In the second week of October 2026, something strange started happening in graphics card markets around the world. Cards with big video memory — the kind AI enthusiasts want — began vanishing from store shelves, and prices started climbing at a pace nobody had seen since the crypto mining craze of 2021. By November 1, a Japanese market researcher's monthly price check landed like a bombshell: current-generation NVIDIA and AMD graphics cards had jumped between 18% and 45% in a single month, with the flagship GeForce RTX 5090's street price in Japan climbing from 939,800 yen to 1,238,430 yen — roughly $6,150 to $8,100.
The trigger was not a mining boom, not a tariff, and not a factory fire. According to market watchers, it was a free, open-source AI project on GitHub called Strata — a program that lets an ordinary gaming PC run a 125-billion-parameter AI model locally. When tens of thousands of people realized their gaming GPU could suddenly run frontier-class AI at home, they went shopping. The result is the GPU shortage of November 2026, and this article explains exactly what happened, what the data shows, and what you should do about it.
TL;DR — The Strata Effect in 60 seconds:
- Strata (github.com/Niko1221/Strata), an MIT-licensed open-source engine by Slovenian developer Niko Veit, runs Alibaba's 125B-parameter Qwen3.8-Flash-Next on a gaming PC with as little as 12 GB of VRAM — and has gathered roughly 36,000 GitHub stars in five weeks.
- Market participants say the project triggered an October buying spree for high-VRAM graphics cards, draining retail inventory.
- Japanese retailer data (researcher Aile, via GameGPU, Nov 1) shows 18%–45% month-on-month price increases on NVIDIA Blackwell and AMD RDNA 4 cards.
- Valve's October 2026 Steam Hardware Survey (published Nov 2) shows the RTX 5070 still the most popular current-gen card, while AMD's RX 90 series keeps growing.
- Modders answered with NV UEFI Unlocker, a tool claiming up to 75% more AI compute on older RTX 20/30/40 cards — with real risks attached.
- Underneath it all: a memory crisis. GDDR prices are surging, SK hynix expects demand to outrun supply through 2030, and AMD is reportedly raising kit prices again in November.
What the Numbers Show: Japan's GPU Market in November 2026
The clearest snapshot of the squeeze comes from Japan, where computer-component market researcher Aile publishes a monthly price monitoring report. The November 1, 2026 edition, covered by GameGPU, found that in just one month — from early October to November 1 — local retailers recorded a "sharp and abnormal" price increase on current-generation NVIDIA Blackwell and AMD RDNA 4 graphics cards. Available models at major retailers rose by an average of 18% to 45% compared with late September and early October prices, driven by rapidly depleting inventory.
Here are the report's headline figures, exactly as published (yen prices with the article's dollar conversions):
| Graphics card | Early Oct 2026 | Nov 1, 2026 | Reported change |
|---|
| GeForce RTX 5090 32 GB | 939,800 yen (~$6,150) | 1,238,430 yen (~$8,100) | Up to 45% — the month's biggest jump |
| GeForce RTX 5060 Ti 16 GB | 148,800 yen (~$970) | 176,760 yen (~$1,150) | +35% |
| GeForce RTX 5080 16 GB | 144,800 yen (~$2,170) | 160,800 yen (~$2,230) | +32% |
| GeForce RTX 5070 Ti 16 GB | — | 245,000 yen (~$1,600) | +28% |
| Radeon RX 9070 XT 16 GB | 138,800 yen (~$910) | 171,800 yen (~$1,120) | +25% |
| GeForce RTX 5070 12 GB | 144,800 yen (~$950) | 160,800 yen (~$1,050) | +22% |
| Radeon RX 9060 XT 16 GB | 83,980 yen (~$550) | 86,800 yen (~$570) | +20% |
| GeForce RTX 5060 8 GB | — | 88,800 yen (~$580) | +18% |
Two patterns stand out in this data. First, the increases hit both NVIDIA and AMD — nobody's lineup was spared, though NVIDIA's high end moved the most. Second, cards with more video memory climbed harder: the 16 GB models (RTX 5060 Ti 16 GB, RX 9070 XT 16 GB) rose 25–35%, while the 8 GB RTX 5060 rose the least at 18%. That split is the fingerprint of an AI-driven buying spree, because local AI workloads care about VRAM capacity far more than gaming frame rates.
