Issue 02: The AI Capex Shock

3. The buildout is near-term inflationary and only potentially disinflationary later. Physical scarcity raises component, construction and electricity costs now. Productivity depends on broad adoption and useful applications, which are developing on a slower clock.

4. Financing risk is moving outside the obvious balance sheets. Hyperscalers can use leases, project companies and long-term capacity commitments. The spending is real even when the asset and debt sit elsewhere.

5. The most dangerous gap is between infrastructure commitments and commercial adoption. Current business AI use remains around one-fifth of US firms. The installed base is being built for a demand curve that still has to prove itself.

This leads to a preference for businesses with visible, diversified orders, control over a hard-to-replicate physical constraint and the ability to turn backlog into free cash flow. It argues for more caution around stories that depend on distant capacity reservations, off-balance-sheet funding or unlimited willingness by regulators and households to subsidize grid upgrades.

What the evidence favors, and what it argues against

Prefer

• Visible, diversified orders
• Control over a hard-to-replicate physical constraint
• The ability to turn backlog into free cash flow

vs

Be cautious

• Distant capacity reservations
• Off-balance-sheet funding
• Unlimited willingness by regulators and households to subsidize grid upgrades

The capital-flow map

Each dollar of hyperscaler capital expenditure enters one of four layers. The economics differ at each layer.

Each capex dollar lands in one of four layers

1. Compute and networkingShortest economic life, fastest upgrade cycles
2. The powered siteConverts equipment into operating capacity
3. Grid equipment and transmissionThe slowest part of the chain
4. Generation and fuelDecides what serves the marginal hour

The economics differ at each layer

Layer 1: Compute and networking

This layer includes accelerators, high-bandwidth memory, central processors, optical components, switches and servers. It receives the most attention because performance is measurable and upgrade cycles are fast.

It also has the shortest economic life in the stack. A substation can operate for decades. An accelerator may remain technically useful for years, but its relative economic value can fall quickly when a more efficient generation arrives. That makes utilization and replacement cost crucial.

Microsoft disclosed that part of its capital expenditure consists of long-lived assets such as buildings, while another part consists of shorter-lived GPUs and CPUs. The distinction matters. If demand exceeds supply, both classes of assets earn attractive returns. If growth slows, the short-lived assets create a faster depreciation and replacement burden.

The constraint within compute can also move. One quarter it is accelerator availability. The next it is advanced packaging, memory bandwidth or network fabric. Investors should avoid turning a temporary shortage into a permanent margin assumption.

Layer 2: The powered site

The site layer includes land, buildings, cooling, uninterruptible power, switchgear, busways and on-campus distribution. It converts equipment into operating capacity.

Vertiv is a clean public indicator. In the first quarter, revenue rose 30 percent and organic sales in the Americas rose 44 percent. Adjusted operating margin reached 20.8 percent, up 430 basis points. The company raised its full-year outlook to $13.5 billion to $14 billion of sales and 29 percent to 31 percent organic growth.

At the end of 2025, Vertiv had $15 billion of backlog, 109 percent more than a year earlier, and a book-to-bill ratio of 2.9 times. Those are extraordinary numbers. They establish strong demand, but they also set a difficult standard. Any evidence that orders are being duplicated, delivery dates are slipping or backlog is converting more slowly would matter more than another record headline.

Vertiv's orders ran at 2.9 times its sales

$15BBacklog, end of 2025Up 109%

2.9×Book-to-bill

+30%First-quarter revenue

20.8%Adjusted operating marginUp 430 basis points

The advantage of the site layer is that density creates technical complexity. Higher-power racks need more advanced thermal management and electrical design. The risk is that customers eventually standardize more of the system or that new capacity attracts competitors.

Layer 3: Grid equipment and transmission

This layer is the slowest part of the chain and therefore one of the most interesting.

Make a guess

In 2019, a distribution transformer took three to six months to arrive. By 2023, what was the upper end of the wait?

