Big Tech AI spending after Q3 2026 earnings: what it means for businesses buying AI

In the last week of October 2026, the companies that rent most of the world's computing power reported their third-quarter results, and Big Tech AI spending reached a new high. Alphabet spent $51.4 billion on property and equipment in the quarter, against $24.0 billion a year earlier. Amazon's purchases of property and equipment came to $77.3 billion, up from $35.1 billion. Microsoft's chief financial officer, Amy Hood, told analysts that "customer demand continues to exceed available supply", according to the company's own call transcript.

Searches for "google stock", "msft stock" and "amazon stock" spiked on Google Trends as the share prices moved. This piece is not about the share prices. It sets out what each company spent and said, where the numbers need a careful reading, and what all of it means for a business that buys AI services and cloud capacity from these same companies.

Nothing here is investment advice. It is about building systems, not buying shares.

What each company reported

Alphabet and Microsoft reported on 28 October and Amazon on 29 October. Meta reported a week earlier, on 21 October, but belongs in the same comparison because it is spending on the same scale.

Sources: Alphabet's release and call, Microsoft's release and call, Amazon's release and call, and Meta's release.

A widely repeated figure says Amazon, Alphabet and Microsoft are together on course for about $590 billion of capital spending in 2026. The parts do add up: $220 billion plus $195–205 billion plus about $170 billion is $585–595 billion. Add Meta and the four together come to roughly $720–740 billion. The next section explains why that total should be read loosely.

Alphabet had the fastest cloud growth. Sundar Pichai said Google Cloud revenue grew 103%, "powered by huge demand for our AI infrastructure and AI solutions", and that the company is "seeing strong demand for our TPU systems", its own AI chips. Google Cloud's backlog, meaning contracted revenue not yet earned, was $561 billion, of which the company expects "just over 55%" to become revenue within 24 months. The finance chief, Anat Ashkenazi, said: "We continue to expect our CapEx to increase significantly in 2027 and will provide further details on our Q4 call."

Microsoft said Azure's growth accelerated to 48% and guided to 49–50% in the current quarter. Its calendar-2026 spending figure fell to about $170 billion from $175 billion, Reuters reported, but not because demand weakened. Hood said spending was lower "as some third-party capacity deliveries we expected in Q1 were delayed", and that "we still have demand that exceeds supply".

Amazon said AWS grew 39% to $46 billion, its fastest growth in nearly five years. Andy Jassy said the AWS backlog "is $578 billion, growing triple digits year over year", and that the company is "installing an AI accelerator about every 14 seconds". He said Amazon expects to "spend meaningfully more in 2027".

Meta sells no cloud services, so all of its spending serves its own products. Its chief executive, Mark Zuckerberg, framed the quarter as progress "toward delivering personal superintelligence for everyone".

Capacity is still the constraint

The clearest message across all four reports is that, for now, the limit on AI is not demand but supply.

Microsoft's new capacity is arriving late. Hood said "we had some supply we were counting on Q2 move out into H2", meaning into the first half of 2027. That is why its spending came in below forecasts, not above.

Google is renting other people's computers. Ashkenazi said that "given the supply-constrained environment, we started utilizing third-party compute capacity at the end of Q3", and warned of "margin headwinds in the near term as we utilize this capacity". A company building data centres as fast as Google is still having to rent capacity to keep up.

Amazon describes chip supply as a ceiling. Jassy noted that "there's only so many chips that provider can provide each year", and said some customers are asking for "hundreds of megawatts or sometimes up to a gigawatt of capacity".

Memory has become expensive. Hood said the spending on servers is "inclusive of higher memory and storage costs". Jassy referred to "heightened component costs like memory", and Amazon's finance chief, Brian Olsavsky, listed "headwinds from component pricing". Amazon's release names "resource and supply volatility, including for memory chips" among its risks. Part of the rise in spending buys the same hardware at a higher price, not more of it.

Where to be sceptical about AI capex in 2026

The numbers are real. Several common readings of them are not.

The totals mix different measures. Alphabet reports purchases of property and equipment. Amazon guides on cash capex, which is a little lower than its gross purchases. Microsoft counts the full value of a finance lease as capex when the lease starts, and in July it began spreading long-term data-centre leases over 25 years rather than 15, which lowered its reported capex while, Reuters reported, its spending plans stayed the same. Adding them into one $590 billion figure is fine as a sense of scale. It is not a precise number.

