The 5090 is the memory crisis in one product: 32GB of GDDR7, the only consumer card serious local-AI users can put in one slot, bought by two markets at once. It fell 42% from its 2025 scalper peak to the $2,300 trough, then doubled and kept going: +$575 on the week, the largest absolute weekly move on the board. The 50-series family's June to September spike (39-41% on the family average) hit it hardest, and the escalation has not paused since. Every week it adds $150-250. This is a card you rent, not buy, until it breaks back under $5,000.

Production ended, so 4090 supply is a depleting finite pool. Every unit either sits in an AI rig or gets bought for export. It held $1,600-1,700 all of 2025, then ran +50% through 2026, and crossed $3,000 this week (+$150 on the week, the largest weekly move since June). At $125/GB it costs 2.2x a 3090 for 7% more bandwidth. The 4090 is now a performance card, not a value card. Buy only if single-slot speed is the priority and the rest of the plan can absorb a finite-pool price.

The 3090 is the card the entire local-AI market rotated into once the 4090 stopped being the value option. The floor moved from $1,300 to $1,400 in a week (+$100), confirming the sub-$1,350 tier is gone: cheap units are evaporating and the market never corrects anymore, each dip closing higher than the last. 96GB for $5,600 (4x) still beats 64GB for $11,950 (2x 5090) on capacity, and the 3090's $/GB is 3.2x better. Repaste before trusting a used unit, but this is still the backbone card of multi-GPU local AI.

The 22GB-modded 2080 Ti quietly became the cheapest route to serious multi-card VRAM: 8 units = 176GB for $4,600, beating the 4x 3090 on $/GB by 2.2x. Three consecutive flat prints = a verified floor, and the baseline is now old enough to be real data. The caveat is the mod: cards are used, VRAM is re-balled or upgraded, and you are buying a Pascal-era card at 616 GB/s. For quantized-model serving where bandwidth tolerates it, this is the best $/GB on the consumer side.

The 5060 Ti 16GB printed its first fall since the 9-20 baseline: $650 to $600, -$50 on the week. The DRAM-cost floor that held it flat cracked one grid step, and the most likely cause is AIB channel pricing softening between board partners rather than end demand. It still carries a $125 premium over the 9060 XT for CUDA and GDDR7, and it is +40% over MSRP. If you are buying for local AI and can live with AMD, the 9060 XT is the better purchase. If this prints $575 next cycle, it is in a downtrend and worth taking.

288 GB/s is the problem: token generation on a 16GB Ada card will be painful for anything above ~7B quantized. This week it repriced up +$75 to $650 while its competitor (the 5060 Ti) repriced down: the dip is not echoing through the 16GB tier. At $40.62/GB it is now the worst $/GB of any 16GB card on the board. Fine as a dash-off node or second slot, wrong as a local-AI primary.

The budget multi-card play: 8x = 96GB for $2,400. Bandwidth (360 GB/s) is the ceiling. This rig runs small and quantized models like MoE fairly well. But at $25/GB it is the cheapest way to get a 96GB local AI box with CUDA, and the $300 floor has held for two consecutive captures and is now confirmed. It moves in $25 steps, never in $100s.

The V100 32GB is the single best $/GB of any 30GB+ card on the board, and 900 GB/s of HBM2 beats the 3090 on raw bandwidth-per-dollar. It appreciates on quiet +$25 prints: $625 to $675 on the week, a stable rising floor, no churn. FP16 only (no FP8/BF16), older driver stack, 300W. Fine for llama.cpp, FP16 serving. 4x = 128GB for $2,700.

At $225 it is an absurd price, but it is a hard-to-find part. Clean units are rare. 8x = 128GB for $1,800. Buy when you see one verified. Do not pay above $250.

No FP16 tensor, needs a power mod (12V pin) and good cooling, but at $275 for 24GB it is the cheapest VRAM-per-dollar in existence. 4x = 96GB for $1,100. The $275 floor held for a second capture: confirmed, not noise. The 8-card P40 rig remains the ultimate budget local-AI box.

The cheapest 16GB of VRAM on the market, but Pascal without FP16 makes LLM inference slow. At $100 it is a tinkerer's card. Perfect for learning, wrong for serving. The price is just money, and it has been rock flat.

64GB of HBM2e on one board is a genuine capacity monster (41.02 $/GB), but the market is a two-seller auction, not a market: $2,700 to $2,800 to $2,625 across the last three captures, then flat. The $2,625 floor has held since 9-24. It is the card to watch, not chase. Set a $2,500 alarm, verify it is not thermally cooked from mining, and be ready to move within the hour when a good unit lists.

32GB of GDDR6 (not GDDR7) at 53.12 $/GB, nearly 3090 pricing, but at 640 GB/s and without CUDA. It crept up $25 on the week, in line with the board. GDDR6 memory means it is less exposed to the worst of the DRAM crisis, so its price has a lower ceiling. Wait for a $1,500 dip rather than paying the current $1,700.

