๐Ÿ“ฆ Already purchased

Estado real do que ja' foi adquirido vs o que ainda esta' por comprar. Updated 2026-05-03 (ticket #504). Por confirmar com o user, lista pode estar incompleta.
ItemVariantStatusWhereDateNotes
Lista a preencher. User precisa de listar items adquiridos via ticket de follow-up ou diretamente neste card.

Sugestao de items a tracker (cobre tudo no shopping list abaixo):
RPi 54GB ou 8GBunknown???
Official PSU 27W USB-CSC1152 branco / SC1153 pretounknown???
Active CoolerSC1148unknown???
microSDEndurance ou ?unknown???
NVMe SSD?unknown???
M.2 HATGeekworm X1004 / Argonunknown???
CaseArgon ONE V3 / V5 / outrounknown???

โญ Final Analysis: 2026-05-01

Snapshot of the buying decision made on 2026-05-01 for labs.jeremytrindade.com and the next decade of personal-infra projects. This is the "if you only read one section, read this" entry. Reasoning per-question lives in the FAQ section below.

Context

The decision sat at the v0.9 โ†’ v1.0 transition for labs.jeremytrindade.com. The infra is ready, SSL is live, auto-deploy works, and the only remaining bottleneck is hardware: the GCP e2-micro free tier is at 73.4% RAM, the v1.9 signal. The buy needed to:

  • Be cheap enough to act on today.
  • Be capable enough to host labs at v1.0 with years of headroom.
  • Slot cleanly into the sovereign-lab framework (Pi rotates across 5โ€“10 projects in a decade).
  • Avoid regret across that decade, not just optimise day 1.

The questions, decided

QuestionDecisionWhy in one line
4GB CeX โ‚ฌ75, or 8GB online โ‚ฌ185โ€“250, or both?4GB CeXWorkload fits 4GB at 17.4%; RAM is soldered so the 8GB premium buys delay of v2.0, not avoidance.
Argon ONE V5 (โ‚ฌ40,80) or V3 M.2 NVMe (โ‚ฌ53,90)?V3The Pi rotates across many projects. V3's HDMI + power button + IR are โ‚ฌ13 of decade-proof options.
Storage path?NVMe via case slotmicroSD wears out under server-class writes within months. NVMe at v1.0, not at v1.1.

Final shopping list (every product has a direct link)

ItemDirect linkPriceNotes
RPi 5 4GB (Grade B used)CeX product page (Grade B) โ†—โ‚ฌ75,00In-store pickup at Glicรญnias. Phone first, only ~2 units typical. 5yr CeX warranty.
RPi 5 4GB (Grade A fallback)CeX product page (Grade A) โ†—โ‚ฌ85,00Same warranty, less cosmetic wear.
Official PSU 27W USB-C white (SC1152)mauser.pt SC1152 โ†—โ‚ฌ12,92Free PT shipping above โ‚ฌ25.
Argon ONE V3 M.2 NVMe PCIe (ASIN B0CVL9BRZL)Amazon.de official listing โ†— ยท mauser.pt PT fallback (โ‚ฌ59,10) โ†—โ‚ฌ53,90Case + NVMe slot + fan + full HDMI + power button. 74 reviews, 4.4โ˜… (verified 2026-05-01).
Kingston NV3 256GB NVMeAmazon.es search (4โ˜…+) โ†—~โ‚ฌ22RPi 5-compatible, โ‰ค3W TDP.
WD SN770 256GB (fallback)Amazon.es search โ†—~โ‚ฌ23โ€“28Equivalent if Kingston NV3 OOS.
Samsung PM991a 256GB (fallback)Amazon.es search โ†—~โ‚ฌ25โ€“30Equivalent, proven with RPi 5.
Total to running Pi with NVMe~โ‚ฌ164~โ‚ฌ102 cheaper than Scenario B (new 8GB + V5).

How this fits the framework

StageWhatWhen
v1.0 (now โ†’ 1โ€“2 weeks)Pi assembled with NVMe, OS installed, labs migrated from GCP.Acts on this buy.
v1.x (months 1โ€“18)Labs grows. Maybe ESP32 satellites are added. RAM stays well below 70%.Pi runs the full software stack.
v1.9 (sustained RAM > 70%)M720q is bought (~โ‚ฌ182 unit + 32GB RAM). The migration trigger fires.Probably 12โ€“24 months out.
v2.0Labs lives on M720q. Pi gets a new SSD (~โ‚ฌ30) and starts hosting Project B at v1.0.The Pi never goes idle.
v2.x โ†’ โˆžPi rotates across Projects B, C, Dโ€ฆ with V3's HDMI + power button + IR available whenever a project needs them.The โ‚ฌ13 V3 premium pays back any time the Pi runs Kodi, a mini desktop, a home hub, or anything with a screen.

What was explicitly rejected and why

  • Buying the 8GB online: โ‚ฌ110+ premium for RAM you cannot upgrade later, when the next step up is x86 anyway.
  • Buying both Pis (4GB + 8GB): ~โ‚ฌ260โ€“325 spent on two identical-capability ARM boards instead of one Pi + the M720q.
  • Argon ONE V5: cheaper today, but lacks HDMI + power button. Regret risk across the decade is not worth saving โ‚ฌ13.
  • microSD as primary storage: wears out under server writes. Saves โ‚ฌ25 today, costs hours of downtime within months.
  • Tier 1 case (cooler-only): skips NVMe entirely and forces a case re-buy at v1.1.

Open items after this buy

  1. Stand up labs on the Pi at v1.0 once the parts arrive (PSU + V3 + NVMe ship from EU).
  2. Migrate the GCP VM workload across once Pi is on the home network.
  3. Free the GCP VM (still useful as a remote canary or bin for a different role).
  4. Watch RAM. The signal that fires next is sustained > 70%, mirroring what Cockpit currently shows.

