The Slope Is the Cycle

We are watching three exponential curves intersect. Capex spending accelerating beyond historical precedent. Revenue growth climbing in ways no software market has seen before. And capability improvement that keeps bending upward with no visible ceiling. They are not independent. They are interdependent. The system is fragile because it depends on all three staying in sync. If one bends while the others keep climbing, the tension snaps.

The number that starts the unease

FY hyperscaler capex

$357.5B

Combined ARR

$55.0B

Latest FY hyperscaler capex is shown against combined OpenAI + Anthropic ARR. This is a scale comparison, not a complete ROI model: total hyperscaler capex is not AI-only capex, and OpenAI + Anthropic ARR is not the whole AI economy. The point is the burden. The buildout has to become monetization, productivity, or durable scarcity before it becomes ordinary depreciation.

There is tension in that gap you see in the chart, and it is not sustainable. A bill this large set against revenue this small is not a steady state. It is a disequilibrium that will eventually resolve.

There are really only two ways out. Either revenue climbs until it justifies the spend — through monetization, productivity, or durable scarcity — or pressure mounts on those bankrolling it: buildouts stall, capacity gets written down, financing strains, and in the worst case some of the players bankrolling this do not survive it. One of those has to happen. The chart is a picture of a system that cannot stay where it is.

Revenue keeps the story alive

• 2025-12 — Anthropic company-stated — $9B ARR
• 2026-01 — Anthropic company-stated — $14B ARR
• 2026-02 — OpenAI reported-leak — $25B ARR
• 2026-04 — Anthropic company-stated — $30B ARR

OpenAI and Anthropic ARR over time. Company-stated and reported-leak figures are tagged on the chart; earlier points are analyst-estimated from public aggregators. This slope is the bullish pressure relief valve against the buildout.

This slope is without precedent. Month-over-month revenue growth at this scale and velocity has no historical analog. You cannot model where exponential growth settles without seeing the inflection. And we have not seen it yet. That is what makes the capex return on investment so hard to model. The revenue is real and breathtaking, but it is growing so steeply that nobody can confidently say where it lands. The future is genuinely wide open. That is what makes the system so hard to plan for.

Log Linear

Snapshot: 2026-04-07 · metr.org YAML

Each dot is one frontier model release; larger dots were state-of-the-art at the time. Log scale shows the exponential as a near-straight line; flip to linear to see the cliff that the same data describes. The dashed trend uses METR's 2023-onward fit (point estimate 129 days, about 4.3 months per doubling), with the all-time fit near 188 days. Source: METR; see the linked YAML for methodology and CI bounds.

The METR Time-Horizon benchmark is what I prefer to look at because it measures against real humans, not abstract scores. It shows how long a software-engineering task a frontier model can complete at an 80% success rate. The unit is minutes or hours of human-equivalent work.

On a log chart, exponential growth looks like a tidy line. Switch to linear and you see what is actually happening: the curve bends upward with no visible ceiling in sight. That gap — between what appears manageable on the log scale and what actually might not stop — is where the uncertainty lives. And that uncertainty is adding fuel to the capex cycle.

One caveat: METR’s suite is heavy on software, ML, cybersecurity, and math tasks — not a complete picture of AI capability across all domains.

Still, the current read is clear enough for this framework. As of this snapshot, the capability axis is Steep.

Demand is real. It is also narrow.

A misleading question to ask would be whether demand exists. It does. Developers are paying, enterprises are experimenting, API usage is real, and compute is still scarce enough that capacity gets absorbed.

The harder question is whether demand stays concentrated or breaks into breadth. Right now it is trapped in one labor pool: Computer and Mathematical work absorbs roughly 38-40% of Claude conversations in the blended series. Everything else sits far below.

That concentration either broadens or it doesn’t. One: the models plateau on capabilities that transfer to other domains — they stay best-in-class for coding and math, but other sectors need breakthroughs in data, training regimes, or new algorithmic approaches before AI becomes useful there. Demand stays narrow and the capex cycle stalls. Or two: capability keeps exploding higher and spills over into operations, healthcare, logistics, finance, compliance — the domains that employ millions — and demand multiplies economy-wide. One of those has to happen.

