When teams request hardware for AI work, procurement usually compares GPUs the way gaming reviews do: by model number and benchmark speed. For generative AI, that comparison optimises the wrong variable.
Speed is a slope; a slower card finishes later. Memory is a wall; a model that does not fit either fails or runs dramatically slower. Buy the wall first.
What the Installed Base Looks Like
Valve's Steam Hardware Survey for September 2026 is the best public view of consumer GPUs. 16 GB is now the most common VRAM tier at 27.21%, narrowly ahead of 8 GB at 26.71%. But 45.89% of systems still have 8 GB or less, and only 7.41% have 24 GB or more. The most common single card is the 12 GB RTX 5070 at 5.86%. If your organisation bought general-purpose machines in the last few years, many of them are likely below the line for the larger models.
What the Models Require
| Workload | Example model | Approximate VRAM |
|---|---|---|
| Lightweight image generation | Stable Diffusion 1.5 | 4 GB |
| Mainstream image generation | SDXL | about 8 GB |
| Video generation | LTX Video | 9.4 GB |
| High-end image generation | SD 3.5 Large | about 12 GB |
| Frontier open image model | FLUX.1 [dev] | 23.8 GB |
The arithmetic is simple. FLUX.1 [dev] is a 12 billion parameter model; at two bytes per parameter its core weights alone are about 24 GB.
The Hidden Cost of "Almost Fits"
Software can offload parts of a model into system memory, and the Diffusers documentation is candid about the trade: the most aggressive form of CPU offloading is "extremely slow" and often impractical. Offload also moves the requirement to system RAM, where the same guide advises twice the model size for its faster streamed variant. An undersized GPU becomes a productivity tax paid on every render.
Procurement Rules
- Start from the models the team needs, not from a GPU shortlist.
- Treat 16 GB as the floor for new AI-capable workstations, and 24 GB for teams that need the largest open models.
- Budget system RAM alongside the GPU. 32 GB is already the most common configuration in the same survey, at 42.22%.
- Compare speed only within a memory tier. A faster card with too little memory loses to a slower card that fits.
- Audit the existing fleet before buying, so you know which machines can serve which workloads.
Why Gaming Benchmarks Mislead Here
Procurement teams reach for gaming benchmarks because they are abundant and easy to compare. They measure the wrong thing for this workload. A game is tuned to fit the memory of the cards it targets, so benchmarks compare speed among cards that can all run it. A generative model does not adjust itself to fit. If it needs more memory than the card has, it either fails or slows dramatically while it borrows system memory.
That changes the shape of the decision. In gaming, a more expensive card buys more frames per second along a smooth curve. In generative AI, there are cliffs: the step from a card that cannot hold a model to one that can is worth far more than any speed difference between two cards on the same side of the cliff.
Plan for the Models You Will Need Next Year
Hardware outlives model generations. The largest open image model in the table above needs roughly three times the memory of the mainstream one, and the direction of travel has been toward larger models, not smaller. A workstation bought today at the minimum for today's mainstream model may be below the line for the model a team wants to adopt within its depreciation period.
That argues for buying memory headroom on machines meant for AI work, even when today's workload does not need it. The incremental cost of a higher memory tier is usually smaller than the cost of replacing a machine early or running a team on offload for its remaining life.
Not Every Seat Needs the Top Tier
Memory headroom is not a case for buying the largest card for everyone. Most organisations have a small group generating heavily and a larger group that occasionally needs AI imagery. Equip the first group for the models they use daily. Give the second group access to shared capacity, or to the lighter models that run on mainstream hardware. Matching tiers to roles keeps the fleet affordable without forcing anyone onto hardware that cannot run their work.
Questions Before Signing a Workstation Order
- Which models will this machine need to run, today and within its expected life?
- How much GPU memory does the largest of those need, and how much headroom does the specified card leave?
- How much system RAM is specified, and is it enough to support offloading if a model does not fit?
- Is the machine for daily heavy use or occasional use, and does its tier match that role?
Don't Forget the Existing Fleet
Before buying, check what you already own. Many organisations have a few machines with capable cards bought for design, video or engineering work. Those may already sit above the memory line for mainstream models, and they can host a pilot while procurement decisions are made. An inventory of GPU memory across the fleet takes little effort and often changes the purchase.
Let Software Tell You What Fits
Fit should be checked before anyone downloads gigabytes of model weights. In Vision Studio, our open-source local AI app, a hardware fit check sizes each model against the machine's GPU and system memory and labels it as fitting, fitting with offload, over budget or CPU-only. That turns a procurement question into a quick inventory exercise.
The Takeaway
For AI workstations, memory is the strategic specification. Buy enough of it, and the speed will be sufficient. Buy too little, and no amount of speed will help.
The full hardware guide, tier by tier is on the Vision Studio blog at vision-studio-x.com/blog/vram-is-the-spec. Vision Studio is a free, open-source, local-first AI generation app built by Strategia-X. If you want a second pair of eyes on how your organisation adopts AI tooling, talk to us at strategia-x.com.
-Rocky
#ITProcurement #GPU #AIInfrastructure #HardwareStrategy #LocalAI #TechnologyPlanning #StrategiaX #VisionStudio