The first question on a hyperscaler earnings call used to be about revenue. This quarter it is about the number that shows up two lines below: capital expenditure. Microsoft, Alphabet, Amazon, and Meta reported results in late April and early May, and on each call executives spent more time defending data centre budgets than selling the cloud growth that justifies them.
In the March 2025 quarter, those four companies spent roughly $70 billion on property, plant, and equipment. The latest filings show a further step up. The investor questions have moved from whether they can afford it to when the money comes back.
A finance director at a 500-person logistics software firm — call him Daniel — sent me his April cloud statement. Two pilot AI features and one fine-tuned model cost $48,000 in a month, on top of a $210,000 annual data platform contract. He spent four days trying to tie that spend to a single new customer. He couldn't.
The base rate: capital spending grows faster than cloud revenue
Start with the numbers from the last quarter before the latest earnings cycle. Microsoft spent $22.6 billion on capital expenditure including finance leases in the March 2025 quarter, up 61% from a year earlier. Alphabet spent $17.2 billion, up 43%. Amazon reported $24.1 billion, a jump of more than 60%. Meta, the smallest spender among the four, reported $6.1 billion, roughly flat while it waited on new data centre deliveries.
Cloud revenue did not move at that pace. Azure grew 33% in the same period. Google Cloud rose 28%. Amazon Web Services grew 17%. Meta's total revenue increased 22%. The gap between those two lines is the scrutiny.
| Company | March 2025 quarter capex | Cloud or total revenue growth |
|---|---|---|
| Microsoft | $22.6 billion | Azure up 33% |
| Alphabet | $17.2 billion | Google Cloud up 28% |
| Amazon | $24.1 billion | AWS up 17% |
| Meta | $6.1 billion | Total revenue up 22% |
The base rate matters because analysts are now comparing capital expenditure against incremental cloud revenue, not against last year's capital expenditure. A dollar of capex that adds 20 cents of cloud revenue looks different from a dollar that adds 50 cents. Capital spending outpaced revenue growth by a wide margin. That arithmetic is tolerable only if the assets keep generating for years. Every new earnings call shortens the market's willingness to wait.
Photo by Kanchanara on Unsplash
DeepSeek changed the argument, not the spending
On January 27, 2025, a single research release from the Chinese lab DeepSeek wiped $589 billion from Nvidia's market value in one session, the largest single-day value loss in the company's history. The model, DeepSeek-R1, claimed training costs far below Western frontier models. The selloff was less about one lab's efficiency and more about the fragility of a supply chain built on advanced chips, a pattern covered in the US-China AI chip export restrictions story.
The market briefly read DeepSeek as a reason to stop building. The hyperscalers read it as a reason to build differently but keep building. Within a week, Meta lifted its 2025 capital expenditure range from $60 billion to $65 billion and then to $72 billion. Microsoft said its next fiscal year would bring another increase. Alphabet held to roughly $75 billion for the year. Amazon guided toward $100 billion.
Investors now press on depreciation schedules, power availability, the percentage of AI compute tied to paying production workloads, and how much capacity remains unrented. Those are the right questions. The earnings calls have produced fewer direct answers than the market would like.
Where the money goes
Inside the companies, the capital spends on accelerators, data centre shells, power equipment, and the physical security that wraps them. Microsoft's March-quarter filing showed the heaviest share going to servers and chips, with data centre construction next. Alphabet's quarterly release pointed to a similar split, with most of its $17.2 billion going to servers followed by data centres.
The money is not abstract. Amazon announced an $11 billion data centre expansion in Georgia in January 2025. Microsoft has committed to a $3.3 billion Wisconsin site. Alphabet and Meta have each signed long-term power purchase agreements with utilities to keep data centres running. Local grid operators, not tech shareholders, often see the bills first.
This is why the revenue question matters. A data centre shell can take 18 to 30 months from groundbreaking to first server. An accelerated chip order can take two to three quarters to arrive. The spending decisions made in 2025 will not show up as cloud revenue until 2027 or later. The market's impatience is grounded in that lag.
Photo by Brian Suman on Unsplash
What this means for the budgets you manage
Daniel's April bill is a small version of the same problem. The hyperscalers can absorb a bad quarter; his firm cannot. The useful question for any organisation buying cloud AI capacity is not whether the technology works. It is whether the spend can be connected to a change in customer behaviour.
Four practices tend to separate teams that control AI costs from teams that don't.
- Track unit economics per successful AI call, not per month of model access. A fine-tuned model that costs $0.40 per useful response may still be a bargain; a $12,000 monthly platform fee with no production traffic is not.
- Separate research spend from production spend in the general ledger. If those two categories sit in one line item, nobody can tell which experiments are working.
- Renegotiate committed-use discounts every quarter, not annually. The list price for one generation of accelerators is not the price you should be paying two quarters later.
- Set an automatic kill switch for any model variant that has not served a paying customer in 60 days. The switch should cut deployment, not just alert someone.
None of this will stop the hyperscalers from spending. It will stop their spending from becoming your consolidated bill.
The question to ask before the next earnings round
Before the next set of Big Tech earnings reports, ask your finance team one question: what is our unit cost per successful AI interaction, and what has it done over the last three quarters? If the answer is a shrug, the capital expenditure scrutiny on Wall Street has already arrived in your own budget. You just haven't noticed it yet.