> ## Content Index > Fetch the complete content index at: https://www.sotkn.com/llms.txt > Use this file to discover other available public pages before exploring further. # What Is AI Capex? Hyperscaler Spend Before AI Hours Bill - URL: https://www.sotkn.com/explained/ai-capex-explained/ - Published: 2026-10-14T01:18:25.000Z - Updated: 2026-10-14T01:18:25.000Z - Description: AI capex is capital spent on GPUs, servers, networking and data-centre build so AI workloads can run — cash that leaves before those hours bill as cloud or ads revenue. Who gets paid first: chips, landlords, power. The stuck step is the lag between spend and billed utilisation. - Author: Sean - Tags: Mechanism, #desk-explained EXPLAINED · Mechanism · Last reviewed: 14 Oct 2026 · Next review: after Microsoft / Alphabet / Amazon / Meta next CapEx prints (Q3–Q4 2026 earnings) AI capex is capital expenditure that hyperscalers and other large AI builders book for accelerators, servers, networking and data-centre build — including related finance leases where the company counts them in CapEx — so they can train and serve AI workloads before those hours bill as revenue. **In short.** AI Capex is the cash hyperscalers spend on GPUs, servers, networking and data halls before AI hours bill. As of late July 2026, Amazon guided about $220 billion of cash CapEx for the year, Alphabet $195–205 billion, Microsoft about $175 billion for calendar 2026 after a lease reclassification, and Meta $130–145 billion including finance-lease principal. Who gets paid first: chip and packaging suppliers, landlords and power. The stuck step is the lag — cash leaves before utilisation converts to cloud revenue or ads return. Why an investor cares: CapEx guidance is the cash leaving the hyperscaler balance sheet for AI kit and halls. Revenue and remaining performance obligation (RPO) show demand later. Free cash flow sits in between. For the company type, see [what a hyperscaler is](https://www.sotkn.com/explained/hyperscaler-explained/). For cash timing of CapEx versus depreciation, see [cash Capex explained](https://www.sotkn.com/explained/cash-capex-explained/). ## What counts as AI Capex? Two buckets show up in IR decks. Short-lived assets: GPUs, other accelerators, CPUs, servers and networking that depreciate in a few years. Long-lived assets: data-centre land, buildings and related finance leases. AI Capex is CapEx aimed at those AI workloads — not a separate GAAP line. Microsoft’s FY2026 fourth-quarter CapEx was $41 billion. Roughly two thirds went to short-lived assets, primarily CPUs and GPUs. Finance leases that quarter were $5.6 billion, mainly large data-centre sites, and cash paid for property and equipment was $35.8 billion. Alphabet’s Q2 2026 CapEx was $44.9 billion of purchases of property and equipment; on the call, about 60% of technical-infrastructure investment that quarter was in servers and 40% in data centres and networking. Meta’s CapEx figure includes principal payments on finance leases. Its 2026 outlook is $130–145 billion on that basis. When you compare guides across names, check whether leases sit inside CapEx or outside. Our page on [Capex vs Opex](https://www.sotkn.com/explained/capex-vs-opex/) covers buy-versus-rent for GPU hours. ## How big is AI Capex in 2026? Late July 2026 prints put four hyperscalers in the low- to mid-hundreds of billions for the year — on their own definitions. | Company | 2026 CapEx guide (as disclosed) | Latest quarterly CapEx print | | --------- | --------------------------------------------------------------------------------------------- | -------------------------------------------------------------------------------- | | Amazon | \~US$220B cash CapEx (raised from \~$200B; memory costs) | Q2 cash CapEx $53.1B; PPE purchases $54.2B (call / release 30 Jul 2026) | | Alphabet | US$195–205B (raised from $180–190B) | Q2 purchases of PPE $44.9B; \~60% servers / 40% DC + networking (22 Jul 2026) | | Microsoft | \~US$175B CY2026 after useful-life/lease reclass (underlying plan unchanged vs prior \~$190B) | Q4 FY26 CapEx $41B; \~2/3 short-lived CPU/GPU; Q1 FY27 CapEx >$50B (29 Jul 2026) | | Meta | US$130–145B incl. finance-lease principal (narrowed from $125–145B) | Q2 CapEx $31.08B (29 Jul 2026) | *Sources: company earnings calls and releases, late July 2026\. Definitions differ (cash CapEx vs leases). Not a tip sheet.