"But ultimately our customers, our partners don't want to buy computers, they want to build AI factories. All of the CPUs until now were created for people. We were the users, we were the users. We were the renters. This is CPU for the agents (AI/LLMs). All the CPUs in the past we built for humans. This CPU is built for agents. And these agents are going to be using the CPUs with very little patience. There'll be a lot more agents then there are people. And then the agents are very impatient."
- Jensen Huang; CEO of NVIDIA, Computex or GTC 2026.
They don't want us to be in control of our own compute anymore because we don't have to pay them subscriptions and they cant dictate what we see and do, they also know that llms can essentially run on anything and it's only a matter if time til hardware gets really good at it and it becomes very viable to run open source models at home that make theirs irrelevant.
I have a small netbook with 64GB of RAM and a Radeon 780M. I can use 55GB of the RAM for running local models (since it's shared VRAM I can put all of that into the GPU and even combine it with an external one). So I've tried qwen3.6 and it's just as useable (and often shitty) as Gemini or ChatGPT, but without the instant response. Depending on the task, I have to wait from thirty seconds to five minutes for the reply.
That was pretty eye-opening for me. If the big players didn't buy up all of them RAM and make GPUs with high VRAM extremely expensive, all of us would already be able to run these models by ourselves.
I'm really waiting for the falllout after the crash, when those Blackwell GPUs will land on eBay for fifty bucks and we can get adapters on Aliexpress to use them at home. It will bring computing back into our own hands and out of the corporations emptying our pockets with subscriptions. And hopefully, it will open up discussions about the actual capabilities of LLMs, because what we're having to endure right now in the media is not based in reality at all.