Tap your guess

Under 9 monthsYour guess9 to 18 monthsYour guess18 months to 3 yearsYour guess3 years or moreYour guess

01234567890123456789 months2023, upper end

Spot on.Close, a little low.Close, a little high.Way off, too low.Way off, too high. It was 30 months.

The Department of Energy says lead times moved from 3–6 months in 2019 to 12–30 months in 2023. Industry participants later described three to four years for large substation units.

2023, upper end30 months

2019, upper end6 months

The Department of Energy says distribution-transformer lead times moved from three to six months in 2019 to 12 to 30 months in 2023. Industry participants in a later DOE convening described one to two years for many transformers and three to four years for large substation units. Demand has grown faster than a supply chain that relies on specialized electrical steel, copper, qualified designs and skilled labor.

Eaton’s first-quarter results show the demand signal. Electrical Americas rolling orders increased 42 percent organically, driven by data centers. Sales rose 20 percent and backlog increased 44 percent. Across Eaton’s electrical segments, backlog rose 48 percent.

GE Vernova provides the broader grid and generation angle. Its Electrification segment booked $2.4 billion of data-center equipment orders in the first quarter, more than all of 2025. Total company backlog increased by $13 billion sequentially to $163 billion. Its gas-turbine backlog and slot reservations reached 100 gigawatts, with management expecting at least 110 gigawatts by year-end.

Grid-equipment orders are running ahead of sales

+42%Eaton Electrical Americas rolling ordersSales up 20%

+44%Eaton Electrical Americas backlog

$2.4BGE Vernova data-center equipment orders, Q1More than all of 2025

$163BGE Vernova total backlogUp $13B in a quarter

These companies are not pure AI trades. That is part of the attraction. Grid modernization, industrial electrification and generation replacement can support demand even if one data-center region cools. Diversification lowers the dependence on a single customer’s model economics.

The valuation risk remains real. Backlog does not equal discounted cash flow. Execution, raw-material costs, warranty obligations and production expansion determine how much of the demand becomes shareholder value.

Layer 4: Generation and fuel

Data centers want reliable power, often on a flat load profile. Renewable additions can supply large amounts of energy, but availability, interconnection and transmission determine what serves the marginal hour.

The Energy Information Administration expects record natural-gas consumption by the electric-power sector next year. It forecasts US gas-fired capacity rising from 494.4 gigawatts in 2025 to 499.7 gigawatts in 2026 and 508 gigawatts in 2027. Solar photovoltaic capacity rises much faster, from 149.8 gigawatts in 2025 to 221.1 gigawatts in 2027, but generation profiles are not interchangeable.

Solar capacity grows 48% by 2027; gas grows 3%

Gas-firedSolar photovoltaic

100200300400500600 GW0

494.4 GW149.8 GW

2025Gas-fired494.4 GWSolar photovoltaic149.8 GW

508.0 GW221.1 GW

2027Gas-fired508.0 GWSolar photovoltaic221.1 GW

20252027

In EIA’s high-data-center-demand case, gas generation rose 7.3 percent from 2025 to 2027, compared with 1.7 percent in the base case. Higher load also slowed the decline in coal generation because new supply could not be built immediately.

This creates a medium-term benefit for gas equipment and certain generators, but not an unlimited one. Gas prices, turbine availability, emissions rules, regional transmission and customer preferences all shape returns. A developer promising nuclear, geothermal or large-scale storage may eventually deliver. The near-term grid will still dispatch the capacity that is physically available.

Make a guess

Data centers used about 4.4% of US electricity in 2023. In the Department of Energy's high scenario, what share could they use by 2030?

Tap your guess

Under 8%Your guess8% to 12%Your guess12% to 16%Your guess16% or moreYour guess

01234567890123456789.0123456789%2030 high scenario

Spot on.Close, a little low.Close, a little high.Way off, too low.Way off, too high. It was 15.3%.

