Power capacity, purchased hardware and usable cloud service are different milestones. A single gigawatt figure cannot establish all three.
Separate the reporting from the measurements
The Guardian reported 2.2 million installed AI chips; Microsoft disputed its calculations.
We have not inspected the underlying internal inventory. The useful next step is to examine what Microsoft's own public disclosures actually measure, rather than treating the disputed estimate as a verified fleet count.
In its July 29 earnings call, Microsoft reported $41 billion in quarterly capex, about two thirds for short-lived assets, primarily CPUs and GPUs. It separately reported $35.8 billion cash paid for property, plant and equipment and $5.6 billion finance leases. These categories are not a purchase receipt for GPUs alone; the two-thirds description also contradicts a simple claim that most spending went to buildings.
Identify the boundary before converting watts
Suppose a fictional campus has 100 MW of total power capacity. Assume 80 MW serves IT equipment, and half of that IT capacity is allocated to accelerator servers. The remaining 40 MW could support at most 4,000 servers at an assumed 10 kW per complete server. With eight accelerators per server, that is 32,000 accelerators under those assumptions.
Dividing the full 100 MW by an individual accelerator's wattage would count cooling, other equipment and non-accelerator workloads as though they were all chips. Even the more careful example remains a capacity model: it does not prove the servers have been purchased, installed, energized or made available to customers.
The cover makes each assumption explicit. None of its numbers describes Microsoft's real fleet, and maximum equipment ratings are not the same as measured average power. An annual energy total in MWh is another quantity again; converting it into installed capacity requires utilization and time assumptions.
Ask for a matching date and operational state
A useful capacity record identifies the site, whether power is contracted or available, the portion assigned to IT, installed hardware and the date customer workloads can run. Region, accelerator type, networking and available quota matter to someone deploying an application.
Keep historical comparisons on the same basis. A previous target is not a measured starting inventory, so subtracting a target from a later reported count does not establish net additions. Similarly, construction spending can precede service delivery. These distinctions help evaluate a cloud expansion without pretending that public totals settle an unresolved inventory dispute.
Keep disputed chip estimate briefly attributed; use Microsoft primary capex disclosure and original explicit power-boundary example, remove unsupported installed-chip subtraction and GPU-only capex conclusions.