A Field Guide to AI Compute AI Compute Guide Cloud Layer
MAY 2026
Menu

LAYER 08 CLOUD LAYER $ / hr-Hopper · depreciation

Last revised · MAY 13, 2026

How hardware becomes cloud revenue.

This layer connects spending on compute infrastructure to rental pricing, utilization, and payback time.

$480 B

AI capex flowing in

$220 B

silicon captured

$57 B

service revenue out

~8×

capital multiplier

Native unit

$ / hr-Hopper · depreciation

What constrains it

Cloud economics depend on capex, utilization, pricing, and how quickly hardware depreciates.

FIG. L08 · SIGNATURE CLOUD CAPITAL → SILICON → REVENUE

Fit overview · pinch to zoom

FIG. 08 · LAYER 08 · CLOUD CAPITAL
$480 billion goes in. $57 billion comes back.
For every dollar deployed at the top of this chart, $0.46 becomes NVIDIA silicon and $0.12 comes back as annual revenue. An 8× capex-to-revenue multiplier — either the largest generational infrastructure bet in history, or the largest capex bubble. Numbers at Q4 2026 run-rate.
CAPITAL · STAGE 01
$480 B
deployed per year into AI compute
SILICON · STAGE 02
$220 B
NVIDIA data-center revenue
REVENUE · STAGE 03
$57 B
actual money earned from compute
BIG-TECH OPERATING CF
$330 B
Microsoft · Google · Meta · Amazon · Oracle — the four-and-a-half-firm club that earns enough on its existing business to self-fund a global GPU buildout. ~68% of all annual AI capital.
SOVEREIGN WEALTH · $45 B
PIF · MGX · GIC · Mubadala · ADIA
PRIVATE CREDIT + DEBT · $55 B
Blue Owl · Carlyle · Apollo · Blackstone incl. GPU-backed asset-finance vehicles
AI-LAB EQUITY ROUNDS · $50 B
OpenAI · Anthropic · xAI · Mistral plus secondaries & employee tender
NVIDIA GROSS MARGIN
$165 B
The single largest claim on the AI dollar. Captured by one company before the chip is even installed in a rack. Data-center gross margin ~75% on ~$220 B revenue.
SILICON COGS · $55 B
TSMC wafers · SK Hynix / Samsung / Micron HBM ASE / Amkor packaging · substrate & test
API revenue · $22 B
OpenAI · Anthropic · et al.
Consumer subs · $12 B
ChatGPT Plus · Claude Pro
Embedded prods · $8 B
Copilot · Cursor · GitHub
Enterprise · $15 B
per-seat & per-call contracts
WHERE THE OTHER 54¢ GOES
Of every $1 of CAPITAL flowing in (above), $0.46 becomes NVIDIA silicon. The other $0.54 — about $260 B annually — buys land, power, networking, copper cabling, real estate, labor, and the debt service on all of it.
orange · NVIDIA gross margin captured
|
grey bands · flow at 1.25 px / $B · heights to scale
FIG. 08 · CAPITAL · Q4 2026 · v1.0
$480 B capital · $220 B silicon · $57 B revenue. ~8× multiplier. Orange marks NVIDIA’s gross-margin capture — the largest single claim on the AI dollar.

What this layer does

Cloud economics translate equipment into a service someone can rent. The important questions are simple: how expensive was the hardware, how often is it used, and how quickly does that investment need to pay back?

How the model works

Read the cloud layer as a timing problem

The machine makes money only after three bets line up.

The signature figure shows capital turning into silicon and then into revenue. The deeper operating question is whether utilization and pricing rise fast enough to outrun the depreciation clock.

01 · Capex first

Cloud capacity is purchased before it is monetized.

A provider commits to land, buildings, networking, chips, and power long before the first accelerator-hour is sold. That is why capex jumps appear before revenue does.

Spend now

02 · Utilization next

Idle accelerators are margin poison.

The same rack can look brilliant or terrible depending on occupancy. Every unsold hour still carries depreciation, financing, and facility overhead.

Keep it busy

03 · Payback clock

Depreciation decides how fast revenue must arrive.

If hardware is expected to earn back its cost over a shorter useful life, pricing pressure rises. If it stays productive longer, the cloud has more room to compound returns.

Beat the clock

The simple mental model

cloud revenue = price / accelerator-hour × hours sold

Revenue lens

Price × occupied hours

Rental economics start with what a customer pays per accelerator-hour, multiplied by how many hours the fleet is actually sold.

Cost lens

Depreciation + power + service

The service must absorb the hardware write-down, the data-center bill, networking, support, and any customer-specific platform costs.

