Nov 6, 2026, 7:01 AMIn my personal opinion, the rightsizing has begun. LLMs won't go away. But they will become smaller and more focused (domain expert LLMs), reducing training and operation costs significantly. This will have a direct impact on demand for compute power. The bubble won't burst, it will fizzle out.
Nov 6, 2026, 7:05 AM@jwildeboer I don't think that's right. Demand for general purpose LLMs will keep growing, not shrink. If the US can't cover it, China will, and certainly not for Germany, so for us this only speeds up our decline.
Nov 6, 2026, 7:08 AM@andre The problem is that they have to make money with those oversized LLMs. Specialised (and thus much smaller and cheaper to train) models are far better positioned to reach a return on investment. The business world is recognising this and are switching from "renting" big models to running their own "good enough" open-weight models that they specialise for their needs.
Nov 6, 2026, 7:11 AM@jwildeboer I think that's an illusion. In theory specialised models are better, sure. In practice you'd have to keep retraining them every time the requirements change, and they change all the time. A big general model just handles the new task, and that's why people keep paying for it.
But will it have the same level of quality? (Honest Question).
I think "non-power users" should be aware that you do not need that 'Astra 6 Cyber Deep Reasoning Self Improving Agentic Coding' model for proofreading or suggestions for your short email. And the media could play their role as well: In my opinion the focus is still to much about size and huge specialised tasks and less about efficiency for specific smaller everyday tasks.
Nov 6, 2026, 7:43 AM@dexternemrod In quite some cases the specialised (or limited, depending on your viewpoint) models deliver better results, as they don't suffer from interference by adjacent knowledge domains. @andre
Nov 6, 2026, 8:03 AM@jwildeboer Maybe for some use cases, I'll gladly believe that. For others, it's just not true.
A small support bot does fine with the usual product questions, but as soon as a customer mixes a billing issue, a contract question, and a bug in one message, it falls flat, while a big model just deals with it.
What's more important for me: can you name some use cases from the praxis? Not research papers, real use cases.
My thoughts went in this broad direction.
Nov 6, 2026, 7:11 AM@andre Another very important reason for that movement is compliance. Many companies, especially here in the EU, fear the legal risks of having to share their confidential information with a handful of "big" LLM providers that are mostly in the US.
Nov 6, 2026, 7:17 AM@jwildeboer I think the compliance problem is getting solved from the other side. Anthropic just announced they'll offer their models on European servers together with Microsoft. Once the data stays in the EU, lots of companies will just keep renting the big models instead of building their own.
Nov 6, 2026, 7:20 AM@andre I'd say we meet again in 6 months and compare results. I don't think a global US (or Chinese) oligopoly of exclusive access to big LLMs will emerge. The technology doesn't lend itself to paywalls, especially not in the corporate environments I talk to. I foresee a swarm approach of domain specific LLMs being combined to deliver results. And most of these "expert" LLMs will run on-prem at companies.
Nov 6, 2026, 7:32 AM@jwildeboer Deal, let's do that, and in the meantime could you share one or two examples of specialized small LLMs from your own work? If you can't, maybe that tells us something too.
Nov 6, 2026, 7:12 AM@jwildeboer Maybe ROI is a problem for the US labs, but China doesn't care about quick returns. They're investing huge amounts in this for strategic reasons. If the US drops out, China takes over, and we won't be the ones getting their good models.