America can pull gigawatts out of thin air through a combination of technology and smartly targeted policy. Let me show you how. …
It is often said that the US electricity grid is under increasing risk of blackouts, price spikes, and other signs of strain. … Most of the time, however, the grid has significantly more power than is needed. This means that the grid can often accommodate, say, a new 750-megawatt data center for the vast majority of the year. There is spare generation capacity available except for those brief periods of high demand. But in the high-demand periods, when all the grid’s generation capacity is required to maintain electricity service, this new data center would require an additional 750 megawatts of electricity generation capacity, and quite possibly also transmission infrastructure upgrades.
This additional investment is only necessary if you assume that the new data center will require all 750 megawatts of electricity during peak-demand periods. Traditionally, this assumption has been true: data center operators rely on extremely high uptime, and grid operators work under the assumption that new electricity demand will be constant during periods of high demand.
If, however, that assumption were not true, and a data center was able to significantly reduce or eliminate its electricity consumption for a small portion of the year (the high-demand period), the calculus changes radically. More power would suddenly become available because the data center can tap into the grid’s existing surplus capacity without requiring investment in net-new capacity on the days when the grid is operating at the limits of its capacity.
How much more power could be unlocked? In a viral paper earlier this year, Tyler Norris and colleagues at Duke University estimated 76 gigawatts if the new users of that power were willing to curtail their electricity demand for 0.25% of the year. In overly simplified terms, this means that America could accommodate 76 gigawatts of new AI data centers today, with no new power generation built, if those data centers were willing to reduce their demand by an average equivalent of roughly 22 hours out of a year.
As it happens, the estimates I trust most about near-term AI-related electricity demand suggest that we will need about 50-75 gigawatts for AI over the coming 5 years—perfectly in line with Norris’ estimates.
And a convenient win-win:
In addition to instantly unlocking more power for AI and other industrial applications, curtailing power at the scale envisioned in the Duke study would achieve other benefits. For example, as Norris observes, more efficient use by industrial customers of existing power generation capacity during non-peak demand periods would result in high utilization rates of existing capital assets, and thus lower prices for consumers.
The result is a win-win for both AI data center operators and average Americans concerned about the affordability and reliability of electricity. The only downside would be that, during periods of peak demand (for example, on a particularly hot day in one region of the country), AI users across America might notice their AI services being slower and less reliable than usual. This seems well worth the cost.
I do wonder how believable this is, given my personal experience that the more familiar I am with a policy claim’s quantitative modelling and the more domain knowledge I have about it the less I trust it generally speaking...
I came across a similar “hack” on LinkedIn from Tom Styer:
“California just pulled off the nation’s largest-ever test of a virtual power plant. This successful test proved VPPs are a fast, low-cost, zero-emissions way to make better use of the clean energy we already have — and to keep the lights on as demand surges from data centers, heat waves, and electrification.”
Basically, they are talking about allowing residential batteries supply the grid during peak demand. I tend to be skeptical about stuff like this because in my own domain, water, there’s a lot of pop science and bold claims that ignore scaling and logistics. I asked a smart fellow in that industry about it https://substack.com/@energycrystals and I thought he gave a good answer that aligns with my experience with water, which is it always come down to implementation: ”The challenge is lining up customer telemetry to incentive structures that matter. With standard demand response products (which some people sell as VPPs), the incentives given to customers don’t pay for the battery and the software admin and API costs to the utilities outweigh the cost savings of a VPPVPPs are vaporware until someone can make the business model pencil and the API integration and UX not suck ass”
So, without knowing more, my prior is that this free capacity is there for a reason, and that utilities aren’t that dumb. On the flip side, I think it’s great that we are thinking this way. Probing our systems and looking for efficiencies are worthwhile. our legacy infrastructure is a mess of path dependent bureaucracy and I’m certain there’s gains to be made in addition to new construction.
Whether this is feasible depends on how concentrated that 0.25% of the year is (expected to be), because that determines the size of the battery that you’d need to cover the blackout period (which I think would be unacceptable for a lot of AI customers).
