AI infrastructure buildout explained: capex, chips, power and financing
How hyperscaler capex, Nvidia GPUs, power limits and new financing structures fit together in the AI buildout, with dated 2026 figures.
Key takeaways
- Microsoft, Alphabet, Amazon and Meta guided to roughly $720 billion to $745 billion of combined 2026 capital spending in their July 2026 earnings reports, most of it for AI data centers and chips.
- Nvidia remains the main supplier: its data center segment brought in $89.0 billion in the quarter ended July 26, 2026, up 117% from a year earlier.
- Power and grid connections, not just chips, are now a binding constraint, which is why bitcoin miners with energized sites have signed multibillion dollar AI hosting contracts.
- Funding is shifting from internal cash flow toward bonds, private credit and vendor financing, which is the center of the current bubble debate.
What the AI infrastructure buildout is
The AI infrastructure buildout is the wave of spending on the physical inputs that train and run AI models: specialized chips (mostly GPUs, or graphics processing units), servers, networking gear, the data center buildings that house them, and the electricity and cooling they consume.
For investors, the key metric is capex (capital expenditure), the money a company spends on long-lived assets. Capex is recorded on the balance sheet and then expensed gradually through depreciation. That timing gap matters: today’s spending shows up in earnings over years, and the assumed useful life of the equipment shapes reported profits.
Hyperscaler capex: the dated numbers
“Hyperscalers” are the largest cloud and platform operators. Their own guidance, as of their most recent earnings reports, is summarized below.
| Company | Latest capex guidance | Given on |
|---|---|---|
| Microsoft | About $175 billion for calendar 2026 (cut from about $190 billion for an accounting change, not a spending cut) | July 29, 2026 |
| Alphabet | $195 billion to $205 billion for 2026 (raised from $180 billion to $190 billion) | July 23, 2026 |
| Amazon | About $220 billion for 2026 (raised from about $200 billion) | July 30, 2026 |
| Meta | $130 billion to $145 billion for 2026 (narrowed from $125 billion to $145 billion) | July 29, 2026 |
| Oracle | Up to $95 billion for fiscal 2027 (ending May 2027), of which $20 billion to $25 billion is expected to be reimbursed by customers | June 11, 2026; reaffirmed capital plans September 10, 2026 |
Two details stand out. Microsoft said it extended the assumed useful life of data centers from 15 to 25 years from fiscal 2027, which shifts some leases out of reported capex. Amazon attributed its increase partly to higher memory prices, and CEO Andy Jassy said the company still would not have enough capacity to meet 2026 demand.
GPUs and Nvidia’s role
Training and running large models requires chips that perform many calculations in parallel. Nvidia dominates this market with its GPUs and the software and networking around them. Its fiscal second quarter 2027 results, released August 26, 2026, showed revenue of $96.2 billion, up 106% year on year, with $89.0 billion from data centers. It guided to about $108 billion for the following quarter.
In practice, a large share of hyperscaler capex becomes Nvidia revenue. Alphabet said about 60% of its second quarter capex went to servers. Custom chips from the hyperscalers themselves and competing suppliers are growing, but Nvidia remains the reference point for the cycle.
Data centers and the power constraint
Chips are only useful once they are housed, cooled and powered. The International Energy Agency (IEA) reported that data center electricity demand rose 17% in 2025 and expects it to roughly double by 2030. Connecting new load to the grid is slow: grid connection waits in parts of Europe are commonly cited at 7 to 10 years, and a Lawrence Berkeley National Laboratory study found only 13% of capacity that requested US interconnection from 2000 to 2019 was operating by the end of 2024.
As a result, developers increasingly build their own generation on site (“behind the meter”), sign long-term power deals, or buy sites that already have grid capacity. Oracle said it delivered 850 megawatts of data center capacity in the quarter ended August 2026.
How the buildout is financed
- Operating cash flow. Historically the largest source. It is being stretched: Meta reported free cash flow (cash from operations minus capex) of $784 million for the second quarter of 2026.
