Amazon's Titus still leans on Nvidia

Business Insider reported on May 18 that internal Amazon documents tied to an AWS initiative called “Titus” describe data-center designs built for more power-hungry, next-generation AI hardware. The documents, reviewed by the publication, show AWS reworking power architecture, liquid cooling, server layouts and construction timelines to support denser AI systems. The report adds to a mixed picture at Amazon, which has spent the past two years promoting its in-house Trainium chips as a lower-cost alternative for some AI workloads while continuing to build around Nvidia-linked infrastructure. Amazon did not publicly release the Titus materials. ### What did the Titus documents reportedly show? Business Insider said the Titus documents outlined a push to accelerate data-center construction while redesigning facilities for larger GPU racks and heavier thermal loads. A slide quoted by the outlet said the goal of the Titus portfolio was to deliver the “next AWS generational Data Center design.” The same report said AWS was changing facility specifications to accommodate hardware that draws more power than earlier generations. Secondary summaries of the article said the project included support for Nvidia’s GB200 systems, though the underlying Business Insider report is the key source for the Titus details. ### How does that square with Amazon’s Trainium push? AWS has continued to market Trainium as a central part of its AI infrastructure strategy. Amazon’s Trainium product page says Trn1 instances offer lower training costs than comparable EC2 instances, and AWS has separately promoted Trainium2 and Trainium3 systems for training and inference. At AWS re:Invent in December 2025, Nvidia said AWS would integrate Nvidia NVLink Fusion into infrastructure for Trainium4 deployment. (africa.businessinsider.com) That announcement showed Amazon’s custom-chip road map still intersecting with Nvidia technology at the networking and systems layer, even as AWS marketed Trainium as an in-house alternative. ### Why does Nvidia still appear in the picture? (aws.amazon.com) Nvidia remains embedded in the AI-compute stack that hyperscalers are building for the largest models. Business Insider’s account of Titus suggests Amazon’s most demanding facility designs are still being shaped by the power, cooling and rack-density needs of Nvidia-era systems, not only by the requirements of Amazon’s own silicon. (developer.nvidia.com) Andy Jassy has publicly argued that customers are diversifying beyond Nvidia. Business Insider reported last month that Jassy said Amazon’s chip business was “on fire” and that a shift had started away from a market where “virtually all AI thus far has been done on NVIDIA chips.” That makes the Titus details notable because they indicate Amazon’s internal buildout still has to account for Nvidia-class infrastructure even as AWS pitches diversification. (africa.businessinsider.com) ### Does this mean Trainium is not being deployed? AWS has already put Trainium at scale inside its cloud. Amazon announced Trn2 availability in late 2024 and said tens of thousands of Trainium chips were already powering Amazon and AWS services. Datacenter Dynamics later reported that AWS had activated Project Rainier, a cluster of nearly 500,000 Trainium2 chips spread across U.S. data centers. (africa.businessinsider.com) What the Titus reporting shows is narrower than that. The documents point to facility requirements for top-end AI infrastructure, not a cancellation of Trainium. They indicate Amazon is building for multiple hardware paths at once: its own accelerators, and systems that still require Nvidia-compatible power, cooling and interconnect assumptions. That is an inference from Amazon’s public Trainium announcements and Business Insider’s description of Titus. (aws.amazon.com) ### What should readers watch next? Amazon’s next public signals are likely to come through AWS product announcements, re:Invent presentations and future disclosures about Trainium4, Trn3 UltraServers and data-center expansion. Nvidia’s partnership materials and AWS infrastructure releases will show whether Amazon keeps tightening the link between its custom chips and Nvidia’s networking and systems ecosystem. (developer.nvidia.com) (africa.businessinsider.com)