AI ROI · Board Defense
The AI Capex Business Case
How CTOs Defend Infrastructure Asks to the Board
Dario Amodei has been repeating a line through 2025 and 2026: if the demand forecasts are off by a year, the frontier labs go bankrupt. TBPN’s Tech Earnings Recap on 2026-04-29 dissected the combined 2026 hyperscaler AI capex guidance with the energy of a sports broadcast — guidance that has continued to step up at each subsequent earnings call. This page is the spreadsheet version of that frame, written for a CTO who has to defend an infrastructure ask to a board that has now read the same coverage and wants to know whether the same logic applies inside the company.
30-SECOND EXECUTIVE TAKEAWAY
- The hyperscaler number is the wrong benchmark. Capex-to-revenue ratios at frontier labs run 20 to 40 percent. For an enterprise running internal productivity AI, the peer band is 1 to 10 percent depending on the business model.
- Stage-gate everything. The Amodei one-year-miss frame applies inside the building too. Break a multi-year capex envelope into 6 to 9 month tranches with re-decision points.
- Kill switches are the gate. Boards reject AI capex cases because the stop condition is missing, not because the upside is too small. Name the metric, the threshold, and the action up front.
Why the 2026 capex conversation is different
In 2023 the AI capex conversation at most enterprises was a few million dollars for a Copilot rollout and a vector database license. In 2026 it is an infrastructure conversation: GPU procurement, power contracts, custom-silicon partnerships at the large end, and serious six-to-nine-figure asks at the mid-market. The reason is not vibes. Model training costs at the frontier moved into the high hundreds of millions per run, inference costs moved with usage rather than dropping as much as raw token prices suggested, and the productivity programs that were supposed to fund themselves through headcount reductions ran into adoption ceilings that turned out to be higher than vendor case studies implied.
The hyperscalers responded by spending more capex than they have ever spent on anything, and they are now defending those numbers quarter after quarter. Microsoft, Alphabet, Amazon, and Meta have each stepped up 2026 AI-related capex guidance call after call, with the combined envelope now at historic highs and still moving. The exact running total shifts each quarter; the more durable point is that capex is now approaching, and in some quarters consuming, a large share of the operating cash flow that funds it. Amodei’s line about the one-year miss collapsing the labs is the same math seen from a different chair.
Boards have read the coverage. The TBPN segment titled “Capex Surge” on Diet TBPN 2026-05-01 is the kind of content a board director now has open in a tab during the budget review. The question they bring to the CTO is whether the company is doing too much capex, too little, or about right, and whether the case in front of them has the same structural defects the frontier-lab cases reportedly have. The defense lines below are the answer.
EIGHT DEFENSE LINES
What a board-ready AI capex case carries
Each line is one paragraph in the final document. The order matters. The kill-switch and stage-gate lines are the ones a serious board reviews first; they are also the ones missing from most cases that get rejected.
Capex as a percent of revenue, with a ceiling
State the percent you are spending and the ceiling you will not cross without board re-approval. Compare against peer benchmarks in the same business model, not against hyperscalers.
Depreciation schedule with stated useful life
24 to 36 months for GPUs, 5 to 10 years for shell and networking. Aggressive schedules reduce reported earnings short term but make kill switches cheap to pull.
NPV with a 24-month horizon
Anything longer is fiction at current model-generation cadence. Apply the adoption haircut and the capture haircut explicitly, and document the discount rate.
Kill-switch triggers at 9 and 18 months
Name the metric, the threshold, and the action. A capex program without a stated stop condition is the program a thoughtful board will not approve in 2026.
Stranded-asset write-down assumption
If we stop the program, what is the recovery value of the GPUs, the shells, the contracts? Build this into the case up front so the board sees the downside math.
Stage-gated commitment, not a single envelope
Break the ask into 6 to 9 month tranches. Each tranche has its own kill criteria. The Amodei frame applies inside the building too: a one-year miss on a single tranche should not blow up the program.
Revenue per dollar of capex, with the source
State the assumed revenue uplift or cost avoidance per dollar of capex deployed, with the data source. If the source is a vendor case study, say so. If the source is internal pilot data, say so.
Power and contract risk explicitly
Power is now the binding constraint at hyperscaler scale and is creeping into enterprise scale. Document the contracts, the durations, and the price exposure separately from the GPU spend.
The Amodei frame, applied inside the company
The Amodei frame is interesting at frontier-lab scale because it is the literal truth: the demand forecast has to be roughly right or the labs run out of money before the revenue arrives. At enterprise scale the literal version does not apply. A regular company is not going to go bankrupt because its internal AI productivity program underperformed. The frame still matters, though, because the structure of the bet is the same: capex commitments now, productivity or revenue capture later, with a wider gap between the two than most finance teams have modeled in a software business.
