At least this time the forecasts for the transformative effects of AI don't seem overblown - so after (in my experience) forty years of AI being the next great thing, it at least seems real. But the question remains of how much the success of AI will really translate into data centre demand.
First a history lesson. At the start of this millenium, global stockmarkets were greeted by the dot.bomb crash. This crash has been well documented in Dot.bomb and Dot.bomb Australia. What's less well known is that a lot of it could have been avoided.
Everyone's favourite statistic from that era was that the capaciy of the internet (then still spelt with a capital 'I') was doubling every hundred days. As was well documented in at least 2002, this 'statistic' was entirely based on a 1998 claim by WorldCom. However, that it was both not true and, indeed, could not be true, was detailed by Andrew Odlyzko and a co-author in a 1998 article The Size and Growth Rate of the Internet. (Odlyzko wrote some other fascinating pieces, among them Content is Not King and Metcalfe's Law Is Wrong, both are worth reading by anyone in the strategy or analysis business).
How could the myth have been believed? Firstly, any network operator who didn't see this growth reflected in their own network traffic reasoned that they were marketing their services wrongly or some other cause meant they were missing out on the growth. Cetainly, every stockmarket analyst was quoting it, and I have observed first hand CEOs whose choice of strategy has been the one that satisfied analysts to put a buy recommendation on the stock rather than the one suggested by all the detailed analysis inside the company.
Secondly the forecast fueled new entrants to the market who were simply riding the market sentiment, not making detailed analysis. It was a traditional bubble. Nobody could afford not to climb on board while the growth phase was on.
Just like the the dark fibre boom, the forecasts of growth in computing capacity cannot be sustainable. Other parallels between the dark fibre and compute booms have been detailed in a piece that is in part written by AI! It is a question that has even been considered by Fortune magazine, and the founder of Open AI even says the AI bubble is about to pop.
By its nature, AI infrastructure (including both the computing requirements and the AI engine itself) has the same economies of scale and demand side economies of scale as other digital platforms. The latter scale economies work exactly like network effects - and even agreeing with Odlyzko that Metcalfe was wrong and the value of a network doesn't increase at the rate of n*(n-1) where n is number of the number of users but only n*log(n) - the effect is significant. That is why these markets tend to monoplies.
An AI platform is continuing to train while it is being used - high usage therefore leads to better capability. In brief, the computing requirements for AI will be nowhere near as large as they seem because there will be less, more efficient, AI platforms.
Students of the history of electricity markets in Australia* will recall that the impetus for market reform in the 1990s was over-investment in generation in the 1980s that was driven by the Fraser Government's desire/plans for an extended mining boom that did not eventuate. It would truly be a shame to replay two historical errors!
As to the 'ban' on nuclear energy, the title suggests that it was at some point a carefully worked out policy rather than just another case of Howard Government misjudgement**. It was never Howard's intention to embed a ban - he was merely outlawing something that no one had ever seriously suggested they wanted to invest in in Australia. And no private capital has wanted to invest in it since.
To create a possibility of such investment requires a heck of a lot more policy work than simply removing the legislative ban. For the new Opposition Energy shadow to make a simple ban removal statement reflects one of the two the stunning consequences of the leadership change in the Liberal Party - on almost every policy position they have moved towards all the policy elements that the suppressed party review said were errors. (The second is that the new leader is virtually invisible - a huge contrast to Sussan Ley who seemed to occupy more media space than the PM).
Maybe, just maybe, the prodigous power of AI might be useful to solve the other technology promise that has underdelivered for just as long as AI - nuclear fusion reactors.
* My chapter in On the Grid provides a useful short history.
** Anyone left with any delusions that Howard was a good Prime Minister or that the country benefited from his Government should read Amy Remikis' Where It All Went Wrong.
*********************************
Life is what happens while you are busy making other plans JWL
No comments:
Post a Comment