DOE scenarios based on Lawrence Berkeley National Laboratory work put the 2030 share between 9.5% and 15.3%. Even the low case asks the grid to absorb several years of demand growth in a short period.

2030 high scenario15.3%

2023 actual4.4%

2030 low scenario9.5%

The power math

Data centers consumed about 4.4 percent of US electricity in 2023. Department of Energy scenarios based on Lawrence Berkeley National Laboratory work put the 2030 share between 9.5 percent and 15.3 percent.

The range is wide because model efficiency, server utilization, cooling performance and construction pace remain uncertain. Even the low case requires the power system to absorb several years of demand growth in a short period.

National averages obscure the problem. The heaviest demand is concentrated in a few regions.

In Dominion’s Virginia zone, summer peak load reached 23,905 megawatts in 2025, up 23 percent from 2019. Winter peak load rose 45 percent over a similar period. PJM projects summer peak demand in the zone to grow 5.4 percent annually over the coming decade.

Peak load in Dominion's Virginia zone is up 23% since 2019

23,905 MWSummer peak load, 2025

+23%Summer peak since 2019

+45%Winter peak, similar period

5.4%PJM's projected yearly summer-peak growth

PJM’s 2025 capacity auction procured 6,517 megawatts less than its target, the first shortfall since 2007, leaving a 14.4 percent reserve margin. The region has generation in development, but projects face interconnection delays, permitting and equipment constraints.

Texas is dealing with a different market design and the same physical reality. ERCOT classifies projects of 75 megawatts or more as large loads and has described expected growth as exceptional. Data-center developers value Texas for land, generation and speed. The grid still has to distinguish credible projects from speculative or duplicated requests.

This is why announced gigawatts are a poor forecasting tool. A developer can inquire about multiple sites while intending to build only one. If every request enters the load forecast, utilities may overbuild. If planners discount too much demand, genuine projects wait for years.

The quality of the commitment matters more than the headline size. We prefer to see deposits, minimum bills, take-or-pay structures, cancellation penalties and clear responsibility for upgrades.

Commitment quality matters more than headline size

Speculative request

• Inquiries at several sites to build one
• Reserves a queue position without serious capital

vs

Credible commitment

• Deposits and minimum bills
• Take-or-pay structures
• Cancellation penalties
• Clear responsibility for upgrades

Regulation decides who keeps the return

Load growth sounds attractive for regulated utilities. More customers can justify more assets, and approved assets earn a regulated return. The problem is allocation.

Suppose a hyperscaler requests a dedicated substation and major network upgrades. Construction begins. Two years later, the customer reduces the campus or cancels it because model economics changed. Someone still has to pay for equipment already ordered and work already completed.

If the utility socializes those costs across households, the political response can be severe. If the developer bears all risk, fewer speculative projects will enter the queue. The efficient answer likely sits between those extremes, with stronger commitments and service structures matched to project certainty.

An abandoned campus still leaves a bill

1. Hyperscaler requests (Substation and upgrades) to Construction begins
2. Construction begins to Campus cut (Model economics changed)
3. Campus cut (Model economics changed) to Households pay (Political backlash)
4. Campus cut (Model economics changed) to Developer pays (Fewer speculative requests)

FERC’s June orders to six regional grid operators put five issues on the table: efficient study of large loads, prevention of cost shifting, co-location and behind-the-meter service, flexible or non-firm service, and coordination between generation and nearby load.

Commissioner David Rosner highlighted cost-recovery agreements so abandoned projects do not burden residential customers. He also discussed flexible service and “bring your own new generation” structures that can reduce the required upgrades.

These rule changes can create winners. A data center willing to curtail during a handful of stressed hours may connect faster. A developer that brings firm generation or pays for upgrades may obtain a clearer schedule. Equipment suppliers benefit from confirmed projects, while utilities gain higher-quality forecasts.

The likely loser is the option-like queue request that reserves a position without a serious capital commitment.

