No other component in a computer stops this abruptly. There is no rattle beforehand and no gradual decline, so the whole of the warning you get is a machine that started perfectly well yesterday throwing an error today where the drive is supposed to be. M.2 sticks and PCIe cards of every description come in from Bracknell households, from gaming builds put together in spare rooms, and from the engineering and software offices scattered along the A329M and out towards Reading.
Anything sent in for NVMe recovery is diagnosed free. The written figure that follows is settled before a screwdriver leaves the drawer: £300 + VAT for any single hard drive or SSD, NVMe included.
No fix, no fee covers logical recoveries. Outside it sit electronic and mechanical failures, chip-level work, DVR jobs and forensic jobs, and physical work is 50% up front. Every band is published on the data recovery cost page.
Reading a symptom back to the failure underneath it is where the work genuinely starts, and this set of thirty accounts for all but a handful of the parcels that reach the Guildford bench. If yours is not on the list, it will still be recognised — describe it on the phone and you will get a straight answer about the odds before you post anything.
Something in the controller or its firmware emptied the bus overnight. Everything written to the module is still present and complete. What has gone is the only part capable of organising it and handing it over.
Four candidates sit behind that: a failed controller, a collapsed power stage, damaged boot code, or dead memory. Dead memory is by a distance the least likely of the four, which means the odds favour recovery, though the module has no way of telling you so.
Translator collapse leaves the module quoting a figure that belongs to no product on sale. It is an unambiguous signpost rather than a mystery, and service-mode repair is the road it points down.
An NVMe module is not one fixed volume. It presents namespaces, and a single administrative command removes one. The blocks stay exactly where they were and the module reports itself as empty. Recovery starts at the raw flash and rebuilds the arrangement from underneath, which is why nothing should be written to the module in the meantime.
The Format NVM command can be told to lay the namespace out at 4,096 bytes a block instead of 512, and doing so throws the entire mapping away in a few seconds. It is quick, it is silent, and what is left looks like a blank module. The contents are still down there and have to be reassembled by hand.
The system module stopped mid-sentence and the machine has nothing left to start from. Lifted out of the laptop and put onto a bench adapter, a surprising proportion hand over their files with very little argument.
Bare modules throttle first, then drop off the bus under load, then scramble their tables after enough hot afternoons in a row. Imaged cold, on a rig with active cooling over the controller, a good number behave themselves.
Whether it came from Windows Update or a vendor tool makes no difference to the outcome: a module that can no longer say what it is. It gets reflashed here where the controller still responds, and read around entirely where it does not.
There is no noise to listen out for and no gradual decline to spot in advance. These modules work perfectly until the moment they stop, which is the entire argument for keeping a second copy of anything that matters.
These are strips of glass fibre less than a millimetre thick and they lose to upgrade levers, stiff envelopes and impatient hands. A short break out near the notch is a track repair under the scope. Anything through the middle of the board goes to chip-off, and chip-level work is 50% upfront.
Surge damage to the power circuitry is rebuilt at board level before anybody attempts to read a single block. Doing those two things in that order is most of the job, and doing them the other way round wastes the one attempt available.
The moment files are deleted the module starts erasing the blocks behind them, and that only pauses when the power goes off. Pull the module out of the machine and the clock stops where it stands. Leave the laptop running while you think it over and the thinking costs files.
Early SN850X firmware would stall under sustained load and then take the module off the bus altogether. WD issued a fix. A great many of these never received it, and they keep the booking sheet busy.
The 980 and 990 generation burned through NAND life considerably faster than intended until an update corrected the behaviour. Modules that never got the update arrive here with most of their rated writes already spent and nothing to show for it.
Kingston has shipped those modules with whichever controller and flash were available that quarter, so two of them wearing identical labels can need completely different handling. Whatever is under the sticker chooses the method, which is why the diagnostic comes first.
Steam Decks and similar handhelds seal short modules into crowded shells with nowhere for heat to go. Tiny format, entirely conventional recovery, and a proper 2230 socket on the bench rather than an adapter stack.
DRAM-less designs borrow a slice of system memory to hold their mapping tables. One bad sleep cycle or one crash and the borrowed copy comes back wrong, at which point the module either mounts as nonsense or refuses entirely. The failure is documented and so is the route out of it.
A fast module running bare sits well above the temperature it was rated for, and it ages accordingly. Dropouts come first, disappearance later, and the imaging window is the gap between the two.
Over-tightened fixings and squashed thermal pads bow the board until joints crack under the controller. It gets reworked under the scope first and imaged afterwards, never the other way about.
After a BIOS update, or in a marginal riser or adapter, the module and the slot stop agreeing terms and the drive appears to have vanished completely. That is proved out in known-good lab sockets rather than by swapping cables at a kitchen table.
One dead die in the stack shows up as a reduced capacity, as errors scattered across the stripes, or as a module that has quietly gone read-only. The imaging works around the dead die and reconstructs what it can from the parity and from the dies that still answer.
Interrupt the background tidying and the map ends up half old and half new. The module then behaves differently on each boot depending on which version it reads first, mounting once and refusing the next time. Rebuilding the tables in service mode settles it for good.
Counterfeit modules get recovered for whatever they actually contain rather than whatever the label claims they contain. The sticker persuades nobody on this side of the bench, and the diagnostic reports what is genuinely inside.
Break the socket latch and the module sits at an angle, losing lanes as it goes. It limps along at a fraction of the bandwidth it should have, then stops appearing altogether. An intact lab socket is where it gets imaged.
These commands are designed to be irreversible and they are extremely effective at it. Run to completion, nothing survives, and you will be told so plainly. Interrupted partway, a great deal often survives. Which of the two happened is the first thing the diagnostic establishes.
