AI & ENTERPRISE
RTX 4090
Twice the memory, all the power.
The first dedicated RTX 4090 VRAM upgrade service in the United Kingdom.
From £900.
Blackstone Repair is the first dedicated provider of RTX 4090 24GB to 48GB VRAM upgrade services in the United Kingdom, to the best of our knowledge at the time of launch. This service is not endorsed by or affiliated with NVIDIA Corporation.
From £900 when the original 12 memory chips are healthy and can be transplanted. Final price depends on parts required and is always agreed before work begins.
Modern AI workloads run out of memory before they run out of compute.
The RTX 4090 has the speed for serious local AI, but NVIDIA gave it only 24GB.
OFF THE CLOUD
AI that runs in the room with you.
A 48GB card turns a desktop into a serious AI workstation, with the models running on the machine in front of you.
The bigger the model, the smarter it is, and the more memory it needs.
A private assistant that reads everything you give it.
Legal contracts and case files, patient records, unpublished university research, client code under NDA. It answers from those documents without a single word leaving the machine.
Image and video generation with no meter running.
Run as many variations as you want, overnight if the job calls for it. The only cost is the electricity.
A coding assistant with the whole codebase in view.
48GB holds a capable model and enough context to keep a whole repository in view at once, not just the file you have open.
70B-class models, the tier where answers get good.
At Q4 quantisation a 70B model needs around 43GB. That tier never fits in 24GB, and it is the reason this upgrade exists.
Why local beats the cloud for this
Private by physics
Nothing is uploaded, logged, or retained by a provider. Data cannot leak from a server it never reached.
A single cost
No per-token billing, no monthly tier, no invoice spike after a heavy week. Buy the memory once, run it flat out.
No terms attached
No rate limits, no model deprecations, no usage policies shifting underneath the work. The models you choose, configured your way, available offline.
Built to multiply
These are 2-slot blower cards, made for dense systems. Four in one workstation is 192GB of VRAM for a fraction of the price of one enterprise card.
If everything you run fits in 24GB, you do not need this. The fit checker below tells you honestly.
What happens to your card
- 1
Register and get your quote
Tell us about your card. We reply personally with a fixed quote, split into parts and our work, plus the full terms.
- 2
Ship it or drop it off
A deposit covering the parts secures your build. Then insured Royal Mail shipping with the cover figure shown before you book, or drop off in Milton Keynes by appointment.
- 3
Teardown and validation
We photograph everything on arrival and check your core first. If it cannot take the upgrade, your deposit comes straight back, before any parts are committed to the build.
- 4
The transplant
Hot air on a controlled preheat curve lifts your core and memory from the old board, with thermocouples tracking every stage. Cleaned and reballed, they move to the new clamshell board alongside the additional modules that double the capacity. Every module is then tested.
TEST PROTOCOL
Tested before it ships
Test protocol v1.0, June 2026
- 5.1
Driver recognition
Standard NVIDIA drivers, no special configuration. The system sees 48GB or the card does not ship.
nvidia-smi reports 48GB at full PCIe link width - 5.2
Memory validation
Industry diagnostic tooling tests all 24 modules, with zero errors tolerated.
every module tested, zero errors tolerated - 5.3
Sustained burn-in
Sustained AI generation workloads at full power draw, with rest and cool-down periods, across a 48-hour test window. Stability under load, not on paper.
AI workloads across a 48-hour test window - 5.4
Thermal mapping
Front and rear module temperatures measured at equilibrium. The clamshell design puts 12 modules on the rear, so we check them all.
memory junction temperatures within spec under sustained load - 5.5
The report
Every card ships with its own test report. It doubles as the warranty baseline: if the card ever comes back, that report is what we measure against.
signed, dated, ships in the box
Balance and insured return
Your deposit covered the parts. The balance, our work, is due only when your card has passed everything. It returns insured, in plain unmarked packaging, with its signed test report in the box.
Every report carries the date, the test versions, and the signature of the engineer who ran it.
As builds complete, their test reports, thermal captures and bench data publish here, card by card. First builds land autumn 2026.
It stays your card. Same core, same silicon, twice the memory.
