordable, and task-targeted model routing can beat dependence on a single frontier system.

Black-equity relevance: Not explicit; falling costs lower the entry barrier for Black-led organizations and firms.

The AI economy is being built in physical infrastructure now. Who supplies, owns, powers, secures, and maintains it is an open economic question.

Black-equity relevance: Not explicit; it surfaces ownership, contracting, workforce, and supplier opportunities.

Transformation has costs: retraining, consulting, workflow change, and measurement. This commentary is a useful skeptic’s check, not a primary report.

Black-equity relevance: Not explicit; it warns against promising productivity gains without funding the transition work.

This is the compounding question in measured form: if AI assistance doubles the pace at which frontier models improve on well-specified tasks, the window for institutions to build skills, ownership, and oversight shrinks at the same rate. METR saw no evidence of significant bottlenecks in algorithmic progress, training compute, experiment compute, or data that would take effect in the next 12 months — while stressing it is not confident in that bottleneck assessment.

Black-equity relevance: Not explicit — a capability assessment at one frontier lab, not a social study. The guide reads it for the timing question: acceleration at the frontier widens the cost of late adoption for smaller institutions, including Black-led firms and nonprofits, unless access and training move just as fast.

Earlier cards priced the buildout and asked who pays for its power (cards 27, 33, and 36). This is the consent layer arriving: in community after community, the default answer at the ballot box is now no — or not without voter approval, community-impact standards, and financial guarantees, as in Perry, Ohio, Issue 16 for hyperscale developers seeking tax incentives. Any institution negotiating over siting, power, or tax breaks is negotiating with voters who have just shown their hand.

Black-equity relevance: Not race-specific — Ballotpedia’s tally is by place and measure type, not by race. The guide reads it for the siting question: the communities asked to host the buildout’s land, water, and power costs are gaining a formal veto, and who holds that veto — and what guarantees they can demand — decides whether host communities share in the buildout or only bear it.

Read with the Canadian hiring-flow study (card 39) and the Danish administrative study (card 31), this is the third large dataset pointing the same way: the early harm is not mass junior layoffs, it is a narrowing junior door while senior hiring grows. The remedy follows from the mechanism — if adopting firms expand at the top and stall at the entry level, institutions preparing new workers have to build the bridge experience that the market has stopped supplying on its own.

Black-equity relevance: Not explicit — a 41-country workforce study, not race-disaggregated. The guide reads it for the entry channel: a senior-tilting hiring shift rewards workers who already hold experience and networks, and penalizes first entrants — a position Black young workers disproportionately occupy after every previous downturn.

This is the capability ceiling measured on real knowledge work rather than exam-style benchmarks, by the people who build the benchmarks. In the researchers’ words, current models “cannot yet replace workers, at least not at Epoch” — and the failure mode matters as much as the verdict: models missed implicit house standards despite ample reference material, and treated results from flawed experiment setups as key findings. That is a deployment rule for any institution in this guide: use agents on well-defined work, keep human judgment on the open-ended parts, and never let a system grade its own homework.

Black-equity relevance: Not explicit — a capability measurement on one research firm’s own tasks, not a social study. The guide reads it for the adoption question: the judgment gap is exactly where smaller institutions, including Black-led firms and nonprofits, should concentrate scarce human oversight rather than automating it away.

This is the corrective card for the entry-level panic: weak youth hiring is real, but in Canada’s flows data it is broadest where AI exposure is lowest, and the coding slide predates the tool blamed for it. Read with the Fed’s Beige Book (card 32) and the Danish administrative study (card 31), the pattern across three countries is consistent — headline pay and headcounts are still, hiring doors are narrowing for other and earlier reasons, and blaming AI alone misdirects the remedy. Institutions preparing young workers should fix the on-ramp, not just warn about the robot.

Black-equity relevance: Not explicit — a Canadian workforce study, not race-disaggregated. The guide reads it for the entry channel: when youth hiring slows across the board, the cost lands first on new entrants with the thinnest networks and credentials, a position Black young workers disproportionately occupy after every previous downturn.

The automation timeline is not set by the technology alone — it runs through the incentives of the people asked to build their own replacement. Any institution deploying AI with its own staff in the loop should expect the training itself to become a negotiation, and should price completion honestly rather than discovering the slowdown later.

Black-equity relevance: Not explicit — a theory model with no race or demographic breakdown. The guide reads it for the bargaining question: whether workers who train automating systems are compensated for that training is a distributional choice made before displacement, not after it.

A downloadable model — one anyone can run, modify, and strip of safeguards — now sits about half a year behind the best closed systems at cybersecurity tasks. Defensive planning can no longer assume capable cyber AI stays behind a provider’s gate; the window between frontier capability and public availability is measured in months.

Black-equity relevance: Not explicit — a capability measurement, not a social study. The guide reads it for institutional readiness: smaller organizations, including Black-led firms and nonprofits without dedicated security teams, face frontier-class offensive tooling on the same six-to-seven-month clock as everyone else.

Previous cards priced the buildout in totals — $31.6 trillion through 2050 (card 27), capex up 92% (card 11). This one prices it by the megawatt, and the anatomy matters: nearly four of every five dollars go to servers and networking, not construction. The supplier opportunity therefore sits mostly in IT equipment, while the remaining fifth — switchgear, generators, cooling — is the industrial layer where new entrants can still compete.