• Management 3.5%
• Business and Financial Operations 3.8%
• Computer and Mathematical 38.3%
• Architecture and Engineering 1.8%
• Life, Physical, and Social Science 4.9%
• Educational Instruction and Library 7.7%
• Arts, Design, Entertainment, Sports, and Media 7.2%
• Office and Administrative Support 12.3%

Source: Anthropic Economic Index raw data releases V1–V5 (Hugging Face: Anthropic/EconomicIndex), task usage aggregated to SOC occupation major groups. Blended series = unweighted mean of Claude.ai and 1P API task usage share where both exist; Claude.ai only for Jan–Mar 2025, before 1P API data was collected. Usage share = percentage of sampled conversations/records whose task maps to that occupation group — not headcount or spending.

That concentration should not be overread. This is Claude usage share by task category, not all AI adoption, not labor-market replacement, not spending share, and not productivity. But it points to the central demand-breadth problem.

Past and Present Capex Cycles.

A capex cycle is what happens when whole economies decide to build big on physical infrastructure like railroads, fiber networks, power plants, data centers. The buildout accelerates, capital flows fast, and usually at some point someone looks out and realizes they’ve built enough or maybe even a little too much. What plays out from there has many diverging paths.

The chart lines up AI data center spending against six of history’s great buildouts.

• AI / data centers 3.08% peak 2028
• UK railway mania 6.7% peak 1847
• US railroads 3.7% peak 1881
• US telephone (Bell) 0.41% peak 1906
• US electric utilities 0.69% peak 1912
• US telecom / fiber 1.18% peak 2000
• US shale / fracking 1.03% peak 2014

AI capex is small as a share of GDP — but it carried most of 2025's GDP growth.

Investment in information-processing equipment & software is ~4% of GDP, yet it was responsible for ~92% of US GDP growth in the first half of 2025; GDP excluding those categories grew at just a 0.1% annual rate (Furman, BEA data). Independently, data-center investment alone added ~1 percentage point to Q1-2025 growth — about the same contribution as all consumer spending (Sløk / Apollo). This is contribution-to-GROWTH, not share of the economy's level, and some of the equipment is imported (which trims the net figure).

Snapshot: 2026-06-05 · 26 cited sources — see the appendix below.

Capex as a percentage of GDP, indexed to years since each buildout began. The AI line is Big Four (MSFT + GOOGL + AMZN + META) fiscal-year capex from SEC filings ÷ US GDP, with 2026–28 from analyst projections — Goldman Sachs, counting only AI-specific capex, puts today's figure nearer 0.8%. Historical cycles use domestic capex ÷ domestic GDP; the UK railway peak is 6.7% (Mitchell 1964 via Campbell), with Odlyzko computing 7.3% on a rounded base. US railroads (1879–1887) are gross capital spending on road and equipment — not the larger flow of securities issued, which ran about 1.7× higher — and peak near 3.7%; US shale (2009–2020) is BEA upstream capex ÷ GDP, peaking near 1% in 2014. US telephone (1894–) and US electric utilities (1902–) are Ulmer (1960) gross capex ÷ GDP, shown over their first decade — both sub-1% buildouts (telephone's monopoly-era plateau ~0.3%; electric ramping, ~0.9% peak in the 1920s). Hover or tap any line to isolate it.

Buildout	The technology	The first generation of investors
UK railway mania (1840s)	Network built; ran for a century	Dividends ~1.9–2.8% against ~3% on risk-free consols — they underperformed cash for a decade
US railroads (1880s)	Network ran for a century	Only 44% of rail shares paid any dividend even in 1892; about a quarter of US rail mileage went into receivership in 1893
US telephone / Bell (1890s)	Network became universal; AT&T ran for a century	Investors won. The independents were bought out or failed, but Vail’s regulated monopoly made AT&T the archetypal stable dividend stock for decades
US electric utilities (1900s)	Universal electrification	Mostly won. Regulated operators paid dividends for decades — but the 1920s Insull holding-company pyramids collapsed in 1932, wiping out their investors → PUHCA 1935
US telecom / fiber (1990s)	Fiber still carries the internet	~$2 trillion of equity value wiped out; WorldCom and Global Crossing went bankrupt
US shale (2010s)	US became the world’s largest oil & gas producer	~−$300B of free cash flow over 2010–19, more than $450B of capital impaired, 230+ producer bankruptcies
AI / data centers (2020s)	Unresolved	Unresolved — the question this post is about