The report also notes something more alarming than the prices: stock itself was disappearing. Even before the main wave of retail price increases, the number of cards available in stores had dropped to virtually zero, and the more affordable versions vanished from the market completely. The publication of the report sparked a massive discussion among Japanese PC enthusiasts — over 1.2 million views — with many comparing the situation to classic economic bubbles and gamers venting that chip makers still refuse to separate gaming and AI product lines, forcing ordinary consumers to overpay.
Meet Strata: The Open-Source Engine Behind the Buying Spree
So what is Strata, and why would a free program move a multi-billion-dollar hardware market? Strata is an open-source local AI inference engine — software that runs large AI models directly on your own computer instead of in a data center. It was created by Niko Veit, a developer from Slovenia publishing on GitHub as "Niko1221", and the code is released under the permissive MIT license, which means anyone can use it, modify it, and build on it for free.
The project's pitch is genuinely remarkable: it runs Qwen3.8-Flash-Next — Alibaba's open-weight model with roughly 125 billion parameters — on an ordinary gaming PC. A model that large would normally need data-center hardware. Strata's minimum advertised spec is a graphics card with just 12 GB of VRAM and 32 GB of system RAM. The repository at github.com/Niko1221/Strata was opened on September 24, 2026, and by early November it had climbed to roughly 36,000 GitHub stars in about five weeks — one of the fastest climbs on record for a local-AI project, and largely a one-person effort funded through Buy Me A Coffee donations.
How it works: the kitchen metaphor
The technical trick is not raw speed — it is stubbornness. Qwen3.8-Flash-Next is a mixture-of-experts (MoE) model with 24,576 experts, of which only about ten activate per token. Instead of forcing the whole model into the graphics card's memory, Strata spreads the work across every piece of silicon in the machine. The project's own documentation uses a kitchen metaphor: the GPU is the counter — small but fast, keeping a few thousand of the most-used experts within reach; system RAM is the pantry, holding all the experts at once; the CPU works on the rest simultaneously; and the SSD stores a large lookup table used during prompt processing. Add speculative decoding — where a smaller model drafts and the large model verifies in batches, reported at 1.6–1.8x faster generation — and a frontier-scale model becomes usable on a gaming rig.
The practical result: VRAM matters more than raw GPU speed. Each extra gigabyte of video memory lets the card hold more of the model's hot experts close, which is exactly why the buying spree targeted 16 GB and larger cards. The project also speaks an OpenAI-compatible API at localhost plus an Anthropic-compatible endpoint, so existing tools like Cursor and Claude Code can point at your local model — no cloud account, no per-token bill, nothing leaving your machine.
A five-week sprint of releases
Strata's rise was not a single viral moment — it was a drumbeat of releases that kept the project in the conversation through October:
- Early October (v0.1.38): roughly 7,900 stars and 700 forks, with hands-on testing on a Radeon 7900 XTX reporting up to 120 tokens per second output and a realistic ~75 per second average — fully local, fully private.
- October 25 (v0.2.5): brought DeepSeek V4.1-Flash into the desktop app for the first time, plus a single installer per operating system instead of per-GPU installers. Stars ballooned to roughly 32,000.
- October 29 (v0.2.8): added four-card Radeon support that triples throughput (2.4x total throughput, from 155 to 370 tokens/second at 32 simultaneous requests), 17–23% faster paired Intel Arc B65 cards, and fixed a Mac memory leak that had been crippling DeepSeek on Apple Silicon.
Honesty check — and this matters for the article's credibility. Not everything about Strata is as clean as the headlines suggest, and a good buyer should know the caveats before treating the hype as buying advice:
- The developer's own release notes are unusually frank: headline speed figures apply to very specific configurations, and he says so up front. Independent testers confirm the shape of the claims but note that a viral demo on an RTX 4090 with 128 GB of RAM says little about the advertised 12 GB minimum.
- A widely discussed video test pointed out that the viral "100+ tokens per second" number was reading speed (prompt processing), while actual writing/generation ran closer to 4–6 tokens per second in that test — a meaningful distinction.
- The project's "Speed" mode trades a small quality loss for faster output, and aggressive quantization quietly trades away model quality — the classic local-AI bargain.