The inflation and interest-rate channel

The June Federal Reserve minutes explicitly connect AI investment to the macro outlook. Officials said the buildout was supporting growth above potential. Many saw continued pressure on technology-product and electricity prices. Some believed AI productivity would eventually be disinflationary, but with a lag.

The timing matters because inflation is already elevated. The latest available personal-consumption-expenditure report at our research cutoff showed May headline PCE inflation of 4.1 percent from a year earlier and core inflation of 3.4 percent. June CPI was scheduled for later on July 14 and had not been released when this issue was completed.

AI investment is not the only reason inflation remains high, and it may not be the largest one. It adds pressure in categories where supply is already slow to respond. Microsoft said higher component pricing added roughly $25 billion to its calendar 2026 capital-expenditure estimate. Meta cited higher component prices and data-center costs when it raised guidance.

The chain is straightforward:

1. Hyperscalers raise orders for servers, electrical equipment and construction.
2. Qualified supply cannot expand at the same pace.
3. Prices and wages rise in the constrained categories.
4. Power demand increases before new generation and transmission arrive.
5. Utility bills and technology prices remain firmer.
6. The Fed sees less room to ease policy quickly.
7. The discount rate applied to long-duration AI profits remains high.

This feedback loop is uncomfortable because the companies spending the most are also valued on distant cash flows. If the buildout keeps rates higher while near-term free cash flow falls, valuation becomes more sensitive to execution.

The loop can reverse. If AI adoption produces measurable productivity, unit costs can fall. If new power and equipment capacity arrives, scarcity pricing can ease. The risk is the interval between the capital outlay and the productivity payoff.

The global trade channel

The Federal Reserve’s analysis of AI infrastructure and trade estimated that US data-center spending would exceed $500 billion in 2025. Supplier economies benefited from exports of servers and critical components.

The physical stack remains globally concentrated. Leading-edge fabrication, advanced packaging, high-bandwidth memory and server assembly connect US spending to Asian production. Domestic construction does not remove those dependencies.

A domestic asset built with a global bill of materials

1. US data-center spending (Over $500B in 2025) to Chips (Fabrication, memory)
2. US data-center spending (Over $500B in 2025) to Assembly (Packaging, servers)
3. Chips (Fabrication, memory) to US data center (One missing input delays it)
4. Assembly (Packaging, servers) to US data center (One missing input delays it)

This has two consequences.

First, the AI investment boom can widen equipment imports even as it strengthens US business investment. A data center is a domestic asset built with a global bill of materials.

Second, geopolitical or logistical disruption can stop a project far from the point of failure. A shortage of one qualified memory product, optical component or transformer input can delay revenue from an entire campus.

The best supplier positions combine scarcity with qualification. A generic commodity can attract new capacity quickly. A component that requires long testing, customer approval and integrated redesign is harder to replace.

The financing is becoming less visible

Hyperscalers still fund much of their investment internally. The scale is pushing more activity into bonds, leases, special-purpose vehicles and private credit.

The BIS says hyperscaler gross bond issuance exceeded $100 billion in 2025. It also estimates that private-credit lending to AI and information technology has quadrupled in five years to roughly 15 percent of private-credit portfolios.

The most important structure is the project vehicle. A separate entity borrows to build a facility. The technology company signs a long-term lease or capacity agreement. The hyperscaler may report less direct capital expenditure, but it still has an economic obligation. The project debt can sit with private lenders or insurers.

Less capex on the books, the same economic obligation

1. Private lenders or insurers (Hold the project debt) to Project vehicle (Borrows to build): Debt
2. Project vehicle (Borrows to build) to Technology company (Long-term lease): Facility
3. Technology company (Long-term lease) to Project vehicle (Borrows to build): Contracted payments
4. Project vehicle (Borrows to build) to Private lenders or insurers (Hold the project debt): Debt service

This arrangement can be efficient. It matches long-lived physical assets with long-term funding and preserves hyperscaler capital for faster-moving compute. It can also obscure leverage and create correlated exposures.

