As ever, what you say makes a lot of sense. But we should never forgot that the falls in price of solar and batteries over time are not automatic, but rather the result of experience - experience of manufacturing above all, and scale. If no-one installs batteries, that falling cost curve will not happen. The price of batteries in 2035 is endogenous to the question of "should we install now or later" - at least at a world scale. The real q for the UK, I think, is "are there useful technologies out there that would be materially helped by the UK adopting early?" Enough of the world is installing batteries (notably in cars) that we don't matter. But we could well matter for tidal stream or direct air capture, or sustainable aviation fuel.
Of course, this is why you are an economist and I am not! But I certainly don't disagree with that - the "when to install batteries" was probably more of a thought experiment than a firm policy recommendation anyway.
Excellent piece. Your bit on academic immigration makes me also think of the the fact that the world is at or around "peak children", there will be active competition for people between countries in the next few decades and that is one competition that the UK could actually win.
I believe that we may have even just reached peak children - I'm sure I saw a tweet from one of the Our World In Data folks last week, but I can't find it now
Interestingly (well, depending on mathematical taste), according to this model, the optimal time to buy the batteries is always –1/log(0.8) = 4.5 years before the deadline – regardless of when that deadline is.
So if you want the most benefit by 2030, you should be buying in July this year, 4.5 years before the deadline; For a 2035 target, that's July 2030, 4.5 years before that deadline. For a 2100 target, you'd wait all the way until July 2095, again 4.5 year's before the deadline, by which time (according to this model) the price of batteries will have gone down by roughly a factor of 7 million from now.
That's neat - of course, there's an issue that we might not be able to reasonably be able to extrapolate prices that far because we'd run up against some of the fundamental laws of physics at some stage, but it's probably no sillier than my answer!
A (maybe the?) big uncertainty is the energy cost of computation. Historically, the cost of one computation has gone down with Moore’s law but people have speculated/estimated that Moore’s law is winding down. Moreover, the complexity of ML models seems to be growing exponentially. OTOH, perhaps ML models will (continue?) to improve and, the energy cost of an inference will decline, perhaps exponentially? This reminds me of McEliece’s 2004 Shannon lecture that compared the impact of Newton (via more energy) vs Shannon (better coding) on the increase in data rates of space communication. IIRC, McEliece concluded Newton and Shannon contributed roughly equally.
Nice, I should dig that out and take a look - seems interesting and very relevant. And yes I agree there's all kinds of tradeoffs (if you have more compute resources you tend to be able to find more stuff to do with it) - and of course if this DeepSeek business that everyone is talking about checks out then maybe this piece is out of date already!
Posted it on BlueSky, which currently feels like early Twitter. I worry that we aren't building enough replacement nuclear power stations for base load, here in the UK. Dull, windless days are surprisingly common!
Energy is life. Success societies have exponentially rising demand for energy and technological improvements lead to more demand not less. Solar is good for sunny countries. The rest need nuclear and lots of it. Not everyone is acknowledging this yet but to quote Doomberg when platitudes meet physics there can only be one winner. Change will come as no politician ultimately wants to be responsible when the power cuts happen, the only question is how difficult and expensive the path is to get there. Make the wrong choices and you end up like Germany having to burn the dirtiest type of coal to replace your carbon free nuclear or have the lights go out.
A key point to consider for datacentres is that many companies are building their own energy sources, rather than relying on the main national grid. This is sometimes called “behind the meter” power
On a ten-year view, you should consider the probability of small modular nuclear facilities onsite at many of the largest DC campuses.
Obviously depends on certification, planning rules etc, but expect the US to lead on this over the next couple of years
Yes that's true - and even the whole "recommissioning Three Mile Island to run a data centre from" thing, which is a scale above that! But yes, I don't think it's impossible we can meet the extra demand (at least in the medium term), it's just surprising to me that the projection doesn't even really seem to think that we need to!
