Research essay · Index v0.1

From AI Capex Headlines to Contract-Backed Cash

The AI Capex Cash Conversion Index asks a harder question than how much companies plan to spend: who has made a firm commitment, who pays before delivery, whether pricing is protected, and who absorbs execution risk?

Zeyuan Li (Amy Li) Founder & Independent Researcher, W-Axis Lab

AI infrastructure is usually narrated through capital expenditure. Hyperscalers announce larger budgets; chip companies secure more supply; utilities, equipment manufacturers, contractors, and data-center landlords report expanding pipelines. The headline number is treated as evidence of demand.

But capex is not demand. Capex is capital placed at risk in anticipation of demand. The quality of that risk depends on a second layer of questions: Is the customer commitment cancellable? How long does it last? Has cash arrived before delivery? Can price move with input costs? Does the supplier carry a fixed-price obligation, a credit backstop, a construction deadline, or a noncancelable capacity commitment?

The AI Capex Cash Conversion Index, or ACCI, is an evidence-first attempt to make that layer visible. Version 0.1 compares eight companies located at different points in the AI infrastructure stack: GE Vernova, Digital Realty, Quanta Services, Eaton, Alphabet, Vertiv, Intel, and NVIDIA.

The index is not a valuation model, an earnings forecast, or a recommendation to buy or sell any security. It measures the strength of publicly disclosed evidence that an AI-capex narrative is being translated into firm contracts, advance customer funding, protected economics, and manageable delivery risk.

The conversion question

Follow the migration of risk

The conventional AI-capex map follows dollars as they move from cloud companies into chips, power, cooling, grid equipment, construction, and real estate. ACCI follows the risk attached to those dollars.

A company with a large backlog may still face cancellation risk. A company with deposits may not disclose the term or pricing of the underlying agreement. A company with extraordinary margins may have committed far more to suppliers than it has disclosed in noncancelable customer orders. A long-duration lease may offer excellent visibility but leave the landlord responsible for construction, power delivery, and uptime.

The index therefore rewards evidence of contracted conversion, not the size or popularity of the underlying AI narrative.

Six-field score

One hundred points, with every assumption exposed

20 points

Contract firmness and duration

  • 0: no disclosed contract evidence
  • 5: narrative demand or cancellable pipeline
  • 10: firm but short-dated or mixed commitments
  • 15: multi-year commitments with a recognition schedule
  • 20: material, noncancelable, multi-year commitments

15 points

Pricing protection

  • 0: terms undisclosed
  • 5: qualitative evidence of favorable pricing
  • 10: disclosed fixed-price, unit-price, or cost-plus mix
  • 15: explicit escalators, pass-throughs, or take-or-pay terms

15 points

Customer funding before delivery

  • 0: none disclosed
  • 5: deferred-revenue proxy with unclear cash status
  • 10: quantified deposits or advance billings
  • 15: material cash advances or slot deposits tied to contracts

20 points

Contract-to-capital coverage

  • Counts firm demand against capital placed at risk
  • Capital exposure includes disclosed supply and capacity commitments, not only property and equipment
  • Scope mismatches cap the score rather than being filled with assumptions

15 points

Margin conversion

  • Gross margin is preferred
  • Operating margin, EBITDA margin, NOI, or cap rate is used only when clearly labeled
  • Margins are assessed within each business model, not compared mechanically across sectors

15 points

Delivery-risk allocation

  • Penalizes open-ended vendor funding, fixed-price exposure, credit backstops, and supplier commitments
  • Rewards customer funding, balanced milestones, cost-plus structures, escalators, and creditworthy leases

Evidence discipline

Known, inferred, and undisclosed are different states

K Known: explicitly stated in a primary company filing or investor release.
I Inferred: calculated from reported figures or represented by a labeled proxy.
U Undisclosed: no public evidence sufficient to support a conclusion.
H / M / L Confidence: High has at least five known fields; Medium has three or four; Low has two or fewer.