If AI demand disappoints, the problem does not vanish because the debt is off the hyperscaler’s balance sheet. The project company still owes creditors. The tenant still owes contracted payments unless it can exit. Private-credit funds may discover that several apparently different loans depend on the same small group of technology customers and the same assumption about data-center utilization.

Spot the red flag

Amazon's free cash flow nearly vanished. Which line shows how it is filling the gap?

Amazon, first-quarter 2026 filing$ billions

Property and equipment purchases, net of incentives, trailing 12 months147.3That is the spending that drained free cash flow, not how it is funded.Free cash flow, trailing 12 months1.2This is the gap itself, down from $25.9B a year earlier.Cash from debt, first quarter59.5Borrowing now funds part of the build, and Amazon said additional financing was expected.Investment in OpenAI, with $35B more committed15.0A model commitment that compounds the exposure, but it uses cash rather than supplying it.

Tap the line you think is the tell.

Spotted it.Not that line. The AI cycle is large enough to change financing behavior even at the strongest companies. This is not a solvency claim: Amazon has a vast operating business.

Explain every lineHide the other explanations

Amazon’s first-quarter filing shows the direct cash pressure. Trailing 12-month purchases of property and equipment, net of incentives, reached $147.3 billion. Free cash flow fell to about $1.2 billion from $25.9 billion a year earlier. The company generated $59.5 billion of cash from debt in the first quarter and said additional financing was expected.

Amazon also invested $15 billion in OpenAI and committed another $35 billion. That commitment is separate from the physical infrastructure, but it demonstrates how model exposure and infrastructure funding can compound.

Again, this is not a solvency claim. Amazon has a vast operating business. It is evidence that the AI cycle is large enough to change financing behavior even at the strongest companies.

Company scorecard

The table separates spending platforms from bottleneck suppliers. It is a monitoring framework, not a list of unconditional buys.

Our relative preference inside this theme is for diversified bottleneck suppliers over pure capacity promises. Eaton and GE Vernova have exposure to data centers without depending entirely on them. Vertiv offers more direct sensitivity and stronger near-term growth, but also greater expectations risk. Among hyperscalers, the quality test is whether existing high-margin businesses can fund the build while AI lifts revenue or productivity before depreciation and financing overwhelm free-cash-flow growth.

Our relative preference inside the theme

Diversified bottleneck suppliersEaton, GE Vernova

• Data-center exposure without depending entirely on it

Direct site supplierVertiv

• More direct sensitivity
• Stronger near-term growth
• Greater expectations risk

HyperscalersMicrosoft, Alphabet, Meta, Amazon

• High-margin businesses must fund the build
• AI must lift revenue before depreciation and financing overwhelm free cash flow

That is a different question for each platform. Microsoft has visible cloud backlog. Alphabet and Meta can use AI inside advertising and consumer products. Amazon combines AWS demand with an unusually heavy funding cycle and a large external model commitment. The right comparison is not capital expenditure alone. It is incremental gross profit and cash generation per dollar of total economic investment.

Three scenarios for 2026 to 2028

Bull case: adoption catches the infrastructure

In the bull case, enterprise adoption broadens from roughly one-fifth of firms. Agentic software moves from experimentation into recurring workflows. Lower inference costs stimulate enough usage to offset efficiency. Hyperscaler backlog converts into cloud revenue, and utilization rises faster than depreciation.

Power projects proceed with stronger customer commitments. Flexible-load agreements allow some campuses to connect without waiting for every network upgrade. Equipment capacity expands, reducing lead times without collapsing supplier margins. Inflationary pressure fades as more generation, transformers and cooling systems enter service.

