The "AI bubble" is probably OpenAI and Anthropic's valuations. Mostly because they don't own their own compute (yet). If you own your compute - and better yet, if you design your own chips - you control the most important layer of generating intelligence. The actual hardware. Companies that can hyper-optimize the chip/compute stack to their models will have the lowest cost to serve intelligence (and fastest to serve as well). That's because these companies will be able to hyper-optimize every transistor on the chip to how their models work. This is especially important over time as the level of intelligence surpasses what humans would need it for, or would appreciate. I'm starting to believe that there's a ceiling for intelligence - aka there's such a thing as "too much" intelligence. How often is it that folks with 100 IQ can appreciate the work of 180 IQ folks? Almost never? This means that at some point, a threshold will be reached where every model will be "intelligent enough" for 80%+ of use-cases that humans would need it for across any field or industry. Once that level is reached, the game becomes cost and speed. Essentially, who can serve the highest intelligence tokens that are "good enough", at the lowest cost, the fastest? The formula would be (Intelligence / Cost) x Speed. And companies that control their entire chip/compute stack would be best positioned to solve for that cost and speed problem, because ultimately that's what people are going to value. I'm willing to bet 90%+ of people would much rather get a 140 IQ result at 10x the speed and 1/10th the cost of a 180 IQ result.