The applied version: assume a one-year miss on the demand or adoption side, and ask what breaks. If the answer is “we depreciate the GPUs faster and re-allocate the shells,” the case is staged correctly. If the answer is “we are committed to a multi-year contract with no exit,” the case is not staged and the board is going to find out the hard way that the kill switch is theoretical. The capex envelope should be broken into 6 to 9 month tranches, each with its own re-decision point, and the contracts behind it should match the tranche structure rather than the total envelope.
This is also where the CFO becomes a co-author rather than a reviewer. The depreciation schedule, the stranded-asset assumptions, and the contract structure are CFO territory; the productivity and revenue capture math is CTO and CAIO territory. A capex case with only one of those signatures is a case the board will not approve in 2026.
BOARD ROOM
Five objections, five answers
These are the actual questions a board director will ask in the room. The answers below are short on purpose. The full document carries the math; the room version is the position.
“The hyperscalers are spending unprecedented capex. Why aren’t we?”
Because we are not a hyperscaler. The capex-to-revenue ratio at MSFT, GOOG, AMZN, META is now in the 20 to 40 percent range. For our business model, the peer benchmark is 1 to 10 percent depending on whether AI is internal productivity, a feature, or the product. We benchmark against peers, not against frontier labs.
“What if we wait 12 months and let prices fall?”
Price per inference token has fallen 5 to 10x annually for two years and will probably continue. But our cost case does not depend on price falling; it depends on capture of the productivity or revenue benefit, which has its own time-to-value. Waiting saves money on the cost line and forfeits the same amount on the benefit line. The right answer is to stage the commitment, not to wait.
“What is the worst-case downside?”
Stranded GPUs at 30 to 50 percent recovery value, broken power contracts at 60 to 80 percent of remaining commitment, and one to two quarters of distraction cost on the team. We have written this into the case. The kill-switch trigger fires before the worst case becomes catastrophic.
“How is this different from the 2000 fiber buildout?”
It is not entirely. There is a real possibility of overbuild, which is why we stage-gate the commitment and depreciate aggressively. The difference is that the productivity benefits of the deployed capacity show up inside the company, not just at the providers, which is a hedge the 2000 fiber companies did not have.
“What is the CFO’s contingency plan?”
A documented stop-or-restructure plan at each kill-switch trigger, an internal asset re-allocation path (training to inference, or inference to internal use), and a written-down recovery value baked into the depreciation schedule from day one. The CFO is co-author on the case.
The 10-20-70 trap
BCG and McKinsey both publish versions of a 10-20-70 rule for AI program spend: 10 percent on models and algorithms, 20 percent on technology and data, 70 percent on people, process change, and adoption. The capex case usually addresses the first 30 percent and waves at the rest. The wave is the failure mode. A capex program that funds the GPUs but underfunds the change management produces stranded utilization, which looks on the books like the model was wrong when the model was fine and the adoption funding was wrong.
For a board defense, the answer is to show the opex line that funds the 70 percent next to the capex line that funds the 30 percent, and to tie the kill-switch metric to adoption rather than to utilization. Utilization without adoption is the worst outcome: GPUs are busy, money is spent, productivity capture is missing. The capex case has to make this visible up front or the program will surface it in the third quarterly review, by which point the conversation is already defensive.
Power, contracts, and the constraint everyone is now talking about
Power has moved from a hyperscaler concern to a mid-market concern in the last 18 months. The first time a CTO discovers that the data center hosting the GPU reservation cannot actually deliver the contracted power density is usually not in the procurement review; it is in the post-mortem on the deployment slip. The capex case should carry power as a separate line item with the contract duration, the price exposure, and the alternate-site fallback if the primary site under-delivers.
Contracts are the other place a capex case quietly under-discloses. A two-year GPU reservation at a hyperscaler at the higher tier pricing looks like an operating expense in the immediate quarter and a capital commitment in the medium term. Boards now ask for the off-balance-sheet view of these commitments alongside the on-balance-sheet capex line. The case that shows both is the case that gets approved. The case that shows only the capex line is the case that gets a follow-up question the CTO did not expect.
Cross-link: this case sits inside a larger CAIO budget conversation
The capex case is one slice of the broader CAIO budget archetype work. A CAIO setting a multi-year envelope picks an archetype (foundation-builder, capability-acquirer, transformation-investor, hedger) and the capex case sits inside that archetype with the right ratios and time horizon. The CAIO hub and the budget-archetypes work cover the envelope; this page covers the defense of the capex line item inside it. The interactive math lives at the AI ROI calculator; the failure-mode framing sits at the AI project failure rate guide.
AI Capex Business Case: Frequently Asked Questions
How do you calculate AI capex ROI?
What is the projected hyperscaler AI capex for 2026?
Is there a formula for capex that boards accept?
What is a good capex-to-revenue ratio for AI infrastructure?
What is the 10-20-70 rule for AI?
How does Dario Amodei’s “go bankrupt if forecasts are off by a year” frame apply to enterprise capex?
How much should AI capex per employee be in 2026?
What is the right depreciation schedule for AI capex?
Continue the AI ROI cluster
Capex is one slice. The business case template, the calculator, and the failure-rate framework cover the rest.