Common Argon V3 misidentifications, what NOT to click on

While shopping for the V3 M.2 NVMe PCIe, three near-identical-looking listings come up that are NOT what you want. Verified live on 2026-05-01:

ASINWhat it actually isPriceWhy to skip
B0D1Y4QJQZ โ†—Argon ONE V3 base (NO M.2 slot)โ‚ฌ37,60The "cheap one" is missing the NVMe slot. To get NVMe you would buy the expansion board separately and lose the price advantage and case cleanliness. Not the same product.
B0D2M8BVWK โ†—Same hardware as the recommended B0CVL9BRZL, on .es by GeeekPiโ‚ฌ55,39โ‚ฌ1,49 more expensive for identical hardware. Plus the .es listing flags "Este produto sรณ pode ser enviado para Franรงa e Mรณnaco", probably will not ship to Portugal.
B0D4YCKHGC โ†—Yahboom clone, NOT an official Argon productโ‚ฌ59,96Third-party brand, only 13 reviews at 3.9โ˜…, partial aluminium (top lid alu, base ABS plastic). Also flagged "France and Monaco only" shipping.
The recommended listing remains ASIN B0CVL9BRZL on Amazon.de at โ‚ฌ53,90, the official Argon product, 74 reviews, 4.4โ˜…, highest rating of the four. If Amazon.de does not deliver to your address, the PT fallback is mauser.pt at โ‚ฌ59,10.
Provenance: All prices verified across mauser.pt, BerryBase.de, Kubii.com, Kiwi Electronics, Tiendatec.es, Amazon.de/es, eBay.de, and CeX PT in April-May 2026. The 4 Argon V3 ASIN comparison was verified via Cowork delegation on 2026-05-01 (raw research at D:\claudeui\.research\2026-05-01-argon-case-variants.md). Reasoning per-question is captured in the FAQ section; this analysis is the consolidated executive summary.

Should I wait for RPi 6?

ModelDateStatus
RPi 5 announcement28 Sep 2023Confirmed
RPi 5 launch23 Oct 2023Confirmed, 4GB $60 / 8GB $80
RPi 6 launchQ4 2026 to early 2027 (estimate)Rumour, no official date
Verdict: does not change today's decision. The Raspberry Pi Foundation does not comment on future products as a matter of policy. The estimate is industry speculation based on the ~4-year cycle between Pi 4 (2019) and Pi 5 (2023). The Pi 5 has been on the market for ~2.5 years and is fully mature.
  • 4GB CeX at โ‚ฌ75 is protected in any scenario. By the time RPi 6 launches, this Pi 5 will already have served v1.0 -> v1.x of labs and will be rotating into Project B. โ‚ฌ75 already discounts the resale risk.
  • 8GB new at โ‚ฌ185-250 would hurt more if RPi 6 lands in 6-12 months, since a new high-end board loses resale value faster than a used low-end one. Another argument for the CeX path.
  • M720q (v1.9 / v2.0) is unaffected. It is x86, a completely different role from any Pi generation.
  • Specific rumour: some sources point to Q4 2026 with a native AI NPU. Even if it confirms, it does not impact the 24/7 home-server role we want from the Pi 5.

1 RPi 5 Board

The reference platform for all v1.0 projects. See the 4GB vs 8GB verdict below before deciding which to buy.

1a โ€” RPi 5 8GB ยท New

RetailerCountryPriceShip to PTStockNotes
PCDiga๐Ÿ‡ต๐Ÿ‡นโ‚ฌ119,90FREEโŒ OOSBest PT price โ€” set stock alert
Farnell PT๐Ÿ‡ต๐Ÿ‡น~โ‚ฌ111โ€“117checkcheckOfficial PT distributor
BerryBase.de๐Ÿ‡ฉ๐Ÿ‡ชโ‚ฌ184,90~โ‚ฌ6โ€“8โœ… 100+Best EU stock depth
Tiendatec.es๐Ÿ‡ช๐Ÿ‡ธโ‚ฌ185,95โ‚ฌ5,03 GLSโœ…Max 1 unit/customer
mauser.pt๐Ÿ‡ต๐Ÿ‡นโ‚ฌ190,70FREEโœ… 2 un.PT invoice + free shipping
Kiwi Electronics๐Ÿ‡ณ๐Ÿ‡ฑโ‚ฌ193,71~โ‚ฌ6โ€“10โœ…NL specialist
PTRobotics.com๐Ÿ‡ต๐Ÿ‡นโ‚ฌ207,13FREE (>โ‚ฌ75)โš ๏ธ orderPT, by order
Kubii.com๐Ÿ‡ซ๐Ÿ‡ท~โ‚ฌ210 โš ๏ธ~โ‚ฌ8โ€“12โœ…FR specialist โ€” verify price
YouGet.pt๐Ÿ‡ต๐Ÿ‡นโ‚ฌ249,90FREEโœ…3-year PT warranty
PcComponentes.pt๐Ÿ‡ต๐Ÿ‡นโ‚ฌ269,99checkโœ…Marketplace โ€” verify seller
Pimoroni๐Ÿ‡ฌ๐Ÿ‡ง~ยฃ75 / ~โ‚ฌ88~โ‚ฌ10โ€“15โœ…UK specialist, great accessories
The Pi Hut๐Ÿ‡ฌ๐Ÿ‡ง~ยฃ75 / ~โ‚ฌ88~โ‚ฌ10โ€“15โœ…UK, excellent accessory range
Best buys: Lowest cost this week โ†’ BerryBase.de ~โ‚ฌ191 total. PT invoice + free ship โ†’ mauser.pt โ‚ฌ190,70. PT 3yr warranty โ†’ YouGet.pt โ‚ฌ249,90. Willing to wait โ†’ PCDiga restock at โ‚ฌ119,90.

1b โ€” RPi 5 4GB ยท New

RetailerCountryPriceNotes
Kubii.com๐Ÿ‡ซ๐Ÿ‡ทโ‚ฌ130,50 โš ๏ธ4GB variant on same product page
BerryBase.de๐Ÿ‡ฉ๐Ÿ‡ช~โ‚ฌ125โ€“135Check stock
Tiendatec.es๐Ÿ‡ช๐Ÿ‡ธ~โ‚ฌ120โ€“130Check stock

1c โ€” RPi 5 4GB ยท Used (CeX Portugal)

GradePriceConditionWarrantyLink
Grade Aโ‚ฌ85,00Functionally perfect, minimal cosmetic wear5 yearspt.webuy.com โ†—
Grade Bโ‚ฌ75,00Fully functional, visible cosmetic marks5 yearspt.webuy.com โ†—
โš ๏ธ CeX Glicinias (Aveiro) only โ€” confirm stock by phone before visiting. Only 4GB available used as of April 2026.

1d โ€” 4GB vs 8GB: Which do you actually need?