Will it fit, and is it worth it
| Model | Quant | Needs | 16GB | 24GB | 32GB | 48GB |
|---|---|---|---|---|---|---|
| 7B classQuick drafts and simple questions | Q4 | ~4GB | fits | fits | fits | fits |
| 13B classA capable day-to-day assistant | Q4 | ~8GB | fits | fits | fits | fits |
| 13B classfull precision, 32GB and up | FP16 | ~26GB | does not fit | does not fit | fits | fits |
| 30B classStrong reasoning and document analysis · tight at long context on 24GB | Q4 | ~18GB | does not fit | fits | fits | fits |
| 30B classhigher precision, needs 48GB | Q8 | ~32GB | does not fit | does not fit | does not fit | fits |
| 70B classNear-frontier quality, the best on one card | Q4 | ~43GB | does not fit | does not fit | does not fit | fits |
| 70B classneeds multi-card | Q8 | ~75GB | does not fit | does not fit | does not fit | does not fit |
7B class
Quick drafts and simple questions
Quant
Q4
Needs
~4GB
13B class
A capable day-to-day assistant
Quant
Q4
Needs
~8GB
13B class
full precision, 32GB and up
Quant
FP16
Needs
~26GB
30B class
Strong reasoning and document analysis · tight at long context on 24GB
Quant
Q4
Needs
~18GB
30B class
higher precision, needs 48GB
Quant
Q8
Needs
~32GB
70B class
Near-frontier quality, the best on one card
Quant
Q4
Needs
~43GB
70B class
needs multi-card
Quant
Q8
Needs
~75GB
70B-class at Q4_K_M quantisation, 8K context, needs about 43GB: fits with headroom on 48GB, does not fit on 24GB. Figures cross-checked against three published 2026 VRAM guides, as of June 2026.
Cloud, per year
£590
Upgrade, one-off
£900
Payback
18.3 months
If you only need occasional bursts, rent. If the card works most days, owning wins.
Cloud rate as of June 2026.
All roads lead to 48GB
The 48GB upgrade
Stock RTX 4090
RTX 6000 Ada 48GB
Cloud rental, 48GB class
DIY kit route
VRAM
48GB
VRAM
24GB
VRAM
48GB
VRAM
48GB, rented
VRAM
48GB, if the build works
Bandwidth
~900 GB/s
Bandwidth
~900 GB/s
Bandwidth
960 GB/s
Bandwidth
960 GB/s class4
Bandwidth
~900 GB/s, same design
Cost
From £9001
Cost
~£1,700 used market3
Cost
£7,172 new, ~£6,300 used3
Cost
~£590 a year at 4h/day4
Cost
~£420 kit, plus modules, plus £2,000+ rework gear5
Warranty
90 days, modules included
Warranty
out of warranty, returns vary by seller
Warranty
manufacturer terms
Warranty
not applicable, rented
Warranty
none5
Note
send a card in, or buy one ready-built1
Note
70B-class will not load2
Note
ECC, pro drivers
Note
data leaves the building
Note
slow, demanding work, one slip can end the core
The 48GB upgrade
VRAM
48GB
Bandwidth
~900 GB/s
Cost
From £9001
Warranty
90 days, modules included
Note
send a card in, or buy one ready-built1
Stock RTX 4090
VRAM
24GB
Bandwidth
~900 GB/s
Cost
~£1,700 used market3
Warranty
out of warranty, returns vary by seller
Note
70B-class will not load2
RTX 6000 Ada 48GB
VRAM
48GB
Bandwidth
960 GB/s
Cost
£7,172 new, ~£6,300 used3
Warranty
manufacturer terms
Note
ECC, pro drivers
Cloud rental, 48GB class
VRAM
48GB, rented
Bandwidth
960 GB/s class4
Cost
~£590 a year at 4h/day4
Warranty
not applicable, rented
Note
data leaves the building
DIY kit route
VRAM
48GB, if the build works
Bandwidth
~900 GB/s, same design
Cost
~£420 kit, plus modules, plus £2,000+ rework gear5
Warranty
none5
Note
slow, demanding work, one slip can end the core
- From £900 on a send-in RTX 4090 with healthy memory. Ready-built units on donor cards we source are quoted per build. The final quote is always agreed before work begins.
- See the fit checker above for what fits where.
- PriceSpy UK best price, eBay UK sold listings, June 2026.
- RunPod RTX 6000 Ada hourly rate, about £0.59 an hour, June 2026.
- One kit vendor’s published price, about £420, June 2026. Their terms state: “PCB Kits are not covered under warranty of any kind.”
Published third-party figures, as of June 2026. Our own measured numbers replace these as our builds complete.
