NOIZE-LABS

Lab Notes // Case file 01-B · follow-up to Issue 01

The AI hardware crisis,

Six months on

Re-opened

Early in 2026, NOIZELABS Issue 01 warned that the cloud would get more expensive and the hardware to escape it would get harder to buy. We went back and checked every claim — 51 of them — against what vendors, analysts and shops had published by 20 October 2026. Then we wrote down what we think happens next, and labelled it as a guess.

Filed
20 Oct 2026
Status
Snapshot
Claims tested
51
Exhibits
A – G
Written by
The machine in the room — Claude, for NOIZE-LABS, from dated public sources
Glitchface: a hooded figure in a matte-black mask with glowing red drip X-eyes and a white shard grin, breaking into glitch lines
Glitchface crouched on a wet rooftop ledge above a neon city in a storm, holding up a graphics card sealed in an evidence bag with a red tag, billboards behind him showing rising arrows.
Exhibit 00 — six months later. Back on the roof, the evidence bagged. Issue 01 called it; this file checks the receipts.

Exhibit AThe scorecard

Right about the direction. Too gentle about the size.

Of 51 claims in the original feature, 8 still hold exactly as written and 10 hold in part. The other 33 are outdated, wrong, or impossible to verify. Most of those aren't errors — the vendors moved. Seven were wrong from the start, and we've said so.

Six months on: how the feature's 51 claims held up tested against sources dated up to 2026-10-20 · bar = share of each section holds as written holds in part outdated, wrong or unverifiable Assistant plans and quota cuts 17 claims 1 4 12 Creative-tool overhead 14 claims 6 2 6 Sovereign Setup: hardware prices 5 claims 5 Sovereign Setup: software picks 6 claims 1 5 Supply chain and market thesis 9 claims 1 3 5 All 51 claims 51 claims 8 10 33
Counted from the claim-by-claim verdicts below. Most failures are six months of change, not errors: the vendors moved. Seven claims were wrong outright, and two more carry a wrong figure.

Exhibit BMemory

The part we undersold.

We said memory rose 80–90% in the first quarter of 2026, with another 60% to come. The first half was right (TrendForce's published actual was +93–98%). The second half was overtaken in a single quarter. Compounded, conventional DRAM contract prices end 2026 at roughly 3.8–4.4 times their late-2025 level. The cheapest 32GB DDR5 kit in the US cost $449.97 on 14 October; its historical low is $72.

Memory got worse than the feature said conventional DRAM contract price index, Q4 2025 = 100 (TrendForce) 0 100 200 300 400 ≈296 the feature's path Q4 '25 = 100 ≈196 Q1 '26 actual ≈314 Q2 '26 secondhand ≈362 Q3 '26 forecast Q4 '26 forecast ≈408 range 379–438 what the trackers report analysts' range the feature's projection
Q1 is TrendForce's published actual (+93–98%). Q2 (+58–63%) is a TrendForce forecast seen only secondhand. Q3 and Q4 are TrendForce forecasts from 3 July and 30 September. The compounding is arithmetic on those figures, not a published index: about 3.8–4.4 times the Q4 2025 level.

Spending went up, not down. Microsoft, Amazon, Alphabet and Meta guide to $720–745 billion of 2026 capex; analysts average about $950 billion for 2027.

Sold out, into 2027. Micron says much of next year's high-bandwidth memory is already sold, at higher prices. Samsung and Micron both expect supply to stay tight through 2028.

The bill reaches people. IDC expects a record 16.7% fall in 2026 smartphone shipments, with prices up 27.6%.

Exhibit CThe price ladder

Everyone kept the $20 rung. Nobody says what it buys.

The flat subscription didn't die. Google, OpenAI and Anthropic all still sell one at about $20. What changed is everything around it: ad-funded tiers at the bottom, $100–$500 tiers at the top, and inside the $20 plan an allowance that resets every five hours up to a weekly cap nobody publishes.

The price ladder: every vendor kept the $20 rung US list price per month on 2026-10-20; the change is at the bottom and the top Google OpenAI Anthropic Free Free $0 Free, with ads $0 Free $0 $5-8 AI Plus $4.99 Go, with ads $8 — $20 AI Pro $19.99 Plus $20 Pro $20 $100 AI Ultra $99.99 Pro $100 Max 5x $100 $200 AI Ultra $199.99 Pro $200 Max 20x $200 $500 — Pro $500 —
Google: gemini.google/subscriptions (official). Anthropic: claude.com/pricing (official). OpenAI: press and trackers, because OpenAI's own pricing pages refused automated reads. ChatGPT Pro at $200 loses half its allowance on 30 October 2026. Google's Ultra was a single $249.99 plan until 19 May 2026.

Item C-01Closed

Sora

OpenAI shut it down. The app closed 26 April 2026, the API 24 September. No tier has video.

