The trap nobody checks before buying
Someone specs a 128GB unified-memory box to run MiniMax M2 locally. Ten billion active parameters, the marketing says, so it should run like a much smaller model. At INT4 with KV headroom, MiniMax M2’s total parameter footprint lands around 157GB (illustrative, from the 0.55GB-per-billion INT4 rule), because inference has to keep every expert resident in memory even though only a fraction fires per token. That’s more than the box has. It starts paging to disk, tokens per second falls off a cliff, and the buyer is stuck either downgrading to a smaller model or eating the cost of reselling a machine that isn’t cheap to ship.
This happens because people size hardware off the “active params” number every MoE model advertises, when the number that actually pays the memory bill is total params. Get that wrong on a fixed unified-memory box and there’s no slot to add capacity. Get it wrong on a GPU rig and you can rent a bigger one by the afternoon.
Rung 1: used 24GB cards
An RTX 3090 or RTX 4090 off the used market is still the best dollar-per-token entry point for anything that fits in 24GB at INT4. GLM-4.7 Flash, 31B total params and 3B active, needs roughly 21GB with headroom: comfortable fit, room left for context. GPT-OSS-20B is easier still, around 14GB. Both are the kind of small, fast MoE this generation was built for, and both run well on a single used card.
Qwen3.6 35B A3B is the one that catches people. Total params sit at 36B, and INT4 plus headroom lands around 25GB, over the edge of a 24GB card. Trim the context window and it squeezes on; leave the defaults alone and you’re offloading the overflow to CPU, which erases the reason you bought a fast card in the first place. On this rung specifically, check total params before active params, because the margin for error is a couple of gigabytes, not a comfortable buffer.
Rung 2: 128GB unified memory boxes, the $/GB winner
Generic 128GB unified-memory mini PCs, the current wave of AI workstation boxes built around unified LPDDR, offer the cheapest capacity per gigabyte of any rung here. GLM-4.5 Air, 106B total and 12B active, needs around 73GB at INT4 with headroom, leaving room for a longer context window or a second small model loaded alongside it. That’s the sweet spot for this rung: models with total params in the 60-110B range and modest active counts.
The tradeoff is memory bandwidth, which is noticeably lower on these boxes than on a dedicated GPU or an Apple Ultra chip. Capacity isn’t the bottleneck here; throughput per active-param read is, and generic unified memory reads slower per token than either a GPU’s HBM or Apple’s memory fabric.
Rung 3: Macs, the bandwidth winner
Same unified-memory capacity class, meaningfully faster memory bandwidth, higher price per gigabyte. For anything where a human is waiting on tokens in real time, a Mac Studio at the same 128GB capacity feels faster than the generic box running the identical model at the identical quant. It won’t out-run a dedicated GPU on raw throughput, but it beats the cheaper unified box on latency per token, and that’s the axis that matters for a coding agent someone is sitting in front of.
Rung 4: RTX PRO 6000 96GB, the startup default
This is the card everyone reaches for first because it’s a single PCIe slot with more VRAM than any other card in the consumer-adjacent tier, backed by real GPU bandwidth. GLM-4.5 Air fits with plenty of room to spare. So does anything else in the sub-90GB INT4 range. MiniMax M2 does not: 157GB doesn’t fit on one RTX PRO 6000. Once a model’s footprint crosses that card’s 96GB ceiling, the only way up is a second card, or a rental.
The rung everyone forgets: renting the same silicon
Every card above this line can be rented by the hour instead of bought. The RTX 3090 page and RTX 4090 page show marketplace tiers running 70%+ under the secure on-demand rate as of this writing, which changes the buy-versus-rent math more than people account for when they’re pricing a used card off a resale listing. If you’re testing whether Qwen3.6 35B A3B or GLM-4.7 Flash fits your workload before committing to hardware, renting the exact card you’re about to buy for a week costs less than the shipping insurance on the used unit.
Past MiniMax M2’s footprint, renting stops being optional. Kimi K2’s total params put its INT4 footprint north of 580GB, which is multi-GPU territory no matter which rung you started on. At that size, price H200 or B200 multi-card quotes on the GPU pages rather than pricing a home rig; nobody is stacking enough RTX PRO 6000 units in a closet to make that math work.
The decision rule
Compute your target model’s INT4 footprint (total params times 0.55GB, plus 25% for KV) before pricing anything. If it clears roughly 90GB, skip the 24GB-card rung entirely and compare the RTX PRO 6000 against a 128GB unified box on the GPU pages. If it clears roughly 150GB, stop pricing single boxes and run the multi-GPU numbers through the calculator instead, because at that size you’re buying or renting a cluster either way. What that calculator can’t tell you is how often your real traffic bursts past whatever concurrency you tested at, and that number comes from your own logs, not from a spec sheet.