Sources for the table — UK rail: Odlyzko, Campbell & Turner. US railroads: Whitten / EH.net, Panic of 1893. Telephone: Ulmer 1960, Table E-1, Potter (2024). Electric: Ulmer 1960, Table D-1, EEI, PUHCA (1935) / Insull. Telecom: Starr, “The Great Telecom Implosion” (2002), Doms (FRBSF 2004). Shale: Deloitte (2020), Haynes & Boone, IEEFA.

The technology almost always turns out to be real and transformative. Railroads genuinely opened the continent. Fiber still carries the internet. Electric utilities powered a century of growth. But here’s the part that matters for this moment: in most of those cycles, the investors who funded the buildout first did not get paid. Rail investors saw a quarter of US mileage go into receivership. Telecom investors lost roughly two trillion dollars. Shale producers burned three hundred billion in free cash flow and went bankrupt by the hundreds.

Electricity is one of the exceptions where the infrastructure worked and the capital actually got returns. Why? Partly regulation, partly that demand really did keep expanding to fill the capacity. But electricity also had time: it built over decades, not years.

AI has it’s own unique schedule, as data center equipment depreciates in under ten years, not decades. The bet is that demand arrives fast enough, and supply stays scarce enough, to pay back the investment before the assets age out.

The map is better than the mood

The usual debate wants a verdict: bubble or not, boom or bust, productivity miracle or expensive hallucination. The better answer is a map.

We are at a genuine crossroads. The future branches three ways — capability could stay explosive or it could slow, compute could arrive tight or adequate or in glut, demand could stay concentrated or spread economy-wide. That creates 27 possible futures. Not all are equally likely.

Of the three axes, capability is the one to watch. Demand broadens gradually, and you can see it coming in API usage and enterprise adoption. Supply arrives early or late, and you can read it in capex burn and data-center commentary. But capability is the only axis where you cannot yet see the ceiling, and that is exactly what makes it the signal that matters most.

It matters for another reason too: capability is not just an input to the system, it is regulated by it. Loose supply that fails to find demand gets redirected into training and experimentation, which pushes capability higher. Tighter supply constrains it. Better capability creates new demand. The system has a thermostat built in.

The scenarios where everything slows together are not really live options. The capex is already committed. Supply will arrive. The real question is not whether it shows up, but whether capability and demand have moved fast enough to absorb it before it becomes a glut problem.

One scenario is worth naming on its own: supply stays tight longer than expected. That is bad for the broader economy — AI stays expensive and concentrated — but it is good for capex ROI. Scarcity pricing keeps revenue high relative to the installed base. The loop closes, but narrowly, and everyone else pays the tax.

The three axes are:

• Capability trajectory: Steep / Slowing / Flat.
• Compute supply: Tight / Adequate / Glut.
• Demand breadth: Narrow / Broadening / Broad.

Today’s cell is Steep / Tight / Narrow. That is amber, not green. The cycle is alive because capability is steep and compute is scarce. The loop is still open because demand remains concentrated.

Interactive view requires JavaScript. Below: all 27 cells, grouped by capability trajectory. Each tile is tinted by its score (0 = capex stressed, 100 = capex clearly justified). Today's cell carries a "●" pin.

Capability trajectory: Steep

Tight

Adequate

Glut

Narrow

Tight · Narrow 58

Today. Revenue grows, internal compute demand floors it. ROI rides on demand broadening before 2027 capacity lands.

● Today

Adequate · Narrow 55

Supply caught up but buyers haven't. Early-adopter core absorbs it; no margin of safety.

Glut · Narrow 18

Overbuild. Capacity dumped, narrow demand, prices crater — depreciation eats earnings.