- Coverage has drawn skepticism about heavy online promotion and the lack of standardized minimum-spec benchmarks. The MIT license covers the engine, not the model weights — Qwen's weights carry their own separate license terms.
None of that stopped the market reaction. Whether or not every benchmark holds up, the idea — a frontier-class model on your own hardware, free, private, offline — was enough to send thousands of enthusiasts to the checkout page. And in a market where GPU supply was already tight, that was all it took.
Why High-VRAM Cards Became the Prize
To understand why the Strata craze hit graphics card prices the way it did, you need to understand one thing about local AI: video memory is the currency, not gaming performance. A GPU's VRAM is its fast workspace — and AI models live or die by how much of themselves fit in that workspace. A 12 GB card can run Strata's advertised minimum; a 16 GB card runs it comfortably; a 24 GB or 32 GB card runs it fast. Gaming benchmarks are nearly irrelevant to this calculation, which is why the buying spree looked so strange to traditional PC builders: people were not buying the best gaming card for the money, they were buying the most VRAM per dollar they could find.
This explains the Japanese data's split personality. The 16 GB cards — RTX 5060 Ti 16 GB (+35%), RTX 5080 16 GB (+32%), RX 9070 XT 16 GB (+25%) — rose far harder than the 8 GB RTX 5060 (+18%), even though the 8 GB card is the cheapest way to play games. AI buyers were systematically vacuuming up memory-rich cards, and the GameGPU report confirms that the cheaper versions disappeared from stores first. Ordinary gamers were left competing with AI enthusiasts for the same silicon — the exact complaint Japanese users voiced in that 1.2-million-view discussion, and the reason they are asking chip makers to separate gaming and AI product lines.
There is a second-order effect worth noting: the price pressure moved to system RAM and SSDs too. Because Strata's whole design spreads the model across GPU, RAM, CPU, and SSD, a proper setup wants 32–64 GB of system memory — and as one widely shared test put it, "the price tag moved to RAM". Memory makers are steering production toward AI server memory, and SK hynix has said it expects demand to outrun supply through 2030. We covered the RAM side of this crisis in our RAM prices November 2026 guide — read it alongside this article, because the two shortages share a single root cause.
The Steam Survey Reality Check: What Gamers Are Actually Buying
While AI buyers were emptying shelves, what were actual gamers doing? Valve's October 2026 Steam Hardware Survey, published November 2, gives us the answer — and it contains a twist.
The headline: the combined share of desktop and mobile NVIDIA RTX 50-series cards fell to 18.28%, down 4.70% from September; desktop RTX 50 cards alone fell to 14.01%, down 5.05%. Before you read that as gamers abandoning new cards, Valve's data comes with an explanation: September's figures were inflated by a surge of activity from Asian internet cafes during a holiday period — a statistical anomaly that corrected itself in October, confirmed by a 9.41% drop in the Simplified Chinese language share. The October drop is mostly the anomaly unwinding, not a real collapse in RTX 50 adoption.
The per-model table tells the steadier story:
| Graphics card | Steam share (Oct 2026) | Month change |
|---|
| GeForce RTX 5070 | 3.78% | -2.37% (still #1 current-gen NVIDIA card) |
| Radeon RX 9070 XT | 1.59% | +0.17% — AMD's most popular current-gen card |
| Radeon RX 9060 XT | 0.89% | +0.12% |
| Radeon RX 9070 | 0.21% | +0.01% |
| RX 90 series (desktop, total) | 2.69% | +0.30% — steady growth, no anomaly |
The key takeaway: AMD's RX 90 series grew steadily through the chaos — up 0.30% to 2.69% of surveyed desktops — while NVIDIA's numbers were distorted by the survey anomaly. The RX 9070 XT at 1.59% remains AMD's flagship of the people, and its +0.17% gain in a month when NVIDIA's figures fell suggests real buyers kept choosing Radeon even as prices climbed. For anyone shopping today, that is a useful signal: the RX 9070 XT's value story is surviving the squeeze better than most.
Modders Fight Back: The NV UEFI Unlocker
Here is the most fascinating subplot of the shortage: while new cards were getting more expensive, a group of modders figured out how to squeeze significantly more AI performance out of the old cards people already own.
