Hyperscalers' GPU Ramp

Microsoft has switched on its Fairwater artificial intelligence datacenter in Wisconsin ahead of schedule, adding a new block of cloud computing built around Nvidia’s Blackwell chips. (blogs.microsoft.com) Microsoft said in September 2025 that the Mount Pleasant site would come online in early 2026 as part of an initial $3.3 billion Wisconsin investment, with another $4 billion committed for a second datacenter over the next three years. (blogs.microsoft.com) Satya Nadella said on April 16 that Fairwater was “going live, ahead of schedule” and would bring together “hundreds of thousands of GB200s into a single seamless cluster,” a scale Microsoft has also described in its own architecture notes. (wccftech.com, blogs.microsoft.com) A graphics processing unit, or GPU, is the chip that does the heavy lifting for training and running large artificial intelligence models; a “cluster” is a giant pool of those chips wired together so they behave like one machine. Microsoft said each Fairwater site uses a flat network that can integrate hundreds of thousands of Nvidia GB200 and GB300 GPUs. (blogs.microsoft.com) That helps explain the spending. Microsoft told investors in its fiscal 2025 fourth-quarter call that capital expenditures for the next quarter would top $30 billion, driven by demand for artificial intelligence infrastructure, and the company separately said it expected to spend $80 billion on AI-capable datacenters in fiscal 2025. (microsoft.com, cnbc.com) Fairwater is not a standard cloud site filled with smaller, mixed workloads. Microsoft said the design pushes about 140 kilowatts per rack and 1,360 kilowatts per row, using closed-loop liquid cooling so the system can run denser rows of AI servers without constant water replacement. (blogs.microsoft.com) The pressure point for cloud vendors is no longer just buying chips. Google said its cloud business ended 2025 at an annual revenue run rate of more than $70 billion, driven by demand for AI products, while analysts at SiliconANGLE said enterprise spending momentum is strongest in Google’s machine learning and AI segment ahead of Google Cloud Next on April 22-24 in Las Vegas. (abc.xyz, siliconangle.com, cloud.google.com) Alphabet also raised its 2025 capital expenditure plan to $85 billion as it worked through cloud capacity backlogs tied to artificial intelligence demand. That puts Microsoft, Google and other hyperscalers in the same position: turn very expensive infrastructure into higher-margin software, cloud services and AI subscriptions fast enough to justify the buildout. (cnbc.com) Microsoft has framed Fairwater as part of a wider “planet-scale AI superfactory” linking Wisconsin, Atlanta and its broader Azure network. The near-term test is not whether hyperscalers can order enough GPUs, but whether customers buy enough AI products to keep those clusters full. (blogs.microsoft.com)