Broadening

Tight · Broadening 82

Strong. Demand widening into big sectors while compute still scarce — premium pricing holds, loop closing.

Adequate · Broadening 85

The clean bull case. Supply and demand both scaling, capability steep — loop closes.

Glut · Broadening 52

Race condition. Glut now, but broadening demand may backfill — timing-dependent.

Broad

Tight · Broad 92

Best case. Broad demand, scarce supply — pricing power maximal, capacity fully absorbed.

Adequate · Broad 95

Goldilocks. Everything scales together, capex fully justified.

Glut · Broad 78

Overbuild rescued. Broad demand soaks up excess capacity; deflation, not distress.

Capability trajectory: Slowing

Tight

Adequate

Glut

Narrow

Tight · Narrow 42

Capability slowing AND compute scarce — the supply-ceiling fear. Revenue growth decays, soft not cliff.

Adequate · Narrow 28

Supply fine, capability fading, buyers absent — little to justify the stock.

Glut · Narrow 8

Worst-feeling cell. Overbuilt, capability stalling, no buyers — writedowns.

Broadening

Tight · Broadening 55

Demand widens but the product improves slower — pricing power erodes gradually.

Adequate · Broadening 52

Demand broadening offsets slowing capability — a longer, flatter payoff.

Glut · Broadening 30

Glut with a fading product — broadening demand may not arrive fast enough.

Broad

Tight · Broad 68

Broad demand carries it even as capability slows — current models monetize fine.

Adequate · Broad 65

Broad adoption of good-enough models — diffusion story, lower altitude.

Glut · Broad 48

Broad demand mops up the glut, but slowing capability caps pricing — thin margins.

Capability trajectory: Flat

Tight

Adequate

Glut

Narrow

Tight · Narrow 22

Capability flat, compute still scarce (inefficient), narrow demand — capex misallocated.

Adequate · Narrow 15

Plateau with no buyers — the spend has nothing to show for it.

Glut · Narrow 3

Total bust. Overbuilt, plateaued, unbought.

Broadening

Tight · Broadening 40

Plateaued models, but demand still widening on good-enough capability.

Adequate · Broadening 42

Good-enough AI diffuses slowly — low-margin, utility-like outcome.

Glut · Broadening 20

Glut meets plateau — broadening demand can't restore pricing.

Broad

Tight · Broad 45

Commodity inference at scale — volume yes, pricing power no.

Adequate · Broad 48

Mass adoption of a commodity — large but unprofitable, like telecom fiber.

Glut · Broad 35

The thesis's key cell: everyone uses AI, it works, nobody pays much — diffusion win, capex bust.

Score 0-100 reflects how cleanly the capex/revenue loop closes in that cell. Today's pinned state is Steep / Tight / Narrow: alive because capability is steep and compute is scarce, unresolved because demand remains concentrated.

What would change the story

Capability slope. Look for METR-style time-horizon evidence, frontier model releases that extend reliable task length, and signs that the log chart is flattening. The slope matters more than any launch headline.

Demand breadth. Watch enterprise production usage, retention, API usage outside coding, and budgets moving from pilots to durable workflows. The clean signal is adoption in large employment sectors and operational functions.

Supply and glut risk. Watch Q2 hyperscaler capex, data-center commentary, chip/memory/power constraints, depreciation expense, and useful-life assumptions. Looser supply will test whether demand is real.

Next quarter’s key question is not whether one number turns green. It is whether today’s cell moves toward Steep / Adequate / Broadening, or whether the slope bends before breadth arrives.

The companion post looks at where AI productivity is actually showing up in sector and workforce data.

Quarterly status snapshot

The automated dashboard is useful context, not the thesis. It catches pressure points that matter to the cycle, but it cannot decide the three-axis map by itself.

Legend GREEN supports the thesis YELLOW is mixed/watch RED is stress Arrows show thesis-direction vs the prior comparable window: ↑ improved, ↓ deteriorated, → flat, — no clean prior.

GREEN ↑

Nonfarm-business productivity

+2.83%

4Q avg annualized

GREEN ≥2.5% and YELLOW ≥1.5% on trailing-4Q nonfarm-business productivity.