Black-equity relevance: Not explicit — a market costing, not race-disaggregated. The guide reads it for the contracting question: at $49 million per megawatt, who supplies the servers, switchgear, generators, and cooling — and whose firms are positioned to bid — decides who participates in the largest buildout of the decade.

The strongest administrative evidence yet says the first two years of generative AI reorganized work instead of eliminating it — and the measurable gain ran through workers who picked up the tools and used them to move. That is the guide's work lane in data form: adoption paired with training is a mobility strategy, and the open risk sits with early-career entrants, whose shrinking openings this study pointedly does not blame on adopting firms.

Black-equity relevance: Not explicit — a Danish workforce study, not race-disaggregated. The guide reads it for the positioning question: mobility gains appeared where workers could choose their own tools and where formal credentials did not bar entry, which is exactly the on-ramp structure Black workers are most often denied.

This is what early displacement looks like in the Fed's own record: not mass layoffs, but attrition left unfilled, hiring delayed, and jobs never created — a channel that barely moves headline employment while quietly narrowing the door for people trying to enter or re-enter work. Read alongside the Brookings null result (card 31), the two sources agree on the surface and differ on the current underneath: pay and headcounts are still, hiring intentions are not.

Black-equity relevance: Not explicit — the Beige Book is not race-disaggregated. The guide reads it for the entry channel: when firms cut by attrition and delayed hiring rather than layoffs, the cost lands first on job-seekers and new entrants, groups in which Black workers are overrepresented after every previous downturn.

Card 27 established that power decides where AI capital lands. This report establishes the fight underneath: who pays for the shared grid that power requires. A community that cannot see the deal (NDAs), cannot verify the jobs (no comprehensive evidence), and shares the upgrade bill through its electric rates is subsidizing the buildout whether it consented or not — which makes cost allocation, not just siting, a question for any institution negotiating with the AI economy.

Black-equity relevance: Not race-specific — the report frames the burden on residential ratepayers, taxpayers, and local communities generally. The guide reads it for the distribution pattern: grid costs spread across all ratepayers fall hardest on households with the least cushion, and Black communities have historically had the least say in siting and rate decisions of exactly this kind.

The number that matters for the guide's work lane is not 11 million — it is 1 in 7. Almost every worker has some route to growing work; the scarce resource is a direct one. That reframes the workforce question from "will there be jobs" to "who gets a passable on-ramp," which is a training-and-credentialing question the audience can answer institutionally.

Black-equity relevance: Not explicit — the report is not race-disaggregated. The guide reads it for the distributional question: the jobs most likely to require transition are the lowest-paid, and the transitions themselves demand credentials 85% of growing jobs require.

The capability ceiling matters for every agent plan in the guide: executing a workflow is not the same as knowing when its conclusions hold up. Epoch's verdict — the agents' discoveries were "underwhelming by the standard of human-led AI research" — is a calibration point for any deployment that asks AI to verify its own work.

Black-equity relevance: Not explicit — the test is capability measurement, not a social study. The guide reads it as a guardrail: oversight stays human, and anyone deploying agents in community-facing systems should keep it there.

The bottleneck moved from chips to power and financing: grid interconnection takes four to ten years in many regions while a data center builds in two to three. Regions with long queues cede share regardless of tax incentives. The infrastructure bet is an energy bet — whoever controls the power map captures the capital.

Black-equity relevance: Not explicit — the report is not race-disaggregated; the guide reads it for the positioning question: who owns the land and power the buildout runs on, and whose communities sit at the front of the interconnection queue.

The first benchmark evidence at the weeks-long-project timescale. The question moved from "can models write code" to "which systems are specified tightly enough for an agent to match behavior exactly" — specification and supervision are the new leverage, not keystrokes.

Black-equity relevance: Not explicit — the benchmark measures capability; the guide reads it for the Black-worker question: who can specify and supervise such agents, and who absorbs the cost when a 99.95% run ships a silent defect.

Two datasets, two stories: a genuine bidding war for talent that directs AI tools, and real wage compression for workers whose tasks overlap with the tools. Which story a worker lands in depends on whether they direct the tools or compete with them — positioning, not exposure, is the variable.

Black-equity relevance: Not explicit — neither dataset is race-disaggregated here; the guide reads the two-tier pattern for what it implies about Black workers' positioning in a reshaped labor market.

The bottleneck is no longer the model — it is power, chips, and organizational change. Whoever owns the infrastructure and does the workflow redesign captures the value.

Black-equity relevance: Not explicit; the finding that gains come from workflow redesign and infrastructure — not access alone — warns institutions that cannot afford failed adoption.

Headline warnings about AI wiping out jobs are running ahead of what firms themselves report. The displacement story is slower and more uneven than the rhetoric — which matters for any workforce plan built on the assumption that the cuts have already happened.

Black-equity relevance: Not explicit. But with 39% of firms still expecting AI-driven cuts ahead, the coming wave lands on workers — and the National Urban League's 2026 data show Black workers already absorbing early losses (see card 21).

Every capability claim in this guide rests on a benchmark somewhere. This is the first standing auditor telling the field which tests can be trusted — and which cannot. Capability numbers should now carry a second citation: the verdict on the test.

Black-equity relevance: Not explicit; but trustworthy measurement is a precondition for any fair evaluation of AI systems deployed in hiring, lending, and housing.