@AndreasSteno Why is this a bad thing exactly? I thought this was where they were aiming at all along

@_Billdozer_ @riteshmjn Robotics is a very good bet imho. Real life AI

@AndreasSteno Compute is constrained by supply. Demand would increase if costs come down.

@ethics3606 That I agree with, but the current pricing also means that people are willing to pay everything for hardware. They wont be willing to do that forever

@realfeaq Its not bad, but it will kill the hardware Trade

@GlobalCollapse @AndreasSteno No. Token pricing goes down —> even more token demand.

More token demand —> more need for hardware and data centers.

@AndreasSteno Covered this topic in detail here:

@AndreasSteno By this logic it was over in february?

@AndreasSteno @MitchMartan98 Does this include or exclude IREN & CIFR/WULF?

@AndreasSteno @realfeaq Token pricing goes down = Hardware demand goes up.

I get where you're coming from but it is nowhere near that simple. It would be true if cost per token was stagnant, but it isnt.

@noel_moore @MitchMartan98 Less sensitive than a lot of other trades, but yes..

@AndreasSteno You are saying: token pricing up = AI trade good; token pricing down = AI trade dead.

It is more like: AI infrastructure demand = token volume × compute intensity per token × price/margins × model mix × latency/SLA requirements

Token pricing compression cuts both ways for the hardware trade. Cheaper inference per token drives volume growth faster than margin compression — AWS and Google are still expanding datacenter capacity even as per-token economics erode. The key ratio is not token price but inference demand growth versus capacity additions. If token prices drop 50% but query volume grows 3x, the hardware trade does not break — it gets repriced at higher utilization rates, with more copper, more power, and more cooling per facility.

@AndreasSteno @carlquintanilla Wait until agents become more popular. This gonna explode 🚀

@AndreasSteno Your only partially right and the part your probably wrong on matters more . “between (a) usage actually contracting, and (b) price deflation merely accelerating faster than volume compounds — and (b) is the Jevons case, which is bullish for inference compute and thus for memory”

@AndreasSteno @riteshmjn I am so curious what the next cycle will be. Robotics? energy? both?

@AndreasSteno Is this cost per token, or an indication of business expenditures on tokens?

@AndreasSteno agree.

@AndreasSteno 😂😂😂😂

@AndreasSteno The models can get more efficient. Or people might migrate to cheaper models. In both cases the volume can pick up. I don’t know man, I think your Korean data and maybe TSMC monthly revenue are better gauges. No?