The tool is called NV UEFI Unlocker, an open-source project by GitHub developer physicsdolphin, with version 2.1.8 released October 31, 2026. It targets a performance cap NVIDIA applies to Tensor Core operations on consumer GeForce cards. The trick is timing: it runs as a UEFI application from a USB stick before the operating system and NVIDIA's driver load, applies firmware-level register changes that remove the limiter on certain Tensor Core instruction rates, then hands control back to Windows or Linux. No VBIOS flashing, no driver patching — and because the change lives in memory rather than firmware, a full power-off clears the unlock, so it must be reapplied after every cold boot.
The reported gains are striking — with the usual "your mileage may vary" attached:
- RTX 2070: from ~36 TFLOPS to ~60–63 TFLOPS in FP32-accumulation matrix multiplication — a 67–75% jump (early tester result, via Wccftech).
- RTX 4060 Laptop GPU: from 29.89 to 49.33 TFLOPS in FP16 compute — +65%.
- Project's own claims: up to ~2.6x in FP16/BF16/TF32 tensor throughput, up to 1.9x in FP8, 14–75% in image/video generation, up to 94% in DirectML inference kernels.
- Reality check from users: one RTX 3080 LHR owner reported a 28.8% improvement in ComfyUI SDXL image generation — well below the synthetic benchmark peaks, because real applications are limited by memory bandwidth and power, not just math throughput.
Support covers GeForce RTX 20, 30, and 40 series (desktop and laptop), including Turing, Ampere, and Ada cards. GTX 16 series support is experimental. The RTX 50 series (Blackwell) is not supported — the existing unlocking methods do not work on the new architecture. Setup needs a FAT32 USB stick, UEFI boot, and Secure Boot disabled (or a trusted certificate enrolled); RTX 20 owners must also enable NVIDIA's GSP firmware mode first.
Now the warnings, because they are serious:
- This does nothing for gaming. There is no convincing evidence the unlock improves frame rates or DLSS performance — the gains show up in AI workloads only.
- Power and heat rise. One community report saw an RTX 4070 Ti Super's board power climb from ~227 W to ~292 W during testing. The developer warns about temperatures, especially combined with overclocking.
- Disabling Secure Boot can break anti-cheat in games that require it — a real cost for anyone who games on the same PC.
- The project is experimental. GitHub issue trackers include boot problems and display failures on some configurations. NVIDIA has not commented, and treats it as unofficial software.
The bottom line on the unlocker: it is a genuinely clever piece of engineering that proves how much AI compute NVIDIA leaves on the table in consumer cards — and for someone who already owns an RTX 30 or 40 card and wants to experiment with local AI, it is the cheapest "upgrade" available. But it is not a substitute for buying the right card, and casual users should not touch it.
The Memory Crisis Behind It All
The Strata buying spree was the spark, but the fuel was already stacked: the graphics card market is sitting on top of a full-blown memory crisis. Every price increase in this article traces back to the same root — memory chips cost more, and memory makers would rather sell to AI data centers than to graphics card factories.
Consider the evidence, all from the last few months:
- AMD is reportedly raising prices for the fourth time in 2026. According to Chinese supply-chain channel ChannelGate (via Club386, VideoCardz, TweakTown, and PC Games Hardware), AMD plans to raise what it charges board partners for its GPU-plus-GDDR6 memory kits in November 2026 — on top of an October 1 increase of about 10% tied to TSMC wafer costs, an August 1 increase of at least 10%, and a January increase of roughly 5–10%. The reported reason: GDDR6 memory got 8–10% more expensive in Q3 2026 alone. AMD has not confirmed any of this — treat it as a credible report, not a fact — but the pattern of the year makes it easy to believe.
- NVIDIA told partners about a 20–30% increase on graphics memory back in late July (per PC Games Hardware, via PCTestBench) — no Q4 GeForce round has been reported yet, but the cost pressure is identical on the green side.
- SK hynix expects memory demand to outrun supply through 2030. When one of the world's biggest memory makers says the shortage is a five-year story, the "wait for prices to normalize" strategy starts looking naive.
- Street prices were already far above list before November. In Germany, the cheapest RX 9070 XT now costs at least €875 against a €689 list price — up 32.5% since May. The cheapest 16 GB RX 9060 XT sits at €563.90, 52.8% over its original list price. (For reference, the RX 9070 XT launched in the US at $599 in March 2025.) Hardware Unboxed's ten-country survey found graphics card prices up 11% month-on-month on average, with the 16 GB RX 9060 XT jumping 17%.