- Corporate bonds. On August 10, 2026, JPMorgan raised its forecast for 2026 technology, media and telecom bond issuance to $540 billion, including $317 billion from hyperscalers. Oracle has said it plans to raise close to $40 billion in debt and equity in fiscal 2027.
- Private credit and joint ventures. Meta’s roughly $27 billion financing with Blue Owl for its Hyperion campus in Louisiana moved a data center into a separate vehicle funded largely by private lenders, keeping much of the debt off Meta’s balance sheet.
- Circular or vendor financing. This describes deals where a supplier invests in or finances a customer that then buys the supplier’s products. Bloomberg reported on July 27, 2026 that Nvidia had announced deals including up to $250 billion of financing to help OpenAI lease computing capacity, reviving these concerns.
Why bitcoin miners are pivoting to AI hosting
Bitcoin miners already own the scarcest input: large, energized sites with grid connections, land and cooling experience. Mining income is volatile and was cut by the April 2024 halving, which reduced the block reward to 3.125 bitcoin. Long-term AI hosting contracts with creditworthy tenants offer steadier, contracted revenue.
The best-known example is IREN, which on November 3, 2025 announced a five-year contract worth about $9.7 billion to provide Microsoft with Nvidia GB300 GPU capacity at its Childress, Texas campus. Other listed miners, including Cipher Mining, TeraWulf, Hut 8 and Core Scientific, have announced similar hosting or lease agreements. The trade-off is that converting a mining site into an AI-grade data center requires heavy new capex and, often, new debt.
The bubble debate
Reasonable observers disagree on whether spending is running ahead of returns.
- The case that demand is real. Cloud providers report capacity shortages, Amazon Web Services grew 37% in the second quarter of 2026, and Oracle reported $664 billion of remaining performance obligations (contracted future revenue) as of September 10, 2026.
- The case for caution. Free cash flow is shrinking at some spenders, more spending is funded with debt, JPMorgan noted spreads widening by about 15 basis points on some hyperscaler debt, and circular deals can make end demand harder to measure.
- Accounting and obsolescence. Useful-life assumptions affect reported profits, and GPUs may lose economic value faster than buildings if newer chip generations arrive quickly.
What to watch
- Late October 2026 earnings, including any first look at 2027 capex plans.
- Free cash flow and net debt trends at the largest spenders.
- Bond spreads and the size of private credit and off-balance-sheet data center deals.
- Nvidia’s guidance and any further customer financing commitments.
- Grid interconnection queues, power prices and on-site generation announcements.
- Whether AI revenue at cloud providers keeps pace with depreciation as new assets come online.
Sources and further reading
- Nvidia second quarter fiscal 2027 results (SEC filing) (released August 26, 2026)
- Meta second quarter 2026 results (July 29, 2026)
- CFO Dive on Microsoft’s capex outlook and useful-life change
- Fortune on Amazon’s raised 2026 capex forecast (July 30, 2026)
- JPMorgan tech bond issuance forecast (Bloomberg via The Star) (August 10, 2026)
- IREN announcement of its Microsoft AI cloud contract (November 3, 2025)
Frequently asked questions
How much are the big tech companies spending on AI infrastructure in 2026?
Based on guidance given in July 2026, Microsoft, Alphabet, Amazon and Meta together expect roughly $720 billion to $745 billion of 2026 capex. Oracle separately guided to up to $95 billion for its fiscal year ending May 2027.
Why does Nvidia matter so much to this cycle?
Nvidia supplies most of the GPUs used to train and run AI models, so a large share of hyperscaler capex becomes Nvidia revenue. Its data center segment generated $89.0 billion in the quarter ended July 26, 2026.
What is circular financing in AI?
It refers to deals where a supplier invests in or finances a customer that then buys the supplier's products. Critics say this can overstate true end demand; supporters say the investments earn their own returns.
Why are bitcoin miners moving into AI hosting?
Miners control energized sites with grid connections, which are scarce. Long-term hosting contracts, such as IREN's roughly $9.7 billion five-year deal with Microsoft announced in November 2025, offer steadier revenue than mining.
This explainer is reviewed and updated as the rules and the market change. Last reviewed September 10, 2026. It is educational content and not financial, legal or tax advice.