โš ๏ธ RAM is soldered on the RPi 5 โ€” you cannot upgrade it later. There is no slot, no SODIMM, no HAT that adds usable system RAM. Whatever model you buy is what you have for the life of the board.
Community research verdict: Across server and Docker use cases, 8GB is the recommended minimum โ€” not the premium option. 4GB appears as "bare minimum" only for light single-purpose workloads (Home Assistant alone, Kodi alone, 3B LLMs). As soon as you add multiple services or Docker containers, 4GB is the floor you'll wish you hadn't hit.

The GCP VM runs Cockpit on 1 GB total and uses 711 MB. A headless RPi OS Lite install idles at ~200 MB โ€” similar OS overhead to a cloud VM. So Cockpit migrates directly to 4GB with no issues on day one. The question is purely about where you end up at v1.3โ€“v1.5 when you start stacking services.

What's runningEst. RAM used4GB โ€” % used8GB โ€” % used
RPi OS Lite idle (no apps)~200 MB5% โœ…2.4% โœ…
+ Cockpit (same as GCP today)~711 MB17% โœ…8.7% โœ…
+ Docker + monitoring + backups (v1.3)~1.5โ€“2.5 GB37โ€“61% ๐ŸŸก18โ€“30% โœ…
+ Local LLM 3B model (v1.5)~2.5โ€“3.5 GB61โ€“85% ๐Ÿ”ด30โ€“43% ๐ŸŸก
+ Local LLM 7B model~5โ€“6 GBout of RAM โŒ61โ€“73% ๐ŸŸก
Honest verdict: 4GB is totally fine for Cockpit on day one โ€” the RPi OS doesn't eat your RAM on startup. The concern only appears at v1.3+ when you stack services. 8GB just keeps that moment further away and removes it as a worry entirely. Since RAM is soldered and there's no upgrade path, the question is: how far do you expect to push this Pi before it graduates to v2.0?
If budget is tight: 4GB CeX at โ‚ฌ75 makes complete sense if you're doing Scenario F (Pi = testing ground alongside M720q). In that role the Pi never needs to carry heavy stacks โ€” it just runs the current idea under test. 4GB is limiting only if the Pi is your sole server for everything.

2 Mandatory Accessories

These are not optional โ€” without them the build either won't work or will fail within weeks.

2a โ€” Official PSU 27W USB-C

RetailerPriceColourDirect Link
mauser.ptโ‚ฌ12,92White (SC1152)mauser.pt โ†—
mauser.ptโ‚ฌ13,30Black (SC1408)mauser.pt โ†—
BerryBase.de~โ‚ฌ12โ€“14Bothberrybase.de โ†—
Tiendatec.es~โ‚ฌ13โ€“15Bothtiendatec.es โ†—

2b โ€” Official Active Cooler SC1148

RetailerPriceDirect LinkNote
mauser.ptโ‚ฌ5,99mauser.pt โ†—Board exposed, no case
BerryBase.de~โ‚ฌ6โ€“8berrybase.de โ†—Board exposed, no case
Tiendatec.es~โ‚ฌ6โ€“8tiendatec.es โ†—Board exposed, no case
๐Ÿ’ก If you're buying a case with an integrated fan (Section 4), skip the cooler above โ€” you don't need both.

3 Storage

Path 1 โ€” microSD
โ‚ฌ8โ€“13
  • Simplest start
  • Wears out under heavy writes
  • Fine for v1.0 learning
  • Upgrade to NVMe at v1.1
Path 2 โ€” NVMe SSD โญ
โ‚ฌ20โ€“30
  • Best performance + lifespan
  • Native PCIe 2.0 on RPi 5
  • Recommended from v1.0
  • Needs HAT or Argon case
Path 3 โ€” USB SSD
โ‚ฌ25โ€“35
  • Works if you own one
  • Slower than NVMe
  • Adds a USB dongle
  • Acceptable compromise

microSD โ€” Endurance variants only

โš ๏ธ Use Endurance variants only. Standard microSDs fail in weeks/months as a server OS drive under constant writes.
ProductSizePriceAmazon.esAmazon.de
SanDisk MAX Endurance32GB~โ‚ฌ9โ€“12search โ†—search โ†—
Samsung PRO Endurance32GB~โ‚ฌ10โ€“14search โ†—search โ†—
For microSD, prices change weekly โ€” the search links above filter to 4-star+ and take you straight to the right category. Pick whichever Endurance card is cheapest that day.

NVMe SSD โ€” RPi 5 compatible models

โš ๏ธ Avoid Samsung 990 Pro โ€” draws too much power for the RPi 5 PCIe slot. Prefer drives rated โ‰ค3W TDP.
ModelSizePriceBuy linkNotes
Kingston NV3500GB~โ‚ฌ25โ€“35Amazon.es โ†—Best value, confirmed RPi 5 compatible
WD Blue SN770250GB~โ‚ฌ28โ€“35Amazon.es โ†—Widely used, reliable with RPi 5
Samsung PM991a256GB~โ‚ฌ25โ€“30Amazon.es โ†—Proven compatible with RPi 5 PCIe
๐Ÿ’ก NVMe prices are volatile. The links above show current filtered results โ€” pick the cheapest one from Samsung, WD or Kingston at that moment. Avoid no-brand drives.

M.2 HAT (only if not using an integrated case)

OptionPriceLinkNotes
Official RPi M.2 HAT+~โ‚ฌ10โ€“15mauser.pt โ†—Official RPi Foundation board
Waveshare / GeeekPi M.2 HAT~โ‚ฌ8โ€“13Amazon.es โ†—Budget option, good reviews
๐Ÿ’ก If you're using an Argon ONE V3/V5 or Argon Neo 5 case, the M.2 adapter is already built into the case. Don't buy a separate HAT.

