Microsoft AICompany disclosed
$37BAnnual revenue run rate in March 2026, up 123% year over year.
MRTNZ / FIELD NOTE 01 / DATA THROUGH JULY 27, 2026
AIInfrastructure
The railroad era's capital cycle is repeating—
with a larger financing loop and a radically compressed clock.
The thesis
AI is the steam engine, factory system, railway network, and power buildout happening at once.
The network can transform the economy while the securities financing it destroy capital. That is the railroad lesson—and the central distinction between being right about AI and being right about AI infrastructure returns.
Revenue vs. capex
The near-trillion-dollar year
Company guidance is firm at the core. The sector total and sector revenue remain estimates and are labeled as such.
Big Four capex plans>$725BCompany guidance · April FT tally, raised again by July
Total AI capex~$870BMorgan Stanley estimate · includes neoclouds · upside risk
Deduplicated gen-AI sales>$175B ARRExponential View estimate · $110B trailing 12 months
David Cahn / Sequoia / July 2026
~$750B
×2estimated 2026 infrastructure capex
~$1.5T
lifetime end-customer revenue required to justify it
His cumulative lifetime revenue requirement for the post-ChatGPT buildout is approximately $3T—not annual revenue.
Named AI revenue run rates
Directional only. These figures overlap and are not additive.
OpenAIEstimate
~$25BAnnualized run-rate estimate for February–March 2026; not GAAP revenue.
AnthropicCompany reported
$47BRun-rate revenue crossed in early May 2026; still not GAAP revenue.
The current counter-signal
AI sales now cover estimated depreciation—barely.
Q1 2026 was the second consecutive quarter above the line, so the current streak began in Q4 2025.
AI sales$25B
Depreciation$21B
Sources: Financial Times · Morgan Stanley · Exponential View · Microsoft FY26 Q3 · Reuters / OpenAI · David Cahn · Bloomberg / depreciation
One stack / two eras
Buildout → diffusion
Select a stage to inspect where rents accrue—and what can break.
Industrial Revolution AI buildout
ThenCoal · iron · capital
NowElectricity · silicon · data · capital
ConstraintMines and steel → generation and transmission
ThenSteam engine
NowAccelerators · foundation models
ConstraintPrecision machinery → GPUs, HBM and packaging
ThenMill · factory
NowData center · AI factory2026 · WE ARE HERE
ConstraintFactory finance → transformers, cooling and interconnect
ThenRail · telegraph · canals
NowFiber · cloud · APIs
ConstraintRights-of-way → power and platform access
ThenMachine tools · mass production
NowCopilots · agents · robots
ConstraintProcess redesign → data, trust and workflow adoption
ThenAbundant mechanical work
NowAbundant cognitive work
ConstraintDistribution → demand, policy and social absorption
Structural analogue—not calendar equivalenceBritain ≈ 1830–45
Factories proven. Network capacity racing to scale.
ThenMill and factory
NowHyperscale data center
Rent captureLow-cost, high-utilization operators
Failure modeOverbuild and stranded capacity
The clock
Generations → investment cycles
The functional sequence compresses from roughly 160 years to a scenario of 20–30 years.
Industrial1712 → 1870
≈160 years
AI scenario2012 → 2030s
≈20–30 years
The railroad ending
The network won. Capital did not.
U.S. rail became permanent economic infrastructure after a brutal transfer of ownership, consolidation, and investor losses.
of the national network entered new receiverships—about 40,000 route miles.
U.S. route miles—the network’s historical peak.
route miles still operating: a smaller, durable freight network.
Durable utility ≠ durable equity value
The investable lesson is not that infrastructure fails. It is that excess capacity, leverage, and weak unit economics can wipe out the first owners before the network reaches its full social value.
Historical scope: U.S. railroads, 1893–present. Route miles are not directly comparable to capacity, ton-miles, or network productivity. EH.Net receivership series · Theodore Roosevelt Center · Federal Railroad Administration
The sharpest rhyme
Circular financing
Vendor capital helps customers buy capacity; those purchases then validate vendor demand. The loop works until customer cash flows become independently sufficient—or financing tightens.
Vendor capitalNVIDIA equity · reported guarantee talks
→Capacity buyersLabs · neoclouds · unconsolidated JVs
→Purchase commitmentsCompute · leases · chips
- $30B
- NVIDIA stake in OpenAI's 2026 round
- ~$1.4T
- Altman-described eight-year commitment envelope
- ~$27B
- Meta Hyperion unconsolidated JV development cost
Vendor financingLucent · Nortel · Cisco
→Capacity buyersCompetitive local carriers
→Equipment ordersBooked revenue · credit risk
- $8.1B
- Lucent credit and loan guarantees
- $5.17B
- Nortel year-end 2000 customer financing
- 47
- CLEC bankruptcies or market exits by January 2003
The structures are analogous, not identical. Today's hyperscalers have much stronger balance sheets than the late-1990s CLECs. NVIDIA's reported $250B guarantee discussions remain early-stage and nonfinal. Meta / Hyperion · Lucent FY2000 10-K · Nortel FY2001 10-K · Network World / CLECs · Reuters / NVIDIA talks · OpenAI financing
The labor split
Construction boom. Thin operating base.
Data centers create large, temporary skilled-trades demand, then operate with a comparatively small permanent workforce.
construction workers
Reported onsite at the Abilene Stargate campus in September 2025.
forecast onsite jobs
Oracle's expectation at full operation after the nearby expansion.
Same campus, different phases. Separately, OpenAI forecast 25,000+ onsite jobs across five additional U.S. Stargate sites without classifying all of them as permanent operations roles. Associated Press / Abilene · OpenAI / five sites
What would weaken the bubble case
Three benchmarks
The thesis should be updated by operating evidence, not by narrative or stock prices.
AI revenue sustains a widening surplus over depreciation for at least three consecutive quarters.
The hyperscalers’ promised 2028 free-cash-flow acceleration appears despite the larger asset base.
Labs and neoclouds fund compute from durable customer economics instead of vendor equity, guarantees, or structured vehicles.
Investor implication
The infrastructure can transform the economy even if many infrastructure investors lose money.
Now Chips · power · networking→Next Utilization · platforms→Later Vertical apps · robotics
Own the bottleneck only when the economics survive normalization. Otherwise, wait for value capture to migrate downstream.
Sources and method
AI capital, revenue, and economics: Financial Times · Microsoft · Reuters / OpenAI · Anthropic · Exponential View
Infrastructure and power: IEA, April 2026 · IEA electricity outlook · Morgan Stanley
Railroad buildout and aftermath: Queen's University Belfast · EH.Net · Federal Railroad Administration
Financing and employment: Meta · OpenAI · Associated Press
Britain is used for the functional Industrial Revolution stack; U.S. railroads are used for the capital-cycle aftermath. Company guidance, company-reported run rates, third-party estimates, and historical statistics are labeled separately. Private-company revenue is provisional until audited filings exist.