Amazon redesigns campuses for Rubin GPUs

Amazon Web Services is redesigning parts of its data-center footprint to handle Nvidia’s next Rubin generation of AI systems, according to a Business Insider report published on May 13. The report said an internal AWS effort known as “Titus” is aimed at higher site power capacity, broader liquid-cooling support and faster construction schedules as Amazon prepares for denser AI hardware. AWS had already said in June 2025 that the company and the wider industry were shifting from air-based cooling toward liquid-based systems for high-density AI servers. Microsoft has been making parallel moves on the power-and-water side of the buildout. Mount Pleasant, Wisconsin, unanimously approved site plans for 15 additional Microsoft data centers on January 26, 2026, while Microsoft said on January 13 that future AI infrastructure would be built under a “Community-First” model that includes paying full power costs and addressing water use in host communities. (businessinsider.com) CoreWeave has emerged as one of the named suppliers tying those pieces together. CoreWeave said on April 9 that Meta expanded its AI infrastructure agreement to about $21 billion through December 2032, with some initial deployments using the Nvidia Rubin platform, and Nvidia said in January that CoreWeave would begin integrating Rubin-based systems in the second half of 2026. Anthropic then signed a separate multiyear cloud agreement with CoreWeave on April 10, according to CNBC. (cnbc.com) ### Why are Amazon and Microsoft changing physical campuses now? Business Insider reported that Amazon’s “Titus” design raises total compute capacity per site to about 68 megawatts from about 58 megawatts and is built for wider liquid-cooling deployment. The report said the redesign is meant to support next-generation Nvidia systems whose heat and power requirements exceed what older air-cooled layouts were built to handle. (coreweave.com) AWS had already described cooling as one of the three core data-center constraints, alongside networking and power. In its June 2025 account of a facility in eastern Oregon, Amazon said the industry was moving to liquid-based cooling because AI servers can overheat within minutes without more aggressive thermal management. ### What is special about Rubin that forces these upgrades? (africa.businessinsider.com) Nvidia said in January that the Rubin platform combines six chips, including the Vera CPU and Rubin GPU, into a new AI supercomputer design. The company said CoreWeave would be among the first cloud providers to offer Rubin and that deployments would begin in the second half of 2026. (aboutamazon.com) Datacenter Dynamics reported from Nvidia’s January launch that Chief Executive Jensen Huang said Rubin would deliver twice the power of Blackwell while being cooled with hot water rather than traditional chillers. That combination — more performance and a different cooling baseline — is why operators are reworking power distribution, piping and building layouts before the systems arrive. (nvidianews.nvidia.com) ### Where do electricity and water constraints show up most clearly? Microsoft’s January 13 policy statement addressed the issue directly. President Brad Smith said the company would pay the full cost of electricity infrastructure needed for its data centers and would work to replenish more water than it uses, framing the commitments as necessary for communities hosting AI infrastructure. (datacenterdynamics.com) Mount Pleasant’s January 26 vote showed the scale of the expansion now moving through local approvals. CNBC reported that the Wisconsin village board approved 15 more Microsoft data centers near an existing facility at the former Foxconn site. ### Why does CoreWeave keep appearing in these reports? CoreWeave’s April 9 agreement with Meta runs through December 2032 and is valued at about $21 billion, the company said. (blogs.microsoft.com) CoreWeave said the capacity will be deployed across multiple locations and include some of the initial deployments of the Nvidia Rubin platform. (cnbc.com) Anthropic added a second large customer announcement a day later. CNBC reported on April 10 that CoreWeave shares rose after the company disclosed a multiyear agreement to provide Nvidia GPU-based cloud resources for Anthropic’s Claude models. ### What should readers watch next? The second half of 2026 is the key date in the current buildout calendar. (coreweave.com) Nvidia said Rubin-based systems will begin shipping into cloud deployments in that period, and CoreWeave and Meta said some of their first capacity under the expanded agreement will use Rubin across multiple sites. (nvidianews.nvidia.com) (cnbc.com)