Confidence is displayed separately from the 100-point conversion score. It is not multiplied into the score. An undisclosed field is not evidence that the economics are poor; it is evidence that a public reader cannot verify them. Version 0.1 caps the relevant component rather than substituting an invented number.

The resulting score is deliberately provisional. It is a structured assessment of disclosed evidence, not a claim that a complex contract can be reduced to a single precise number.

Provisional results · July 24, 2026

Contract-backed conversion, not stock quality

86 / 100 · High confidence

GE Vernova · GEV

Equipment and service obligations extend from two years to beyond fifteen years. Q1 remaining performance obligations were $163.276 billion, while contract liabilities and deferred income were $31.949 billion. Q2 backlog reached $176 billion. Advance funding and slot reservations support conversion, although exact pricing protections remain undisclosed and delivery obligations are substantial.

Q1 2026 filing Q2 2026 results

83 / 100 · High confidence

Digital Realty · DLR

Hyperscale leases signed in Q1 carried a weighted-average term of approximately thirteen years. The Q2 signed-but-not-commenced backlog represented $1.9 billion of annualized GAAP rent at 100 percent. A specific June portfolio has fifteen-year leases and 3.6 percent annual escalators, but that pricing term must not be generalized to the entire portfolio.

Q1 2026 filing Q2 2026 results Lease evidence

79 / 100 · High confidence

Quanta Services · PWR

Remaining performance obligations were $26.242 billion, with 69 percent expected within twelve months and most of the balance in the following twenty-four months. Contract liabilities reached $3.840 billion. The score uses firm obligations, not the larger $48.471 billion headline backlog, because roughly 45 percent of that backlog reflects estimated master-service-agreement volume that customers can generally terminate on short notice.

Q1 2026 filing

71 / 100 · Medium confidence

Eaton · ETN

Eaton reported $22.8 billion of firm customer commitments, with 68 percent scheduled within twelve months. Deferred revenue was $1.113 billion, and Q1 added $1.233 billion of customer deposits and billings. The evidence supports contract conversion, but exact escalators, pass-throughs, and other pricing protections are not publicly disclosed.

Q1 2026 filing

67 / 100 · Medium confidence

Alphabet · GOOGL

Q1 revenue backlog was $467.6 billion, including $462.3 billion in Google Cloud, with just over half expected within twenty-four months; cancellable contracts were excluded. That demand evidence sits against $35.7 billion of Q1 capex, $332.4 billion of purchase and contractual obligations, and material infrastructure backstops. The scope and risk on each side are not perfectly matched, so the score does not treat backlog as automatic capex coverage.

Q1 2026 filing

60 / 100 · Low confidence

Vertiv · VRT

Vertiv disclosed $2.607 billion of deferred revenue and a Q1 deferred-revenue cash-flow benefit of $651.2 million, alongside a calculated Q1 gross margin of 37.7 percent. But no consolidated firm-backlog amount was disclosed in the Q1 filing or release. Deferred revenue is therefore treated as a contract-liability proxy, not silently relabeled as a cash-funded hardware backlog.

Q1 2026 filing Q1 2026 results

52 / 100 · Medium confidence

Intel · INTC

Intel disclosed $1.7 billion of contractually enforceable deposits under long-term customer arrangements, with cash expected in Q2. Exact tenor, pricing, customers, and scope were not disclosed, and receipt should not be asserted until verified in a subsequent filing. Q2 gross capex was $2.652 billion, while the Q3 GAAP gross-margin outlook was 41.0 percent.

Q1 2026 filing Q2 2026 results

48 / 100 · High confidence

NVIDIA · NVDA

NVIDIA's 74.9 percent Q1 GAAP gross margin demonstrates extraordinary conversion after sale. ACCI is measuring a different asymmetry: $119 billion of manufacturing, supply, and capacity commitments—$95 billion due in fiscal 2027—against $2.6 billion of disclosed remaining performance obligations on contracts longer than one year, largely support, cloud, license, and development arrangements rather than a disclosed hardware-order backlog.