* Amazon raised cash CapEx to about $220 billion for 2026 from about $200 billion, citing higher memory costs, and still said capacity would not meet all 2026 demand and likely not all 2027 demand. Alphabet raised full-year CapEx guidance to $195–205 billion from $180–190 billion to accelerate capacity delivery. Microsoft kept underlying calendar-2026 investment expectations unchanged but reported an adjusted CapEx expectation of about $175 billion after extending useful lives of data centres and office buildings from 15 to 25 years — a change that shifts more future leases from finance (inside CapEx) to operating (outside CapEx). ![AI Capex — 2026 CapEx guidance for Amazon, Alphabet, Microsoft, Meta with cash-before-hours stuck step](https://storage.ghost.io/c/cd/10/cd10e222-131f-4ba4-b4a9-ca4214e8c12d/content/images/2026/10/ai-capex-chart.png) Company CapEx definitions differ. Amazon call 30 Jul · Alphabet call 22 Jul · Microsoft call 29 Jul · Meta release 29 Jul 2026. ## Who gets paid when AI Capex is spent? Chip designers and foundries, memory and packaging suppliers, server and networking vendors, landlords and construction contractors, and power suppliers. They get paid when tools, packages, racks, halls and megawatts are delivered or leased — often quarters before a cloud customer’s token invoice. Hyperscalers pay that CapEx from operating cash flow, debt and equity raises. Alphabet’s June 2026 equity raise stated proceeds would support CapEx to scale AI infrastructure. Neoclouds and other GPU renters sit beside this lane — they often lease kit rather than own the full stack. See [neocloud vs hyperscaler](https://www.sotkn.com/explained/neocloud-vs-hyperscaler/) and [neocloud explained](https://www.sotkn.com/explained/neocloud-explained/). ## Where is the stuck step? Cash leaves before hours bill. CapEx and PPE purchases print first. Free cash flow can turn negative even when revenue grows. Amazon’s trailing-twelve-month free cash flow as of 30 June 2026 was a $7.6 billion outflow, driven primarily by a year-over-year increase of $66.1 billion in purchases of property and equipment net of sales and incentives, reflecting AI investments. Demand can still exceed supply after CapEx rises. Microsoft said customer demand continued to exceed available Azure capacity; commercial RPO grew 84% to $678 billion. Alphabet’s Google Cloud backlog was about $514 billion on the Q2 call. Those backlogs are contracts not yet recognised as revenue — see [remaining performance obligation explained](https://www.sotkn.com/explained/remaining-performance-obligation-explained/) and [RPO vs take-or-pay](https://www.sotkn.com/explained/rpo-vs-take-or-pay/). A second bind is the hall and power clock: CapEx for servers without energised space still waits. Facility definition: [what is an AI data center](https://www.sotkn.com/explained/what-is-an-ai-data-center/). Layer door: [AI Data Centers](https://www.sotkn.com/stack/ai-data-centers/). Hyperscale versus colo rent: [hyperscale vs colocation](https://www.sotkn.com/explained/hyperscale-vs-colocation-data-center/). ![AI Capex — where cash goes and where it sticks: short-lived kit, halls, who gets paid, cash before hours](https://storage.ghost.io/c/cd/10/cd10e222-131f-4ba4-b4a9-ca4214e8c12d/content/images/2026/10/ai-capex-infographic.png) Stuck step: cash before billed hours. Sources: company late-July 2026 materials. ## How