If it happens in a single few days then this makes sense, buying 22GWh of batteries for a 1GW dataset is still extremely expensive (2B$ for a 20h system at 100$ / kWh plus installation, maybe too expensive for reliability for a 1GW datacenter I would expect, assuming maybe 10B revenue from the datacenter??). If it’s much less concentrated in time then a smaller battery is needed (100M$ for a 1h system at 100$/kWh), and I expect AI scalers would happily pay this for the reliability of their systems if the revenue from those datacenters
Demand response could be done by covering the data center with battery energy or not. Demand response and batteries can stack: if the grid is really stressed, a data center can both turn off and discharge its battery into the grid.
Economically, it makes sense to accept some true downtime to avoid months-long delays in data center construction. This is clearly true for training workloads which are very important but don’t have live demand. But downtime for even inference clusters is acceptable: you can reduce the compute demand by temporarily slowing down token generation, or use dynamic rate limits. And any curtailment would almost certainly be isolated to one region, so inference data centers in other places would still be operational.
In any case, the paper says the curtailments would last about two hours each:
The average duration of load curtailment (i.e., the length of time the new load is curtailed during curtailment events) would be relatively short, at 1.7 hours when average annual load curtailment is limited to 0.25%, 2.1 hours at a 0.5% limit, and 2.5 hours at a 1.0% limit
I’ve heard a rule of thumb that if you can avoid buying power off the California power grid’s spot pricing in the 1% most expensive times, you can roughly get a 50% discount on power.
Dean Ball’s Out of Thin Air: A proposal for the grid is the most “big if true” thing I’ve read on AI infrastructure recently:
And a convenient win-win:
I do wonder how believable this is, given my personal experience that the more familiar I am with a policy claim’s quantitative modelling and the more domain knowledge I have about it the less I trust it generally speaking...
I came across a similar “hack” on LinkedIn from Tom Styer:
“California just pulled off the nation’s largest-ever test of a virtual power plant.
This successful test proved VPPs are a fast, low-cost, zero-emissions way to make better use of the clean energy we already have — and to keep the lights on as demand surges from data centers, heat waves, and electrification.”
Basically, they are talking about allowing residential batteries supply the grid during peak demand. I tend to be skeptical about stuff like this because in my own domain, water, there’s a lot of pop science and bold claims that ignore scaling and logistics. I asked a smart fellow in that industry about it
https://substack.com/@energycrystals
and I thought he gave a good answer that aligns with my experience with water, which is it always come down to implementation:
”The challenge is lining up customer telemetry to incentive structures that matter. With standard demand response products (which some people sell as VPPs), the incentives given to customers don’t pay for the battery and the software admin and API costs to the utilities outweigh the cost savings of a VPPVPPs are vaporware until someone can make the business model pencil and the API integration and UX not suck ass”
So, without knowing more, my prior is that this free capacity is there for a reason, and that utilities aren’t that dumb. On the flip side, I think it’s great that we are thinking this way. Probing our systems and looking for efficiencies are worthwhile. our legacy infrastructure is a mess of path dependent bureaucracy and I’m certain there’s gains to be made in addition to new construction.
Whether this is feasible depends on how concentrated that 0.25% of the year is (expected to be), because that determines the size of the battery that you’d need to cover the blackout period (which I think would be unacceptable for a lot of AI customers).
If it happens in a single few days then this makes sense, buying 22GWh of batteries for a 1GW dataset is still extremely expensive (2B$ for a 20h system at 100$ / kWh plus installation, maybe too expensive for reliability for a 1GW datacenter I would expect, assuming maybe 10B revenue from the datacenter??). If it’s much less concentrated in time then a smaller battery is needed (100M$ for a 1h system at 100$/kWh), and I expect AI scalers would happily pay this for the reliability of their systems if the revenue from those datacenters
Demand response could be done by covering the data center with battery energy or not. Demand response and batteries can stack: if the grid is really stressed, a data center can both turn off and discharge its battery into the grid.
Economically, it makes sense to accept some true downtime to avoid months-long delays in data center construction. This is clearly true for training workloads which are very important but don’t have live demand. But downtime for even inference clusters is acceptable: you can reduce the compute demand by temporarily slowing down token generation, or use dynamic rate limits. And any curtailment would almost certainly be isolated to one region, so inference data centers in other places would still be operational.
In any case, the paper says the curtailments would last about two hours each:
I’ve heard a rule of thumb that if you can avoid buying power off the California power grid’s spot pricing in the 1% most expensive times, you can roughly get a 50% discount on power.