As ever, what you say makes a lot of sense. But we should never forgot that the falls in price of solar and batteries over time are not automatic, but rather the result of experience - experience of manufacturing above all, and scale. If no-one installs batteries, that falling cost curve will not happen. The price of batteries in 2035 is endogenous to the question of "should we install now or later" - at least at a world scale. The real q for the UK, I think, is "are there useful technologies out there that would be materially helped by the UK adopting early?" Enough of the world is installing batteries (notably in cars) that we don't matter. But we could well matter for tidal stream or direct air capture, or sustainable aviation fuel.
Of course, this is why you are an economist and I am not! But I certainly don't disagree with that - the "when to install batteries" was probably more of a thought experiment than a firm policy recommendation anyway.
Excellent piece. Your bit on academic immigration makes me also think of the the fact that the world is at or around "peak children", there will be active competition for people between countries in the next few decades and that is one competition that the UK could actually win.
I believe that we may have even just reached peak children - I'm sure I saw a tweet from one of the Our World In Data folks last week, but I can't find it now
This is very good - the problem that getting your head around exponentials is hard is definitely relevant here
Thanks!
Interestingly (well, depending on mathematical taste), according to this model, the optimal time to buy the batteries is always –1/log(0.8) = 4.5 years before the deadline – regardless of when that deadline is.
So if you want the most benefit by 2030, you should be buying in July this year, 4.5 years before the deadline; For a 2035 target, that's July 2030, 4.5 years before that deadline. For a 2100 target, you'd wait all the way until July 2095, again 4.5 year's before the deadline, by which time (according to this model) the price of batteries will have gone down by roughly a factor of 7 million from now.
That's neat - of course, there's an issue that we might not be able to reasonably be able to extrapolate prices that far because we'd run up against some of the fundamental laws of physics at some stage, but it's probably no sillier than my answer!
A (maybe the?) big uncertainty is the energy cost of computation. Historically, the cost of one computation has gone down with Moore’s law but people have speculated/estimated that Moore’s law is winding down. Moreover, the complexity of ML models seems to be growing exponentially. OTOH, perhaps ML models will (continue?) to improve and, the energy cost of an inference will decline, perhaps exponentially? This reminds me of McEliece’s 2004 Shannon lecture that compared the impact of Newton (via more energy) vs Shannon (better coding) on the increase in data rates of space communication. IIRC, McEliece concluded Newton and Shannon contributed roughly equally.
Nice, I should dig that out and take a look - seems interesting and very relevant. And yes I agree there's all kinds of tradeoffs (if you have more compute resources you tend to be able to find more stuff to do with it) - and of course if this DeepSeek business that everyone is talking about checks out then maybe this piece is out of date already!
Posted it on BlueSky, which currently feels like early Twitter. I worry that we aren't building enough replacement nuclear power stations for base load, here in the UK. Dull, windless days are surprisingly common!
Not sure if "like early Twitter" is a good or a bad thing TBH. But yes, totally agree on nuclear
Energy is life. Success societies have exponentially rising demand for energy and technological improvements lead to more demand not less. Solar is good for sunny countries. The rest need nuclear and lots of it. Not everyone is acknowledging this yet but to quote Doomberg when platitudes meet physics there can only be one winner. Change will come as no politician ultimately wants to be responsible when the power cuts happen, the only question is how difficult and expensive the path is to get there. Make the wrong choices and you end up like Germany having to burn the dirtiest type of coal to replace your carbon free nuclear or have the lights go out.
A key point to consider for datacentres is that many companies are building their own energy sources, rather than relying on the main national grid. This is sometimes called “behind the meter” power
On a ten-year view, you should consider the probability of small modular nuclear facilities onsite at many of the largest DC campuses.
Obviously depends on certification, planning rules etc, but expect the US to lead on this over the next couple of years
Yes that's true - and even the whole "recommissioning Three Mile Island to run a data centre from" thing, which is a scale above that! But yes, I don't think it's impossible we can meet the extra demand (at least in the medium term), it's just surprising to me that the projection doesn't even really seem to think that we need to!