What Alphabet’s AI spending can tell us about returns
Google Cloud revenue rose sharply as Alphabet more than doubled company-wide spending on property and equipment in the first nine months of 2026.

Alphabet spent US$131.953 billion on property and equipment in the first nine months of 2026, more than double the year-earlier figure, according to its third-quarter results. The spending and cloud growth show the scale of the investment, while the company’s filings leave the return on individual AI projects unclear.
Alphabet’s spending jumped on a consistent measure
Alphabet reported US$131.953 billion of cash purchases of property and equipment during the first nine months of 2026, compared with US$63.596 billion in the equivalent 2025 period. The cash-flow statement presents these purchases as cash outflows. The chart displays their absolute amounts, in billions of U.S. dollars, over matching nine-month periods.
Alphabet describes investment in technical infrastructure such as servers, network equipment and data centres. The spending line covers the whole company. The filing does not disclose a stand-alone return on AI investment.
Alphabet cash purchases of property and equipment, first nine months
Alphabet’s company-wide cash purchases of property and equipment more than doubled versus the same period in 2025.
Alphabet also reported that Google Cloud revenue reached US$30.7 billion in the third quarter of 2026, up 103% from the year before. The company attributed growth to enterprise AI infrastructure and solutions alongside core cloud services. Cash spending and revenue tell different parts of the story.
The capital number is for the whole company and covers nine months. The revenue figure is for one segment and one quarter. Investors who want to compare the economics should use consistent periods and account for expenses that take effect as assets enter service.
What counts as a return
For the cloud provider, the return depends in part on how long equipment earns revenue, how often customers use it and how much it costs to keep it running. A provider may need to reserve capacity before a contract starts. Underused capacity still ties up capital. Conversely, if an operator can use the same equipment for several customers or shift between workloads, it has another way to make use of the investment. A quarterly filing can give clues through segment margins, depreciation and commitments. An internal investment model would also track utilisation and revenue over the asset’s useful life.
Alphabet’s September 30 filing reported US$560.8 billion of cloud-related remaining performance obligations. This disclosure gives investors a view of contracted business the company expects to recognise as revenue over time, subject to how the contracts and reporting definitions work. It helps explain why the provider builds ahead of some delivery.
Revenue recognition, the cost of fulfilling those contracts and the resulting profit will unfold on their own schedules. One audit of the growth story is to track where reported revenue and operating income move together over successive quarters. Another is to see how fast spending and depreciation grow alongside them.
An AI agent needs a measurable job
For an enterprise buyer, the benchmark can be much smaller. Imagine a service team that receives requests, classifies them, looks up records and drafts a response. An agent might handle some steps across systems under assigned permissions.
Before the pilot, record the time it takes for a request to reach an accepted resolution, the error rate and the share of cases that need a person to intervene. Run the same measures during the pilot. Count supervisor review, fixes and the time required to maintain the workflow. The comparison can reveal whether faster first responses also improve the end result.
A technology manager in Canada can use the same test as a U.S. counterpart while still checking rules and vendor terms that apply to the Canadian operation. Statistics Canada found that 25.2% of Canadian businesses planned to use AI to produce goods or deliver services over the coming 12 months when surveyed in the third quarter of 2026. That gives a snapshot of intended adoption. For a particular workplace, the decision depends on the kind of records the system sees, who can approve decisions and whether staff can stop or correct a mistaken action before it reaches a customer.
The arguments for and against deployment can both be tested. One manager may expect staff to resolve more routine cases and spend longer on unusual ones. Another may expect the reviewers to spend the saved time checking plausible but incorrect answers.
An evaluation with real cases and a clear definition of a completed task can account for both outcomes. For work involving customer records or binding decisions, the organisation can begin with limited permissions and keep high-impact steps under review. Calculate cost per accepted case so that faster output with more rework becomes visible in the same metric.
Custom chips change the comparison
Cloud providers can buy accelerators or design hardware around workloads they expect to run at scale. Google’s documentation describes Ironwood as part of its TPU lineup and lists the supported machine series and accelerator characteristics. Nvidia reports a business premised on selling accelerated computing systems and describes export restrictions in its filings as a risk to sales and services. A buyer comparing hardware options has several costs to examine: how much useful work a region can deliver, which software needs adapting and how much capacity is available at the desired time. A peak specifications table is a starting point for that test.
A better comparison measures the same workload and output quality. For training, run the job to a defined target and record the time and total cost, including engineering time or failed runs. For inference, compare the price to serve requests within the response-time and quality requirements, including quiet periods when reserved capacity may be idle. Those measurements can lead to different hardware choices for different workloads. A lower cost per chip-hour can coexist with a higher cost to deliver an accepted result if the job runs longer or requires more integration work.
Rights, operating costs and the next budget cycle
The test should also reflect how procurement actually works. A company may pay a platform fee, an amount for usage and a separate integration bill. Put those costs next to the cost of the work the system is meant to improve. If the tool saves time in one department and sends more work to a second, the extra time belongs in the comparison. A promising pilot can warrant a limited expansion while management waits for more evidence.
At that stage, agree on when to review the results, which tasks will still need approval and what cost would make the project uneconomic. That converts a broad claim about productivity into a decision the finance team can revisit.
Copyright and licensing questions extend the cost comparison. Canadian Heritage reported that participants at its summit identified copyright, licensing and protected works in AI training as significant issues for the cultural sector. The U.S. Copyright Office published a pre-publication report on generative AI training as part of its AI work. Those sources capture policy discussion and analysis in two jurisdictions. A Canadian purchaser dealing with an American platform should ask which party provides assurances about training data, what rights it receives to use outputs and which terms govern the use of customer material. Legal advice may be necessary for a specific deployment.
The coming budget cycle offers two distinct tests. Follow Alphabet’s spending, operating income and disclosures about cloud commitments to judge how the infrastructure business develops. For a specific company buying an AI service, follow the time and money required for accepted work, with oversight and integration in the calculation.
Investors and enterprise buyers are both asking whether today’s costs lead to durable benefits. Their evidence will arrive at different intervals: in company filings for the cloud provider and in measured work outcomes for the business that signs the contract.
Key takeaways
- The chart compares cash purchases of property and equipment over equal nine-month periods.
- Alphabet’s spending line covers the entire company.
- Cloud revenue and customer obligations are different measures with different time horizons.
- Enterprise buyers can assess agents by measuring cost and quality per completed task.
- Same-workload testing makes comparisons of chips and services more informative.
- Training and output rights can affect the cost and risk of adopting a tool.