Nvidia Corp. (NVDA) shares closed at $221.97 on Wednesday, up nearly 3 percent—the stock's best day in over a month—as announcements from GTC Berlin and analyst commentary reinforced the company's dominant position in enterprise AI infrastructure.
Deepwater Asset Management managing partner Gene Munster said investors underestimate Nvidia's integrated strategy. "Software, libraries and hardware form a full stack that raises switching costs and supports recurring upgrade cycles," Munster said. The observation is material: once a customer commits to Nvidia's CUDA software layer, the cost and complexity of switching to a rival's hardware architecture becomes prohibitive. That lock-in effect underpins Munster's forecast that physical AI—autonomous agents consuming 10 to 15 times more compute tokens than chatbots—will eventually contribute one-third of Nvidia's revenue.
Concrete evidence arrived Wednesday. Google Cloud announced general availability for its Vera Rubin-powered A5X instances in November, with European expansion to follow, eventually scaling to 960,000 Rubin GPUs across sites. Dell Technologies expanded its AI Factory with Nvidia to more than 6,500 customers, up from 5,000 last quarter. Hewlett Packard Enterprise broadened enterprise AI offerings. CrowdStrike is integrating Falcon Guardian with Nvidia OpenShell and plans Sentry integration on BlueField-4 processors to enforce hardware-level security. Nutanix and Netris disclosed new infrastructure work for AI cloud providers.
CEO Jensen Huang detailed the efficiency gains from Nvidia's factory-assembled DSX Pod, which integrates computing, networking and cooling. The system cuts time from equipment arrival to token production from four to five months to approximately one month. Huang said this cuts wasted idle compute and idle capital expenditure, boosting customer profitability. Vertiv modeled that its DSX-based approach yields up to 30 percent more cumulative tokens in the first 12 months under the same installed capacity. Polar Data Centers is implementing the approach for a 41-megawatt development phase in Norway.
Huang confirmed that the Vera Rubin architecture is in full production and claims 30 times Blackwell's performance in large language model processing. Nvidia's nearly $10 billion investment in Europe over the past year and more than 12,000 regional partnerships reinforce the company's stranglehold on infrastructure supply.
The partnership announcements and product debuts validate Munster's core thesis: Nvidia's switching costs—embedded in software, trained on proprietary libraries, locked into hardware—are rising. At current valuations, the market has not fully priced the durability of that moat or the upside from physical AI consuming multiples of today's compute demand.
