Everyone's worried $NVDA is "losing share" to custom chips They're missing the point. UBS has Nvidia going from 7.5 GW of AI rack capacity in 2025 to 50.5 GW in 2030 That's ~7x in 5 years Yes, Nvidia's share shrinks on this chart. It goes from ~65% of all AI racks in 2025 to ~48% by 2030, as custom ASICs (Google TPUs, Amazon Trainium) grow from ~4 GW to 45 GW But share is the wrong number to watch. The total market goes from 11.5 GW to 104.5 GW, roughly 9x A slightly smaller slice of a pie that's 9x bigger is still a monster. Nvidia adds ~11 GW in 2027 ALONE, which is about the size of the entire AI rack market in 2025 And the competitors growing doesn't mean they're catching up on the product SemiAnalysis just benchmarked Vera Rubin on agentic inference. On one test it delivered ~30x the throughput per dollar of AMD's MI355X, and even beat Nvidia's own GB300 by 1.4-3x at the speeds most providers actually serve models Hyperscalers build ASICs to lower costs on their own workloads. Everyone else (the labs, the neoclouds, sovereign AI, enterprises) still buys Nvidia because it's the best product you can get And here's the kicker: even 104 GW won't be enough SpaceX alone is targeting 10-20 GW of compute by the end of 2027. That's equal to 20-40% of all the AI racks UBS expects the world to deploy in 2027, for one company. Now add OpenAI, Anthropic, Meta, Google, Microsoft and every agent we haven't even launched yet Demand isn't the constraint here, supply is $NVDA remains a core holding inside my Milk Road PRO portfolio. If you want to see what else I hold, you can check out my real-time portfolio here: Don't forget to give me a follow @kylereidhead for more insights on AI and markets