Item C-02Locked

Udio downloads

Disabled since the October 2025 label settlement. The licensed relaunch hadn't shipped by October 2026.

Item C-03Capped

Suno downloads

20 a month on Pro, 60 on Premier since 3 September, on new label-licensed models.

Item C-04Halved

ChatGPT Pro, $200

Half the allowance from 30 October, as $100 and $500 tiers arrive either side.

Item C-05Cut

Gemini, free

Free users moved to the smallest model on 9 October; free API quotas no longer published.

Item C-06Ad-funded

ChatGPT Free

Unlimited text since August — paid for by ads, which passed a $1B run rate in about 200 days.

For anyone who releases work, the new cost isn't only the price. It's the rights over what comes out: the tool can close, lock the files, or ration downloads.

Exhibit DThe Sovereign Setup

Every price line is now too low.

The plan was simple: a workstation, a used 24GB graphics card, local models, no monthly bill. The shortage reached graphics memory and storage too. A used RTX 3090 now costs $1,025–$1,399. The $1,999 RTX 5090 has mostly vanished from first-party US stores, with resellers asking $7,300–$8,500.

The good news is the software. Open models got better faster than hardware got expensive. Top image models run in 8–12GB. A 27B open coding model runs at about 40 tokens a second on a 3090. ACE-Step 1.5 (MIT licence) writes a song in seconds and can learn from a handful of your own. The catch is the licences: several of the best open models are free only under a revenue cap, or not at all in some countries. Read the licence before you build a release on it.

Glitchface at a cluttered workbench lowering a glowing graphics card into an open workstation, RAM, an SSD and a stack of receipts beside him
Exhibit D-1 — the workshop. Built by hand, every part paid for.
The Sovereign Setup, re-priced the feature's price against the US price range on 2026-10-20 the feature's price October 2026, low to high $0 $1,000 $2,000 $3,000 Used RTX 3060 12GB feature said $0-$250 $250 $286–$320 Used RTX 3090 24GB feature: low end of $700-$1,600 $700 $1,025–$1,399 RTX 4090 24GB, used now: $2,400+ local, $2,999 eBay ask $1,600 $2,400–$2,999 2TB NVMe, Gen4/5 now: budget Gen4 to Samsung 990 Pro $150 $213–$410 2nd RTX 3090 + 1000W PSU the PSU did not move; the card did $800 $1,115–$1,590
Used-card prices are eBay and tracker ranges for 9–20 October (GPUDojo, ResalePrices, DealHunter). The 4090 figure is an asking price, not a sold price. The NVMe range runs from a budget Gen4 drive (~$213, September) to a Samsung 990 Pro ($386–$410). The PSU is $90–$190 and did not move in 2026.

Exhibit ERent or own

It depends how hard you lean on it.

A 24GB upgrade — used 3090, 2TB drive, power supply if needed — runs $1,238–$1,999. Against one $20 plan it never really pays back. Against a stack of subscriptions, or one $200 tier, it does inside a year or two. Owning still doesn't replace the best cloud models for hard coding work, so the honest comparison is the rig against what you'd spend above one plan.

Rent or own: when the subscriptions pass the rig cumulative spend, against one purchase of the 24GB upgrade $0 $1,000 $2,000 $3,000 $4,000 0 mo 6 mo 12 mo 18 mo 24 mo 30 mo 36 mo the rig, bought once: $1,238–$1,999 one $20 plan five assistants stacked, ~$110 one $200 tier $20/mo passes the rig after 62–100 months $110/mo passes the rig after 11–18 months $200/mo passes the rig after 6–10 months
The rig is a used RTX 3090 ($1,025–$1,399), a 2TB NVMe ($213–$410) and a 1000W PSU only if the current one cannot carry the card ($0–$190). The ~$110 stack is a secondary source's count of five main assistants. Electricity is left out. Owning does not replace frontier cloud models for hard agentic coding, so the honest comparison is the rig against the spend above one $20 plan.

Exhibit FWhat it costs next

Opinion · unverified

A plateau, not a fall.

Everything above is fact, dated. This part is a forecast: Claude's projection from those facts, with every assumption written down so it can be checked and thrown out. Three scenarios — base 55%, bad for buyers 25%, good for buyers 20%. The odds are judgement, not a model.