AI Capex shows up on the income statement later After purchase, short-lived kit hits depreciation and cost of revenue as it is used. Longer halls hit depreciation over many years — Microsoft’s useful-life extension to 25 years changes that timing, not the cash already spent. GPU-specific depreciation clocks: [GPU depreciation explained](https://www.sotkn.com/explained/gpu-depreciation-explained/). Some buyers convert CapEx pressure into opex by leasing GPUs or halls. Take-or-pay GPU contracts lock spend without owning the chip: [take-or-pay GPU contracts](https://www.sotkn.com/explained/take-or-pay-gpu-contracts-explained/). ## What to check next Next CapEx guides and free-cash-flow prints from Microsoft, Alphabet, Amazon and Meta through Q3–Q4 2026\. Watch whether short-lived (GPU/CPU) mix stays dominant, whether lease classification moves CapEx up or down without changing the underlying build, and whether RPO growth still outruns live capacity. Earnings-season line checklist: [which hyperscaler CapEx lines to watch](https://www.sotkn.com/explained/hyperscaler-capex-lines-to-watch/). ## Common misconceptions - **People assume AI Capex is a separate GAAP line.** Actually it is CapEx aimed at AI workloads. Companies still report CapEx and PPE purchases; the AI label is in the commentary. - **People assume CapEx guidance is cash leaving this quarter.** Actually annual guides span the year. Quarterly PPE and cash CapEx prints show the pace. - **People assume higher CapEx means open GPU supply now.** Actually CapEx buys kit and halls. Qualified, energised capacity can still lag demand. - **People assume all four hyperscalers use the same CapEx definition.** Actually finance leases, useful lives and “cash CapEx” labels differ. Compare footnotes, not just headlines. ## FAQ **What is AI Capex?** AI capex is capital expenditure for the hardware and facilities that run AI — GPUs and other accelerators, CPUs, servers, networking gear, and data-centre land and buildings — booked before those assets earn cloud, ads or other revenue. Hyperscalers disclose it in CapEx guidance and cash-flow tables. **How is AI Capex different from ordinary CapEx?** The accounting line is still CapEx. The AI label marks spend aimed at training and serving models: short-lived servers and accelerators, plus halls and power that host them. Microsoft said in July 2026 that roughly two thirds of a $41 billion quarterly CapEx print went to short-lived assets, primarily CPUs and GPUs. **Who pays AI Capex — and who gets paid?** Hyperscalers and other large AI builders pay the CapEx. Chipmakers, OSATs, server and networking vendors, landlords, contractors and power suppliers get paid as tools, packages, racks and halls are bought or leased. Cloud and ads customers pay later, when hours or impressions bill. **Why does AI Capex hit free cash flow before revenue?** Cash leaves when property and equipment are purchased or when finance-lease principal is paid. Revenue arrives only after capacity is live and customers consume it. Amazon’s trailing-twelve-month free cash flow was a $7.6 billion outflow as of 30 June 2026, driven mainly by higher AI-related property and equipment purchases. **What CapEx numbers are hyperscalers guiding for 2026?** As of late July 2026 prints: Amazon about $220 billion cash CapEx; Alphabet $195–205 billion; Microsoft about $175 billion for calendar 2026 after a lease-classification shift (underlying investment plan unchanged); Meta $130–145 billion including finance-lease principal. Each company defines CapEx slightly differently — read the footnotes. **Is this investment advice or a buy/sell list?