4 Cases & Enclosures

Tier 1 โ€” Cooling only
~โ‚ฌ12โ€“16
  • Official Active Cooler โ‚ฌ5,99
  • + basic case โ‚ฌ6โ€“10
  • No NVMe support
  • Best for: microSD or USB SSD
Tier 2 โ€” NVMe + Cooling โญ
~โ‚ฌ40โ€“45
  • Argon ONE V5 or Neo 5
  • NVMe slot included
  • Fan included
  • Best value all-in-one
Tier 3 โ€” Full mini PC
~โ‚ฌ54โ€“59
  • Argon ONE V3 M.2 NVMe
  • Full-size HDMI ports
  • Physical power button
  • Most polished build

Tier 2 options

ProductPriceDirect LinkNotes
Argon ONE V5 (newest)โ‚ฌ40,80kubii.com โ†—Newest Argon, NVMe PCIe included โš ๏ธ verify specs
Argon Neo 5 M.2 NVMe PCIe~โ‚ฌ42โ€“44Amazon.de โ†—Lower profile, same NVMe + cooling

Tier 3 โ€” Argon ONE V3 M.2 NVMe PCIe

RetailerPriceStockDirect Link
Amazon.deโ‚ฌ53,90โœ…amazon.de/dp/B0D1Y4QJQZ โ†—
mauser.pt (42 units)โ‚ฌ59,10โœ… 42 un.mauser.pt โ†—
What's inside the Argon ONE V3: Aluminium heatsink body + 30mm PWM fan (controlled by RP2040) ยท M.2 NVMe PCIe slot (Key M/B+M, no SATA) ยท 2ร— full-size HDMI Type A ยท 2ร— USB 3.0 + 2ร— USB 2.0 + Gigabit Ethernet ยท Physical power/reboot/shutdown button ยท GPIO access via magnetic lid ยท NVMe heatsink included in box.

5a What is an ESP32?

๐Ÿ”Œ
ESP32 โ€” A tiny WiFi-connected brain for physical things
A microcontroller, not a computer. It runs one program, connects to WiFi, costs โ‚ฌ3โ€“8.

Your RPi 5 is a full computer โ€” it runs Linux, Docker, databases, web servers. But it lives in a box on a shelf, connected to power and ethernet. It can't "feel" the room temperature, detect if a door opened, or switch on a physical relay to control a lamp.

An ESP32 is a tiny chip โ€” smaller than a matchbox โ€” that bridges the gap between the software world (your RPi 5) and the physical world. It has pins you can connect to sensors, buttons, relays, lights, motors, and anything electronic. It also has built-in WiFi and Bluetooth, so it can talk to your RPi 5 wirelessly.

The key difference: an ESP32 doesn't run Linux. It runs one single program โ€” read temperature every 10 seconds and send it over WiFi to the RPi 5. That's it. It boots in milliseconds, uses ~0.25W of power, and costs โ‚ฌ3โ€“8.

๐ŸŒก๏ธ ESP32 (Temperature sensor) โ†’ WiFi/MQTT โ†’ ๐Ÿ–ฅ๏ธ RPi 5 (stores data, sends alerts) ๐Ÿ’ก ESP32 (Controls LED strip) โ†’ WiFi/MQTT โ†’ ๐Ÿ–ฅ๏ธ RPi 5 (gives commands) ๐Ÿšช ESP32 (Door sensor) โ†’ WiFi/MQTT โ†’ ๐Ÿ–ฅ๏ธ RPi 5 (logs events, triggers automations)

The RPi 5 is the brain and hub. The ESP32 is the sensor and actuator at the edge. One RPi 5 can talk to dozens of ESP32 nodes simultaneously, each doing something different in a different room or location.

What can you actually do with an ESP32?

๐ŸŒก๏ธ
Environmental sensors
Temperature, humidity, air quality, COโ‚‚ โ€” read every few seconds, send to RPi 5 to log and alert.
๐Ÿ’ก
Smart lighting
Control LED strips, smart bulbs, relays. RPi 5 sends commands, ESP32 switches the physical circuit.
๐Ÿšช
Presence & security
Door/window sensors, PIR motion detectors, reed switches โ€” trigger automations when someone enters.
โšก
Relay control
Turn real 230V appliances on/off (via relay board). Automations that control your physical home.
๐Ÿ“ก
MQTT data nodes
Publish sensor data to a broker on the RPi 5. Home Assistant reads it. Everything talks to everything.
๐Ÿ”‹
Battery-powered nodes
Deep sleep between readings = months on a battery. Place sensors anywhere, no wiring needed.
ESP32 vs Arduino: Arduino is older, simpler, no WiFi. ESP32 has WiFi + Bluetooth built-in, is faster, cheaper, and is now the industry standard for new projects. Unless you're following a specific Arduino tutorial, start with ESP32.

How to program it: Connect via USB to your computer, write code in MicroPython (easiest โ€” it's just Python) or Arduino C++, click upload. The ESP32 immediately starts running your program every time it powers on.

5b Satellite Nodes โ€” Where to Buy

ESP32 โ€” Recommended first satellite

ProductPriceWhereDirect LinkNotes
ESP32-WROOM-32 DevKit (generic)โ‚ฌ3โ€“6Amazon.esAmazon.es top-rated โ†—Buy 2 โ€” one to test, one to deploy
AZ-Delivery ESP32~โ‚ฌ8โ€“10Amazon.deAmazon.de top-rated โ†—German brand, premium quality, excellent docs
ESP32-S3 DevKit (newer chip)โ‚ฌ5โ€“9Amazon.esAmazon.es โ†—More RAM, USB native, better for future projects
mauser.pt~โ‚ฌ10โ€“15Portugalmauser.pt โ†—PT stock, no shipping wait
PTRobotics.com~โ‚ฌ8โ€“12Portugalptrobotics.com โ†—PT specialist

RPi Zero 2W โ€” Lightweight Linux satellite

For satellites that need a full Linux OS โ€” camera streams, complex daemons, anything that needs pip install.

RetailerPriceDirect Link
Amazon.de~โ‚ฌ18โ€“22amazon.de/dp/B09KLVX4RT โ†—
BerryBase.de~โ‚ฌ18โ€“20berrybase.de โ†—
mauser.pt~โ‚ฌ20โ€“25mauser.pt โ†—
Tiendatec.es~โ‚ฌ18โ€“22tiendatec.es โ†—

Arduino โ€” Physical control & learning

Prefer ESP32 for new projects. Arduino is useful if you're following a specific tutorial or need real-time analog I/O.
ModelPriceAmazon.esBest for
Arduino Uno R3~โ‚ฌ20โ€“25Amazon.es โ†—Learning, prototyping
Arduino Nano~โ‚ฌ15โ€“20Amazon.es โ†—Compact builds
Arduino Mega 2560~โ‚ฌ30โ€“40Amazon.es โ†—Many I/O pins

6 v2.0 Hardware โ€” Lenovo M720q Tiny

When the RPi 5 becomes the bottleneck. Triggers: CPU sustained >70%, ARM64 Docker incompatibilities, need for >8GB effective RAM, or dual storage requirement.

vs GCP VM โ€” how much bigger is this upgrade?

