BofA dropped their August AI Tracker and the divergence is wild:
Model layer is in freefall:
• AI Token Price Index: -9% MoM
• LLM Token Expenditure Index: -27% MoM
Frontier models are cannibalizing each other on inference costs. Race to zero.
Meanwhile GPU rental prices? Still rock solid:
• B200: -2% MoM but +7% YoY
• H100: +2% MoM, +33% YoY
• A100: flat MoM, +17% YoY
Memory is even worse:
• DRAM: +8% MoM, +483% YoY
• NAND: flat MoM, +432% YoY
The takeaway: AI is getting dirt cheap to USE. But building it? Still insanely expensive.
Pricing pressure hits models first. Infrastructure stays tight.
For tradoors: This setup still favors memory plays like $MU and $SNDK. The picks and shovels thesis isn't dead.
For AI labs: The real question is who survives when performance converges and pricing power evaporates. Commoditization is coming fast.
Customers win. Model providers bleed. Infra suppliers print.
Model layer is in freefall:
• AI Token Price Index: -9% MoM
• LLM Token Expenditure Index: -27% MoM
Frontier models are cannibalizing each other on inference costs. Race to zero.
Meanwhile GPU rental prices? Still rock solid:
• B200: -2% MoM but +7% YoY
• H100: +2% MoM, +33% YoY
• A100: flat MoM, +17% YoY
Memory is even worse:
• DRAM: +8% MoM, +483% YoY
• NAND: flat MoM, +432% YoY
The takeaway: AI is getting dirt cheap to USE. But building it? Still insanely expensive.
Pricing pressure hits models first. Infrastructure stays tight.
For tradoors: This setup still favors memory plays like $MU and $SNDK. The picks and shovels thesis isn't dead.
For AI labs: The real question is who survives when performance converges and pricing power evaporates. Commoditization is coming fast.
Customers win. Model providers bleed. Infra suppliers print.