Not for gaming. Games do not use 48GB.
Not for sim racing, video editing, or general use.
Not for workloads that already fit in 24GB. The extra memory would sit idle.
Built in the UK
No borders, no customs
Your card never leaves the country. No import VAT on the service, no duty, no seizure risk.
UK consumer law applies
The Consumer Rights Act 2015 covers this work. Your statutory rights survive everything in our terms.
Insured both ways
Shipping cover is shown in pounds at the point of booking, before you commit. The return leg is insured to match.
You talk to the engineer
The person who quotes your card is the person who builds it and signs its report.
Sending your card abroad, or buying a ready-modified one from overseas, typically costs hundreds of pounds more once shipping, VAT, duty and the markup are added. It also means weeks of waiting while the card sits in transit and at customs checkpoints.Ours is a tracked Royal Mail parcel, there and back.
Sending your card abroad, or buying a ready-modified one from overseas, typically costs hundreds of pounds more once shipping, VAT, duty and the markup are added. It also means weeks of waiting while the card sits in transit and at customs checkpoints.Ours is a tracked Royal Mail parcel, there and back.
PRICING
What the price includes
From £900.
Two ways to get one: send a 4090 in for the upgrade, or have us build one on a donor card we source and test. Ready-built units are quoted per build.
A £400 deposit covers your parts and starts the build. The £500 balance is for our work, due only when your card passes.
- Teardown, core and memory validation, photographed on arrival
- New clamshell board, the latest 3.0 revision, in a 2-slot blower form factor built for multi-card systems
- 12 additional Micron D8BZC modules fitted alongside the original 12
- VBIOS configured for the doubled memory
- The full test protocol and a signed report in the box
- 90-day warranty covering our work and the modules we fit
The only thing that can add to the price is replacement memory, at our published chip prices, added to the parts deposit, and only with your agreement after diagnosis.
Card faulty? Most faults live on the board rather than the core, and core failures are rare. The upgrade moves your core and memory to a new board anyway, so a faulty card is often a perfect candidate. We repair and upgrade in a single job.
Warranty at a glance
Returned insured, in plain unmarked packaging.
Register your interest
We reply personally within two working days.
Also available: the RTX 4080 SUPER 16GB to 32GB upgrade.
See the 4080 SUPER upgradeQuestions, answered properly
No. The card runs on standard NVIDIA drivers and reports 48GB out of the box.
No. Your core and bandwidth are unchanged, and published benchmarks show parity with stock across AI workloads. Your report shows your own numbers.
The 48-hour burn-in exists to answer this. For business-critical deployments we recommend a spare card, as we would with any hardware.
Very low, as far as we can tell, but we have just started and will not invent a number before the builds add up. What stands behind every card instead: every module tested, the 48-hour test window before shipping, and the 90-day warranty. Real figures publish here as they accumulate.
Often ideal, oddly enough. Most faults live on the board, which is replaced anyway, and core failures are rare. If diagnosis finds something deeper, we tell you first and your deposit comes back in full.
Yes, any board-level shop with the right tooling can. We are the best placed though, since we built the card and hold its baseline test report.
You can, and the comparison above prices that route too. It needs the board, twelve matched modules, BGA rework equipment, and the skills to use them, with no warranty if it goes wrong. If you have all four, you do not need us.
No. Games rarely use anywhere near this much VRAM. The upgrade is for memory-bound AI and compute work.
Yes, it is your card. Our terms ask one thing: describe it accurately as a modified card when you sell it. The signed test report helps your sale, since the next owner gets the same proof you did.
Cover figures are shown in pounds next to every shipping option before you book. High-value returns are insured to the card's upgraded value. Full details are in the shipping section of our terms.
Three or more cards quote as a batch with scheduled build slots. A £500 reservation deposit secures the slots and comes off the final invoice. Use the form below and tell us how many.
Only a bad core can stop it, since memory just gets replaced. If your core cannot take the upgrade, your deposit comes back in full and you choose between a return at your shipping cost or donating it to us.
Twice. A parts deposit when you accept the quote, then the balance for our work once your card passes testing. If we cannot proceed after diagnosis, the deposit comes back in full.
Heat, mainly: the transplant takes controlled heat cycles, with a small but never-zero risk to the core and the memory. We publish the honest picture, risk by risk, on its own page:
What can go wrong, and what we do about it →Same approach, 16GB to 32GB, from £750. The full page is linked at the bottom.