HardwareRAMRAM vs GCPCPUCPU noteCockpit usage
GCP e2-micro (baseline)969 MBโ€” baseline2 shared vCPUs (fractional, throttled)baseline โ€” shared, burst-only73.4% ๐ŸŸก
RPi 5 4GB4096 MB+323% ยท 4.2ร—4ร— A76 @ 2.4GHz4 dedicated cores โ€” no throttling17.4% โœ…
M720q + 32GB RAM32768 MB+3,280% ยท 33.8ร—6ร— i5-8500T @ 3.5GHz boost6 dedicated cores โ€” significantly faster<3% โœ…
The M720q with 32GB RAM is 33.8ร— the RAM of the current GCP server. CPU comparisons are harder to state as a clean multiplier โ€” the e2-micro uses shared fractional vCPUs that throttle under sustained load, while both the RPi 5 and M720q give you dedicated cores that run at full speed all the time. In practice: RPi 5 feels noticeably snappier for real workloads; M720q is in a different league entirely. Your entire Cockpit stack (711 MB) would use under 3% of M720q memory. This is the machine where you stop thinking about hardware constraints.

6a โ€” M720q Unit

RetailerConfigPriceShippingDirect LinkNotes
eBay.de โ€” enter-lesznoi5-8500T / 8GB / 256GB / Win11โ‚ฌ177โ€“187FREE EUebay.de โ†—โญ Best price
eBay.es โ€” yournextiti5-8400T / 8GB / 256GB / Win11โ‚ฌ206,80checkebay.es โ†—Good backup
PCDiga ๐Ÿ‡ต๐Ÿ‡นHP EliteDesk G4 i5-8500 / 16GBโ‚ฌ279,90FREEpcdiga.com โ†—โš ๏ธ Non-T CPU (65W TDP)

6b โ€” RAM Upgrade 32GB

โš ๏ธ The M720q BIOS caps RAM at 2666MHz regardless of kit rating. Buy DDR4-2666 SO-DIMM to avoid paying for speed you won't use.
KitPriceWhereDirect LinkNotes
Crucial 2ร—16GB DDR4-2666 SO-DIMM~โ‚ฌ45โ€“60Amazon.de/esAmazon.es โ†—Best value, confirmed M720q compatible
Kingston 2ร—16GB DDR4-2666 SO-DIMM~โ‚ฌ45โ€“60Amazon.esAmazon.es โ†—Alternative brand

6c โ€” Storage for M720q

UseProductPriceDirect Link
OS + services (SATA 2.5")Samsung 870 EVO 500GB~โ‚ฌ45โ€“55Amazon.es โ†—
Fast workloads (NVMe M.2)WD Blue SN770 500GB~โ‚ฌ35โ€“45Amazon.es โ†—

7 Complete Build Scenarios

Scenario A ยท Minimal
v1.0 โ€” Cheapest possible start
~โ‚ฌ111
RPi 5 4GB Grade B (CeX)โ‚ฌ75,00
Official PSU 27Wโ‚ฌ12,92
Official Active Coolerโ‚ฌ5,99
SanDisk Endurance 32GB microSD~โ‚ฌ9
Basic case (~โ‚ฌ8)~โ‚ฌ8
โš ๏ธ microSD wears out under server load. Plan to add NVMe at v1.1.
Scenario B ยท Solid โญ Recommended
v1.0 โ€” The build you won't regret
~โ‚ฌ266
RPi 5 8GB (mauser.pt)โ‚ฌ190,70
Official PSU 27Wโ‚ฌ12,92
Argon ONE V5 (case+NVMe+fan)โ‚ฌ40,80
Kingston NV3 500GB NVMe~โ‚ฌ30
Runs 24/7 without issues. Argon ONE V5 is all-in-one โ€” no separate HAT needed.
Scenario C ยท Premium
v1.0 โ€” Full mini PC experience
~โ‚ฌ280
RPi 5 8GB (mauser.pt)โ‚ฌ190,70
Official PSU 27Wโ‚ฌ12,92
Argon ONE V3 NVMe (Amazon.de)โ‚ฌ53,90
Kingston NV3 500GB NVMe~โ‚ฌ30
Adds full-size HDMI ports + physical power button vs V5.
Scenario D ยท v1.0 + Satellite
Scenario B + first ESP32 node
~โ‚ฌ278
Everything in Scenario B~โ‚ฌ266
ESP32 DevKit ร—2~โ‚ฌ8โ€“12
Buy 2 ESP32s โ€” one to prototype, one to deploy. You'll always want two.
Scenario E ยท v2.0 Migration
When the RPi 5 hits its limits โ€” Pi gets re-deployed, not discarded
~โ‚ฌ430
Lenovo M720q i5-8500T (eBay.de)~โ‚ฌ182
32GB RAM 2ร—16GB DDR4-2666~โ‚ฌ50
Storage (NVMe + SATA)~โ‚ฌ80โ€“100
The RPi 5 from this project becomes the v1.0 host for the next project. Hardware is never wasted.
Scenario F ยท Dual Setup โšก Jeremy's Plan
Buy both simultaneously โ€” Pi as permanent testing ground, M720q as production from day one
~โ‚ฌ380โ€“400
๐Ÿ–ฅ๏ธ RPi 5 4GB โ€” Testing Ground (~โ‚ฌ103)
RPi 5 4GB Grade B (CeX)โ‚ฌ75,00
Official PSU 27Wโ‚ฌ12,92
Official Active Coolerโ‚ฌ5,99
SanDisk Endurance 32GB microSD~โ‚ฌ9
๐Ÿ–ง Lenovo M720q โ€” Production (~โ‚ฌ282)
M720q i5-8500T / 8GB / 256GB~โ‚ฌ182
32GB RAM DDR4-2666~โ‚ฌ50
NVMe SSD 500GB~โ‚ฌ35โ€“45
How this works: New project idea โ†’ always starts on the Pi (v1.0). Once proven and stable, it moves to the M720q (v2.0). The Pi is never idle โ€” it's permanently the experimentation machine. You live the full framework from day one, with both stages running in parallel. This is a smart setup.

FAQ Why not skip the RPi 5 and start on the M720q directly?

