Frank Downing
Director of AI & Cloud Research at ARK Invest
Nvidia can raise prices 15% and still leave customers plenty of margin because Vera Rubin delivers 2-30x more throughput per megawatt than Blackwell Ultra, depending on interactivity. A megawatt of Vera Rubin racks should generate at least twice the revenue of Blackwell racks, and much more in some configurations. SpaceXAI also appears to be a new standalone Vera CPU customer; Nvidia says Vera completes tasks up to 1.8x faster than x86 CPUs. x.com x.comÂ
Nvidia's hardware lock-in is deepening as model and chip co-design grows tighter, and the GPT-5.5 release shows that co-design yielding more efficiency. Per @GavinSBaker, labs used to port models freely across GPUs, Trainium and TPUs, but true portability is eroding as system architectures diverge in scale-up topology, memory hierarchy and networking. Blackwell's scale-up domain, for example, is 9x larger than AMD's MI355. x.comÂ
Nvidia can capture more AI infrastructure spending by supplying Vera CPUs and networking alongside GPUs. Oracle outlined an initial supply of 150,000 Vera CPUs, and OpenAI says it is putting them into production for agent tool execution; reducing CPU waits could improve cost per task and GPU-memory use. Benefits from independently scaling CPUs and GPUs still need validation in utilization and completion times. x.comÂ
Better chips explain falling AI token prices, while rising GPU rents reflect demand for Nvidia's hardware. B200 rentals rose 28% in the first eight months of 2026 even as token prices more than halved since May. A Vera Rubin rack is estimated at ~$7.8M, almost twice a Blackwell GB300 rack's price, but Nvidia says Rubin cuts cost per token by up to 10x versus Blackwell; its September agentic-coding benchmark puts the gap at up to 45x. x.comÂ
Nvidia's AI profits are real, but its compute demand depends heavily on fragile venture funding. Per Chamath, 1% of firms account for 90% of AI spending, and 72 cents of each dollar comes from unprofitable VC-backed firms. Venture money passes through startups to Nvidia and cloud suppliers; if the next raises fail, concentrated orders could contract together. x.com x.com x.comÂ
A rising value of compute per watt should benefit AI cloud, chip, connectivity and power suppliers. Elon Musk says that value will rise over time. Nebius has Microsoft and Meta deals; Marvell benefits from custom silicon demand; Credo connects GPUs; Vistra has long-term nuclear supply deals with Amazon and Meta; Nokia's Nvidia partnership could help it capture data-center networking growth. x.com x.comÂ
Nvidia keeps making more money from each gigawatt even if power becomes AI's bottleneck, as Elon Musk says. Its revenue opportunity per gigawatt has grown from about $18 billion for Hopper to $25 billion for Blackwell and $40 billion for Vera Rubin. x.comÂ
Nvidia and other accelerator makers now face a finite semiconductor profit pool and must fight their supply chains for incremental gross margin, pushing down dollars per GW after years of inflation. Economic surplus needs to move from silicon to users; otherwise the inflation from semis and backlash against data centers will lead rate markets and policy to end this trade. Expecting $150-175B of AI lab ARR to justify $1T+ of capex sounds insane. x.com x.comÂ
SpaceX's reported deployment pace adds concrete demand for Nvidia GPUs. Musk plans a 440,000-unit Vera Rubin training cluster online by Q2, apparently 2027; SpaceX says it is turning on 220,000 Nvidia GB300s a month, with Musk expecting that pace to accelerate next year. x.com x.comÂ
SpaceX's large Nvidia GPU deployments underpin a $400 Nvidia stock call. Elon Musk said SpaceX is deploying roughly 200,000-220,000 Nvidia GB300 GPUs per month, expects to transition primarily to Vera Rubin hardware in 2027 and discussed a planned training cluster of 440,000 Vera Rubin units. x.comÂ
Nvidia's push into AI models and agent software looks weak even though its GPUs and racks are great. Its Nemotron model is mediocre and its LangChain-based agent stack won't make it. If Nemotron were capable of frontier performance it would be there already, and Nvidia doesn't seem to know how to get there. x.com x.comÂ
Nvidia would be the biggest winner of SpaceX's proposed compute spending because its GPUs are too good for SpaceX to replicate in-house. If SpaceX builds all 13 GW by end-2027, about $470B of capex over five quarters could yield $414B of Nvidia revenue from one customer, with an estimated $310B in gross profit and $200B in net income. Nvidia's share of each dollar rises from about 83% with Blackwell to 89% with Vera Rubin, expected to ship at scale from Q2 2027. x.comÂ
