Zhang Pinpin · Hands-On AI Coach
← Back to home

Don't Blow Your Budget on a GPU—Top-Tier AI Tools Don't Need a Top-Tier Machine

Long-Form Video · EP0058 July 12, 2026 6:40
What this episode covers

Top-tier AI tools require a maxed-out machine with a flagship GPU? That's the misconception I got asked about most this week, and the one I most want to kill. I'm a hardware guy myself—I've bought and used a 3090, a 4090, all kinds of flagships, and I sold them all recently, because for personal AI work that GPU is close to wasted.

Large-model inference happens in the cloud. What your local machine needs isn't a GPU, it's RAM and CPU cores: they set how many windows you can keep open and how many sub-agents run in the background. This episode covers the buying logic, three machines, the worst traps, and the minimum VPS spec for running CC in the cloud.

It’s been a while since I posted, and a lot of new people added me this week. I was in Shanghai teaching two days of in-person classes to Owl Academy’s 19th cohort, on putting WorkBuddy and Claude Code to work on real jobs. I also dropped into Dedao New Business School’s livestream to talk about building a short-video creator matrix with AI.

The question I got asked most this week: I’m about to start using Claude Code, Codex, and WorkBuddy—what kind of computer should I get to make the most of these tools?

The biggest misconception: top-tier AI means a maxed-out machine

I keep running into one big misconception—that using top-tier AI tools requires a maxed-out machine. Plenty of people spend thirty thousand, even forty or fifty thousand yuan on a MacBook, or on a flagship laptop with a top-end gaming GPU. None of that is necessary.

I’m a hardware guy myself. The 3090, the 4090, the top flagship laptops—I bought them, used them, put them through their paces, and sold every one of them over the past stretch. Hardware prices had climbed, so I even made money on the way out.

I did once think about running a local model on local compute to keep my day-to-day costs down. In practice it doesn’t hold up. Only a handful of video, image, and audio jobs work on local models. One or two discrete GPUs, top-end or not, can’t unlock everything an open-source model can do. And on complex tasks they aren’t remotely close to the closed-source labs’ models, on speed or on quality of result. So if you’re using AI as an individual, don’t waste money on a GPU.

The real logic of speccing an AI machine: money goes to RAM, not the GPU

In a traditional build, the discrete GPU often eats 40% or more of the budget. But if large-model inference never happens locally and runs entirely in the cloud, the one component worth real money is RAM—part of why RAM prices have climbed the most in the current hardware market.

RAM and CPU set how many tasks you can run at once and how many parallel sub-agents you can keep in the background. Take Claude Code: each window you open takes roughly 400–600MB of RAM, and long-context windows take more. The concurrency ceiling is generally thread count minus 2. Since most of the computation happens elsewhere, CPU threads aren’t the pinch point either. Add the resident memory that MCP tools and small local services need, plus Windows’ own footprint, and a comfortable setup starts at 16G.

RAM is expensive right now, so if you can’t get to 32G or 64G in one shot, pick a machine with upgradable memory: run 16G now and add an identical stick later. What matters is capacity, not frequency—lower-frequency, higher-capacity RAM is the better buy.

For the screen I still recommend 4K. It draws more power, but when you’re reading model output across several windows, the sharpness and eye comfort beat 2.5K or 3K. Everything past that is preference—portability, battery life, even looks.

Three machines worth buying

If you’re buying new, there are three good options:

  • ThinkBook 16+ Core edition—great value, and it’s the version with replaceable RAM;
  • Lenovo Xiaoxin Pro 16 GT;
  • Redmi Book Pro 16-inch—better value at the same size, thickness, and weight.

If you want more performance, go with the ThinkBook 16p 2025 edition: high CPU core and thread counts, priced past ten thousand and sometimes up to twenty thousand, but far cheaper than the laptops packing high-end desktop-class GPUs.

On a tight budget and open to used, the Dell XPS 15 9510 is a good pick too. It shipped in a huge number of configurations, so make the seller spell out exactly which one it is—the spread runs from just over three thousand to eight thousand.

The traps that catch people most often

  • RAM is the worst thing to buy right now—the prices are brutal, so used is usually the better deal;
  • A certain “pride of domestic tech” in-house OS can’t run Node or CC at all;
  • Don’t buy the wrong version—the same model number ships as a 2025 and a 2026 edition, with big gaps in RAM and CPU performance and big gaps in price, and some sellers will pass the old one off as the new one;
  • Know the difference between RAM and storage—a salesperson may tell you “this has 256G of memory,” but they mean storage, how many files you can keep. That’s not the memory that sets how many programs you can run at once. The difference is enormous.

Putting CC in the cloud: how much VPS do you need

Plenty of people, me included, install CC in the cloud rather than on a local machine. So how do you pick a spec when you buy a VPS? If you’re not trying to run seven or eight windows at once and just need two or three for day-to-day work, two cores and 2G of RAM is enough. The official recommendation is two cores with 4GB. If you want long sessions with headroom, four to eight cores and 16G is best. To be straight with you, I run cloud CC with a wrapper on top, three sessions in parallel on a two-core 2G box, and it holds up fine.

Don’t buy a monster rig just to “play with AI”

Whatever you do, don’t fall for the online pitch: “I’m getting into AI, so I need a maxed-out machine worthy of AI’s high tech.” It doesn’t work that way. We’re using AI, not playing games—you don’t need high clock speeds, you don’t need the newest GPU. You need more RAM and more CPU cores. It’s closer to speccing a server.

When people come asking, I point them at a used Acer Edge 16: 16-inch 4K screen, 16G of RAM, thin and light, right around a kilogram. Claude Code, Codex, day-to-day work—it handles all of it.

Updates should pick up pace from here. See you next episode.

We're using AI, not playing games—you don't need high clock speeds, you don't need the newest GPU. You need more RAM and more CPU cores. It's closer to a server spec.