Capex vs Opex: Buying GPUs vs Renting GPU Hours

EXPLAINED · Versus · Last reviewed: 07 Oct 2026 · Next review: after CoreWeave and Nebius Q3 2026 results

Capex vs opex is the split between buying an asset that lasts years and paying a running cost as you go. Capex, or capital expenditure, is cash spent on things like GPU servers and buildings. Opex, or operating expense, is the cost of running the business now, like rented GPU hours, power and staff.

Why an investor cares: one company’s opex is another company’s revenue, and its capex. When an AI lab rents GPU hours, someone else has already paid for the GPUs. For how that up-front cash works, see cash capex, explained. For how the hours are sold, see GPU as a service, explained.

What is the difference between capex and opex?

Capex buys something that lasts. Opex pays for something you use up now. That is the whole split.

The difference is timing. Capex cash leaves on the day the kit is paid for, but it reaches profit slowly, as depreciation. Opex reaches profit in the same period you pay it. So two companies can spend the same cash and report very different profits.

Capex vs opex examples in AI compute

Buying. Say a company spends $1 billion on GPU servers this year. All $1 billion of cash leaves now. CoreWeave depreciates its technology equipment over six years, its FY2025 10-K says. On a straight line, that is about $167 million of expense a year ($1 billion ÷ 6, no salvage, our arithmetic). Microsoft gives its servers and network equipment two to six years. For why the life matters, see GPU depreciation, explained.

Renting. Nebius lists an NVIDIA H100 at $4.50 per GPU-hour on demand from 1 October 2026, up from $3.85. One GPU running every hour of a year is 8,760 hours. At list price that is $39,420 of opex (8,760 × $4.50, our arithmetic). Contract buyers pay what their contract says, not the list, so treat this as a list-price example, not a typical bill. See Nebius’s October price rise.

The renter has no asset to depreciate and no up-front cash. It also has no GPU at the end. The bill stops when the hours stop, unless a contract says otherwise.

Is renting GPUs capex or opex?

Mostly opex, but not always. Plain cloud bills are a running cost. Long leases are different. Under US accounting, a lease of more than a year goes on the balance sheet.

Meta shows the opex side. Its Q2 2026 10-Q, filed 30 July 2026, says cost of revenue rose 33% to $11.33 billion, “primarily due to higher infrastructure expenses related to our data centers, technical infrastructure, and third-party cloud services.” That is rented compute landing straight in costs. Meta also lists $349.31 billion of non-cancelable commitments, mostly third-party cloud capacity plus its own servers, network gear, data centres and consumer hardware.

Microsoft shows the lease side. In its FY2026 10-K, filed 29 July 2026, operating leases cost $6.97 billion as one running line. Finance leases cost $7.95 billion, split into $5.40 billion of amortisation and $2.55 billion of interest. A finance lease looks a lot like owning, paid for over time.

The big number is the one not started. Microsoft says it had $329.1 billion of additional leases, primarily for datacenters, that had not yet commenced at 30 June 2026. They start between fiscal 2027 and fiscal 2033, with terms of one to 20 years. Its purchase commitments of $194.1 billion also include take-or-pay contracts.

Who carries the capex when you rent?

The provider does. Renting moves capex off the renter’s cash-flow statement and onto the neocloud’s. That is the business model.

CoreWeave is the clearest print. Its Q2 2026 10-Q, filed 12 August 2026, shows $14.1 billion of property and equipment bought in the first half of 2026, up from $3.9 billion a year earlier. Revenue in the same six months was $4.65 billion. Capex was about three times revenue (14.117 ÷ 4.653 = 3.0, our arithmetic). See what CoreWeave does.

Nebius shows the same pattern, more sharply. In Q2 2026 it reported revenue of $582.3 million and spent $5.66 billion on property, equipment and intangibles. That is about 9.7 times revenue in one quarter (5,657.4 ÷ 582.3, our arithmetic). Its operating cash flow was still $2.25 billion positive in the quarter.

The capex is paid for with debt and with contracts. CoreWeave had $35.6 billion of debt principal at 30 June 2026 and $1.18 billion of net interest expense in the half. Its 10-K says committed contracts run one to six years and are “take-or-pay,” paid whether the hours are used or not. In Q2, 98% of its revenue came from those commitments. See take-or-pay GPU contracts and who finances GPU clusters.

Who gets paid, and where is the stuck step?

The renter pays the neocloud as opex. That opex is the neocloud’s revenue. The neocloud has already paid the chip, server and network makers, as capex, on their invoices. Lenders are paid interest along the way. Vendors are paid first. The neocloud is paid last, hour by hour.

The stuck step is the gap between cash out and hours billed. The neocloud carries it. CoreWeave reports $103.7 billion of remaining performance obligations, with 41% expected as revenue in the 24 months to June 2028. That is work promised, not cash received. Until the halls are live and the hours bill, the kit is paid for and the revenue is not.

The risk also sits in few hands. CoreWeave’s largest customer was 36% of Q2 2026 revenue and the second was 26%. A renter can shift its spend; the owner still holds the servers and the debt.

Why rent GPUs instead of buying them?

Speed and cash. Renting gets a lab compute without building a hall or waiting for servers. It keeps cash on hand and turns a large up-front bill into a running cost.

The trade-off is control and price. On-demand rates can move, as Nebius’s did on 1 October 2026. Long contracts fix the supply but lock in the payments. Meta, for example, has contingent obligations to buy up to $14.72 billion of cloud capacity over five years, reduced only if the provider can sell it elsewhere.

Which is cheaper depends on how busy the GPU stays. An owned GPU that sits idle still depreciates. A rented hour you do not use costs nothing, unless you signed take-or-pay. For the bigger version of this choice, see neocloud vs hyperscaler.

Common misconceptions

• People assume renting GPUs means no capex. Actually the capex moves to the provider. CoreWeave bought $14.1 billion of kit in six months to rent out.
• People assume every rental is opex. Actually a long lease goes on the balance sheet. Microsoft’s finance leases cost $7.95 billion, split into amortisation and interest.
• People assume capex hits profit when it is spent. Actually it reaches profit slowly, as depreciation. A $1 billion buy over six years is about $167 million a year.
• People assume opex is always flexible. Actually take-or-pay contracts bill whether the hours run or not. Meta lists $349.31 billion of non-cancelable commitments.

FAQ

Related terms

• Cash capex — the cash that leaves before AI hours bill.
• GPU as a service — how rented GPU hours are sold.
• GPU depreciation — how bought servers reach profit over time.
• Take-or-pay GPU contracts — rented hours that bill even when idle.

Stack doors

Layers: Neoclouds and Hyperscalers. Markets: how CoreWeave’s contracts become cash. Analysis: when GPUs are ordered before renters sign.

Get the Wire — short daily notes when GPU prices, contracts, or capex prints move.

Sources