Nvidia's AI Chip Demand Is Outpacing Supply 12 to 1, Says Dan Ives

Wedbush analyst Dan Ives said on CNBC on July 27 that demand for Nvidia's AI chips is outpacing supply by 12 to 1, a figure he has since raised to 15 to 1. The shortage is driven by a bottleneck in TSMC's CoWoS packaging capacity and HBM memory supply from SK Hynix, Samsung, and Micron, with no meaningful new capacity expected until late 2026. Wedbush has a base-case price target of $250 for Nvidia by the end of 2026, with a bull case of $275.

Wedbush analyst Dan Ives just put a number on the AI chip shortage everyone's been guessing about: Nvidia demand is running twelve times ahead of supply, and that's before robots even enter the picture. Dan Ives said it plainly on CNBC on July 27: "Demand to supply today is 12 to 1" for Nvidia's AI chips. He wasn't hedging. The Wedbush senior managing director has spent the past year insisting the AI buildout is nowhere near finished, and this is the clearest number he's put behind that claim yet. Why the chips can't get made fast enough The bottleneck isn't really about Nvidia's chip design. It's about who gets to package the chips and who gets to supply the memory that sits next to them. Nvidia has locked up roughly 60% of TSMC's advanced CoWoS packaging capacity for 2026, according to industry trackers watching the foundry's allocation. Every Blackwell chip has to pass through that packaging step before it ships, and TSMC only has so many lines running. Memory is worse. HBM capacity at SK Hynix, Samsung and Micron is booked solid for 12 to 15 months out. Micron has confirmed its 2026 HBM output is fully committed. SK Hynix, which controls close to half the HBM market, wrapped up its 2026 supply negotiations months ago. Samsung is in the same spot. None of the three can just add a line and catch up. HBM fabrication takes years to scale, not quarters. That's the mechanical reason the shortage is expected to run through at least the third quarter of 2026, with no meaningful new packaging or memory capacity arriving until late in the year. The numbers keep climbing, not falling Ives has cited even bigger numbers since. Fifteen to one, he said in more recent commentary, up from the 12-to-1 figure he gave just weeks earlier. That's not what you'd expect if Nvidia's investment cycle were topping out. Bears have spent much of 2026 warning about an AI bubble. Ives keeps pointing at backlog data instead. Supply, not demand, is the constraint, he argues. He calls this the "third inning" of the AI revolution, and he means it. Wedbush has set a base case of $250 for Nvidia by the end of 2026, with a bull case running as high as $275, arguing Wall Street is still underestimating how much compute the hyperscalers need to buy. None of that 12-to-1 figure even counts physical AI. Ives has been explicit that robotics demand, the chips headed into humanoid robots, autonomous factory equipment and delivery machines, hasn't started showing up in Nvidia's order books in any meaningful way yet. When it does, the gap between what customers want and what TSMC and the memory makers can produce gets wider, not narrower. Who actually feels the squeeze For startups trying to rent GPU time, this is the part that actually bites. Cloud providers building out AI capacity are competing for the same constrained CoWoS slots and the same booked-out HBM supply as everyone else, and that scarcity shows up downstream as higher prices and longer waitlists for compute. A startup training a model in 2026 isn't just paying for silicon. It's paying a scarcity premium set by a packaging bottleneck at a fab in Taiwan and a memory allocation SK Hynix locked in a year ago. Supply doesn't catch up overnight. That's why Microsoft, Amazon, Google and Meta have all locked in multi-year capex commitments rather than betting they can buy chips on the spot market when they need them. The companies without that kind of clout, smaller AI startups and mid-size cloud resellers, are the ones left waiting. Whatever happens to AI valuations from here, the physical constraint is real and it's dated. TSMC isn't expected to bring meaningful new CoWoS capacity online until late 2026. Until then, the twelve-to-one ratio, or whatever it climbs to next, is the ceiling everyone in the AI supply chain is building around. Also read: A Hacker Turned DeepSeek Into an Autonomous Weapon Against 460 Servers https://startupfortune.com/a-hacker-turned-deepseek-into-an-autonomous-weapon-against-460-servers/ • AMD's MI355X Undercuts Nvidia's B300 on Cost to Run China's Kimi K3 https://startupfortune.com/amds-mi355x-undercuts-nvidias-b300-on-cost-to-run-chinas-kimi-k3/ • Anthropic Admits Its Own Bugs Broke Claude Code After Weeks of Denial https://startupfortune.com/anthropic-admits-its-own-bugs-broke-claude-code-after-weeks-of-denial/