The numbers behind the arms race
Three Q3 2026 earnings reports landed in the same week at the end of October, and together they describe a capital spending pace that has no real precedent in enterprise technology. Microsoft's purchases of property and equipment hit 39.4 billion dollars for the quarter, more than double the 19.4 billion dollars it spent a year earlier. Alphabet's capital expenditures reached 51.4 billion dollars, up from 24.0 billion dollars a year prior, a jump of well over 100 percent. Amazon's spend was the largest of the three by a wide margin at 77.3 billion dollars, more than double its year-ago figure of 35.1 billion dollars, and large enough on its own to rank among the biggest quarterly capital programs in the history of corporate America.
Add the three together and the total comes to roughly 167.8 billion dollars in a single quarter, aimed overwhelmingly at data centers, custom silicon and the power infrastructure needed to run both. Amazon's own disclosure made the driver explicit, attributing the increase primarily to investments in artificial intelligence and reaffirming guidance of approximately 220 billion dollars in cash capital expenditures for the full year 2026. That annual figure is larger than the total revenue of most Fortune 500 companies, and it represents a single line item inside one firm's budget for a single calendar year.
Growth that is outrunning the spending, for now
The capex is not speculative in the sense of chasing demand that has not materialized. Cloud revenue grew fast enough in the same quarter to make the spending look rational on paper, at least by the metrics each company chooses to disclose. Microsoft Cloud revenue rose 30 percent to 63.9 billion dollars, with Azure revenue specifically up 48 percent year over year. Google Cloud revenue climbed 103 percent to 30.7 billion dollars, with TPU systems sales cited by Alphabet as a meaningful contributor to that growth. AWS segment sales rose 39 percent to 46.0 billion dollars, a slower growth rate than its two largest rivals but on a larger absolute revenue base.
The backlog metric is the more telling number for anyone trying to gauge how long this pace of spending continues. Microsoft's commercial remaining performance obligation, the contracted revenue not yet recognized, grew 73 percent year over year to 677 billion dollars. That figure represents Microsoft's own customers, enterprise and otherwise, committing today to multi-year spend against capacity Microsoft has not yet fully built. It is a genuine demand signal, and it is also a structural reason the three hyperscalers have little incentive to compete that locked-in capacity away on price in the near term.
The cost inputs nobody priced in a year ago
Buried in the same earnings cycle were warnings that the cost of building this capacity is itself rising faster than most forecasts assumed twelve months ago. Qualcomm told investors in July that input costs are climbing across the semiconductor industry, covering wafer fabrication, advanced packaging and memory, and said it is taking active steps to pass those costs through in its own pricing to customers. Amazon separately named supply volatility, including for memory chips specifically, as a factor that could affect its financial results in coming quarters, a notable admission from a company guiding to 220 billion dollars in annual capex.
That matters because memory and advanced packaging are inputs to every server a hyperscaler racks, not just the GPU-heavy training clusters that generate most of the industry's headlines. When the chip suppliers selling components into AWS, Azure and Google Cloud are telling their own investors that costs are rising industry-wide, it is a reasonable expectation that at least some of that cost gets reflected in what enterprises eventually pay for compute, storage and managed services over the next several quarters of contract renewals.
Capacity, not generosity, is the strategic variable
It is worth being precise about what this capex race actually buys each company, because the framing shapes how enterprise buyers should respond to it. The spending is not primarily a bet on winning new logos through aggressive pricing. It is a bet on having enough physical capacity, power and silicon on hand to serve the backlog each company has already contracted, plus the next wave of AI workloads that executives expect to land over the following several quarters. Reading the capex figures as evidence of generous future pricing misunderstands the purpose of the spending entirely.
The capex figures are better read as evidence of how seriously each hyperscaler takes capacity risk relative to pricing risk right now. A vendor spending tens of billions per quarter to add capacity is signaling that running short of capacity, and disappointing a backlog worth hundreds of billions of dollars, ranks as the outcome it fears most. That changes how negotiation leverage works for any enterprise buyer sitting on a renewal this year, because the vendor's biggest fear centers on having too few racks to serve everyone who already signed, well ahead of any concern about losing a workload to a competitor on price.
What this means for your 2027 cloud contract
For a CIO or CFO negotiating cloud commitments heading into 2027, the practical read is that capacity, not price, is now the primary lever the big three are managing. A hyperscaler spending 50 to 80 billion dollars a quarter on infrastructure is optimizing for utilization of that infrastructure, and multi-year committed-use discounts increasingly look like the mechanism for securing guaranteed capacity rather than simply a price break on commodity compute. Enterprises that treat those discounts purely as cost savings are missing the more important benefit, which is priority access during periods when capacity is genuinely constrained.
The practical move is to treat capacity commitments and pricing commitments as two separate negotiating lines rather than one bundled discussion. Lock in volume and priority access where your roadmap genuinely needs GPU or TPU capacity over the next two to three years, but push back hard on being bundled into long-dated price commitments justified by the vendor's own capex story. Their capex problem is not automatically your pricing problem, and the backlog numbers above suggest they can afford to compete harder for your specific workload than their public capex narrative would otherwise imply.
