there's a $3,700 GPU with more VRAM than a $4,700 RTX 5090 — and almost nobody running local AI has heard of it

An RTX 5090 is currently running $4,300-4,700 on the street for 32GB of VRAM. A Radeon PRO W7900 — a real, current, buyable workstation card — is going for around $3,700-3,999 new, and it has 48GB. More memory, lower price, and almost nobody shopping for local AI hardware seems to know it exists. That inversion is the most useful thing I found pricing out every rung of the local-AI hardware ladder this week, from a $4,300 gaming flagship to a $16,000 professional card.

The four rungs, priced today

Tier 1 is the obvious default because it's also a genuinely great gaming card — if this machine needs to double as your gaming rig, the RTX 5090 earns its premium. As a pure local-AI purchase, though, you're paying roughly $147 per gigabyte of VRAM for the privilege, and CUDA compatibility is the only thing you're buying beyond raw memory.

The tier that breaks the pattern

The Radeon PRO W7900 is where this ladder stops behaving the way you'd expect. It's a current-generation, ECC-capable, 48GB professional card, and as of this week it's selling new for less than an RTX 5090 while offering 16GB more memory. The tradeoff is real: it's not a gaming card in any meaningful sense, and AMD's ROCm software stack still lags CUDA on out-of-the-box compatibility for a lot of local inference tooling — a gap I've measured directly on the AMD-vs-Nvidia side before. But for anyone running Ollama, vLLM, or llama.cpp against open-weight models and not gaming on the same box, that software friction is a one-time setup cost, not a permanent tax.

Skip the $16,000 card unless someone else is paying

The RTX PRO 6000 Blackwell is the clearest example of a tier that's priced itself out of home use. Tom's Hardware confirmed it launched at $8,565 in March 2025; by June 2026, Nvidia's own marketplace had it at $13,250; by this August, it's listed at $16,000. Nvidia points to GDDR7 supply constraints, since the card's 96GB clamshell design uses more of that memory than almost anything else on the market. Whatever the cause, the number now speaks for itself, and I've made the case in detail for why it's a hard sell outside enterprise budgets.

If 96GB of memory is genuinely the requirement, unified-memory systems get you there for a fraction of the cost, just with lower bandwidth. The NVIDIA DGX Spark ships 128GB of LPDDR5X for $4,699, having risen from $3,999 in February citing the same memory supply pressure. Ryzen AI Max+ 395 mini PCs hit the same 128GB figure for less — Framework's desktop starts at $3,449, GMKtec's EVO-X2 at $3,649.99 — on Windows or Linux rather than Nvidia's DGX OS. And on the Apple side, the newly announced Mac Studio M5 Ultra starts at $5,499 for 96GB, with the jump to 256GB adding $4,000 (about $9,499 total), confirmed directly in Apple's own launch coverage. A 512GB configuration is coming in October, pricing still unannounced, with the current top configuration already at $18,299.

My actual pick

None of these tiers are wrong, exactly — they're just priced for different jobs. If you're still narrowing down a specific machine rather than a whole category, our broader local-AI desktop leaderboard and the prebuilt-vs-DIY breakdown are the next two things worth reading before you spend anything.