People are having trouble following the logic in my posts so let me explain 1) top AI spenders are spending less on AI 2) even if overall AI token usage is going up, spend is not going up in line. Their spend = AI related revenue 3) LLMs becoming more efficient = less revenue for more usage 4) the amount companies spent and continue to spend on AI capex is fixed and expects certain amount of revenue to pay it back 5) even with LLMs getting cheaper, and usage going up, companies are spending less because they are cost-conscious and must spend on tokens responsibly. Recent trend shows reigning in costs. Also, note that this is during the "VC subsidized" era of AI compute. Meaning if the real pricing was passed through, companies would be using a lot less as they could not afford/justify the bill. You can be bullish AI token usage but skeptical on the sustainability of the capex given the expected revenue. It isn't economical. Trillions spent on capex, todays use cases will not pay it back for some time. Revenue from top 1% is already declining, and Ant ARR slower than est which lines up. People will mention open source models, but fail to understand those are even cheaper options which means even less revenue generated in the ecosystem. Also, much of the top 1% spenders are early stage start ups that require even more VC funding to keep going and burn the cash. If money dries up, people are underestimating the near term impacts on AI compute demand. You can extend this one step further to the frontier labs who are driving a ton of the overall compute demand for internal model development (which does not have tangible ROI to justify/foot the bill). These companies are not able to fund this tens of billions of dollars of spend themselves. This again requires tons of external capital. What happens if that capital is also fleeting? This is also where all the S&P earnings growth is coming from (AI infra, $MU, $NVDA) - aka contagion and earnings could roll over. The capital is key. It's not infinite. It's not being paid back fast enough to replenish spend. The true demand is being subsidized. The top users are already showing declining spend. This has a big air pocket until the next wave of demand hits (like physical AI). If you can't fund the model training at frontiers, or start ups can't raise as much, we could possibly go from compute constrained to supply glut. Big risk, and you're talking trillions in capex + leverage in the ecosystem. And again the potentially profitable AI revenue from profitable companies is already being optimized for spend (lowering costs).