The M720q is โ‚ฌ177 and more powerful than a โ‚ฌ190 RPi 5 8GB. It's a fair question โ€” here's the full honest answer.

๐Ÿ“ 2026-05-01 live decision: 4GB at CeX Aveiro today, 8GB online for double, or buy 4GB now and 8GB later (two Pis)?

Verdict: drive to CeX Glicรญnias today and buy the 4GB at โ‚ฌ75. The 8GB online does not buy migration runway you actually use, and buying both Pis is the only path that wastes money. Decision asked while standing up labs.jeremytrindade.com.

Why the 4GB is enough for labs.jeremytrindade.com

Labs is a software-first project alongside Cockpit-class workloads. Cockpit, the bigger of the two, runs on a GCP e2-micro with 1 GB total RAM and uses 711 MB. Move that exact workload to a RPi 5 4GB and it occupies 17.4% of memory with 3.3 GB free. Labs is lighter than Cockpit. There is no measured workload currently in the stack that 4GB cannot host comfortably.

Side-by-side: 4GB CeX vs 8GB online

Factor4GB CeX โ‚ฌ758GB online โ‚ฌ185โ€“250
Priceโ‚ฌ75โ‚ฌ185โ€“250
Stockโœ… today, in personโœ… days of shipping
Warranty5 years (CeX PT)1โ€“3 years
Headroom for labs today3.3 GB free7.3 GB free
Headroom in 2 yearsTight at v1.5+ if running 7B LLMComfortable
Migration target when maxedโ†’ M720q (32GB, x86)โ†’ M720q (32GB, x86), same
RAM is soldered. The Pi cannot be RAM-upgraded later regardless. So when the Pi becomes the bottleneck, the framework's answer is the M720q, not a bigger Pi. The โ‚ฌ110+ premium for an 8GB Pi is buying a delay of v2.0, not avoiding it. The right next step is M720q at ~โ‚ฌ182, which gives you 33.8ร— the GCP VM RAM in one jump, not a marginal 2ร— upgrade on the same constrained ARM platform.

Why buying both Pis is the worst option

Two Pis at โ‚ฌ75 + โ‚ฌ185โ€“250 = ~โ‚ฌ260โ€“325 leaves you with two 4-core ARM boards. What you actually want as your second machine is the M720q (x86, 32GB, no ARM Docker incompatibilities) for roughly the same money. Two Pis is the only path that spends real money without unlocking a single capability you don't already have.

The honest catch

4GB is permanent. If you later decide to run a 7B local LLM on the Pi specifically (not on the M720q), you cannot, that needs ~5โ€“6 GB just for model weights. The framework's answer to "I want big LLM inference" is M720q with 32GB. So the only real loss from choosing 4GB is: you lose the option of running heavy Docker stacks plus LLMs on the Pi tier. Given that the v2.0 path already exists in the framework, that loss is not a real loss.

Which case to pair with the 4GB

CeX only sells the bare board. The case is a separate online order. The framework's whole point is that the Pi gets re-deployed across 5โ€“10 projects in a decade, so the case decision needs to survive role-changes, not just optimise for the current role.

TierCasePriceExtras over V5Verdict
1Active Cooler + basic case~โ‚ฌ12โ€“16(none, no NVMe slot at all)โŒ skip, microSD wears out
2Argon ONE V5โ‚ฌ40,80(baseline)Cheapest case-with-NVMe, no HDMI/power button
3 โญArgon ONE V3 M.2 NVMeโ‚ฌ53,902ร— full-size HDMI, physical power button with soft-shutdown, IR receiver, magnetic GPIO lidโœ… regret-proof for the decade
The regret math (V3 over V5): for labs.jeremytrindade.com today the V3 extras are unused (headless server, SSH in). But the framework's whole architecture is "this Pi rotates across many projects". When labs migrates to v2.0, the Pi gets a new SSD (~โ‚ฌ30) and goes to Project B, then C.

Likely future projects: which want V3's extras?

Future projectWants V3 extras?
Pi-hole / DNS nodeNo
Camera / surveillance nodeNo
Edge AI inferenceNo
Home AssistantMaybe (power button for safe reboot)
Sandbox / dev boxMaybe (HDMI is convenient)
Retro gaming / media (Kodi, RetroPie)Yes, HDMI + IR + power button
Mini desktop occasionalYes, HDMI + power button
Home hub with screenYes, HDMI

Probability that at least 1 of the next 5โ€“10 projects benefits from V3's extras across a decade: very high (>80%).

Cost of being wrong each way:
โ€ข Buy V5, regret it later โ†’ buy V3 anyway (โ‚ฌ53,90), V5 ends up in a drawer or sold at loss. Effective: โ‚ฌ94,70 in cases + redundant hardware.
โ€ข Buy V3, never use the extras โ†’ paid โ‚ฌ13 for unexercised options. No redundant hardware.

The V3 is โ‚ฌ13 more in exchange for zero regret risk across the framework's intended decade-long lifecycle.

Final shopping list with the case

ItemWherePrice
RPi 5 4GB Grade BCeX Glicรญnias todayโ‚ฌ75,00
Official PSU 27Wmauser.ptโ‚ฌ12,92
Argon ONE V3 M.2 NVMe (case + NVMe + fan + HDMI + power button)Amazon.deโ‚ฌ53,90
Kingston NV3 256GB NVMeAmazon.es~โ‚ฌ22
Total to running Pi with NVMe~โ‚ฌ164
~โ‚ฌ102 cheaper than Scenario B (which uses the new 8GB Pi at โ‚ฌ190,70) for better case longevity (V3 over V5), same NVMe path, same v1.0 to v1.9 longevity, plus regret-proof extras for whatever the Pi rotates to next. If Kingston NV3 is out of stock, WD SN770 or Samsung PM991a are equivalent picks (all โ‰ค3W TDP, all confirmed compatible with the RPi 5 PCIe slot).

Recommended sequence from this decision

1. Today: Drive to CeX Glicรญnias, buy 4GB Grade B at โ‚ฌ75 (or Grade A at โ‚ฌ85). Phone first, only 2 units typical.
2. This week: Order PSU 27W (โ‚ฌ12,92) at mauser.pt + Argon ONE V3 M.2 NVMe (โ‚ฌ53,90) at Amazon.de + Kingston NV3 256GB (~โ‚ฌ22) at Amazon.es. Total to running Pi with NVMe: ~โ‚ฌ164.
3. v1.0 โ†’ v1.x: Stand up labs.jeremytrindade.com on the 4GB. Watch RAM. Same v1.9 signal as Cockpit: sustained >70%.
4. When v1.9 hits: Buy M720q (~โ‚ฌ182) + 32GB RAM, NOT a bigger Pi. The 4GB Pi becomes the permanent testing ground.

