Jensen on AGI, power and Colossus
Jensen Huang told Lex Fridman AGI is forming here but power and compute are the bottlenecks — he also name‑checked xAI’s Colossus building 200K GPUs in four months and said HBM is forcing industry changes. That’s a public push to frame NVIDIA as central to scaling AI compute. (x.com)
xAI’s Colossus page lists a 122‑day build for the first phase and says the cluster doubled to 200,000 GPUs after an additional 92 days, with published specs including 194 petabytes/sec total memory bandwidth and more than 1 exabyte of storage. (x.ai) The Colossus site in Memphis is tied to a 150 MW supply from the local grid and supplements that with roughly 150 MW of Tesla Megapack battery capacity, with a second substation planned that would bring total capacity toward 300 MW. (tomshardware.com) NVIDIA has been explicit about a shifting constraint: at GTC executives argued the industry is becoming “power‑limited,” and highlighted a projected global data‑center capex surge that could hit about $1 trillion in the coming years to support large‑scale AI. (fdiintelligence.com) At CES and in subsequent briefings NVIDIA introduced the Vera Rubin (Rubin) platform and said it is the initial beneficiary of next‑generation HBM4 memory, while HBM4 production partners such as Micron report high‑volume runs tailored for Rubin. (nvidianews.nvidia.com) Industry reports flag a concrete production risk: several outlets noted Rubin’s ramp could be pushed by HBM4 supply constraints and competing demand for advanced packaging and TSMC capacity from other AI chipmakers. (msn.com) At GTC 2026 NVIDIA quantified customer demand, saying backlog and purchase‑order visibility for its newest Blackwell and Rubin families now amount to at least $1 trillion of infrastructure demand through 2027. (cnbc.com)
