The graphics card market is the most deliberately incomparable consumer hardware market in existence. Every card has a different amount of memory, a different amount of compute, a different memory bandwidth, a different target workload, and a model number engineered to prevent you from doing simple math. This is not an accident. Incomparability is a pricing strategy — when buyers cannot easily compare two products, sellers capture more margin.
The antidote is the same as it is for any obscured market: normalize to the unit that matters, then sort. For GPUs there are two such units, because GPUs are bought for two different reasons.
- If you are buying a GPU for memory — to hold a large AI model, a big dataset, high-resolution assets — the unit is dollars per VRAM gigabyte ($/GB).
- If you are buying a GPU for compute — raw math throughput for training, rendering, simulation — the unit is dollars per teraflop ($/TFLOP).
Get clear on which axis you are actually buying, compute the right ratio, and the market stops being able to hide value from you.
Why the sticker price is even more useless for GPUs than for storage
With hard drives, at least the capacity is standardized — a terabyte is a terabyte. GPUs have no such courtesy. Two cards at the same price can differ by 3x in memory and 2x in compute, and the marketing will emphasize whichever spec makes the more expensive card look good.
Consider the variables a single sticker price is hiding:
- VRAM capacity — how much memory, in gigabytes. Determines the largest model or dataset the card can hold at once. For AI workloads this is often a hard wall: either the model fits or it does not.
- Compute throughput — floating-point operations per second, measured in teraflops (TFLOP). And even this splinters: FP32 for general compute, FP16/BF16 for AI training, INT8/FP8 for inference. A card can be mediocre at one and excellent at another.
- Memory bandwidth — how fast data moves between VRAM and the compute cores. A card with plenty of memory and plenty of compute can still be bottlenecked here.
- Architecture generation — newer silicon does more work per watt and per dollar, so a lower raw spec on a newer architecture can outperform a higher spec on an older one.
A single price cannot express any of this. Which is exactly why you must decompose it.
Unit one: dollars per VRAM gigabyte
For a growing share of GPU buyers — anyone running local AI models, large scientific datasets, or memory-hungry creative work — VRAM capacity is the binding constraint. A model that needs 24GB of memory will not run on a 16GB card at any speed. It simply will not load. Compute is irrelevant if the data does not fit.
For these buyers, the honest unit is dollars per VRAM gigabyte:
$/VRAM-GB = card price / VRAM capacity in GBThis single ratio reorders the entire market. A flashy card with blazing compute but modest memory looks expensive per gigabyte. A workstation card with enormous memory but middling compute suddenly looks like a bargain — if memory is what you need.
| Card (illustrative) | VRAM | Price | $/VRAM-GB |
|---|---|---|---|
| High-clock gaming card | 16GB | $800 | $50.00/GB |
| Mid gaming card | 12GB | $500 | $41.67/GB |
| Memory-heavy workstation card | 48GB | $1,800 | $37.50/GB |
| Prior-gen high-memory card | 24GB | $700 | $29.17/GB |
Ranked by $/VRAM-GB, the “premium” gaming card is the worst value for a memory-bound workload, and a prior-generation high-memory card is the best. The sticker prices told the opposite story. That reversal is the entire point of unit pricing. You can compare every GPU by price per gigabyte of VRAM on our live board, which applies exactly this ranking across current and prior-generation cards.
The caveat: memory type and bandwidth still matter. GDDR6, GDDR6X, and HBM are not interchangeable; HBM delivers far more bandwidth at a higher cost. For a pure capacity constraint (“does the model fit?”), raw $/VRAM-GB is the right first filter. For memory-bandwidth-bound workloads, treat it as the first cut, then compare bandwidth within your shortlist.
Unit two: dollars per teraflop
The other reason to buy a GPU is raw math. Training runs, rendering farms, physics simulation, any embarrassingly parallel numerical workload — these are bottlenecked by compute throughput, and the unit is dollars per teraflop:
$/TFLOP = card price / compute throughput in TFLOP (in the relevant precision)The critical phrase is “in the relevant precision.” A GPU’s teraflop rating is not a single number. It varies by data type:
- FP32 (single precision) — traditional scientific and graphics compute.
- FP16 / BF16 (half precision) — the workhorse of AI training. Cards often deliver 2x or more their FP32 throughput here.
