Creator Prakash agrees that winner-take-all competitions will boost frontier model demand.
Users are reacting to the claim that AI value flows mainly to chips and implementation rather than models, with positive users affirming the thesis while negative users worry non-frontier strategies face losing dynamics and Coasean risks.
No Digg Deeper questions have been answered for this story yet.
@lugaricano Agree! I wrote the same 3 years ago - should have made some investments on this basis, hah! That said, all 3 leading labs are making bets either on downstream complements or on RSI giving a sufficient lead, so my read is that folks do get this.
I'm generally in favour of adoption-focused strategy, but a little worried about fully betting on this thesis for two reasons:
first, it seems like the divergence between labs and fast followers is widening due to structural forces that will get much stronger. fast-following was decently easy when no one had big AI datacenters online, flywheels hadn't started, and distillation and research was still a wild west. that means you could build a fairly efficient fast follower, and neither the capability gap nor the efficiency gap matters much.
but all of that is changing - the asymmetries in compute between leading US developers and the world are just starting to kick in, and they are baked into at least the next four years of global compute trajectories because of the lead times on buildouts and chip procurement. crackdowns on distillation, as well as the fact that leading developers get much more revenue to reinvest into talent and compute and much more data to further widen the gap with makes it much worse. I wouldn't be surprised if the gap to fast followers extends to 1.5 years in an otherwise accelerating environment, and I also wouldn't be surprised if that gap also meant that the frontier labs will be able to offer much more efficient sub-frontier models than any pure fast follower competitor.
and second, I'm not sure that 'how many tasks can you do with non-frontier models' is the relevant question, it's 'how prohibitively important are the tasks for which you do need frontier models because you're engaged in some directly competitive dynamic, e.g. between attackers and defenders or directly competing wrappers around frontier systems'.
to the extent that the latter become economically vital and the ability to run them remains scarce, you could just imagine a lot of value (and especially leverage) concentration with those that can provide comparatively leading models even in the world where most tasks are carried out by sub-frontier models, because value asymmetrically accrues around the tasks that do require access to comparatively-frontier instead of absolutely-good models. that's definitely a contentious prediction, but it seems devastating to this strategy if true.
@lugaricano These results from CAISI seem to be a narrative violation to the idea that open models are at a fixed lag behind the frontier - the gap seems to be increasing. https://www.nist.gov/news-events/news/2026/05/caisi-evaluation-deepseek-v4-pro
@AlexTPet Thanks Alex. Agreed, more GPT-5 than GPT-5.5, so eight months behind rather than six. But our argument only needs a tier or two back, good enough to keep the labs from rationing or implicitly colluding on price.
@lugaricano As a matter in fact one of the reasons behind many OpenAI moving to Anthropic over the past few months isn’t that their latest model is better—in fact, it isn’t—but rather how they’re shaping the “harness,” the wrapper that makes the model far more capable to apply in practice.
@lugaricano which layer captures value then?
@lugaricano I unfortunately think that deciding to go for non-frontier models is a losing proposition. Empirically, the practical usefulness of increased AI capabilities seems massive. AI is not about to get commoditised. I think Anton Leicht has better suggestions https://writing.antonleicht.me/t/geopolitics.
@lugaricano makes sense! I do wonder what is driving the gap though, and whether it will continue to grow.
@AlexTPet @lugaricano Anton Leicht believes import agreements (https://writing.antonleicht.me/p/import-imperatives) and datacenters in non-US soil (https://writing.antonleicht.me/p/can-you-poach-a-frontier-lab) are what will provide geopolitical leverage and prevent being cut off (https://writing.antonleicht.me/p/cut-off). But he believes access to frontier AI is necessary.
@Noahpinion defense already proved this - capability gap of one generation ends the fight before it starts
@Pablomorecasa @lugaricano I would agree with Anton that access to the frontier will still matter one some things that the open pre-frontier won't be able to compete on. Things like R&D, military capabilities.
@AlexTPet @lugaricano I am worried about betting on non-frontier models. It's good to have bargaining chips in open-weight models, but the returns to marginal capabilities seem very significant. It'd be good to see a more concrete formalisation of this question, but my intuition is that it won't work.
@AlexTPet @lugaricano Separately, it is non-obvious to me that Coasean dynamics won't kill a strategy based on diffusing AI in the economy. Model providers could build mostly automated companies if they achieve sufficient capabilities.
@Noahpinion yup, every winner take all game will require frontier models
@CaesarIpse Completely agree.
@Afinetheorem Indeed!!
@Noahpinion You are right that in adversarial contexts the best model wins and takes all. The question is what share of the economy looks like this.
@stalmico I would say hardware layer and implementation. We explain it in the piece! I hope you will like it!
@Noahpinion The very best model isn’t essential, but it does fairly easily pass a cost/benefit analysis. Even if gains are marginal, until cost is more than a small fraction of the cost of the time of the human interacting with it, it’s just so obviously better to use the better model.
@Noahpinion I don't think that's the case; it's certainly not true in high frequency trading. If we look at the HFT market, not every participant firm has class-A models, short-run infrastructure, the fastest switches/hardware, etc.
There's a gradient. And there will be here, too.