Compute now a business weapon

Eli Lilly did not brag about a new app. It showed off a machine room with 1,016 NVIDIA Blackwell Ultra graphics processors and said the system can deliver 9,000 petaflops for drug discovery, which is the kind of number companies used to leave to national labs. (youtube.com 1) (youtube.com 2) That shift is the story: compute is starting to look less like back-office information technology and more like a factory line. Deloitte wrote in December 2025 that firms moving artificial intelligence from pilot projects into production are finding their old infrastructure plans do not fit recurring, large-scale workloads. (deloitte.com) The reason is simple: once a company uses artificial intelligence every day, every customer query, model call, and automated step burns compute the way every truck delivery burns fuel. Deloitte said some enterprises are already seeing monthly artificial-intelligence bills in the tens of millions of dollars. (deloitte.com) That is why Lilly and NVIDIA announced a co-innovation lab on January 12, 2026 with a plan to invest up to $1 billion over five years in talent, infrastructure, and compute. The lab is meant to put Lilly biologists and chemists beside NVIDIA engineers so they can generate data and build models in one loop instead of handing work from one department to another. (investor.nvidia.com) In drug research, a wet lab is the room with pipettes, cells, and chemicals, and a dry lab is the room with models and simulations. NVIDIA said the new setup is designed to connect Lilly’s wet labs to computational dry labs for 24/7 artificial-intelligence-assisted experimentation. (investor.nvidia.com) When Lilly says 9,000 petaflops, the point is not the unit. The point is that a drug company now wants enough computing power to explore huge biological and chemical search spaces before making a single molecule in the physical lab. (investor.nvidia.com) (youtube.com) Manufacturing is moving the same way. In May 2025, Foxconn and NVIDIA said they would build an artificial-intelligence factory supercomputer in Taiwan with 10,000 NVIDIA Blackwell graphics processors so researchers, startups, and industrial companies could use shared compute as infrastructure. (nvidianews.nvidia.com) That project matters because Taiwan Semiconductor Manufacturing Company, the world’s biggest contract chipmaker, said its researchers planned to use the system for research and development. A factory company and a chip company are both now treating compute capacity as something that speeds product cycles, not just office software. (nvidianews.nvidia.com) Finance is a quieter version of the same race. IBM said on February 5, 2025 that only 8% of banks were developing generative artificial intelligence systematically in 2024, while 78% still had a tactical approach, which means most of the industry was still in the pilot stage. (newsroom.ibm.com) When banks move past pilots, the bottleneck stops being a clever demo and becomes reliable capacity, cost control, and speed. IBM said banks were shifting toward a strategic enterprise approach, while Deloitte said organizations are rethinking where workloads run because cost, latency, data sovereignty, and intellectual property protection all sit inside the compute decision. (newsroom.ibm.com) (deloitte.com) So the Lilly number is useful less as a record than as a signal. When a pharmaceutical company, a manufacturer, and banks all start talking about compute the way they talk about plants, supply chains, and capital budgets, artificial intelligence has stopped being a software feature and become industrial capacity. (investor.nvidia.com) (nvidianews.nvidia.com) (deloitte.com)