Amongst the many concerns being raised about AI, one of the most persistent is that there is going to be some firm that “wins” the AI economy. I see that not so much as a concern but an inevitability. When has it ever been the case that some firm didn’t “win” the <insert new technology here> economy?
When you think about it for two seconds, the “real” concern is that the winner will be an existing large firm and that, if we know that, we can stop them from winning. While many people have just presumed that an existing incumbent will win, a new paper by Pierre Azoulay, Josh Krieger and Abhishek Nagaraj does the hard analysis to establish that outcome as a prediction. (Here is a link to the gated NBER working paper. Unfortunately, none of the authors have put out the paper on their websites.)
Here’s their argument. First they note that winning requires two things: (i) that you can actually appropriate the value of an innovation by preventing some form of imitation and (ii) that there isn’t someone else holding a key complementary asset that you need access to. This is a framework put forward decades ago by David Teece although in any respects, I have always thought of it as about one thing: value capture. But I digress (or do I?)
Second, with respect to those two things, they first dismiss appropriability as being a good thing for current AI innovators to lean on. Some Googlers wrote a leaked memo saying that Google didn’t have a moat and neither did OpenAI. And they point to Meta’s open source releases as evidence that these foundation models are unlikely to be excludable. That is, there will be enough providers of these for this not to be a source of power in the market. I agree with that assessment.
But what about complementary assets? If the foundation models are going to be free flowing, they will need stuff to actually run. There’s data (notably training data) and there is what is called “compute” (hardware to do the training and then to run “inference” which actually allows the AI systems to be used). Those are some serious complementary assets and the authors predict that whomever controls them will control the AI economy. And I can hardly fault them for that: after all, the entire point of my book, The Disruption Dilemma, was to note that incumbents weren’t nearly as vulnerable to disruption as people thing precisely because they control complementary assets.
Now the authors stop short of naming the current incumbents as the likely winners but they do note that those firms seem to currently hold all of the cards and that in each case there are early mover advantages to all of this. Put two and two together and you see what is going to happen.
This also leads to some conclusions. One is a discussion about “hope.” Hope that some maverick incumbent will just give everything away. I have to admit that I just don’t buy it. The other is a set of regulatory interventions, many of which are designed to make it easier for AI to be developed by many, but one stood out: fractionalisation of AI infrastructure. I read that to mean regulating or subsidising entry or something like that; and by “read” it is was pretty much what they were saying. And yes such action is the logical consequence from their conclusions.
My issue here is that I am not convinced about their conclusions. Namely, I don’t think the complementary assets they identify will necessarily be the ones that will be “key” to the whole AI economy. In forming this view, I rely on history.
I’m now old enough that this isn’t my first major technology rodeo. Back in the 1990s and 2000s, the Internet (remember that) was taking off. There was a massive amount of regulatory consternation about that. It was clear to many — myself included — that the winners of the Internet economy were going to be the companies that owned the pipes — specifically, the last mile to users. And the pipes were all controlled by telecoms which were large and not particularly competitive incumbents. Thus, there was a massive push to regulate these companies with a view to fractionalising the Internet infrastructure. And that regulatory push was largely successful. But by the time it was, no one saw those telcos as the winners any more.
Instead, who “won” the Internet? The answer, mostly, was Google. Not even Microsoft — the other target for regulatory attention to protect the Internet — featured. Facebook and Apple were winners as well but for other stuff, not the core Internet. The point here is that Google didn’t exist when we were running around protecting the Internet from monopoly.
So why did Google win? The answer is it won by doing the things that all winners of a new technology economy do: they solved the biggest problem with that new technology. In the case of the Internet, it was information overload. The Internet dumped a ton of information into public availability but I even remember back in 1994 struggling to work out where the good and relevant stuff was. Lots of others tried to organise the Internet but Google really solved it. Without Google Search, the Internet was barely usable as an information source.
None of the current identified complementary assets solve the biggest problem with AI. And what is that problem? Buggered if I know. If I did, I wouldn’t be writing a frigging substack that’s for sure. There are plenty of contenders. There’s hallucinations, safety, AIs talking to one another, etc. All of those are problems to be solved. But which one is THE problem? History tells us that firm will win the AI economy. History also tells us that right now, it is not obvious what the winning problem, let alone the solution, is.
This doesn’t rule out one of the current players winning the AI economy. But history tells us they are unlikely to be the winners either. There is something about critiquing your own investments and rationales that makes it hard for them to see the real problem let alone implement the solution. Remember even the firms like Yahoo and LookSmart who were dealing with information overload, didn’t understand that their role was to help people find information and then let them leave their website. The right answer conflicted with the money that was flowing in.
It may well be a good idea to pursue the policy recommendations based on current complementary assets being important regardless. Just don’t expect it to mean that there won’t be a winner. It is a pretty sure bet there’ll be one.