Some module and platform combinations, Linux laptops above all, have well-documented trouble with the deepest sleep states: the module goes down and never comes back up. It reads as a hardware failure. Usually the silicon is fine and the tables want rebuilding after the crash that followed.
A module with nothing spare cannot reorganise itself, because moving a block requires an empty one to move it into. Writes slow to a crawl, errors accumulate, and some modules go read-only rather than carry on. Difficult to live with, and rather convenient to image.
M.2 slots carry different signals from one machine to the next, and a module that fits mechanically may not be wired to anything capable of talking to it. Nothing appears, nothing is broken, and an entire afternoon disappears into a fault that never existed. Rule that out before writing the module off.
PM981a, BG5 and similar modules are fitted at the factory and carry firmware built for one manufacturer's machines. They fail like any other module, but the service-mode routines have to match the OEM variant rather than the retail one, which is a question of having the right profiles to hand.
BitLocker and OPAL are common on NVMe because the hardware supports them without any speed penalty, so plenty of laptops are encrypted whether or not the owner knew. Recover the module and you recover ciphertext. With the recovery key from the Microsoft account, the domain or a printed sheet, that unlocks at the end. Without it, no lab on earth opens it, and we say so on day one.
An NVMe module sits directly on the PCIe lanes and answers the processor without anything mechanical standing in between. No motor, no actuator arm, no platter. That arrangement is where the speed comes from and it is also why there is no warning period at the end. Behind the connector, a controller runs a substantial firmware layer across stacked NAND and keeps a running record of where every block of your data has been relocated to. Should the controller stop responding, or should that record become inconsistent, the module simply goes quiet on the bus while the contents behind it remain completely intact. Two ways in exist from there. Service mode, provided by the manufacturer, will still get a conversation out of many controllers. Where it will not, the flash packages are read directly and the translation layer is rebuilt from first principles. Neither route benefits from the machine being switched on and off another twenty times, however strong the temptation.
If it unclips and fits in a padded envelope, it can be worked on. That covers 2280, 2242 and 2230 sticks, add-in cards for a PCIe slot, and everything from Gen3 to Gen5, wearing whichever badge you bought: Samsung, WD, Crucial, Kingston, SK hynix, Kioxia, Sabrent, Solidigm. What is refused as a matter of course is flash soldered onto a logic board, which is outside the scope of this service altogether. Heat is worth its own note. Drive a bare module hard with nothing sitting over it and the first response is throttling, the second is dropping off the bus entirely, and enough of those incidents will leave the mapping tables in a state. Small-form-factor builds and thin laptops account for a steady dribble of exactly that through the summer. Whatever set your fault going, the free assessment names it and attaches a figure to it before you have spent anything at all.
Controller repair, firmware repair and chip surgery at PCIe speeds. The bench is specified so that no job ever sits waiting on equipment:
Silent modules are brought up in the maker's own service mode, the firmware is corrected, the translator rebuilt, and the image taken before anybody starts gambling with the memory.
A module that stalls or wanders about gets captured behind timeouts set in hardware, with retry behaviour controlled at a level ordinary software has no access to.
Every length from a 2230 stub to a 22110, plus add-in cards and enterprise U.2, connected natively with a fan over the controller for modules that misbehave once they warm up.
Where the controller will not respond in service mode, the packages are removed and read individually. It is the slow road to the same files, and it is quoted at 50% upfront.
Interleave, XOR and error correction worked out from the raw dumps until the mapping stands up again and the file system can be rebuilt on top of it.
Power stages rebuilt component by component, bowed and cracked boards reworked under magnification, and nothing imaged until that work is finished.
Every reputation on that list was earned rather than invented. SN850X modules stalled under sustained load until WD published a fix, and the ones that never received it are still turning up here. Samsungs of the 980 generation that missed their update arrive with most of their rated writes already gone. An NV3 is a lottery, and whatever sits under the label picks the method. Gen4 and Gen5 modules installed where somebody decided a heatsink was optional cook themselves gradually. Modules fitted at the factory in Dell, HP and Lenovo machines need OEM service profiles rather than retail ones, which is a question of holding the right library. Every length from 2230 to 22110, add-in cards, enterprise U.2 and every generation from Gen3 onwards are workable, at £300 + VAT a module, with a free diagnostic closing two working days after booking in. The only refusal is a module soldered to the mainboard. Phones and tablets are not taken at all.
Nearly every job here arrived as a parcel. Tracked, insured post is the calmest way to move a drive that is already struggling, and something handed over in Berkshire, Surrey or London is normally on the Guildford bench the next working day.
Is the storage still bolted into a machine — laptop, tower, iMac, MacBook, rack server, a DVR under the till? Free it first and send the bare unit. Stripping hardware is not something this lab does, though it is ten minutes' work for any repair shop on your high street. There is a single case with no way round it: memory chips soldered flat onto the mainboard, which is how Apple Silicon machines and certain ultrabooks are built. Where the storage cannot be unbolted, there is no parcel to make up.
↓ Print the shipping & booking-in form (PDF)
The name on the parcel wants to be Guildford Data Recovery. Driving it over from Bracknell is roughly forty minutes on the A322 then the A3; posting it costs you a stamp and a day. Either way, a message goes out to you as soon as it is logged onto the system, and two working days later the diagnostic is finished.
Unsure whether something should go in the box? Ring 0800 689 0668 while the lid is still open, or work through the free online diagnostic and let it tell you.
Nothing to pay for the diagnosis, one written figure before any work begins, and the band on this page is £300 + VAT for any single hard drive or SSD, NVMe included.