Margin lens

Why utilization matters so much

High occupancy spreads fixed cost across more billable work. Low occupancy makes even premium hardware look overbuilt.

The supporting view

FIG. 8.2 LAYER 08 DEPRECIATION CURVES

Fit overview · pinch to zoom

FIG. 8.2 · LAYER 08 · DEPRECIATION · SUPPORTING
Three years of accounting runway.
A NeoCloud books a Hopper across 4 years. A hyperscaler books the same chip across 7. The accounting choice slides break-even from year ~2.6 to year ~4.6 — and turns the same $1.50/hr rental into a 21% IRR instead of a 10% IRR.
BOOK VALUE OVER TIME · STRAIGHT-LINE DEPRECIATION
BOOK VALUE
YEARS SINCE PURCHASE
100%
75%
50%
25%
0%
0
1
2
3
4
5
6
7
8
BREAK-EVEN · 35% BOOK · $1.50/hr × 35% GM
year 2.6
year 4.6
4-year · NeoCloud
7-year · hyperscaler
+ 3 years of runway
the gap between the curves is the hyperscaler's accounting edge — three extra years of rental runway above break-even
IRR CONSEQUENCE · PER GPU PER YEAR
GROSS RENTAL REVENUE
$1.50/hr × 8 760 hr × 80% util = $10 500
GROSS PROFIT · AFTER DEPRECIATION · IRR
4-YEAR · NEOCLOUD
depreciation $7 500/yr
profit $3 000/yr
IRR
≈10%
7-YEAR · HYPERSCALER
depreciation $4 300/yr
profit $6 200/yr
IRR
≈21%
WHAT THIS BUYS YOU
Same chip, same rental, same utilization. Stretch the depreciation schedule and you book $3 200 more profit per GPU per year. Across an Anthropic-scale fleet of one million Hoppers, that is roughly $3 B of accounting alpha — the difference between a NeoCloud and a hyperscaler.
orange · 7-year curve · hyperscaler accounting
|
grey · 4-year curve · NeoCloud accounting
FIG. 8.2 · LAYER 08 · DEPRECIATION · v1.0
Same chip, same rental rate, same utilization. A longer depreciation window pushes break-even later and leaves more accounting room for the operator.

The cloud is increasingly pre-sold

Backlog now finances the hardware. Oracle’s FY2026 Q3 release puts remaining performance obligations at $553 billion, up 325% year over year, against roughly $50 billion of capex guidance for the same year. CoreWeave reported a $99.4 billion revenue backlog at Q1 FY2026 against a full-year revenue guide of $12–13 billion.

These are multi-year, take-or-pay commitments signed before the racks energize. Operators raise capital against contracted paper rather than against speculative utilization, and the payback clock starts at signing, not at first watt.

The risk shifts accordingly. A counterparty default or a falling spot price for compute does not slow the depreciation curve — the hardware still ages on its own schedule.

The NeoClouds burn cash to convert backlog into racks

The other side of pre-sold compute: converting paper into power costs cash up front. CoreWeave’s Q1 FY2026 release shows $7.7 billion of property and equipment purchases against $2.08 billion of revenue — capex at 3.7× the top line, and free cash flow of negative $4.7 billion. Nebius spent roughly $2.5 billion in the same quarter to support $399 million of revenue, itself up 684% year over year.

The hyperscaler-vs-NeoCloud distinction is not about backlog. Both sides pre-sell capacity. The split is who can absorb the negative free-cash-flow gap during the buildout.

That is why each new NeoCloud quarter brings a new financing event. CoreWeave’s $3.1 billion DDTL 5.0, disclosed in a May 18, 2026 8-K, is the latest — capital raised against contracted backlog and pledged GPUs, drawn to bridge the gap between signing and first watt.

The Chinese cloud trio runs the same buildout in RMB

Same architecture, different chips, different currency, same pace. Alibaba’s March-quarter FY2026 release reports RMB 126 billion of full-year capex against Cloud Intelligence revenue of RMB 158 billion. Tencent’s Q1 2026 release shows RMB 31.9 billion of capex in a single quarter — already 40% of full-year 2025 spending.

Demand is pulling the spend forward. Baidu’s Q1 2026 release puts AI Cloud Infra revenue at RMB 8.8 billion, up 79% year over year, with GPU Cloud revenue up 184%. The acceleration mirrors what the US hyperscalers print in dollars.

The constraint inside China is which silicon clears export control, not whether the workload exists. Two AI infrastructure builds now run in parallel, on opposite sides of the export line, at comparable intensity.

Cloud pricing is where hardware decisions become business outcomes.