What we would expect to see:

• Census business AI use moves decisively above 25 percent and keeps rising.
• Cloud growth reaccelerates while capex growth begins to moderate.
• Hyperscaler free cash flow recovers despite higher depreciation.
• Vertiv, Eaton and GE Vernova convert backlog into revenue without cancellations or margin erosion.
• Transformer lead times shorten because supply expands, not because demand disappears.
• Energized megawatts rise and reserved capacity becomes utilized capacity.

In this outcome, the value chain remains broad. Platforms, equipment suppliers and power infrastructure all participate, though the strongest returns may rotate as bottlenecks ease.

Base case: the bottleneck moves, rates stay firmer

Our base case is a slower conversion. AI demand remains strong, but physical constraints limit how quickly spending becomes revenue. Enterprise adoption improves, though not fast enough to justify every announced campus. Strong projects proceed while speculative requests are withdrawn.

Electrical and cooling suppliers retain good pricing through 2027 because backlogs remain large. Hyperscalers continue to grow cloud revenue, but depreciation, component prices, leases and interest expense absorb more cash. The Fed sees the productivity benefit as real but delayed, leaving policy tighter than technology investors would prefer.

Utilities receive more disciplined large-load rules. Developers post deposits, accept minimum bills or bring generation. Household cost shifting becomes a larger political issue, slowing projects with weak commitments.

Expected signals:

• Business adoption rises gradually but remains below the infrastructure narrative.
• Cloud backlog grows, while revenue conversion is uneven.
• Capex stays high through 2027, but guidance becomes more selective by project.
• Equipment backlogs remain healthy while book-to-bill normalizes.
• Gas generation rises in constrained regions as transmission and clean firm power lag.
• Private financing expands and disclosure quality becomes a valuation differentiator.

This case favors suppliers with strong execution and multiple end markets. It is less friendly to companies valued on unlimited spending growth or immediate rate cuts.

Bear case: reservations do not become workloads

In the bear case, business adoption stalls and AI applications fail to generate enough recurring revenue. Model and hardware efficiency reduce required compute faster than new uses increase it. Customers reserve capacity but do not utilize it fully.

Data-center developers cancel or defer campuses. Grid operators discover that load forecasts included duplicate requests. Equipment orders are pushed out, and high book-to-bill ratios reverse. Hyperscaler depreciation continues rising after revenue expectations fall.

Financing becomes the transmission mechanism. Project vehicles face lower valuations and weaker refinancing conditions. Private-credit funds mark down loans that share the same customers and utilization assumptions. Hyperscalers remain viable but renegotiate contracts and reduce new commitments.

Expected signals:

• Business adoption plateaus below 25 percent.
• Cloud companies discuss optimization and digestion more often than shortage.
• Announced megawatts materially exceed energized and contracted capacity.
• Backlog cancellations or deferrals rise at site and grid suppliers.
• Hyperscaler free cash flow remains compressed while revenue growth slows.
• Credit spreads widen for data-center project vehicles and heavily exposed lenders.
• Regulators impose larger deposits after canceled projects threaten household rates.

This scenario would hurt the most crowded direct beneficiaries first. Diversified electrical suppliers would still have secular demand, but estimates and multiples would fall.

Three scenarios for 2026 to 2028

Bull caseAdoption catches the infrastructure

• Business AI use moves above 25%
• Cloud growth reaccelerates as capex growth moderates
• Free cash flow recovers

Base caseThe bottleneck moves, rates stay firmer

• Adoption rises gradually
• Backlog grows, conversion is uneven
• Capex stays high through 2027

Bear caseReservations do not become workloads

• Adoption plateaus below 25%
• Announced megawatts exceed energized capacity
• Credit spreads widen for project vehicles

What would change our view

We would become more constructive on the broad AI infrastructure complex if three developments occurred together:

1. Adoption widened materially. We want evidence that recurring enterprise workloads are growing outside the largest firms and technology sectors.
2. Cash returns improved. Cloud and AI revenue should begin to rise faster than depreciation, leases and replacement capital.
3. The power queue became more credible. Stronger deposits, clear cost allocation and a higher ratio of energized to announced megawatts would reduce stranded-asset risk.

