2026–2028 - the agentic transition and the capex test#

← Part IV · Index

Models become reliable enough for multi-hour autonomous work in verifiable domains - the qualifier doing real work there, per the data asymmetry. What this period is not is a general-purpose agent economy. The commercial threshold crosses first where ground truth is cheap: software, customer support, parts of finance and insurance, first-pass legal. Everywhere else, the same models still sit behind a human checkpoint because the cost of being wrong has not fallen.

The rate-limiting factor is capital and reliability, not intelligence. The intelligence is already mostly here for the domains that will move; the question is whether revenue tracks the capex that bought it.

Why "multi-hour" is the right unit: autonomy has a compounding arithmetic. A system that is 99% reliable per step fails roughly one task in three by step 40; the same system with a cheap verification checkpoint every few steps resets the error budget and runs indefinitely. That is why the commercial frontier tracks verification cost rather than raw capability - domains with ground truth get checkpoints for free (the code compiles, the ticket closes, the claim reconciles), and domains without it pay for checkpoints in human review, which caps autonomy at whatever length a reviewer will tolerate. The threshold this period crosses is not a model getting smarter; it is checkpoint density in a handful of domains falling below the cost of the human it replaces. The failure mode of this framing: if long-horizon reliability turns out to improve faster than verification-cost logic predicts - models that self-correct mid-task without external ground truth - the qualifier dissolves and the general-purpose agent economy arrives inside this window rather than after it. That would show up first as autonomy lengths growing in verification-expensive domains, which B-family instrumented measures would catch.

What changes#

1. The purchasing unit shifts from seats to outcomes#

Per-seat pricing assumes a human operator per license. Outcome pricing does not. That break is the commercial signature of agentic reliability, and it is already visible in vertical tools that bill against tickets closed, claims processed, or pull requests merged rather than against named users.

The consequence for software business models is structural: seat-based SaaS was priced against headcount, and headcount is the thing that stops growing first. Firms that cannot reprice will watch ARR per customer fall while usage rises - a death spiral that looks like growth in the usage dashboards and contraction in the P&L.

2. Entry-level hiring in knowledge work contracts further#

This is the first period where the apprenticeship gap moves from prediction to measurable fact. The mid-2026 data already shows the leading edge: entry-level share of large-tech hires ~7%, junior software postings down 60%+ from peak, recent-graduate unemployment elevated against its long-run average. → Game 4

The discriminating test lands inside this window. If junior hiring fails to recover when aggregate white-collar hiring does - roughly 2027–28 - the substitution share of the decline was large, and the gap is structural rather than cyclical. That single observation is worth more than any amount of further cross-sectional analysis. → B1

Demography does not rescue this period. Aggregate labor-force shrinkage helps absorption later; it does not restore the junior rung if the junior tasks are gone.

3. Energy becomes the acknowledged bottleneck#

Electricity prices rise in datacenter-dense regions. The political backlash is a state-level force, faster than federal reform and harder to reverse. Utility commissions impose special rate classes, siting conditions, and in some markets de facto moratoria. That is Uncertainty 2 converting from abstract constraint into lived politics.

The state-level channel deserves the emphasis because its clock speed is different in kind. Retail rates are set in state rate cases, not federal ones; a commission responding to constituent bill complaints can impose a large-load tariff class in a single proceeding measured in months, while federal permitting reform moves in multi-year legislative cycles and grid-scale transmission in decades. And the asymmetry is one-directional: a rate class or siting condition, once granted, creates a settled expectation among ratepayers that no commission reverses cheaply. The practical consequence for the labs is that datacenter siting fragments into a fifty-jurisdiction map where the binding variable is local politics, which is exactly the environment that pushes hyperscalers toward owned generation and bilateral deals - the mechanism behind the energy 2029 prediction.

By late in this window the US rate-limiting step on frontier AI is closer to environmental review and interconnection than to chip supply - and no export control addresses it. The US is constrained by the US. → Energy, Bipolar

4. Capital markets test whether revenue is tracking capex#

The central question from Compute and Capital gets its first real answer. Hyperscaler 2026 capex guidance of ~$700–725B (mid-2026 company guidance; up from earlier ~$635–670B prints) against still-small external AI revenue is a bet that revenue compounds at 45–55% for years. Enterprise software has never done that at scale. Investor scrutiny of the spend is already rising in 2026 earnings seasons - the political economy of the capex test arrives before the financial event.

Probability of a significant AI-sector correction in this window: ~40%

The form matters more than the occurrence. Expect consolidation and culling of secondary labs, not a halt to capability progress. The technology does not un-invent itself; the cap table changes. And the proximate trigger is ~65% likely a credit event in AI-adjacent structured finance rather than an earnings miss at a hyperscaler - because the financing mix is migrating from operating cash flow toward private credit, SPVs, and vendor financing, and that structure turns a demand disappointment into a financial event. → A2, A3

A correction is differentially destructive. It culls commercially financed labs and leaves standing those financed strategically - Gulf sovereign wealth, Chinese state-directed capital, increasingly industrial-policy money. The post-correction map is more state-adjacent than the pre-correction one. → Game 2, Gulf

Why a correction would not slow capability much: the relevant precedent is telecom in 2001, where the equity holders of the fiber overbuild were destroyed and the fiber itself kept transmitting - the overbuilt asset repriced and became the cheap input for the following decade's applications. Trained models and datacenters have the same property: the marginal cost of running an already-trained model is small against what was spent creating it, so a financial event transfers ownership of capacity without retiring it, and inference prices fall into the downturn. The claim has a limit worth stating: unlike dark fiber, frontier capability requires continued training runs to advance, and those are exactly the discretionary spend a post-correction board cuts first. So the correction pauses the frontier race more than it pauses diffusion - deployment of existing capability may even accelerate on cheaper compute while the next generation slips.

5. Regional stress becomes visible where the export ladder is thinnest#

India's IT-services and BPO sector is the cleanest public test of Game 4 anywhere: revenue growing while headcount flattens, reported quarterly. That pattern solidifies in this window. Europe writes more of the rulebook while converting little of it into industrial capacity. Neither story is decisive yet; both become legible.

What does not happen in this window#

• No general-purpose humanoid at economic scale. Structured robotics (warehouses, ports, agriculture) advances; unstructured does not. The data problem has no shortcut. → Robotics
• No binding international compute agreement. Coordination remains event-driven; the event has not arrived. → Game 2
• No resolution of the Taiwan assumption. Everything downstream still inherits the ~90% modelling assumption that leading-edge fabrication continues. A disruption here invalidates the period entirely. → Bipolar, Uncertainty 4

What to watch#

Failure modes for this period's claims#

• If revenue clears ~$200B/yr run-rate by 2028 without a correction, the capex-wall argument was too pessimistic and the whole capital story softens. → A2
• If junior hiring recovers with the cycle, the substitution share was overstated and Game 4's sharpest claim needs revision.
• If permitting reform compresses interconnection materially, Uncertainty 2 resolves toward faster and the US capacity story improves relative to the base case.
• If the first salient AI-attributed incident lands early (2027 rather than later), the regulatory architecture of the 2030s gets written during this window, not after it - and every subsequent period inherits a different institutional regime.

Evidence family for this period is financial. Score revenue, financing mix, and junior hiring - not TFP prints or humanoid unit sales. Wrong evidence family is a category error (timelines hub).

Related: Capital · Game 4 · Indicators A · Part V row 6

View markdown source