Our buddy Casey Handmer (@CJHandmer) of Terraform Industries (@TerraformIndies) sat down with Dwarkesh to drop fat knowledge on this AI-vs-electricity situation.
I pulled out some great datapoints that help me understand this much better 🧵:
I pulled out some great datapoints that help me understand this much better 🧵:
0: All of the below is w/r/t the AI business IN PARTICULAR.
AI cares about power availability, NOT PRICE.
If $1 of AI electricity makes $100 of value for someone, then the electricity can cost anything. If it's there.
AI cares about power availability, NOT PRICE.
If $1 of AI electricity makes $100 of value for someone, then the electricity can cost anything. If it's there.
[So we’re ALREADY in a “human zoo reservation” type situation. Otherwise, it’s much cheaper to “delete” competing electric loads — like employment, AC and HEAT — to free up kWhr’s for better-paying AI.]
1: It’s not that solar+batteries is cheap or clean that's tasty about it, but SCALABLE, and quickly.
Forget the “green” angle. That’s a (mostly honest) puppet show for Karen.
Solar+batteries is many identical, solid state, reliable and mass-made sub-elements, plugged together.
Forget the “green” angle. That’s a (mostly honest) puppet show for Karen.
Solar+batteries is many identical, solid state, reliable and mass-made sub-elements, plugged together.
Gas, OTOH, is a few gigantic & ornate elements that take 5-10 years to actually GET.
Even if the fuel was free, the power plants to burn it just aren’t coming in fast enough for OpenAI/xAI/Jeff-B/Zuck to NOT get skunked while waiting for them.
Even if the fuel was free, the power plants to burn it just aren’t coming in fast enough for OpenAI/xAI/Jeff-B/Zuck to NOT get skunked while waiting for them.
2: Solar+batteries is getting cheaper FASTER than any other power source.
And like I said, in this special case of AI in particular, the power’s COST doesn’t matter anyway.
Just speed-scaling IRL.
And like I said, in this special case of AI in particular, the power’s COST doesn’t matter anyway.
Just speed-scaling IRL.
3: Ramping up gas turbine manufacturing is iffy.
10x’ing the GW/year production of these machines is a near-generational investment.
10x’ing the GW/year production of these machines is a near-generational investment.
...and meantime, will solar+batteries skunk gas turbines out of business before that investment pays off? Or before the AI bubble pops and the extra turbines aren’t needed anymore? Who knows?
[And WHAT gas, BTW? Gas where, and for how long? At least sunlight is fixed-cost.]
[And WHAT gas, BTW? Gas where, and for how long? At least sunlight is fixed-cost.]
3A: BTW, paying off a present-vintage gas turbine eats up ~4 cents/kWhr alone.
…if you can order one, that is. Which you can’t.
…if you can order one, that is. Which you can’t.
4: No one is taking coal seriously for hyper-scale AI (regardless of CO2).
That stuff takes even longer to get and install than gas turbines do. All that water-boiling heat must be conducted THROUGH a heat exchanger, and those things cost a fortune and take forever.
That stuff takes even longer to get and install than gas turbines do. All that water-boiling heat must be conducted THROUGH a heat exchanger, and those things cost a fortune and take forever.
[OTOH, using peaking-power battery tricks to run our present-day gas/coal/nuclear gear at 100%, 24-7-365, is an obvious low-hanging fruit.]
5: The chips are where the expense is, not their power supply.
Nvidia H100’s cost ~$50/watt. That’s a lot!
Neither solar+batteries, nor gas turbines, nor windmills, nor nuclear, cost anything like $50/watt. But! Which of them can you actually GET, enough of AND fast enough?
Nvidia H100’s cost ~$50/watt. That’s a lot!
Neither solar+batteries, nor gas turbines, nor windmills, nor nuclear, cost anything like $50/watt. But! Which of them can you actually GET, enough of AND fast enough?
6: 1MW of “4-nines reliability” solar+battery AI in outback West Texas requires roughly…
~1000 H100’s
~25 MWhr’s of battery
~10MW of raw solar
spread over ~10 acres
~1000 H100’s
~25 MWhr’s of battery
~10MW of raw solar
spread over ~10 acres
So at $0.10/raw-watt, those solar panels cost ~$1 per watt of 4-nines 24-7 AI action.
While the chips cost $50 per watt! 50x as much!
Ergo, 200% tariffs on Chinese raw solar panels are no big deal.
While the chips cost $50 per watt! 50x as much!
Ergo, 200% tariffs on Chinese raw solar panels are no big deal.
7: At $3000/acre for useless TX desert land, those 10 acres cost $30,000 = just $0.03 per “4-nines” watt!
Incredible.
So forget about what the land costs.
Incredible.
So forget about what the land costs.
[Ergo, AI can also out-compete nearly all AGRICULTURE for land too, so without more “human zoo protectionism”, we’re all screwed a second way, with no hot water, heat OR FOOD!]
8: Just forget about transmitting all this AI power over the public power grid.
In the Anglosphere at least, it’s hopeless. All big AI farms will need their own power plants, and will be "off-grid" little countries.
In the Anglosphere at least, it’s hopeless. All big AI farms will need their own power plants, and will be "off-grid" little countries.
[OR, put them next to existing power plants, like right on-site (= no transmission), and run on their night-time base-load surplus, via batteries.]
9: BTW, the word is there are ~10M “H100 equivalents” of AI compute out there now, and in 2028 there'll be ~100M of them, 10x as many.
If each H100 uses 1kW, then 100M of these beasties would need ~100GW of new/scrounged electricity.
If each H100 uses 1kW, then 100M of these beasties would need ~100GW of new/scrounged electricity.
[And now some questions and comments from me:]
10: Note that China CAN do all this, even with crapper and less-efficient chips, if they have enough electricity for them.
Needing 2-3X as much power as Jensen's Nvidia chips, “per token”, isn’t that terrible, as long as they actually HAVE that electricity. (Which they do?)
Needing 2-3X as much power as Jensen's Nvidia chips, “per token”, isn’t that terrible, as long as they actually HAVE that electricity. (Which they do?)
11: Solar+battery AI farms imply buttloads of water-chilling HVAC.
Who knows anything about the cost, availability and speed-scaling of all this HVAC?
HVAC isn’t as high-tech as gas turbines, but they’re not paperweights either.
Who knows anything about the cost, availability and speed-scaling of all this HVAC?
HVAC isn’t as high-tech as gas turbines, but they’re not paperweights either.
At least HVAC can be many “smaller” machines, one shipping container each, plugged and plumbed together on-site.
[Big upside here, I bet, to make this scaling story actually work out, especially in rare-earth-free way.]
[Big upside here, I bet, to make this scaling story actually work out, especially in rare-earth-free way.]
12: If hyper-scale solar+battery AI will be off-grid, then 60Hz AC needn’t even be involved at all.
Lots of fast-switching DC-to-DC tricks to try here.
[I bet a rare-earth-free treatment of this need would get some serious play.]
Lots of fast-switching DC-to-DC tricks to try here.
[I bet a rare-earth-free treatment of this need would get some serious play.]
13: Maybe gas turbines actually CAN scale-up quickly, if we use crappier turbines?
Remember that super-efficiency doesn’t help here, just the GW/year production rate, as long as the gas is there.
So maybe smaller, older-technology and less-efficient turbines, and lots of them?
Remember that super-efficiency doesn’t help here, just the GW/year production rate, as long as the gas is there.
So maybe smaller, older-technology and less-efficient turbines, and lots of them?
@threadreaderapp unroll
• • •
Missing some Tweet in this thread? You can try to
force a refresh