Q1 fiscal 2027 filing

The NVIDIA result

A low score can reveal an information asymmetry

NVIDIA's position at the bottom of this version is not a judgment about its products, competitive position, valuation, earnings quality, or expected stock return. It is not a bearish call.

It reflects a public-disclosure asymmetry. NVIDIA reports enormous supplier-side commitments needed to secure manufacturing and capacity. Its disclosed remaining performance obligations for contracts longer than one year are much smaller and cover a different mix of support, cloud, licensing, and development obligations. Public filings do not provide an equivalent noncancelable customer hardware backlog with duration and pricing terms.

The index refuses to fill that gap with the popularity of the product, current revenue growth, or analyst estimates. Strong demand may exist. ACCI asks what portion of it a public reader can verify as contractually committed against the capital exposure already assumed.

What the first ranking says

Four different ways capex becomes cash

Equipment plus service

GE Vernova

Long-duration equipment and service obligations combine with customer down payments and slot reservations. The strength is funding visibility; the counterweight is execution and performance risk over many years.

Lease-backed infrastructure

Digital Realty

Long-duration leases and explicit escalators can convert development capex into contracted rent. Construction, power delivery, uptime, and tenant concentration remain the risk-bearing layer.

Milestone-funded delivery

Quanta and Eaton

Firm orders, deposits, milestone billings, and a visible delivery schedule support conversion. Fixed-price work can still move inflation, labor, and delay risk back to the supplier.

Capacity secured before full disclosure

Alphabet, Vertiv, Intel, and NVIDIA

Each shows real demand evidence, but at least one critical bridge—pricing, cash status, hardware backlog, scope matching, or delivery allocation—remains incomplete in public disclosures.

Reading rules

What this index refuses to do

First, it does not call a pipeline revenue. Backlog is treated as firm only when the company defines it as a contractual commitment or remaining performance obligation and explains cancellation treatment.

Second, it does not call every deferred-revenue balance a cash prepayment. When the filing does not establish cash receipt, the balance is labeled a contract-liability proxy.

Third, it does not compare a chip designer's gross margin mechanically with a construction contractor's gross margin or a real-estate company's cap rate. Margin evidence is assessed within the economics of the business model and every substitute is named.

Fourth, it does not let property-and-equipment capex stand in for total capital risk. Noncancelable supply, manufacturing, capacity, power, and credit-backstop commitments belong on the liability side of the conversion question.

Finally, it does not mistake missing disclosure for proof of weak demand. Missing disclosure lowers confidence and caps the relevant score. It creates a diligence question.

Known limitations

Version 0.1 is a map of public evidence

Cross-sector

Business models are not interchangeable

A turbine service agreement, a data-center lease, a cloud commitment, and a chip supply arrangement allocate risk differently. The framework creates a common language without claiming economic identity.

Timing

Reporting periods differ

The snapshot uses the latest primary filings and releases reviewed through July 24, 2026. Some companies had reported Q2 results; others remained on Q1 disclosures. Each source is linked at the company level.

Scope

AI exposure is not perfectly separable

Backlog, capex, and margins often cover an entire segment or company rather than AI-specific activity. Scope mismatches are disclosed and constrain the contract-to-capital score.

The first non-consensus conclusion is not that the highest-scoring company is the best AI investment. It is that the infrastructure stack contains different regimes of risk transfer.

Some companies receive deposits before building. Some lock in long-duration rent. Some secure capacity far ahead of a disclosed customer commitment. Some operate with exceptional margins but provide limited visibility into contract duration and price. The capex headline compresses all of these regimes into one number.

The more useful research question is where capital becomes contract-backed cash—and where the balance sheet is still carrying the future on behalf of the customer.

Research update layer

Have a primary-source contract signal that changes the index?

Route the evidence

Research commentary only. Provisional evidence-based scores as of July 24, 2026. Nothing on this page is investment, legal, tax, or financial advice. Company names and tickers are used for research identification; W-Axis Lab is not affiliated with the companies discussed.