YELLOW —

Hyperscaler capex YoY

+80.5%

latest matched quarter; total capex

Capex still accelerating (>30.0% YoY); watch consecutive count.

RED ↑

Frontier-lab ARR / hyperscaler capex

0.15

ARR / latest FY capex

Per the cycle thesis, sustained <0.20 = capex/revenue gap too wide to sustain on faith.

YELLOW ↓

Real interest rates supportive

2.13%

10Y TIPS, latest FRED

AI capex is long-duration; real rates >2.5% raise the hurdle materially.

GREEN ↑

Credit conditions (IG BBB OAS)

94 bps, -5 bps 90D

Spreads tightening or flat — risk appetite intact, financing available for capex.

Five automated indicators graded against the cycle thesis. The productivity tile is nonfarm business, not sector productivity; sector-level productivity is handled in the companion diffusion post. The arrows are thesis-direction arrows, so a lower spread can point up even when the raw OAS value fell.

Source and data appendix

Unless noted otherwise, data are through the May 25, 2026 build. Frontier-lab ARR runs through 2026-04; hyperscaler capex runs through Q1-2026 filings; Anthropic Economic Index releases run through February 2026; METR Time-Horizon snapshot is from 2026-05-25.

SEC EDGAR hyperscaler capex

• Source type: Official SEC API data.
• Endpoint: https://data.sec.gov/api/xbrl/companyfacts/CIK{cik}.json.
• Fetched at: 2026-05-22T13:48:27.891645+00:00.
• Notes: PaymentsToAcquirePropertyPlantAndEquipment, or PaymentsToAcquireProductiveAssets for filers that use that tag. Total capex, not AI-specific. Oracle excluded because its fiscal year ends May 31 and Q3 capex is reported YTD, not per-quarter.

Company	Latest matched quarter
MSFT, Q3 FY2026 ending 2026-03-31	$30.88B
GOOGL, Q1 FY2026 ending 2026-03-31	$35.67B
AMZN, Q1 FY2026 ending 2026-03-31	$44.20B
META, Q1 FY2026 ending 2026-03-31	$19.00B