Depreciation arrives later, and it is growing. Spending goes on the balance sheet and is charged against profit over the life of the asset. Alphabet's depreciation rose to $8.4 billion in the quarter from $5.6 billion a year earlier, and Ashkenazi said the investment "will continue to weigh on the P&L in the form of higher depreciation expenses". The mix matters: about 65% of Alphabet's infrastructure spending went on servers and networking, with the rest on data centres that "have useful lives of seven to 40 years". At Microsoft, "roughly three quarters" went on "short-lived assets, primarily CPUs and GPUs". Short-lived assets are written off far sooner than buildings, so the cost of this year's spending will show up in profits for years.

Backlog is a promise, not cash. A $561 billion or $578 billion backlog is a strong signal, but it is contracted revenue spread over years, not money already earned. Google expects to recognise just over half of its backlog within two years.

Investors are already pricing in doubt. Alphabet's shares fell about 10% in extended trading after it reported negative free cash flow of $9.07 billion, while Microsoft's rose about 4% after spending less than analysts expected. Amazon's fell as much as 7% after the bell; one fund manager told Reuters that "operations generated plenty of cash but their capex just absolutely swallowed it". Meta's closed down 6.98% the day after its results. This quarter, the market rewarded restraint.

The executives have a stake in the answer. When Jassy says Amazon has "a clear line of sight to attractive financial returns", he is the person whose plan depends on it. That does not make him wrong. It means the returns are a forecast, not a result.

None of this settles whether AI as a whole is overvalued. That wider question is covered in our earlier piece on whether AI is a bubble and what it means for business, and China has already shown what a sharp correction in AI stocks looks like.

What Big Tech AI spending means for a business buying AI and cloud

If you buy AI models, GPU time or ordinary cloud services from Google, Microsoft or Amazon, these results tell you a few practical things.

Plan for capacity to be rationed. When the providers say demand exceeds supply, it shows up for customers as quotas, rate limits, regions where the newest hardware is unavailable, and waits for dedicated capacity. If an AI feature is important to your business, find out your actual limits before launch, ask for increases early, and design the system to slow down gracefully rather than fail when a limit is hit.

Do not assume infrastructure gets cheaper every year. Even if the price of calling a model keeps falling, the hardware underneath is getting more expensive, with memory named by both Microsoft and Amazon. Providers carrying rising depreciation and higher component costs have reasons to protect their margins. Budget with some headroom, and re-read the pricing pages of the services you depend on each quarter.

Commit carefully to multi-year contracts. Those backlogs are made of commitments: customers agreeing to spend a fixed amount over several years in return for a discount or guaranteed capacity. That can be a sensible trade if you are sure of your usage. Commit to what you would use anyway, keep the term no longer than you can forecast, and check what happens to unused commitment and whether it can move between services, such as from GPUs to storage.

Design for more than one provider. Microsoft's own chief executive, Satya Nadella, said "the number of customers building agentic solutions with multiple models increased 5X year over year", and asked aloud whether a customer can "put multiple of the frontier models, because these have very different prices". Keep the model behind a thin layer of your own code, store your prompts and evaluation tests outside any one vendor's console, and test a second provider before you need it. When capacity is short, the customer who can move a workload has more choice.

Be clear about what is locked in. Model APIs are relatively easy to swap. Proprietary agent platforms, vector databases, fine-tuned models and data stored in one cloud are harder. Lock-in is not always bad, but it should be a decision made knowingly, not a side effect.

Use spare capacity where the work allows it. Not every AI job is urgent. Overnight document processing, back-office classification and test runs can often wait. Where your provider offers interruptible capacity or batch processing at a lower rate, putting deferrable work there frees budget and quota for the work customers actually see.

The short version

Alphabet, Microsoft, Amazon and Meta are each on course to spend more on AI infrastructure in 2026 than ever before, and each says it will spend more again in 2027. For investors, that raises hard questions about cash flow and returns. For a business buying AI, the useful conclusion is simpler: capacity is tight, the cost of the hardware is rising, and the providers are asking customers to commit for longer. Build systems that can work within limits, keep a second option open, and sign only the commitments you would be comfortable holding if the market changed.