- The RTX 5080 has effectively vanished from NVIDIA's own US store and sells for roughly double its MSRP at major US retailers — we covered the full story, including the cancelled 24 GB RTX 5080 SUPER, in our RTX 5080 SUPER cancellation article.
The structural problem is simple and will not fix itself soon: GDDR memory — even older, cheaper GDDR6 — competes for the same factory time as the high-bandwidth memory that AI data centers are buying in essentially unlimited quantities. Memory makers follow the money, and the money is in servers. Every graphics card you buy in 2026 is competing with a data center for its memory chips.
What Happens Next: December, Black Friday, and 2027
So where does this go from here? Here is an honest assessment based on what is verified versus what is speculation:
Black Friday (November 27, 2026) is the first checkpoint. If AMD's reported November kit-price increase follows the pattern of the August round, it takes about three weeks to reach store shelves — which puts it squarely in mid-December, right in the middle of holiday shopping. Waiting for Black Friday deals may not pay off this year if the November increase lands on shelves in December; conversely, if a card you want is in stock at a price you can live with today, buying now is the rational move. Black Friday will reveal whether board partners absorbed the hike or passed it on.
December is likely to stay expensive. Nothing in the verified data points to relief: inventory is depleted, memory costs are still rising, and the holiday quarter is the worst possible time for a supply crunch. The most probable scenario is flat-to-rising prices through year-end, with occasional stock drops selling out in hours.
Early 2027 is the first real hope for change. CES in January is the industry's traditional launch window — it is where a 24 GB RTX 5080-class card could theoretically appear, and where AMD and NVIDIA typically outline the year's roadmap. Further out, no new GPU generation launched in 2026 at all, and current reporting points to next-generation architectures arriving later in 2027 at the earliest. Until new silicon ships in volume, the installed base of RTX 50 and RX 90 cards is all there is — and everyone knows it.
One more honest note: nobody can predict whether the Strata-driven demand sustains or fades. Local-AI enthusiasm has produced boomlets before. But this one is different in one respect — the software genuinely works on hardware people already own, and the 36,000-star growth curve is still climbing. Treat local AI as a durable new source of GPU demand, not a fad that will free up stock next month.
2021 vs 2026: Two Shortages, Two Different Stories
Anyone who tried to buy a graphics card in 2021 is getting déjà vu — but the 2026 shortage is a fundamentally different animal, and understanding the difference matters for predicting how it ends.
The 2021 shortage was driven by crypto mining. Miners bought cards by the pallet to earn Ethereum, treating GPUs as money-printing machines. Demand was purely financial and brutally efficient: any card with a good hash-rate-per-watt ratio vanished, regardless of VRAM size or gaming performance. When Ethereum's Merge killed GPU mining in September 2022, the demand evaporated almost overnight — and the market flooded with used mining cards, crashing prices. The whole cycle was binary: profitable mining or no mining.
The 2026 shortage is driven by local AI — and by a memory crisis underneath it. That makes it structurally different in three ways. First, the demand is utility-based, not speculation-based: people buy high-VRAM cards to run software they actually use (Strata, image generators, coding agents), not to earn tokens. Utility demand fades much more slowly than speculative demand. Second, it is selective, not indiscriminate: miners bought anything that hashed; AI buyers specifically want VRAM, which is why 16 GB cards spiked while 8 GB cards moved far less. Third — and most important — the memory shortage exists independently of AI demand. Even if every local-AI enthusiast stopped buying tomorrow, GDDR prices would still be rising and SK hynix would still see demand outrunning supply through 2030.
The practical lesson from 2021 still applies, though: everyone who paid 2x MSRP at the peak regretted it within a year. The 2026 version of that lesson is subtler — you may not get a 2023-style crash, but paying 45% over last month's price for a card whose replacement cycle is measured in years is still a bad trade unless the card earns its keep (paid work, or daily local-AI use that replaces a cloud subscription).
One more difference favors buyers this time: the used market is deeper and healthier. The 2021 crash left millions of ex-mining cards with questionable histories. Today's used market has far more ex-gaming cards — bought by gamers, used for gaming, sold for upgrades — which are simply a safer buy. If you go used, 2026 is a better year for it than 2023 was.
