Compute becomes a moat
Compute capacity is turning into a strategic asset as non‑tech firms build large clusters for domain workloads — the media discussion highlighted Eli Lilly’s reported 9,000 petaflops figure as evidence that vertical players are betting on owned or reserved compute. That shift makes infrastructure deals and guaranteed capacity key procurement questions, not just model selection. (youtube.com)
A year ago, most companies shopping for artificial intelligence asked one question: which model should we use. In 2026, companies are asking a different one first: where do we get guaranteed graphics processor capacity, and can we keep it for years. (nvidia.com) That shift got a clear example in March, when Eli Lilly said its new “LillyPod” system delivers more than 9,000 petaflops of artificial intelligence performance. NVIDIA said the machine uses 1,016 Blackwell Ultra graphics processors and was assembled in four months in Indianapolis. (nvidia.com) A petaflop is a way to measure how many math operations a machine can do each second. 9,000 petaflops means Lilly is buying computing power the way an airline buys gates or a factory buys assembly lines: as core capacity, not rented overflow. (nvidia.com) Lilly is not building this to run a chatbot for employees. NVIDIA said the system is meant to compress drug discovery timelines, with workloads in genomics, molecular design, and personalized medicine. (nvidia.com) That is the important change: the buyer is a drug company, not a cloud platform or a social network. NVIDIA described LillyPod as the largest and most powerful artificial intelligence factory wholly owned and operated by a pharmaceutical company. (nvidia.com) Once a company decides the models are available to everyone, the scarce thing stops being the model. The scarce thing becomes access to chips, networking, power, cooling, and the engineers who can keep 1,000-plus graphics processors running as one machine. (nvidia.com) That is why infrastructure vendors are selling “superclusters” instead of just servers. Oracle says its cloud supercluster can scale to 131,072 graphics processors, and NVIDIA markets DGX SuperPOD as a full data center platform with storage, networking, software, and deployment services bundled together. (oracle.com) (nvidia.com) The deals are starting to look less like software subscriptions and more like industrial procurement. Oracle said in September 2024 that it was taking orders for cloud clusters with up to 131,072 Blackwell graphics processors, which tells you customers now reserve giant blocks of future capacity before the hardware is even common. (oracle.com) NVIDIA is packaging that same idea for enterprises that do not want to wait in the public cloud line. Its March 2025 launch of Blackwell Ultra DGX SuperPOD pitched an “out-of-the-box” artificial intelligence supercomputer, and Equinix was named as an early provider of preconfigured space for customers that need fast deployment. (nvidianews.nvidia.com) For companies in medicine, finance, manufacturing, and defense, this changes the buying checklist. The first question is no longer only whether one model scores 2 points better on a benchmark; it is whether your supplier can promise enough chips, power, and uptime to run your domain workload when everyone else wants the same hardware. (oracle.com) (nvidia.com) Lilly pushed the logic even further in January 2026, when NVIDIA and Eli Lilly announced a co-innovation lab for drug discovery and manufacturing. Once a company pairs reserved compute with internal data and a workflow built around that compute, the advantage starts to look like a moat, not a feature. (nvidianews.nvidia.com)