The card that started the week's only sustained downtrend: $500 to $450, then back to $475 on the last two days. The dip paused, it did not reverse: day-vs-week divergence means the last 48 hours moved opposite to the trend. At 29.69 $/GB it is still the best $/GB of any new 16GB card, and the GDDR6 base keeps it ahead of the crisis. 4x = 64GB for $1,900 on ROCm. Buy the dips below $450, do not chase above $475.

Hovering at MSRP last week, now $950: +$50 on the week, the AMD flagship is repricing with the board (see the Arc B70 jump in 02). At 39.58/GB it is actually below the 3090's 58.33/GB, with matching bandwidth (960 vs 936 GB/s). The catch is the ecosystem: ROCm is good now, but CUDA still owns local-AI tooling. If you are committed to AMD, this is the buy. If you want the safest software path, the 3090 remains it.

Matches the V100 32GB on price/GB (20.31 $/GB) and comes with an actual support channel, but at half the bandwidth. Try to find dips.

The Arc 32GB line was the least-affected family of the crisis all year, and this week it repriced +$175 in two days to $1,475: +54% over MSRP, 46.09 $/GB. Intel's GDDR6 supply is now priced like everyone else's. It is priced worse than the 3090 and the 7900 XTX, and the XPU software stack remains the weak link. Wait for DRAM relief to do the work, not for a dip.

Held its $1,200 baseline this week while the B70 jumped: the tier below the flagship has not repriced yet. Same XPU story as the B70. If you are testing the XPU path for local AI, the B65 at $900 or less is the entry ticket. Above that, the 3060 12GB ($300) and P40 ($275) both beat it for pure inference capacity.

The one Arc that is priced sanely: 24GB at 27.08 $/GB, the closest-to-MSRP card in the Intel line, and it held flat this week. If you want 24GB of CUDA-alternative capacity cheap and accept the XPU stack, this is it. Otherwise it is outclassed by the 3090 and 7900 XTX.

The 128GB unified-memory box is genuinely impressive on $/GB (26.56), and it is the cheapest 128GB single box on the board, but 256 GB/s of bandwidth means it generates tokens at a crawl compared to anything with HBM or GDDR6/7. Decent hybrid desktop box. It held its $3,400 baseline this week while the DGX Spark repriced $400 above it. Watch for price drops, best for x86 AIO.

Micro Center sold out at $4,900 for a day and has thin restock. Expect a reprice. $5,300 is the typical new online price (+32.5% over MSRP). At 41.41 $/GB and 273 GB/s it is an expensive 128GB box

The M1 Max just did what no other AIO on the board is doing: it fell. -$125 in two days to $1,200, as refurb supply (retired enterprise and trade-in units) caught up to the DRAM tax. At 18.75 $/GB with 400 GB/s of unified bandwidth, silent, ~155W, and -40% of MSRP, it is now the best $/GB box on the board above 32GB and the first AIO in this cycle that is a clear buy. Confirm the $1,200-1,250 floor holds next week; if it does, the AIO tier has a new center of gravity. It out-specs the M4 Pro (48GB) for $50 more.

The M4 Pro fell $50 to $1,250, in step with the M1 Max: the Apple AIO tier is repricing as a unit, which points to refurb supply setting the market rather than DRAM cost. At 26.04 $/GB it is still 16GB short of the M1 Max for $50 less. Buy only if you need the M4 CPU. For VRAM-per-dollar the M1 Max wins today, on both price and capacity.

$186.72/GB is 3.2x the 3090, +199% over MSRP, and the memory crisis keeps ratcheting new-card bills of materials up. It added $575 on the week and shows no deceleration. Forecast peak ~$6,650 by Dec 2026, then a slow decline toward $5,100-5,700 by Q4 2027. Buy only on an emergency single-slot 32GB need, and even then price it against 4x V100 32GB ($2,700 for 128GB) first.

End-of-life supply is depleting. It crossed $3,000 this week (+$150 on the week) and is +87.5% off its 2024 price, forecast to hit ~$3,450 by Dec 2026. At $125/GB it is a speed purchase, not a value purchase. If 24GB speed is your actual requirement, the post-holiday Jan-Feb 2027 window is the first reasonable re-entry point.

Still the value king at $58.33/GB with 936 GB/s and NVLink. The floor moved $1,300 to $1,400 in a week: the sub-$1,350 tier is gone, so start your multi-unit set now at the current floor and buy remaining units before mid-November. Expect ~$1,675 by December. Repaste every unit.

$26.14/GB, floor confirmed on a third consecutive print. The cheapest serious multi-card VRAM on the board (8x = 176GB for $4,600). Verify the 22GB mod and run VRAM stress on every unit. The only reason to wait is under $550. The only reason to skip is if 616 GB/s cannot serve your models.