Capex-cycles buildout comparison

• Source type: Mixed — live FRED/BEA annual data (US shale), live SEC EDGAR (the AI line, documented above), and human-curated cited points (UK railway mania, US railroads, US telephone, US electric utilities, US telecom/fiber).
• Methodology: Each line is annual capital spending ÷ GDP for that economy, indexed to years since the buildout began. The ratio is nominal-over-nominal and therefore unit-free, so no inflation adjustment is applied (deflators cancel). Domestic capex ÷ domestic GDP per cycle.
• US shale / fracking: E318RC1A027NBEA (private fixed investment in mining exploration, shafts & wells — overwhelmingly oil & gas wells; current $B, BEA) ÷ GDPA (nominal US GDP, current $B), fetched live from https://fred.stlouisfed.org/graph/fredgraph.csv each build. Charted 2009–2020; peak ≈ 1.0% of GDP in 2014. The “how it ended” figures — roughly −$300B of sector free cash flow over 2010–2019, more than $450B of invested capital impaired, and 230+ North American producers (~$152B of debt) in bankruptcy from 2015 — come from Deloitte (2020), the Dallas Fed Energy Survey, Haynes & Boone’s Oil Patch Bankruptcy Monitor, and IEEFA.
• US railroads: gross capital expenditure on road + equipment, steam railroads (Melville Ulmer, Capital in Transportation, Communications, and Public Utilities, NBER 1960, Table C-1) ÷ nominal US GDP (MeasuringWorth; pre-1929 on Gallman 1966). Charted 1879–1887 (the peak window of the 1879–1893 boom, trimmed so the long arc doesn’t dominate the axis); peak 3.7% of GDP in 1881. We plot capital spending, not securities issued: railroad securities over 1880–84 reached ~4.8% of GNP (Tufano 1997) — about 1.7× actual outlays — validated against gross fixed capital formation in Gallman & Rhode (2020). Per MeasuringWorth’s terms only the derived ratio is published here, not their GDP table.
• US telephone (Bell) and US electric utilities: gross capital expenditure from Melville Ulmer (Capital in Transportation, Communications, and Public Utilities, NBER 1960) — telephone Table E-1, electric light & power Table D-1 — ÷ nominal US GDP (MeasuringWorth). Shown over their first ~decade (telephone 1894–1906; electric 1902–1912). On Ulmer’s single-industry gross-capex measure these are sub-1% buildouts — telephone runs ~0.16–0.41% (a ~0.3% plateau through 1950), and electric ramps 0.33%→0.69% with its all-series peak only ~0.9% in 1924 — well below the ~1–2% sometimes assumed. Electric’s 1902–1912 points are interpolated from census benchmarks (Ulmer’s own method; ~16% single-year error, slope reliable) and flagged approximate; the line stops at 1912 to avoid the interpolated→reported method seam. Annotation figures: AT&T assets ~$5B (1939) → ~$74B (1974), and 1960–73 telephone capex ~$70B vs Apollo’s ~$26B (Potter, Construction Physics, 2024); Samuel Insull’s Middle West Utilities collapse (1932) → PUHCA 1935; modern EEI-member capex ~$178B (2023), ~$208B (2025), >$1.1T projected 2025–29 to power AI/data centers (EEI). Per MeasuringWorth’s terms only the derived ratio is published, not their GDP table.
• UK railway mania and US telecom / fiber: unchanged human-curated cited points (Mitchell 1964 via Campbell & Turner, and Odlyzko, for UK rail; Doms, FRBSF 2004, for telecom).
• Investment-performance figures (the “how the capital did” comparison): UK rail dividends ~1.9–2.8% vs ~3% consols — Odlyzko / Campbell & Turner; US railroads (44% of shares paid a dividend in 1892; ~a quarter of mileage in receivership in 1893) — Whitten, EH.net and the Panic of 1893 record; US telecom ~$2T equity loss — Starr, “The Great Telecom Implosion” (2002); US shale free cash flow / impairments / bankruptcies — Deloitte (2020), Haynes & Boone, IEEFA.

Frontier-lab revenue

• Source type: Human-curated YAML; company-stated, reported-leak, and analyst-estimated rows labeled individually.
• Endpoint: posts/the-slope-is-the-cycle/pipeline/data/frontier_lab_revenue.yaml.
• Fetched at: 2026-05-25T00:00:00+00:00.
• Notes: OpenAI and Anthropic ARR with citations and source dates. Run-rate ARR is not audited GAAP revenue.

Recent cited points used in the chart:

Lab	Period	Value	Type	Source
OpenAI	2025-12	$20.0B ARR	company-stated	OpenAI
OpenAI	2026-02	$25.0B ARR	reported-leak	The Information
Anthropic	2026-01	$14.0B ARR	company-stated	Anthropic
Anthropic	2026-04	$30.0B ARR	company-stated	Fortune

Earlier analyst-estimated OpenAI and Anthropic points come from The AI Corner and are labeled as analyst estimates.

METR Time-Horizon benchmark

• Source type: Public YAML from METR.
• Endpoint: https://metr.org/assets/benchmark_results_1_1.yaml.
• Fetched at: 2026-05-25T13:32:46.247412+00:00.
• Notes: Frontier-model task-completion horizons, p50 and p80, calibrated to humans timing themselves on the same software tasks. Values are in minutes of human-equivalent work. Doubling-time block carries METR’s fitted exponential trend.

Anthropic Economic Index

• Source type: Anthropic Economic Index raw data releases V1-V5.
• Endpoint: Hugging Face: Anthropic/EconomicIndex.
• Notes: Usage share is the percentage of sampled conversations or records whose task maps to an occupation group. It is not headcount, spending, revenue, or productivity.

BLS and FRED scorecard sources

The collapsed scorecard uses BLS Productivity & Costs / Industry Productivity data from https://api.bls.gov/publicAPI/v2/timeseries/data/ and FRED daily macro data from https://fred.stlouisfed.org/graph/fredgraph.csv. Daily FRED macro latest observations are through 2026-05-21; BLS nonfarm productivity is through 2026 Q01.