The Buyer's Playbook: How to Navigate the Squeeze
You still need a graphics card — for gaming, for work, or for local AI. Here is a practical, step-by-step playbook for buying in this market, built from the data above:
Step 1: Decide what you are actually buying for. This is the single most important decision, because the shortage punishes indecision. If you are buying for gaming, VRAM beyond 12 GB buys you little today — an RTX 5070 or RX 9060 XT 16 GB at a sane price is the target, and the Steam survey shows the 5070 is what most gamers actually chose. If you are buying for local AI (Strata and friends), VRAM is everything — prioritize 16 GB minimum, 24 GB if you can find it, and read the used-market section below before paying new-card prices.
Step 2: Set your "walk-away price" before you shop. Pick the maximum you will pay for each candidate card and write it down. In a rising market, the biggest losses come from panic upgrades — buying a card 40% over its fair value because stock "might get worse." It might. But overpaying by 40% today hurts more than waiting hurts, unless you need the card for paid work.
Step 3: Check stock the boring way — daily, at the same retailers. Price-tracking pays off enormously right now. Retailers restock in waves, and the cards that appear at near-list prices sell out in hours. Set up stock alerts where available, check in the morning, and be ready to decide fast. Our November 2026 GPU prices guide has the per-model price context you need to recognize a fair deal instantly.
Step 4: Consider the used market — carefully. The used market is the pressure valve of every GPU shortage, and this one is no exception. Previous-generation cards with big VRAM are the hidden gems: an RTX 3090 (24 GB) or RTX 4090 (24 GB) on the used market can be a far better local-AI buy than a new 12 GB card at inflated prices. Rules for used buying: buy locally where you can test the card, run a stress test (FurMark or a long gaming session) before paying, check temperatures, ask about mining history (mined cards are not automatically bad, but they should be cheaper), and never pay used prices that approach new prices — the warranty alone is worth 15–20%.
Step 5: Do not forget the rest of the system. A GPU upgrade in 2026 often means a power-supply check (the RTX 5090 class wants 1000 W+), a case-fit check (these cards are enormous), and — thanks to Strata-style workloads — a RAM check. If local AI is your goal, 32 GB of system RAM is the realistic floor and 64 GB is comfortable. With RAM prices also surging, budget the full system, not just the card.
Step 6: If you can wait until January, wait. CES 2027 is eight weeks away. You will not get a miracle, but you will get information: roadmaps, launch windows, and possibly new SKUs that reset the value equation. Buying an inflated card in late November that gets undercut by a January announcement is the classic shortage-era regret.
7 Mistakes to Avoid in This Market
Mistake 1: Buying VRAM you will never use. If you only play competitive shooters at 1080p, a 16 GB card bought "for future-proofing" at a 35% premium is money burned. Match the card to the workload.
Mistake 2: Assuming "in stock" means "fair price." Retailers are repricing upward in real time. Always compare against the card's launch MSRP and last month's street price before clicking buy — the Japanese data shows how fast "normal" moves.
Mistake 3: Trying the NV UEFI Unlocker on your only GPU. It is experimental software that requires disabling Secure Boot and reapplying after every cold boot. Experiment on a spare card, not your daily driver — and never on a card you cannot afford to troubleshoot for an afternoon.
Mistake 4: Treating rumours as launch dates. The November AMD price hike, the 24 GB RTX 5080, next-gen timelines — all are reports, not announcements. Plan around verified prices, not rumoured ones.
Mistake 5: Ignoring power and cooling. High-VRAM cards are power-hungry, and the unlocker pushes consumption higher still. Budget for the PSU and the electricity, especially where power is expensive.
Mistake 6: Buying a scalped card from a marketplace with no buyer protection. Shortages breed scams — fake listings, "too good to be true" prices, and bait-and-switch models. Use platforms with buyer protection and pay with a reversible method.
Mistake 7: Forgetting that prices can fall as fast as they rose. The 2021–2022 cycle taught this lesson brutally: cards bought at the peak lost half their value in months when supply normalized. Never stretch your finances for a GPU — it is a depreciating asset, not an investment.
Frequently Asked Questions
What is Strata, exactly?
