Vault
analysis

Four Mountains — Splitting $3T by Why It Is Being Spent, and Who Pays

Created

Four Mountains — Splitting $3T by Why It Is Being Spent, and Who Pays

Builds-on: what-three-trillion-has-to-earn, five-years-of-capex-already-signed Related: the-shadow-balance-sheet-nikkei-1-65t-and-the-spv-layer, compute-as-collateral-the-residual-value-wrap, the-mask-that-eats-what-it-hides, ai-circular-financing-and-banking-exposure-audit, ai-infrastructure-endgame-indicators, cyclical-20-and-the-ai-capex-mask Led-to: ai-capex-watchlist-check-august-20-2026

what-three-trillion-has-to-earn priced the whole $3T as if every dollar had to be repaid in sold tokens, then corrected itself: Meta sells no tokens and its ad lift already covers its capital charge. That correction was left half-finished. This doc finishes it by splitting the $3T four ways according to why the money is being spent and who the counterparty is, then running the token test separately on each.

The result changes the shape of the problem. The aggregate requirement falls — but the distribution is much worse than the aggregate implied, and the guarantees turn out to be written on the wrong mountains.

Model: workspaces/ai-capex-mask/model/mountains.py.

The four mountains

The sorting rule is a single question: is the spend contracted to a paying counterparty, and does that counterparty have external revenue?

Allocation

Firm Off-BS ($B) M1 M2 M3 M4 Anchor for the split
Meta 420 210 — — 210 Sells no cloud. Ad lift is real; 5GW campuses are not for ad ranking.
Alphabet 500 225 175 — 100 $811B obligations haircut for non-AI. Own lab, so no counterparty risk.
Amazon 600 90 300 150 60 AWS external book is genuinely diversified; Rainier is the M3.
Microsoft 700 70 245 315 70 OpenAI = 70% of FY26 AI revenue, 45% of the $625B backlog. $329.1B of leases not yet commenced.
Oracle 273 — 68 191 14 RPO $553B, of which ~$300B (54%) is Stargate/OpenAI. No internal AI business.
Nvidia / Broadcom / AMD 400 — 60 240 100 RVG and platform exposure, wrapped around neocloud and lab borrowers.
SpaceX / other 107 — 11 11 86 xAI-adjacent, sovereign, largely uncontracted.
Total 3,000 595 859 907 639
Share 20% 29% 30% 21%

The firm-level percentages are judgement calls and the honest place to attack this. The anchors are disclosed facts; the splits between them are mine. Meta's 50/50 is the softest — its ad lift genuinely comes from the new compute, so the line between "ads infrastructure" and "moonshot" is not clean. Alphabet's $811B figure certainly includes non-AI obligations and is haircut accordingly. Change these and the mountain sizes move by a few hundred billion; the ordering and the conclusions do not.

What each mountain must earn

Applying the cost-to-serve ratio from what-three-trillion-has-to-earn (21.7% of commitment per year, being the amortization plus opex) and the 1.8x model-layer markup for the merchant tiers:

M1 — Defensive · $595B commitment · $129B/yr cost to serve

Customer: advertisers and users who already pay. Tokens required: zero.

This is the only mountain with a customer that exists, is already transacting, and is not being asked to change behaviour. The test is whether incremental product revenue exceeds the carry. Meta alone: $47.3B of 2026 ad revenue growth against a $14.7B capital charge — passes at any plausible AI attribution. Google defending 5% of $239.5B of search advertising is worth $12B/yr.

Verdict: passes today. No token forecast required, and none of the analysis in this thread threatens it.

M2 — Merchant, diversified · $859B commitment · $187B/yr cost to serve

Customer: enterprises spending IT budget. Required end-user revenue: $336B/yr.

That is 24% of global enterprise software ($1.40T) and 5% of total worldwide IT spending ($6.37T). Demanding, and not absurd — this is the tier where the ordinary software-budget story actually works, because the customers are diversified, contracted, and creditworthy.

Verdict: hard but bounded. The constraint is IT-budget growth, which is a known quantity that grows around 10–14% a year.

M3 — Merchant, lab-dependent · $907B commitment · $197B/yr cost to serve

Customer: OpenAI and Anthropic. That is the entire customer list.

Required: $355B/yr of end-user revenue from the two labs, against roughly $110B today. 3.2x in five years — a 26% CAGR.

Taken by itself that is the least alarming number in this entire thread. Both labs have grown far faster than 26% recently. The problem is not the growth rate; it is that thirty percent of a three-trillion-dollar buildout has a customer list with two names on it, both private, both loss-making, and both partly funded by the same firms selling them the compute. Microsoft's own disclosure is the proof: $24.1B from OpenAI, 70% of its FY26 AI revenue, 45% of the backlog.

Verdict: plausible on growth, unacceptable on concentration. This is a credit problem wearing a demand problem's clothes.

M4 — Speculative · $639B commitment · $139B/yr cost to serve

Customer: none identified.

A fifth of the entire buildout — $639B of commitments carrying $139B a year — is contracted against a counterparty that does not exist yet. This is not a criticism of anyone's intentions. Option value is a legitimate reason to build when the option is cheap. This option is not cheap.

The independent cross-check supports the size, though the sharpest version of it is contested. Sightline Climate counts 190 GW announced across 777 facilities since 2024, 16 GW slated for 2026, and only 5 GW under construction, with 30–50% of the 2026 pipeline judged unlikely to materialize.

