Gavin Baker sees a path where AI operators generate more money from each gigawatt of power, provided that improvements in how many tokens models can process per watt beat the ongoing drops in what each token sells for. Smarter models that unlock higher-value tasks could help sustain or even lift end-user pricing, turning energy constraints into a revenue lever rather than a hard ceiling.
"If token throughput per watt rises faster than price per token falls, revenue per gigawatt can net expand. This is compounded by models getting smarter, which unlocks higher value tasks for end users [justifying increased costs]"
Well said @ShanuMathew93 and @downingARK
Whether projected efficiency jumps will actually outrun price erosion remains unclear, as does the question of how much added model intelligence will translate into real pricing power for new tasks.
David Lee notes that search ad prices fell for years without crushing revenue once efficiency rose and higher-value inventory appeared, suggesting token markets could see a parallel outcome if value creation accelerates.
Users agree token efficiency gains outpacing price drops will boost AI revenue per gigawatt because they expect exploding yields from model improvements and broader enterprise adoption.
No Digg Deeper questions have been answered for this story yet.
Original post:
Great walk-through & visuals.
Key point: "If token throughput per watt rises faster than price per token falls, revenue per gigawatt can net expand. This is compounded by models getting smarter, which unlocks higher value tasks for end users [justifying increased costs]"
@GavinSBaker @ShanuMathew93 @downingARK @ShanuMathew93 if you have not check out the Carl Shulman on Dwarkesh ep, lays so much of this out in amazing detail 3+ years ago and talks about adding researches (labor capacity) through machines capacity / RSI
@GavinSBaker @ShanuMathew93 @downingARK revenue per gigawatt assumes you can site one. 69 US jurisdictions have blocked new data centres
@GavinSBaker @ShanuMathew93 @downingARK analogy i think of is when CPC declines for Google. Mainstream media thought that was a bad thing for awhile.
"If token throughput per watt rises faster than price per token falls, revenue per gigawatt can net expand. This is compounded by models getting smarter, which unlocks higher value tasks for end users [justifying increased costs]"
Well said @ShanuMathew93 and @downingARK
@GavinSBaker @ShanuMathew93 @downingARK Effective Token Yield will explode once the models can build heuristic layers as part of the agentic flows.
And then on top of that just adding "Rewrite this prompt if it can be done more efficiently" to some people's flows would be another 10x
@GavinSBaker @ShanuMathew93 @downingARK 🎯
@GavinSBaker @jaminball @ShanuMathew93 @downingARK I like the framework Gavin, but I think it needs to account for the cost.
Throughout per dollar in costs per watt.
Otherwise we don’t account for material costs, shortages, etc.
@GavinSBaker @ShanuMathew93 @downingARK Price per token will inevitably fall significantly when the vast majority of non-tech enterprises shop for open source inference at scale? Frontier not needed for vast majority of gains there…. seems inevitable
@GavinSBaker @ShanuMathew93 @downingARK @GavinSBaker, token throughput per watt is the main thing. With GLM-5.2 now on the Pareto frontier, signaling marginal cost will go to zero due to opensouce, do you think future TTPW gains will come more from the model layer or the chip layer?
I think this could be a long term theme but looking at the price history of various commoditized products I would not be surprised to see price declines overwhelm throughput growth over any 12-18 month period. New GPU's & TPU's take time to engineer and smarter models take time to code.
@GavinSBaker @ShanuMathew93 @downingARK I've spent ~15 yrs in various growth/bd/revenue roles in tech companies and Fable with a skilled operator is more capable in most tasks than me.
Another year and assuming the "skilled operator" part is less of a barrier, a lot of unit economics make a lot of sense.
@GavinSBaker @ShanuMathew93 @downingARK very much agree. this is the way forward for broadened adoption by enterprises and users.
@GavinSBaker @ShanuMathew93 @downingARK What if tokens are commodities and token providers don’t have pricing power?
@GavinSBaker @ShanuMathew93 @downingARK If China helps drive token prices toward marginal cost, why should exponential token growth translate into exponential profits?
@boo @GavinSBaker @downingARK Carl's a good dude! Agreed, he's been clear on his view (& successful with their $ bets)
@GavinSBaker @ShanuMathew93 @downingARK Or by hyperscalers owning & optimizing the harness management on the enterprise stack through their own compute. Both bets imply tokens aren’t the unit to gauge AI on.
@orionintx @GavinSBaker @ShanuMathew93 @downingARK Throughout PER WATT. Which can mean "same watts, but more tokens on those watts". E.g. fuel efficiency.
GPUs are a scarce resource, and China does not have many. That means Chinese open-source models are largely being run on US clouds, not being exported from China to rest of world.
The US clouds have healthy & improving margins, and thus little incentive to use their scarce GPUs to serve open models at lower than company level margins.
@GavinSBaker @ShanuMathew93 @downingARK In theory yes, in practice it's not linear
@GavinSBaker @ShanuMathew93 @downingARK This is actually the Bear Case that makes sense to me. Token yield could rise faster than Jevon's Paradox could keep up with.