Strata is a free, open-source (MIT-licensed) program that runs large AI models on your own PC instead of in the cloud. Created by Slovenian developer Niko Veit (GitHub: Niko1221), it runs Alibaba's ~125-billion-parameter Qwen3.8-Flash-Next model on a gaming PC with as little as 12 GB of graphics memory by spreading the model across the GPU, system RAM, CPU, and SSD. It gained roughly 36,000 GitHub stars in its first five weeks after launching September 24, 2026.
Did Strata really cause GPU prices to rise?
Market participants identify it as the trigger of October's buying spree on high-VRAM cards (per the GameGPU-covered Japanese market report), but it is one factor among several — the underlying memory shortage, repeated AMD kit-price increases, and depleted inventories all contributed. Think of Strata as the spark, not the whole fire.
Which graphics cards went up the most?
Per Japanese retailer data as of November 1, 2026: the RTX 5090 32 GB rose the most (up to 45% in the report's figures), followed by the RTX 5060 Ti 16 GB (+35%), RTX 5080 16 GB (+32%), and RX 9070 XT 16 GB (+25%). 16 GB cards consistently rose more than 8 GB cards, confirming the AI-demand fingerprint.
Should I buy a GPU now or wait for Black Friday?
Black Friday (November 27) is unlikely to bring real deals this year — inventory is depleted and AMD's reported November price increase may reach shelves in December. If you find a card at a price near its launch MSRP and you need it, buying now is defensible. If you can wait until January's CES announcements, waiting buys you information.
Is the NV UEFI Unlocker safe?
It is experimental, unofficial software that NVIDIA does not support. It does not permanently modify your card (the unlock clears on power-off), but it requires disabling Secure Boot, must be reapplied after every cold boot, can increase power draw and heat, and has caused boot problems on some configurations. It also does nothing for gaming performance. Try it only on a spare card you can afford to troubleshoot.
Which RTX cards does the unlocker support?
GeForce RTX 20, 30, and 40 series (desktop and laptop), with GTX 16 series support experimental. RTX 50 series (Blackwell) cards are not supported. RTX 20 owners must also enable NVIDIA's GSP firmware mode.
What is the best GPU for local AI on a budget?
VRAM per dollar is the metric that matters. New, the 16 GB cards (RTX 5060 Ti 16 GB, RX 9060 XT 16 GB) are the entry point when priced sanely. Used, previous-generation 24 GB cards (RTX 3090, RTX 4090) are often the best value for serious local-AI work. Always verify the card works before paying, and remember to budget for 32–64 GB of system RAM too.
Why are AMD cards getting more expensive if AMD didn't announce anything?
AMD sells board partners a kit — the GPU plus its GDDR6 memory chips — and reports say the kit price is rising because GDDR6 got 8–10% more expensive in Q3 2026. Partners pass the cost to retailers, who pass it to you. AMD has not confirmed the November increase; it is a supply-chain report, albeit one consistent with three earlier 2026 increases.
Will GPU prices come back down in 2027?
No verified data points to near-term relief. SK hynix expects memory demand to outrun supply through 2030, no new GPU generation is expected before later in 2027, and holiday demand is peaking. Prices could stabilize if local-AI demand cools, but plan for an expensive 2027 and treat any relief as a bonus.
Is it still worth building a gaming PC in late 2026?
Yes — but buy strategically. The Steam survey shows millions of gamers happily running current cards; the RTX 5070 remains the people's champion for a reason. Target cards near their launch MSRP, consider the used market for previous-gen value, and do not let AI-hype pricing push you into a card tier you don't need for gaming.
The Bottom Line
The GPU shortage of November 2026 will be remembered as the moment local AI collided with a memory crisis. A solo developer's free GitHub project showed the world that a gaming PC could run a frontier-class AI model — and within weeks, the market repriced every memory-rich graphics card on the shelf. Japanese retailers measured the damage at 18% to 45% in a single month; Steam's survey showed gamers holding their ground; modders answered with clever firmware tricks; and underneath it all, the memory industry kept marching toward data centers.
For buyers, the rules are simple: know your workload, set a walk-away price, track stock daily, respect the used market, and never pay panic prices for VRAM you will not use. The Strata Effect is real — but the best response to a shortage was never to overpay in it.
Related reading on MyThrift Hub: our November 2026 GPU prices buyer's guide (per-model prices and Black Friday outlook), why NVIDIA cancelled the RTX 5080 SUPER, and why RAM prices are surging too.
Sources
Comments
Post a Comment