Decision tree updated for this question

QuestionAnswer
Will the 4GB run labs.jeremytrindade.com today?Yes, with 3+ GB headroom
Will the 4GB run labs in 12 months as it grows?Yes, until you stack 3+ heavy Docker containers or local LLMs
Is the 8GB worth a โ‚ฌ110+ premium?No, the next step up is M720q anyway
Is buying both Pis a good plan?No, the second machine should be x86
What is the right buy today?4GB CeX โ‚ฌ75
โšก The M720q uses 7โ€“10ร— more electricity, every single day

The RPi 5 draws 5โ€“15W under normal load. The M720q Tiny with i5-8500T draws 35โ€“65W. For a home server running 24/7, that difference compounds fast.

๐Ÿฅง RPi 5 โ€” ~10W average
  • ~88 kWh/year
  • ~โ‚ฌ18/year (at โ‚ฌ0.20/kWh PT)
  • ~โ‚ฌ54 over 3 years
  • Silent even under load
๐Ÿ–ง M720q โ€” ~50W average
  • ~438 kWh/year
  • ~โ‚ฌ88/year (at โ‚ฌ0.20/kWh PT)
  • ~โ‚ฌ264 over 3 years
  • Fan noise under load

Over 3 years, you pay ~โ‚ฌ210 more in electricity for the M720q. The M720q "only โ‚ฌ177" becomes "effectively โ‚ฌ387" if it runs a project that could fit on a Pi for 3 years. The Pi's cost savings are not just upfront โ€” they compound daily.

๐Ÿ”Œ The M720q has no GPIO pins โ€” it cannot connect to the physical world directly

The RPi 5 has a 40-pin GPIO header. You can directly wire sensors, relays, LEDs, motors, and displays to it. The M720q is a standard PC โ€” it has USB ports and that's it.

Every project in sovereign-lab is designed to potentially extend into the physical world (sensors, automation, presence detection). The RPi 5 can do that directly at v1.x. With the M720q you'd always need an ESP32 or Arduino as a middleman even for simple things โ€” the Pi can sometimes just do it natively.

๐Ÿง  Starting on constrained hardware makes your software better

When you build on a Pi, every architectural decision is under pressure. "Does this need to run as a container? Can I use a lighter stack? Is this query efficient?" โ€” these questions get asked naturally when you only have 8GB RAM and an ARM chip.

When you graduate to the M720q, you already know the software works. You get the speed boost as a reward, not a crutch. If you build exclusively on x86 with 32GB RAM from day one, you tend to accumulate bloat you never notice until you try to migrate or replicate.

The framework's v1.0 stage on RPi 5 is intentionally constrained. The constraint is the feature.

โ™ป๏ธ The Pi pays for itself again at v2.0 โ€” and again, and again

When your project moves to the M720q at v2.0, the Pi isn't sold or put in a drawer. It becomes the v1.0 host for the next project, which costs only ~โ‚ฌ25โ€“35 for a new SSD. The RPi 5 you buy today will run 5โ€“10 projects over a decade.

If you go M720q-only from day one: you need another M720q for your second project, and another for your third. Or they all share one machine (which works, but you lose hardware isolation and the low-power "always testing" benefit).

The RPi 5 is not "the cheap starter kit you'll throw away." It is a permanently useful node in your lab. The Pi runs all new ideas. The M720q runs proven production workloads. They are complementary, not interchangeable.
โœ… So what should I actually do? (Jeremy's specific situation)

Your instinct is right: buy the RPi 5 4GB at โ‚ฌ75 (CeX) + the M720q at ~โ‚ฌ182 together. This is Scenario F above and it's the smartest setup in this guide.

Here's why it works perfectly for you:

๐Ÿฅง RPi 5 4GB โ€” Your lab bench
  • All new project ideas start here
  • v0.1 โ†’ v0.9 โ†’ v1.0 always on Pi
  • โ‚ฌ75 with 5-year warranty (CeX)
  • Low power: always-on is fine
  • GPIO available if needed
  • The "is this worth building?" stage
๐Ÿ–ง M720q โ€” Your production rack
  • Proven projects live here (v2.0+)
  • 32GB RAM, x86, fast NVMe
  • No ARM64 Docker issues
  • Multiple services, no resource pressure
  • Runs while Pi tests the next thing
  • The "this is infrastructure" stage

The only thing to avoid: don't skip the Pi stage and go M720q-first for new ideas. That defeats the purpose of the framework. The Pi's constraints are what force good decisions. Once something is proven and depended upon โ€” then it earns the M720q.

๐Ÿ’ธ Money is tight right now โ€” should I go straight to the M720q and buy a RPi 5 later?

Yes โ€” this is a legitimate and smart path. Buying the M720q first and adding the RPi 5 later still gets you to the exact same destination: both machines in the lab, working the way the framework intends. You just arrive in reverse order.

M720q first (~โ‚ฌ182 now)
  • Run v1.0 AND v2.0 projects on one machine
  • No need to migrate hardware later
  • Full x86, 32GB RAM potential from day one
  • When RPi arrives: it slots as the permanent testing node
  • Framework still works โ€” M720q plays both roles
Downsides to know about
  • ~โ‚ฌ88/year electricity vs โ‚ฌ18/year on Pi
  • No GPIO pins โ€” physical world needs an ESP32 middleman
  • You miss the "build lean on constraints" experience until Pi arrives
  • All projects share one machine until you get the Pi

The M720q can run v1.0 projects perfectly fine โ€” it just doesn't have the low-power always-on charm of a Pi, and no GPIO for direct sensor wiring. For software-first projects (self-hosted apps, Docker services, Home Assistant, etc.) it handles everything.

When you later buy a RPi 5: the 4GB from CeX (~โ‚ฌ75) is all you need. By the time you can afford it, your first projects will already be on the M720q, and the Pi slots in perfectly as the permanent "new ideas" testing ground โ€” which is exactly the role it's designed for in the framework.