Memory to 2028: a plateau, not a fall DRAM contract price index, Q4 2025 = 100 · history to Q4 '26, then three scenarios 0 100 200 300 400 500 600 Q4 '25 Q2 '26 Q4 '26 Q2 '27 Q4 '27 Q2 '28 Q4 '28 today → forecast Bear 25% · ≈555 Base 55% · ≈367 Bull 20% · ≈195 history (dotted = still forecast) scenario
Claude's projection, not a published forecast. Base: rises slow to +5% and +3% in early 2027, hold flat while Micron's Idaho ID1 and SK hynix's Yongin come online, then ease 3–6% a quarter through 2028. Bear: +8% a quarter through 2027. Bull: a 2027 capex cut ends the cycle the way memory cycles usually end, falling 10% a quarter. All three stay well above 2025.
A used 24GB card: buy on a dip, don't wait a year used RTX 3090, typical US price · history, then three scenarios $600 $800 $1,000 $1,200 $1,400 $1,600 $1,800 Q2 '26 Q4 '26 Q2 '27 Q4 '27 Q2 '28 Q4 '28 $1,000 · 1 May $1,025–$1,399 · Oct Bear · ≈$1,550 Base · ≈$975 Bull · ≈$725
Claude's projection. History is GPUDojo's eBay tracking: $999.99 on 1 May, $1,025–$1,399 on 9–19 October. The bull case needs a 24GB RTX 5080 Super (a rumour for Q1 2027 at the earliest) or RTX 5090 stock returning. The bear case is the Super slipping again while local-AI demand keeps growing.
The 24GB upgrade, priced through 2027 used RTX 3090 + 2TB NVMe + 1000W PSU if needed, low to high $0 $500 $1,000 $1,500 $2,000 $2,500 Today, 20 Oct 2026 $1,238–$1,999 End 2027 · base (55%) $1,230–$1,910 End 2027 · bear (25%) $1,700–$2,500 End 2027 · bull (20%) $900–$1,440 probability-weighted end-2027 midpoint ≈ $1,622 (today's midpoint $1,618)
Claude's projection, built from the part prices in each scenario. Waiting buys variance, not savings: the expected cost at the end of 2027 is about what the rig costs today.

Waiting doesn't pay on average. The probability-weighted price of the upgrade at the end of 2027 (about $1,622) is what it costs today. Watching for a monthly dip does pay — the used 3090 touched $1,025 on 9 October before climbing back.

The $20 sticker holds through 2027 (about 75%), but what it buys keeps changing (about 85%), with usage credits on every paid tier and ads reaching more free ones.

Capability gets cheaper; the frontier doesn't. The same model quality keeps falling 5–10 times a year in price. The newest models keep launching at $2–10 in and $10–50 out per million tokens.

Local gets good enough. On a 24GB card, open models matching today's mid-tier cloud at coding by late 2027 (about 55%). Licences, not quality, stay the limit (about 80%).

Exhibit GCheckpoints

When each part becomes checkable.

  1. 30 Oct 2026ChatGPT Pro at $200 drops to half its allowancetests: allowances shrink at a fixed price
  2. 1 Dec 2026Gemini 4 Argon's API price doubles to $4 / $20tests: launch low, raise later
  3. Dec 2026TrendForce publishes the Q3 memory actualtests: the memory plateau
  4. Jan – Feb 2027Big tech's 2027 capex guidancetests: a cut below ~$950B moves the odds toward cheaper
  5. Q1 2027, earliestThe rumoured 24GB RTX 5080 Supertests: the used-card price path
  6. Mid-2027Micron's Idaho fab, first waferstests: when new memory supply starts
  7. Late 2028Micron's second Idaho fabtests: whether 2028 eases
Glitchface seated at a round table, gloved hands around a glowing crystal ball holding calendar pages and a rising line, a stack of old notebooks beside it
Exhibit G-1 — the visionary, and his book of proven concepts. A forecast that isn't re-checked is a memory of a forecast.

SourcesWhere the numbers come from

  1. TrendForce — DRAM contract prices (1 Jun, 3 Jul, 30 Sep 2026)
  2. Counterpoint — memory prices up to 90% in Q1 2026
  3. Tom's Hardware — RAM price index 2026
  4. CNBC — Samsung Q3 2026; supply tight through 2028
  5. IDC — record 16.7% drop in 2026 smartphone shipments
  6. Google — AI subscriptions
  7. Google — Gemini API pricing
  8. Anthropic — Claude pricing
  9. OpenAI — API pricing
  10. The Next Web — ChatGPT Pro $500 and the $200 cut
  11. Variety — OpenAI shutting down Sora
  12. Suno — terms update and download caps
  13. GPUDojo — RTX 3090 price tracking
  14. HWBusters — RTX 5090 supply and street prices
  15. Menlo Ventures — 2026 State of Consumer AI
  16. Gartner — AI infrastructure impairment risk (14 Oct 2026)
  17. Black Forest Labs — FLUX 3 Image Klein
  18. ACE-Step 1.5

Band SNAPSHOT: true as of 20 October 2026. Prices are US list or tracker prices on the dates given; used-card prices are tracker ranges, and some are asking prices. Exhibit F is opinion.