We would become more cautious if any two of the following appeared:

1. Hyperscalers lowered capacity guidance while keeping spending high because existing projects could not be stopped.
2. Equipment suppliers reported rising deferrals, weaker deposits or a sharp fall in book-to-bill.
3. Private-credit or project-finance spreads widened without a broader credit shock.
4. Data-center load forecasts were revised down because duplicate requests were removed.
5. Business AI adoption stopped rising while model efficiency improved quickly.

These are falsifiable conditions. The thesis should not survive merely because AI remains culturally important.

Catalyst calendar

The next several weeks contain useful checkpoints.

• July 14: June consumer-price data, released after this issue’s research cutoff.
• July 22: GE Vernova second-quarter results. Watch Electrification orders, turbine reservations, backlog conversion and free cash flow.
• July 28 to 29: Federal Reserve meeting. Watch whether officials continue to discuss AI investment as a source of growth and near-term inflation.
• July 29: Microsoft fiscal fourth-quarter results. Watch Azure capacity, calendar 2026 capex, component pricing, depreciation and cloud-backlog conversion.
• July 30: June PCE inflation. The composition matters more than one headline print.
• August 11: EIA Short-Term Energy Outlook. Watch gas generation, power demand and capacity additions.
• September 7: Comment deadline for the DOE’s draft National Transmission Needs Study. Utility, grid and large-load responses may clarify the coming rule fights.

Alphabet, Meta, Amazon, Eaton and Vertiv reporting dates should be verified against their investor-relations calendars before trading around results.

The final judgment

AI infrastructure is not a single trade. It is a sequence of constraints funded by a sequence of obligations.

Compute created the first scarcity. Power availability, electrical equipment and site engineering are creating the next one. Grid rules and financing structures will determine how much announced demand becomes operating capacity. Commercial adoption will determine whether that capacity earns an adequate return.

The most attractive position in the chain is not necessarily the company with the loudest AI story. It is the company that sells a hard-to-replace input, has qualified capacity, can deliver on time and does not need every speculative campus to proceed.

The biggest risk is not that AI disappears. It is that the market confuses real technological demand with unlimited economic returns. A useful technology can still produce overcapacity. A booming supplier can still lose pricing power. A world-class platform can still spend too much before customers are ready.

We will keep following four conversions: capital to energized megawatts, megawatts to utilization, utilization to revenue and revenue to free cash flow. That is the sequence that turns an industrial vision into an investment return.

Four conversions turn a buildout into a return

1. Capital
2. Energized megawatts
3. Utilization
4. Revenue
5. Free cash flow

Each layer must convert into the next

Everything else is a reservation.

Source ledger

• Federal Reserve: Minutes of the June 2026 FOMC meeting
• BEA: Personal Income and Outlays, May 2026
• EIA: Server electricity use in commercial buildings
• EIA: Annual Energy Outlook 2026
• EIA: July 2026 Short-Term Energy Outlook
• EIA: High data-center demand and generation
• EIA: Virginia data-center load growth
• Department of Energy: Draft 2026 National Transmission Needs Study
• Department of Energy: Transformer supply-chain analysis
• FERC: Large-load integration orders
• Federal Reserve: Global trade effects of the AI infrastructure boom
• Federal Reserve: Monitoring AI adoption
• US Census Bureau: AI use among businesses
• BIS Annual Economic Report 2026
• BIS: Financing the AI investment boom
• Microsoft fiscal 2026 third-quarter earnings
• Alphabet 2025 fourth-quarter earnings call
• Meta first-quarter 2026 results
• Amazon first-quarter 2026 Form 10-Q
• Eaton first-quarter 2026 results
• GE Vernova first-quarter 2026 results
• Vertiv first-quarter 2026 results
• PJM 2026 Load Forecast Report