Graphics card pricing has been abnormal for most of the past decade, and the reasons are largely external to gaming.
The manufacturing base
Modern graphics processors are made at a small number of foundries using leading-edge processes.
Capacity at those nodes is limited, expensive and allocated between customers competing for it.
Which means graphics card availability depends on allocation decisions made against demand from every other user of the same capacity.
Wafer costs have risen substantially at each new node, which pushes up the floor price of any chip made on them.
Die size and yield
Larger chips cost disproportionately more.
Manufacturing defects are distributed across a wafer, so a larger die is more likely to contain one, which reduces the proportion of usable chips.
Which means a chip twice the area costs considerably more than twice as much to produce.
Binning — selling partially defective chips as lower-tier products with sections disabled — recovers value and is why product stacks are structured as they are.
Memory
Graphics memory is a separate market with its own supply cycle, and it is a substantial share of card cost.
Which is why memory capacity is used as a segmentation lever, and why cards at similar prices can differ substantially in it.
Memory bandwidth matters as much as capacity for performance, and the two are frequently conflated in marketing.
The external demand shocks
Cryptocurrency mining consumed enormous quantities of graphics cards during two separate periods, since the same parallel computation suited both purposes.
Prices rose far above list, availability collapsed, and secondary markets filled with used cards afterwards when the demand evaporated.
The subsequent shift of the largest cryptocurrency away from mining removed much of that demand permanently.
Machine learning demand then arrived, and it is larger and less cyclical, drawing on the same manufacturing capacity and in some cases the same products.
Segmentation
Manufacturers sell into gaming, professional visualisation and data centre markets with products sharing underlying architecture.
Data centre products command far higher margins, which affects how capacity is allocated when it is constrained.
Which means gaming cards compete for supply against products sold at multiples of their price, and the outcome of that competition is predictable.
What has actually improved
Performance per unit price at the lower end has stagnated more than at the high end, which is the reverse of the historical pattern.
Upscaling technologies, rendering at lower resolution and reconstructing, have delivered substantial effective performance gains without new hardware.
Which has become a significant part of how performance is marketed, and comparing figures that use it against figures that do not is not meaningful.
Buying sensibly
Determine the resolution and frame rate you actually want, which sets the requirement.
Compare at your target settings rather than at benchmark defaults.
Check memory capacity against the games you play, since it is the specification most likely to become limiting.
And consider the previous generation and the used market, since depreciation in this category is steep and the performance gap between generations has narrowed at several price points.
Power and cooling
A practical constraint that has grown into a purchasing consideration.
Power draw at the high end has risen substantially, which affects power supply requirements, case airflow and room temperature.
Connector standards have changed to carry the load, and a specific connector design attracted attention after reports of melting under partial insertion.
Which is worth checking on installation rather than assuming, and full insertion with a straight cable run addresses the reported cases.
Performance per watt
A more useful comparison than raw performance for anyone concerned about running costs or heat.
Efficiency has improved substantially across generations, and it is frequently obscured because manufacturers push power limits higher to gain the last few percent of performance.
Reducing the power limit manually frequently costs a small proportion of performance and saves a large proportion of power, which is among the better-value adjustments available.
Integrated graphics
Now capable enough for a substantial range of games at moderate settings, which was not true a few years ago.
Which has changed the calculation at the low end considerably, and it is worth checking before assuming a discrete card is required.
The used market
Substantial and worth considering, with the caveat that heavily used cards from mining operations circulated widely after each cycle.
Checking warranty transferability, and running a sustained stress test immediately, covers most of the risk.
Board partners
Most cards are built by third parties using the chip designer's reference specification.
Which produces variation in cooling, power delivery, clock speeds and warranty at different prices for the same underlying chip.
The differences in gaming performance between partner cards are generally small, and the differences in noise and temperature are not.
Longevity
Memory capacity is generally the specification that dates a card first, since requirements rise with texture resolution faster than processing requirements rise.
Which is worth weighting more heavily than a small difference in benchmark performance.