** No. Second Order Explained pages define how the money moves. They do not give buy or sell calls. ## Related terms - [Cash Capex](https://www.sotkn.com/explained/cash-capex-explained/) — when CapEx cash leaves versus when depreciation hits. - [Capex vs Opex](https://www.sotkn.com/explained/capex-vs-opex/) — buying GPUs versus renting hours. - [Remaining performance obligation](https://www.sotkn.com/explained/remaining-performance-obligation-explained/) — contracted revenue not yet recognised. - [Hyperscaler](https://www.sotkn.com/explained/hyperscaler-explained/) — who books this CapEx at scale. - [AI tokens](https://www.sotkn.com/explained/ai-tokens-explained/) — the unit hyperscalers often bill AI by. ## Stack doors Layer: [Hyperscalers](https://www.sotkn.com/stack/hyperscalers/). Also: [AI Data Centers](https://www.sotkn.com/stack/ai-data-centers/). Markets: [United States](https://www.sotkn.com/markets/us/). Related Explained: [CapEx lines to watch](https://www.sotkn.com/explained/hyperscaler-capex-lines-to-watch/). [Get the Wire](https://www.sotkn.com/#/portal/signup) — short daily notes when CapEx guides, RPO or free cash flow move. **How we check this** Figures come from Microsoft’s, Amazon’s, Alphabet’s and Meta’s own late-July 2026 earnings materials listed under References. Each CapEx guide and quarterly print was re-checked against the saved source on 14 Oct 2026\. Company CapEx definitions differ (cash CapEx, finance leases, useful-life changes); we keep their wording. Where we judge how the clocks fit together, the wording says “our read”. This page explains how the money moves. It does not give buy or sell calls. **Last reviewed:** 14 Oct 2026 · **Next review:** after Microsoft / Alphabet / Amazon / Meta next CapEx prints (Q3–Q4 2026 earnings) Spot an error? [Tell us](mailto:sean@thekopinotes.com) and we will correct it and note the change here. [Our editorial standards →](https://www.sotkn.com/editorial-standards/) ## References 1. [Microsoft — FY2026 Q4 earnings conference call (29 Jul 2026)](https://www.microsoft.com/en-us/investor/events/fy-2026/earnings-fy-2026-q4?ref=sotkn.com) — Q4 CapEx $41B (\~2/3 short-lived CPU/GPU); CY2026 CapEx \~$175B after useful-life/lease reclass; Q1 FY27 CapEx >$50B; commercial RPO $678B 2. [Amazon.com — Q2 2026 earnings release (30 Jul 2026)](https://s2.q4cdn.com/299287126/files/doc%5Fearnings/2026/q2/earnings-result/AMZN-Q2-2026-Earnings-Release.pdf?ref=sotkn.com) — TTM free cash flow −$7.6B; Q2 purchases of property and equipment $54.2B; AWS sales +37% to $42.2B 3. [Amazon — Q2 2026 earnings call transcript (Motley Fool, call 30 Jul 2026)](https://www.fool.com/earnings/call-transcripts/2026/08/07/amazon-amzn-q2-2026-earnings-call-transcript/?ref=sotkn.com) — Cash CapEx guide \~$220B for 2026 (from \~$200B; memory costs); Q2 cash CapEx $53.1B; capacity still short into 2027 4. [Alphabet — Q2 2026 earnings release (22 Jul 2026)](https://s206.q4cdn.com/479360582/files/doc%5Fnews/2026/Jul/22/attachments/2026q2-alphabet-earnings-release.pdf?ref=sotkn.com) — Q2 purchases of property and equipment $44.9B; Google Cloud revenue $24.8B (+82%) 5. [Alphabet — Q2 2026 earnings call transcript (22 Jul 2026)](https://s206.q4cdn.com/479360582/files/doc%5Fevents/2026/Jul/22/2026%5FQ2%5FEarnings%5FTranscript.pdf?ref=sotkn.com) — FY2026 CapEx guidance $195–205B (from $180–190B); \~60% technical infrastructure in servers, 40% data centres and networking; Cloud backlog cited 6. [Meta Platforms — Q2 2026 results (29 Jul 2026)](https://investor.atmeta.com/investor-news/press-release-details/2026/Meta-Reports-Second-Quarter-2026-Results/?ref=sotkn.com) — FY2026 CapEx incl. finance-lease principal $130–145B; Q2 CapEx $31.08B S **[Sean](https://www.sotkn.com/author/sean/)** Writes Second Order by TKN — a plain-English map of AI infrastructure and semiconductors for non-specialist investors. Focus: who gets paid, where money sticks, and which step is stuck. Covers AI infrastructure, semiconductors, and second-order stack reads. [Author page →](https://www.sotkn.com/author/sean/) · [Editorial standards →](https://www.sotkn.com/editorial-standards/)