Listings checked on October 5, 2026.
Running a model through an API for every routine task was starting to feel expensive, so I began looking at a local setup. First question: which GPU could I actually buy for it?
I started with 24 GB of VRAM as a shopping filter. More room for model weights and context was the appeal. The exact model, quantization, and context length would still determine what fits, but this gave me a concrete place to start: an RTX 3090 with 24 GB.
I could have asked an LLM for “the best GPU for local AI” and received a very confident top-five list. Useful for learning the specs, less useful when the card in that list is sold out, the price belongs to a different seller, or nobody can tell you exactly which board you're getting.
At 2Captcha, I used our MCP for web search and structured extraction from product pages, then checked the promising page myself. I wanted to get from a search query to an offer I could verify. Fewer tabs and, ideally, fewer expensive guesses.
A surprisingly unhelpful five results
I searched for a 24 GB RTX 3090 in the US with search_web. It returned five links. The first was RTX Corporation. As in aerospace and defense. Another was Raytheon. One was an RTX stock quote. NVIDIA's general RTX page made an appearance too.
Not a single product listing.
That was a pretty good reminder that five search results are not five offers. An LLM will happily compare snippets if you ask it to. The comparison may even look convincing. It still won't tell you whether a specific GPU is available to buy.
So I used search for reconnaissance and moved to direct retailer pages for the actual checks. I chose those URLs myself and passed them to parse_marketplace through the same MCP. The fields I cared about were deliberately boring: exact model, price, currency, stock status, condition, and the source URL. Boring is good when you're about to spend four figures.
From search results to a shortlist
The Walmart listing for a refurbished ASUS RTX 3090 had an attractive-looking $1,139.05 price. The structured result also said out of stock. That's not a deal, that's a number attached to an unavailable card.
On Newegg, MCP surfaced a refurbished Founders Edition with a price of 2,395. I confirmed the $2,395 price on the retailer page. Some pages returned fewer fields or restricted access, so I kept comparing candidates.
The listing I kept for a closer look was a refurbished MSI RTX 3090 Gaming X Trio 24 GB at GotoDirect. Its MCP result included part number G3090GXT24 in the title, plus $1,654.74, USD, and InStock. I opened the retailer page on October 5 and saw the same part number, price, stock status, and refurbished condition.
There's a catch visible right on that page: “Image may differ from original.” So a matching product feed still isn't a substitute for checking the actual unit, warranty, returns, and final delivered price. I wouldn't buy a refurbished GPU on the strength of a stock photo and a green IN STOCK label alone. I'm trying to run a local LLM, not start a small collection of expensive surprises 🙃
A part number beats a top-five list
The MSI listing made my shortlist because it matched the 24 GB filter and gave me an exact part number, price, and stock status I could check against the store. At $1,654.74 refurbished, price and warranty would still need scrutiny before checkout. What I had settled was which offer deserved that scrutiny: G3090GXT24, with a source behind every field.
The first search took me to aerospace pages. The second pass led to G3090GXT24. I'll take the part number.
Top comments (0)