- FP8 / INT8 — AI inference. Newer architectures post enormous numbers in these low-precision modes.
A card can be poor value at FP32 and excellent value at FP16, or vice versa. If you compute $/TFLOP using the wrong precision for your workload, you will draw exactly the wrong conclusion. Match the precision to what you actually run, then compute the ratio.
This is also where architecture generation reasserts itself. Newer silicon typically delivers more teraflops per dollar and dramatically more per watt. Since electricity is a real and recurring cost for compute-heavy workloads, the honest buyer eventually extends the analysis from $/TFLOP to total cost per teraflop-hour, folding in power draw over the expected life of the card. A card that is cheaper to buy but hungrier to run can lose on a multi-year horizon.
Which unit are you actually buying?
The single most common GPU buying mistake is optimizing the wrong axis — paying for compute you will never use because your workload is memory-bound, or buying a big-memory card when you needed raw throughput. Diagnose first:
- “Will my model / dataset / scene fit?” -> You are memory-bound. Rank by $/VRAM-GB. Compute is a secondary filter.
- “How fast will it finish?” and it already fits -> You are compute-bound. Rank by $/TFLOP in your precision. Memory is a secondary filter.
- “Both — it barely fits and it needs to be fast.” -> You are on the frontier, where you pay the early-adopter premium on both axes. Here you accept worse unit pricing because the alternative is not doing the work at all.
Most buyers are more memory-bound than they think, especially for local AI. The models grew faster than consumer VRAM did, so capacity became the wall long before compute did. If you are in that camp, our GPU rankings sorted by $/VRAM-GB surface the cheapest capacity first — and they flag unified-memory systems separately, since shared memory is not the same as dedicated VRAM.
Why GPU markets distort — and how to read the distortion
GPUs sit at the intersection of three demand shocks that rarely align in other hardware: gaming, cryptocurrency mining cycles, and AI compute. When any one of these surges, it pulls supply away from the others and warps the entire pricing curve.
The AI demand wave is the structural story. When compute for training and inference becomes the scarce resource, the same inversion that hits high-capacity hard drives hits high-memory, high-compute GPUs: the most capable cards get allocated to datacenter and enterprise buyers, retail supply thins, and prices climb across the board. Secondary markets — used and prior-generation cards — become disproportionately attractive on a unit basis, because they are the supply that datacenters have already moved past.
Reading the distortion is the same discipline as before. Compute $/VRAM-GB and $/TFLOP across new, used, and prior-generation cards. When the curve inverts, the best unit value often sits in the used or last-generation tier rather than the current flagship. And as with storage: “expensive but available” is a wait signal; “vanishing from every retailer” is a buy signal, because scarcity of availability is a worse problem than scarcity of good pricing. Watching the best $/GB VRAM available today against its own recent history is the cleanest read on which regime the market is in.
Mark’s Take: GPU pricing is the purest example of a market that profits from incomparability. The model numbers are noise, the marketing emphasizes whichever spec flatters the price, and the specs are structured so no two cards line up cleanly. The instant you decompose the sticker into $/VRAM-GB and $/TFLOP-in-your-precision, the fog clears and the value ordering often reverses. The discipline is not GPU-specific. It is the universal move for any market built to confuse you: find the unit that determines whether the thing does its job, divide price by that unit, and sort. The market wants you comparing sticker prices. Compare unit prices instead.
The bottom line
A GPU is not one product priced one way. It is memory and compute sold together under a model number designed to keep you from separating them. Buyers who pay the sticker price are paying for a bundle they cannot evaluate.
Decompose it. If you are memory-bound, rank the market by dollars per VRAM gigabyte. If you are compute-bound, rank it by dollars per teraflop in the precision you actually run, and extend to cost-per-teraflop-hour if power matters. Diagnose which axis binds you before you spend a dollar, because optimizing the wrong one is how buyers overpay by a full tier.
The card the marketing wants to sell you is rarely the best value on the axis you actually need. The unit price tells you which card is. Do the division.
MarketCrystal provides trend analysis and market commentary for informational purposes only. Nothing in this publication constitutes financial advice or purchasing recommendations. Hardware prices and specifications change constantly; always verify current data before buying. Past pricing does not guarantee future results.
Follow the plumbing.