Printed its first dip this week ($650 to $600): a good early signal, not yet a trend. It is +40% over MSRP and the 9060 XT does 16GB for $125 less. Revisit only if you need CUDA specifically, and only if it prints a second downward step ($575) or touches $550 on a sale.

Repriced up to $650 this week and 288 GB/s is a token-generation bottleneck. It is now the worst $/GB of any 16GB card on the board. Fine as a dash-off node or secondary, wrong as a local-AI primary. In the forecast it eases to ~$525-600 by 2027 but never becomes competitive.

The $300 floor held for a second capture and is confirmed. 8x = 96GB for $2,400 is the easiest budget CUDA rig that exists. Bandwidth (360 GB/s) caps you at small and quantized models. Buy it for MoEs, avoid for Dense.

$21.09/GB, 900 GB/s HBM2, 94% under its original MSRP. The best $/GB of any 30GB+ card, and the crisis cannot inflate its cost structure (2017 memory). The board is drifting it up slowly (+$50 week). 4x = 128GB for $2,700. FP16-only and older drivers, fine for llama.cpp and FP16 serving.

$14.06/GB is absurd, but clean units are rare and this is an auction, not a market. 8x = 128GB for $1,800. Strike immediately when a verified unit appears at $250 or less. Do not pay above that.

$11.46/GB, the cheapest VRAM-per-dollar on the board. Floor confirmed on a second capture. Needs a power mod and good cooling. No FP16 tensor. 4x = 96GB for $1,100 is the ultimate budget local-AI box.

$6.25/GB is essentially free, but Pascal without FP16 makes real LLM inference slow. A $100 learning and rack-filler card. The right price for what it is, the wrong card for serving.

64GB HBM2e at $41.02/GB is a genuine capacity monster, but the market is a two-seller auction ($2,700 to $2,800 to $2,625). The $2,625 floor has held since 9-24. Set a $2,500 alarm, verify mining wear, and be ready to buy within the hour. Do not chase a $2,800 listing.

32GB at $53.13/GB is 3090 pricing without CUDA and with less bandwidth. GDDR6 means lower crisis exposure, so it has a lower ceiling than the NVIDIA equivalents. No urgency to pay $1,700. If you are all-in on the AMD pro path, $1,500 is the buy.

The only card on the board with a sustained downward trend (-$50 week), now paused at $475 after a +$25 two-day. $29.69/GB, and the GDDR6 base keeps it ahead of the crisis. 4x = 64GB for $1,900 on ROCm. Buy the dips below $450, do not chase above $475.

$39.58/GB actually beats the 3090 on paper, with matching bandwidth, but the CUDA moat still owns local-AI tooling. It is now +$50 on the week, repricing with the board. If your stack is ROCm-first, $950 near MSRP is a fair buy today. Otherwise wait for $850 or less, or take the 3090.

Matches the V100 32GB on price/GB with vendor support, but at half the bandwidth. The wrong pick at this price for bandwidth-hungry serving.

Repriced +$175 in two days to $1,475 (+54% over MSRP): the crisis reached Intel's channel this week, and the 'closest to list' status is gone. At $46.09/GB it is priced worse than the 3090 and 7900 XTX, and the XPU stack is the weak link. Wait for DRAM relief to do the work for you.

Held its $1,200 baseline while the B70 jumped, but the tier above it has repriced and this card is next. The XPU entry ticket at $900 or less. Above that, the P40, 3060, and 9060 XT all beat it for inference capacity-per-dollar.

The sanely priced Arc: 24GB at $27.08/GB, the closest to MSRP in the Intel line, held flat this week. If you specifically want to evaluate the XPU path for local AI, this is the entry point. For everything else, outclassed.

Best $/GB of the 128GB single boxes ($26.56), but 256 GB/s caps token generation hard. It is a hybrid CPU+GPU box, not an inference server. It held $3,400 this week while the DGX Spark repriced $400 above it. If single-box 128GB convenience matters more than bandwidth, $3,200 is the buy and $3,400 is a watch.

The $4,900 ghost price is resolved: $5,300 is now the confirmed street price (+32.5% over MSRP), rebounding faster than the 2-6 week forecast assumed. At $41.41/GB and 273 GB/s it is a convenience box. 4x V100 32GB gets 128GB for $2,700. Buy only if a flash listing reappears under $5,100.

The correction winner: -$125 in two days to $1,200 on refurb supply, and now the best $/GB box on the board above 32GB ($18.75) at 400 GB/s unified bandwidth, silent, ~155W, -40% of MSRP. If MPS is acceptable for your stack, $1,200-1,250 is a clear buy. Confirm the floor holds next week; out-specs the M4 Pro (48GB) for $50 more.

Fell $50 with the M1 Max: the Apple AIO tier is repricing as a unit. At $26.04/GB it is still 16GB short of the M1 Max for $50 less. Buy only if you specifically need the M4 CPU. For VRAM-per-dollar, the M1 Max wins today on both price and capacity.