On "a decent RPi 5 or higher": As of 2026, the RPi 5 8GB is the top of the RPi lineup โ€” there is no higher model. For your testing-ground role, the 4GB from CeX at โ‚ฌ75 is genuinely sufficient. The 8GB is worth it if you ever want to run heavier workloads directly on the Pi, but for a lab bench that's just testing ideas before they graduate to the M720q, 4GB is plenty.

Decision tree โ€” which order to buy

Can afford ~โ‚ฌ380 now? โ†’ Scenario F (buy both) โ€” best setup
Can afford ~โ‚ฌ182 only? โ†’ M720q first โ†’ RPi 5 CeX โ‚ฌ75 when possible
Can afford ~โ‚ฌ103 only? โ†’ RPi 5 CeX first (Scenario A) โ†’ save for M720q
Can afford ~โ‚ฌ75 only? โ†’ RPi 5 CeX 4GB + microSD + PSU โ†’ start learning today
๐Ÿ“Š Is the RPi 5 4GB actually enough for a real app? My current GCP VM says so.

The Cockpit personal dashboard (Flask + SQLite + AI queue + email processing) runs on a GCP e2-micro with 1 GB of RAM. Measured live on 2026-04-30:

๐Ÿ“ GCP e2-micro (current)
  • RAM: 711 MB used / 969 MB total
  • Memory usage: 73.4%
  • CPU load (1m): 0.04 โ€” idle
  • Disk: 6.12 GB / 29.3 GB used
  • 2 shared vCPUs (throttled)
  • Uptime: 10.9 days stable
๐Ÿฅง RPi 5 4GB (target)
  • RAM: 4096 MB available
  • Same workload = ~17% memory
  • 3.3 GB headroom to grow
  • Quad-core A76 @ 2.4GHz โ€” dedicated
  • NVMe storage (vs cloud disk I/O)
  • Always-on, ~โ‚ฌ18/year electricity

The RPi 5 4GB has 4ร— the RAM of the current production VM and a significantly faster CPU. The entire Cockpit stack uses 711 MB โ€” that leaves 3.3 GB free on 4GB. Even if the app grows 3ร— in memory usage, it still fits comfortably.

The 8GB is not needed for this workload. It makes sense if you plan to run many heavy Docker containers simultaneously โ€” but for a personal app stack (Cockpit + email + SQLite + AI queue), 4GB has years of headroom.

Watch this number: Memory at 73.4% on the 1GB GCP VM is approaching the yellow threshold (85%). As Cockpit grows โ€” more contacts, heavier AI queue, more concurrent features โ€” this will push toward the migration trigger. This is exactly the "v1.9 approaching" signal from the framework. The RPi 5 4GB migration removes this ceiling entirely.

8 EU Retailer Reference

All retailer links go directly to the RPi 5 8GB product page, not the homepage.

Specialist RPi / Maker Retailers

RetailerCountryShips to PTRPi5 8GBDirect Product LinkStrengths
mauser.pt๐Ÿ‡ต๐Ÿ‡นFREEโ‚ฌ190,70RPi 5 8GB โ†—Best PT all-rounder, broad stock
PTRobotics.com๐Ÿ‡ต๐Ÿ‡นFREE (>โ‚ฌ75)โ‚ฌ207,13RPi 5 8GB โ†—PT components specialist
YouGet.pt๐Ÿ‡ต๐Ÿ‡นFREEโ‚ฌ249,90RPi 5 8GB โ†—3-year PT warranty
Farnell PT๐Ÿ‡ต๐Ÿ‡นcheck~โ‚ฌ111โ€“117RPi 5 8GB โ†—Official PT distributor
BerryBase.de๐Ÿ‡ฉ๐Ÿ‡ช~โ‚ฌ6โ€“8โ‚ฌ184,90RPi 5 8GB โ†—Best EU stock depth, fast
Tiendatec.es๐Ÿ‡ช๐Ÿ‡ธโ‚ฌ5,03 GLSโ‚ฌ185,95RPi 5 8GB โ†—Spain, fast GLS delivery
Kubii.com๐Ÿ‡ซ๐Ÿ‡ท~โ‚ฌ8โ€“12~โ‚ฌ210 โš ๏ธRPi 5 8GB โ†—FR specialist, great accessories
Kiwi Electronics๐Ÿ‡ณ๐Ÿ‡ฑ~โ‚ฌ6โ€“10โ‚ฌ193,71RPi 5 8GB โ†—NL specialist, reliable
Pimoroni๐Ÿ‡ฌ๐Ÿ‡ง~โ‚ฌ10โ€“15~ยฃ75RPi 5 8GB โ†—UK, unique accessories
The Pi Hut๐Ÿ‡ฌ๐Ÿ‡ง~โ‚ฌ10โ€“15~ยฃ75RPi 5 8GB โ†—UK, excellent accessory range

Retailers to Avoid

RetailerReason
chip7.ptDoes not carry RPi or maker hardware โ€” confirmed zero results
electrofun.ptNo RPi 5 boards โ€” accessories only
PcComponentes marketplaceโ‚ฌ269,99 vs โ‚ฌ190 elsewhere โ€” no justification for the premium

9 Future Projects Cost

When this project moves to v2.0 and the RPi 5 is freed, starting the next project costs almost nothing.

โ™ป๏ธ Cost to start the next project after v2.0 migration

~โ‚ฌ25โ€“35
ItemStatusCost
RPi 5 boardโœ… re-usedโ‚ฌ0
PSU 27Wโœ… re-usedโ‚ฌ0
Cooling / caseโœ… re-usedโ‚ฌ0
New NVMe SSD 256โ€“500GBโŒ new~โ‚ฌ25โ€“35
One RPi 5 + accessories serves 5โ€“10 projects over a decade. Each re-deployment costs only the price of a new SSD โ€” ~โ‚ฌ25โ€“35 โ€” and you have a fully capable server ready to run.
Project A: RPi 5 โ†’ v1.0 โ†’ v1.x โ†’ v1.9 โ†’ v2.0 (moves to M720q)
                                         โ†“ Pi freed โ†’ ~โ‚ฌ30 for new SSD
Project B:              uses same Pi โ†’ v1.0 โ†’ v1.x โ†’ v2.0
                                                              โ†“ Pi freed again โ†’ ~โ‚ฌ30
Project C:                                 same Pi โ†’ v1.0 โ†’ ...