Opinion · from the person who rented all of it · Updated July 2026
Here's the thesis this whole site's buying advice runs on: every product in capitalism gets commoditized over time, and compute is a product. Gaming compute already went through it. AI compute is still scarce, limited VRAM, limited power, real waiting lists, but Blackwell is the beginning of the same curve. If that's true, it changes what you should buy, what you should rent, and what you should refuse to pay for in 2026.
Gaming already got commoditized. Look at the numbers. A 20-series, 30-series or 40-series card games just fine. That's not an opinion, it's our benchmark tables. The 40-to-50-series gaming jump is going from about 110 FPS to about 130: you'll see it, but it's nothing like the old 30-to-60 upgrades that actually changed how games felt. When every tier of a product does the job and upgrades stop feeling like events, that's what a commoditized market looks like. Which tells you what the 50 series is really for. If you're buying Blackwell, you're buying it for AI, the inference gains over the 40 series are the real generational story. The gaming improvements are the footnote.
AI compute is still scarce, and that's exactly why the pricing is weird right now. We're in a compute-scarce environment: VRAM is the binding constraint, power is the other one, and the frontier models keep growing faster than consumer cards do. Scarcity is why a used RTX 3090 that sold for $700 now pushes $1,100, an aging architecture appreciating because of what it can hold, not what it can compute. And it's why you can't fix an old card by imagining more VRAM onto it: a 30-series card with 96GB would still lack the compute to run a frontier model at usable speed. Capacity decides what loads; compute decides whether you'd want to use it. But scarcity produces strange, temporary discounts too. The cloud GPU providers are pricing rentals below what the hardware economics suggest, because they're fighting for the developer market, at the exact moment the AI subscription services are tightening token limits on consumers. Free credits, ~$1/hour Blackwell, a 288GB B300 for about $7/hour. There's a circular-financing flywheel behind it, and for once the end user is the beneficiary. That window won't stay open forever.
The 2026 vs 2030 argument. The compute you can access in 2026 will not resemble what you can access in 2030, and it should be dramatically cheaper by then. That's the whole case against big silicon purchases right now: buy $50,000 of workstation cards today and you own a depreciating slice of a curve that's bending against you. Put a fraction of that into rentals and you ride the curve instead, every hardware generation the providers buy becomes your upgrade, automatically, with the old stuff discounted beneath it. So the ownership question stops being about specs and becomes about utilization: you'd have to be running AI agents at a genuinely extreme, always-on level before owning beats renting, the point where the meter simply never stops. Short of that, ownership in 2026 is paying a premium to hold depreciation.
What this means practically. Buy the commoditized thing, rent the scarce thing. Concretely, from our own bench: buy a 4080 16GB-class card for gaming plus light AI, best value per dollar of anything I measured, or a 4060 Ti 16GB on a budget, a used 3090 if the price is right for its 24GB, or a 4090 if you find one at sane money (it's the last consumer flagship whose ownership math still works). Rent everything above that: the 5090-and-up tier, the workstation cards, all the datacenter silicon. Have an objective before you spin up serious horsepower, because rented compute rewards a plan. And keep one eye on the curve. Blackwell is the start of commoditizing AI compute the way Pascal-through-Ada commoditized gaming. When that finishes, when every tier runs your models fine and upgrades stop feeling like events, the buy math flips back. We benchmark the whole stack continuously, so this site will say so when it does.
Compute is a product, products commoditize, and AI compute is mid-curve while gaming compute is post-curve. Rent the scarce tier while providers are subsidizing it, buy only the commoditized tier, and don't lock five figures into 2026 silicon that 2030 will embarrass. That's the thesis, and every buying recommendation on this site is downstream of it.