AI Data Centers Run 24/7: Chip Depreciation Is 5 to 10 Times More Expensive Than Electricity
In AI inference, chip depreciation is five to ten times the cost of electricity; in Bitcoin mining, it's only 1.5 to three times. This cost ratio difference explains why AI data centers run continuously rather than turning on and off to save electricity like mining rigs.
The video won't play here. Listen to the audio instead:
The argument · timestamps estimated from transcript position
Fifteen-fold LNG surge proved Japan's power shortage warnings were real
Russia's invasion of Ukraine combined with European energy policy errors sent LNG prices up roughly 15-fold in 2022. Japan's grid input costs spiked as a result, creating genuine blackout risk during Tokyo's evening peak demand. This was why SoftBank singled out residential air conditioning in its conservation alerts.
— Patrick McKenzieStranded power plants funded miners first, then supplied AI factories
Industrial users often build on-site power plants rather than connect to the grid. When those users shut down, the plants become ‘stranded generation’—power still flows but has no buyer. Bitcoin miners were the ideal customer: flexible location, unlimited demand. CoreWeave rented GPUs to Ethereum miners using existing power contracts. When Ethereum demand waned, the same H100-filled data centers pivoted to AI inference and training, where the profit per megawatt was far higher.
— Patrick McKenzieElectricity flows too fast to store; grids survive on engineering workarounds
Power storage is prohibitively expensive relative to its value—not a limitation of battery chemistry alone, but physics itself. Electricity travels at roughly 0.6 times light speed through ordinary transmission lines. Power generators, especially nuclear reactors, need hours to ramp output up or down. These timescales are fundamentally mismatched. Electric grids survive only because of engineering compromises.
— Patrick McKenzieAt 93 percent capacity, temperature swings alone can crash Tokyo's grid
Electricity demand peaks in afternoon and evening, driven on weekdays mainly by residential air conditioning and heating. Tokyo's grid needs about 20,000 MW just to keep the city running—this is its ‘base load’. Between that baseline and maximum capacity sits an 18,000 MW gap. The grid normally operates at roughly 90 percent of peak capacity, which is safe. But once demand approaches 93 to 97 percent, even a small temperature swing can push residential AC demand over the limit, risking shortage.
— Patrick McKenzieNuclear fission stops instantly; residual steam means the plant never does
What power plants can be flexibly dispatched is determined by physics and economics. Nuclear reactors can halt their chain reaction in seconds, but residual heat keeps steam driving turbines for hours—this locks nuclear into base-load operation. Coal plants face the same economic lock-in. Solar, wind, and some natural gas plants offer flexibility but at higher cost; they occupy a middle tier. Peaker plants, mostly fossil-fueled, have the worst marginal cost but can start and stop instantly. They might run only a few hours per day, or a few weeks per year.
— Patrick McKenzieA contract to save power is economically equivalent to a power plant
An operator promising customers 40 megawatts but authorized to reduce service to 30 megawatts on five minutes' notice is economically equivalent to owning a hidden 10-megawatt power plant. SoftBank applies this logic to residential customers: it forecasts peak demand 24 to 48 hours ahead, calculates each household's typical use, and offers payment if that household uses one kilowatt-hour less during the predicted peak. There's no penalty for using more. Users click once to enroll. SoftBank's payment company delivers the reward in near-real-time at nearly zero marginal cost.
— Patrick McKenzieBitcoin miners held the spare capacity ERCOT desperately needed in crisis
During Texas's 2023 heat wave, ERCOT's grid was in acute crisis. Riot Bitcoin Mining held a forward contract giving it power at a fixed rate. It sold that power back to ERCOT at the peak spot price, earning tens of millions in a single transaction. The public blamed Riot for worsening the shortage. The author argues the opposite: miners represent ‘deferrable load’—they normally consume power that would go to waste. In an emergency, they become spare capacity the grid can activate instantly.
— Patrick McKenzieAI chip costs are 5x electricity costs; Bitcoin miners pay 1.5x
For Bitcoin miners, chip depreciation is only 1.5 to three times the cost of electricity. This makes them hypersensitive to power prices—they search globally for stranded power and play sophisticated demand-response games to capture every marginal gain. For AI inference, chip depreciation is five to ten times the cost of electricity. Power is a minor fraction of total operating cost. The default strategy is simple: connect as much electrical capacity as available and run continuously. This makes data centers behave like base-load demand rather than flexible virtual power plants.
— Patrick McKenzieIn their own words · checked verbatim
The first thing to understand about the energy markets is that everything written about them is a lie.
Patrick McKenzie2:58
These are fundamentally incompatible timescales and yet we somehow make it work anyway
Patrick McKenzie6:35
This is economically equivalent to having an invisible 10 MW power plant powered by clean-burning unicorn flatulence.
Patrick McKenzie13:34
because SoftBank invested in a payments company which offers zero-marginal-cost sends, they'll happily instantly transfer me 17 yen.
Patrick McKenzie13:34
And so ERCOT paid Riot several tens of millions of dollars to buy power.
Patrick McKenzie13:34
The depreciation is somewhere on the order of five to ten times more than the electricity costs.
Patrick McKenzie21:05
That is not primarily because Singapore is an excellent place, but because it allows you to smuggle inference into China.
Patrick McKenzie21:05
Figures
| LNG price increase (2022) | Approximately 15-fold | 1:03 |
| Tokyo grid base load | Approximately 20,000 MW | 7:57 |
| Tokyo dynamic load gap | 18,000 MW | 7:57 |
| US data center electricity share | 5% to 10% of total | 21:05 |
| AI inference chip depreciation vs. power cost | Approximately 5 to 10 times | 21:05 |
| Bitcoin mining chip depreciation vs. power cost | Approximately 1.5 to 3 times | 21:05 |
| AI provider bulk API discount | 50% | 13:34 |
| Smart meter reduction in meter readings | Over 99% | 19:50 |
| SoftBank demand response incentive | 17 yen | 13:34 |
| Riot's revenue from selling power to ERCOT | Tens of millions of dollars | 13:34 |
Glossary
- Demand response
- Grid notifies users in advance to reduce or defer electricity use in exchange for payment.
- Virtual power plant
- Aggregates flexible electricity consumption through contracts and software to function like a physical power plant.
- Base load
- The minimum electricity demand a grid must satisfy continuously, every hour of every day.
- Peaker plant
- Operates only during peak demand periods; has the highest marginal cost but can start and stop instantly.
- Stranded generation
- Generating equipment that loses its intended industrial customer and cannot find a new buyer.
- CapEx / OpEx
- CapEx = one-time capital investment in equipment; OpEx = daily operational expenses.
How to listen
Investors, founders, and engineers concerned with AI data center power constraints and grid economics.
The meter-reading history segment around 19:50 has low information density and can be skipped.