Alphabet, Meta, Microsoft and Amazon more than doubled their combined cash spending on property and equipment in Q3 2026, to USD 201.9bn from USD 97.3bn a year earlier. Their combined operating cash flow grew 10% to USD 175.5bn over the same period. A USD 61.7bn surplus after those purchases in Q3 2025 became a USD 26.4bn shortfall. Investment at the four companies is now growing about ten times faster than the cash their operations generate, which makes funding the gap the central question for the next round of AI spending.
Cloud revenue gives the companies a reason to keep spending. Google Cloud revenue rose 103% YoY to USD 30.7bn in Q3, with its operating margin reaching 36.2%. Microsoft’s Azure revenue grew 48%, while Amazon’s AWS grew 39.3% to USD 46.0bn. Oracle’s cloud infrastructure revenue rose 121% in its August quarter, and CoreWeave more than doubled June-quarter revenue to USD 2.6bn. Demand for computing already generates sizeable sales. How much additional capacity these customers will use, and at what price, will decide the return on the new investment.
Microsoft sells computing alongside enterprise software, and Amazon combines AWS with its retail and advertising businesses. Alphabet can earn from cloud services while using Gemini and its own AI processors to support Search. Meta funds its Muse models and agents primarily from advertising, where AI can improve recommendations and ad performance. At these platforms, AI can lift revenue in the existing businesses without appearing as a separate line in the accounts.
Alphabet’s group revenue grew 25% YoY and Meta’s 26%, supported by more ad impressions and higher prices. The additional cost shows up in the margins. Meta’s operating margin fell from 40.1% to 20.1%. Excluding only the USD 9.89bn litigation charge raises it to 35.4%, still below last year. Costs grew faster than revenue even after that adjustment, although the results do not isolate how much of either was attributable to AI. For Alphabet, preserving advertising economics as queries shift towards AI answers is another part of the investment case.
Revenue and operating margin, Alphabet, Meta, Microsoft and Amazon (Q3 2025 vs Q3 2026, USD bn, %)
Source: Company data, InterCapital Research
Operating cash flow less gross cash purchases of property and equipment left a Q3 surplus of USD 18.6bn at Microsoft, against shortfalls of USD 9.1bn at Alphabet, USD 1.2bn at Meta and USD 34.8bn at Amazon. The comparison excludes finance-lease principal and asset-sale proceeds. The figures cover each group as a whole, including non-AI activities, and do not measure AI profitability or the total financing requirement.
Operating cash flow and cash purchases of property and equipment (Q3 2026, USD bn)
Source: Company data, InterCapital Research
A shortfall during construction can be consistent with a profitable project, since equipment has to be installed before it earns revenue. Cash receipts can also arrive early. Oracle’s USD 23.1bn of August-quarter operating cash flow included USD 11.4bn of customer prepayments with a significant financing component. Those customers help finance capacity before the service is delivered. Depreciation, meanwhile, spreads equipment costs over future reporting periods, so current margins do not show whether the investment will earn enough over the equipment’s useful life.
Nvidia’s July-quarter revenue rose 106% YoY to USD 96.2bn, at a GAAP operating margin of 66.2%. AMD competes in processors and accelerators, while Broadcom’s custom chips and networking generated USD 16.7bn of AI semiconductor revenue in its August quarter. Their customers have an incentive to reduce computing costs, including through custom silicon. A growing market can therefore raise suppliers’ sales and put pressure on their margins at the same time.
TSMC, which manufactures chips, and memory producers Micron, SK hynix and Samsung must also finance production capacity. A chip designer’s strong orders can support a factory investment, but the factory still needs customers when it opens. A cloud provider’s purchase is revenue for a supplier, and a model developer’s computing bill is revenue for the cloud provider. The same investment can therefore support revenue at several stages without representing several independent sources of final demand. Households, advertisers and businesses ultimately have to keep paying for the service for the chain to earn a return.
Apple can put AI into devices and services, using its own silicon and models developed with Google. Palantir connects models with enterprise data and workflows, and SpaceX’s xAI operation combines Grok with access through X. These positions offer ways to earn from adoption without winning every model benchmark. The company map groups each participant by its principal commercial route and by how many layers of the AI chain it covers.
Selected AI companies by principal commercial route and breadth across the AI chain (November 2026)
Source: Company data, BIS, OECD, InterCapital Research
OpenAI and Anthropic have to fund expansion without the same history of diversified operations. OpenAI announced USD 122bn of committed capital in March at a USD 852bn post-money valuation, and is reported to be seeking around USD 20bn at a USD 1.25tn valuation, a round the company has not confirmed. Anthropic publicly filed for an IPO on 26 October, with the price and offering size still to be determined. ChatGPT, Codex and Claude already generate commercial revenue, and the funding needs of both companies depend on how quickly customer receipts cover more of their computing and research costs. Amazon and Nvidia participated in OpenAI’s March round, linking the capital supplied by investors with companies that also sell into the AI expansion.
Chinese groups compete across the same chain. Alibaba combines Qwen, cloud services and custom chips, and Huawei combines hardware, cloud and Pangu models. DeepSeek offers models and developer access, while ByteDance, Tencent and Baidu bring established distribution. If cheaper models attract more users, application providers can benefit even as prices come under pressure. For infrastructure suppliers, the outcome depends on whether the extra usage outweighs the reduction in spending per task.
The US accounts for roughly 72% of estimated cumulative AI-compute arrivals from 2022 through Q2 2026, mainland China for 10% and the EU for 5%. These estimates include Chinese domestic production and express hardware in equivalent computing performance. They describe arrivals, not live installed capacity or model quality, and they do not measure where profits accrue. Taiwan and South Korea supply much of the hardware, while Europe’s ASML supplies chipmaking equipment and Mistral develops models.
Estimated cumulative AI compute arrivals by region (Q1 2022 – Q2 2026, m H100-equivalents, left; share of total, %, right)
Source: Epoch AI , InterCapital Research.
We would watch whether cash coverage improves over several quarters as the new capacity enters service, whether advertising margins withstand the additional costs, and whether the laboratories fund more computing from customer receipts. Slower investment alone would be ambiguous, as it could reflect better efficiency or weaker demand. The more useful combination would be sustained sales growth and stronger cash generation relative to investment. The industry already produces substantial revenue, and the next few quarters will show how much of the capital now being committed that revenue can support.