(Qualified 2026-08-20. SemiAnalysis argues Sightline's 5 GW under-construction figure is "off by multiples" — the top two hyperscalers alone exceed it — and that the "half of 2026 capacity is cancelled" headline is an artifact of taking announcements at face value. Their read: the real bottleneck is not cancellation but interconnection queues of 7–10 years in major metros and 1–5 year equipment lead times, and their own 2026 forecast has moved about 1% in six months. This mostly supports the M4 sizing while relocating its meaning. If the announcement layer is full of what SemiAnalysis calls "phantom demand" — they count 311 GW of it in Texas alone — then a large speculative tier is real; it simply never had financing attached, so it strands as abandoned optionality rather than as defaulted debt. Independent confirmation arrived August 3–14: ERCOT's interconnection queue holds ~1,800 projects and 474 GW, roughly 90% data centers, and Texas paused new approvals pending an audit of up to 300 of them. A 474 GW queue against a national buildout measured in tens of GW is phantom demand quantified by a grid operator.)

The 21% estimate here is if anything conservative.

Verdict: fails, by construction. There is no test to run.

The waterfall — what happens if the payroll pool never opens

How much revenue arrives without labor substitution — on tooling diffusion alone, more firms buying more tools at rising but sub-payroll intensity? Gross global enterprise AI spend is ~$419B today, of which roughly 30% reaches the compute layer. Grow that at 25% a year for five years, add $104B of consumer, and the ceiling is $488B/yr. Serve the mountains in order of claim strength.

Mountain Needs Gets Covered Commitment stranded
M2 merchant/diversified $336B $336B 100% —
M3 lab-dependent $355B $152B 43% $517B
M4 speculative $250B $0B 0% $639B
Total stranded ~$1.16T

(Revised 2026-08-19. The first version used a $289B ceiling built on what-three-trillion-has-to-earn's hand-set $35/seat/month, which that doc's correction section shows was 3–5x below measured spend per employee. On the corrected base M2 clears rather than falling short. The direction of the finding is unchanged; its severity is about a third lower. Sensitivity is real: at 15% tooling growth the ceiling is $357B and M2 barely clears; at 35% it is $668B and M3 reaches 94%.)

The corrected statement is narrower and holds better. Tooling diffusion alone funds the safest merchant tier and under half the lab tier. It funds none of M4, and leaves roughly $1.16T of commitment without a payer. Labor substitution is not required for the buildout to be partly fine. It is required for M3 to survive intact and for M4 to be anything other than a write-off.

The Ramp distribution is what makes this tractable rather than speculative: the top 1% of adopters already spend ~$90,000 per employee per year — 95% of a knowledge-worker salary. The payroll-scale price point exists in production today. The question is diffusion speed, not whether the market can exist, and diffusion speed is exactly what decides how much of M3 and M4 gets paid for.

Where the guarantees sit

Now put the allocation next to compute-as-collateral-the-residual-value-wrap.

The vendor residual-value exposure — Nvidia's up-to-25% support, Broadcom's AI XPV wrap, the ~$70B of disclosed off-balance-sheet guarantees — allocates 0% to M1, 15% to M2, 60% to M3, and 25% to M4.

85% of the wrap sits on the two mountains with the weakest or absent customer. None of it sits on the only mountain that passes its test today.

That is not an accident and it is not a scandal — it is exactly what guarantees are for. Meta's ad-ranking capex does not need a credit enhancement because Meta's advertisers already pay. Oracle's Stargate build and the neoclouds need one precisely because their counterparty is two private companies and, in M4's case, nobody at all. The wrap exists in proportion to the weakness of the underlying demand.

Which means the residual-value guarantee is a fairly precise map of where the demand risk is concentrated, drawn by the parties with the best information. Read it that way and it stops being a scandal and becomes an instrument reading. The instrument says: the risk is in M3 and M4, and it is $1.5T of commitment.

flowchart TB
  A[3T of commitments] --> B[M1 defensive 595B<br/>20%]
  A --> C[M2 merchant diversified 859B<br/>29%]
  A --> D[M3 lab-dependent 907B<br/>30%]
  A --> E[M4 speculative 639B<br/>21%]
  B --> F[Customer: existing advertisers<br/>tokens needed: zero]
  C --> G[Customer: enterprise IT budgets<br/>needs 336B per year]
  D --> H[Customer: OpenAI and Anthropic<br/>needs 355B per year]
  E --> I[Customer: none identified]
  F --> J[PASSES today]
  G --> K[Hard but bounded]
  H --> L[Depends entirely on<br/>labor substitution]
  I --> L
  M[Residual-value guarantees<br/>85% wrapped here] -.-> D
  M -.-> E

Revised aggregate

Stripping M1, which needs no tokens at all, the token-funded requirement falls from $1,170B/yr to $941B/yr — 358 quadrillion paid tokens a year at proprietary pricing, 1,775 quadrillion at open-weight pricing.

The aggregate got better. The picture got worse. A $1.17T requirement spread evenly across nine well-capitalized firms is a growth problem. A $941B requirement in which $595B of the stack is already fine, $336B lands on IT budgets that can nearly bear it, and $605B lands on two private companies plus a tier with no customer at all — that is a concentration problem, and concentration problems resolve through credit events rather than through slow disappointment.

For cyclical-20-and-the-ai-capex-mask: the capex line keeps printing regardless, because M1 and M2 are healthy and M3/M4 are contractually pre-committed per five-years-of-capex-already-signed. What breaks is not the GDP contribution. It is the $1.5T of M3+M4 commitment sitting under 85% of the guarantees.

What would change this

Sources

Backlinks