GTC pushed DLSS, GPUs, price gaps
NVIDIA’s GTC showcased DLSS 5 for real‑time rendering and flagged a full‑stack push — commentators note ~1M GPUs deployed globally and 1.7GW of AI capacity even as high‑end inference hardware (GB300) lands at roughly $170K, highlighting a hyperscaler focus that leaves many enterprises scrambling. ( )
Several major studios, including Ubisoft and Capcom, told reporters they were unaware their titles were used in NVIDIA’s GTC demos, a reaction that developers said left them blindsided. (pcworld.com) Public reaction on forums and social coverage skewed negative for the new neural‑rendering demo, with Wired reporting gamers and some developers described the results as uncanny or off‑putting. (wired.com) A Reuters account relayed comments from NVIDIA VP Ian Buck that a multi‑year cloud supply plan will see shipments begin this year and run through 2027 as part of a deal to put a very large number of NVIDIA chips into AWS data centers. (money.usnews.com) NVIDIA told GTC attendees its cloud partners now account for more than one million GPUs and about 1.7 gigawatts of AI capacity, a jump from roughly 400,000 GPUs and 550 megawatts a year earlier that Silicon Report calculated as roughly a 3.1× year‑over‑year increase. (siliconreport.com) GB300 NVL72 rack systems are sold as 72‑GPU, 36‑CPU liquid‑cooled units that draw about 132–140 kW per rack according to HPE’s product listing, and industry reporting has put full‑rack GB300 price estimates in the multi‑million‑dollar range while separate analyses flag liquid‑cooling add‑ons of around $50,000 per rack. (buy.hpe.com) (techpowerup.com) (barrons.com) Cloud providers and rental marketplaces list GB300 GPU hours from roughly $1.02/hr up to tens of dollars per hour across providers, while vendors are also shipping deskside GB300 workstations — MSI pricing around $97,000 was cited in industry coverage — underscoring the gap between cloud access and on‑premise capital costs. (computeprices.com) (computerworld.com) NVIDIA framed GTC as a full‑stack “AI Layer Cake” strategy with multiple chips and rack products built for hyperscalers, and analysts at the show warned that the focus on hyperscaler deployments shifts the bottleneck from silicon to site engineering, power delivery, and integration for smaller enterprise buyers. (blogs.nvidia.com) (siliconreport.com)
