AMD and Intel are dipping their toes into AI at your expense
TL;DR: I bought AI-branded workstation cards from both Intel and AMD this year. Intel archived its own AI software while I owned the card. AMD shipped a flagship model its own "AI PRO" card can't run, then announced an architecture change that strands it. Neither company staffs the software for the cards they actually sell you, and the perverse result is that NVIDIA, a company nobody would call generous, is the only vendor whose product respects your purchase. If Intel and AMD won't support these cards, the least they owe us is the firmware, so we can support ourselves.
I run ModDog, an AI Discord moderation bot, on a self-hosted GPU fleet. This year I gave both of NVIDIA's competitors real money on the strength of their AI marketing: an Intel Arc Pro B70, and two AMD Radeon AI PRO R9700s. I wanted them to be good. GPU prices are obscene, VRAM-per-dollar on both cards embarrasses NVIDIA, and I liked the idea of not feeding the monopoly.
Here's what that goodwill bought me.
Intel: the card outlived the software
The short version of my B70 postmortem: under sustained inference load the card crashes with engine resets that survive every driver update, firmware update, workaround flag, and serving stack I could throw at it. The evidence points at the closed firmware. Reports are filed upstream. Silence.
But the part that belongs in this article is what happened to Intel's software while the card was in my rack. In January 2026, Intel archived ipex-llm, their LLM library. In March, they archived intel-extension-for-pytorch, their PyTorch backend. That's a workstation card marketed for AI whose vendor shut down two of its own AI software projects within my first months of ownership, leaving one upstream path that crashes the card.
I didn't buy a lemon. I bought a product line the company was quietly walking away from while the marketing was still running.
AMD: the card their own model can't run
AMD sells the R9700 under the literal brand "AI PRO." Last week AMD released Instella-MoE, their flagship open language model, trained end to end on AMD hardware. A genuinely impressive release. I downloaded it the day I heard about it, because running AMD's own model on AMD's own AI card is exactly the workflow this product was supposedly built for.
It runs at 5.6 tokens per second. Not because the card is slow, but because no fast runtime on earth can execute this model on this card. AMD's recommended inference stack for their own model does not support the architecture family they sell to workstation buyers. The kernels that make it fast exist only for their datacenter GPUs, the ones that cost more than my car. On my card, the model that AMD's marketing celebrates falls back to the slowest possible execution path, the one every GPU on earth supports out of politeness.
FP8 is broken on this card too. And the fix on the horizon makes it worse, not better: AMD's announced UDNA architecture merges their gaming and datacenter designs into one, which means every future kernel gets written for silicon I don't own. Unification is great news for whoever buys the next card. For the people who bought this one, it's an orphaning notice with a press release.
The toe-dip pattern
Put the two stories side by side and the shape is identical. Both companies want the AI story: the earnings-call slide, the "Pro" and "AI" branding, the headline about competing with NVIDIA. Both companies priced their cards aggressively to get that story. And both companies declined to fund the unglamorous part, the years of kernel engineering and driver support that make a GPU purchase worth anything, for the specific cards they sold to people like me.
That's what dipping a toe in looks like from the buyer's side. The vendor risks a marketing budget. You risk a thousand dollars. When the water turns out to be cold, the vendor pulls the toe back, archives a repo, pivots an architecture, and you're standing there holding the hardware.
The uncomfortable NVIDIA part
I want to be clear about what I am not saying. NVIDIA's pricing is hostile. The VRAM segmentation is cynical. The market position they hold is bad for everyone, including me. This is not praise.
But here's the fact my rack keeps rubbing my face in: the cheapest, oldest card I own, a used $500 RTX 3080 Ti, two generations old, runs every model that drops. Day one. Whatever lands on Hugging Face this morning, in whatever new architecture some lab invented, gets CUDA kernels first, usually before the paper is out. That card will keep doing this for years, because NVIDIA has kept decade-old GPUs running new software as a matter of routine.
That's not ethics. Nobody at NVIDIA loves you. It's incentives: CUDA compatibility is the moat, so supporting your old card is their business model. AMD and Intel's business model is selling you the next card. The irony I can't get past is that the most ruthless company in the market is the only one whose incentives point in the same direction as mine, and that alignment does more for me than either competitor's goodwill ever did.
"Ethical by comparison" is a hell of a thing to say about a near-monopoly. The comparison is what's damning.
The one honorable exit
If Intel and AMD aren't serious about supporting these cards, there's still one ethical move on the table: open them up and let us support ourselves.
Both companies love their open-source credentials, and parts of the stack genuinely are open. That's exactly how I know where the bodies are buried. My B70 investigation upgraded the open kernel driver: no change. Upgraded the open compute runtime: no change. Every open layer got ruled out one by one, and the evidence funneled down into the closed GuC firmware blob, the single component I'm not allowed to see, patch, or even read the errata for. The community cannot fix this card. Not for lack of skill, the Linux GPU community has plenty, but because the one layer where the bug lives is a sealed box. AMD's situation rhymes: ROCm is open, the kernel driver is open, and the firmware and the staffing decisions about which silicon gets kernels are not things a contributor can pull-request.
The current arrangement is the worst of both worlds: closed enough that owners can't fix it, unsupported enough that the vendor won't. Pick one. If you're going to walk away from a product people paid a thousand dollars for, then publish the firmware, the documentation, and the errata on your way out, and let the people who bought the silicon finish the job you didn't. There's precedent for hardware living a long second life the moment the vendor gets out of the way; every homelab full of decade-old enterprise gear proves it. Abandonware with sealed firmware isn't a product. It's e-waste with a marketing budget.
What I'd tell a buyer
One rule, learned twice at retail price: buy the software that exists today, not the roadmap. If the runtime you need doesn't run on the card right now, it does not exist. If the vendor's support history for previous generations is bad, that history is the roadmap. Marketing that says "AI" on a box tells you what the company wants investors to believe, not what its engineers are staffed to deliver.
My AMD cards will keep serving the mainstream models they serve well, and I'll run them until the wheels fall off. They owe me nothing at this point; they've earned their keep on duty that works. But the next dollar I spend on AI hardware goes to the company I trust the least and rely on the most, because in this market, that's somehow the same purchase.