A technician using a bench microscope to closely inspect a circuit board component held with tweezers, verifying its markings in an electronics lab

In short: A relabeled, remarked or grey-market GPU is usually a real, working chip with a falsified or unverifiable identity, not a dead fake that fails on power-up. It boots, it installs drivers, and it will often run a workload for a while before a specification mismatch shows up as a reliability or performance problem. The checkpoint that actually catches this is intake verification — serial number pedigree, firmware cross-checks, and sustained-load testing against the claimed specification — not a visual inspection or a successful first boot.

The Checkpoint Procurement Documents Don't Cover

A growing share of knowledge-center coverage on this site addresses GPU procurement from the outside in: export control compliance, semiconductor supply chain resilience, and allocation scarcity during periods of high demand. Those are upstream problems, about whether hardware can be obtained at all and through which channel. This article is about the checkpoint that sits downstream of all of that and gets comparatively little attention: the physical intake process at a colocation facility, the point where hardware that has already been procured, shipped and unboxed is about to be racked, powered and connected to a shared power and cooling infrastructure on a customer's behalf.

That checkpoint matters because procurement paperwork and physical hardware can diverge. An invoice, a packing slip and a purchase order describe what was supposed to be bought. They do not, by themselves, confirm what actually arrived in the box. For most IT hardware categories this gap rarely matters in practice. For GPUs specifically, where a single high-end accelerator can carry a price difference equivalent to an entire server chassis of other components, the incentive to misrepresent what is inside the box is considerably higher than it is for a network switch or a storage drive.

What "Not Genuine" Actually Means for a GPU

It helps to separate two distinct problems that intake verification is meant to catch, because they call for different checks.

Counterfeit or relabeled hardware is a real, functioning chip whose reported identity has been falsified. The most common pattern is a lower-tier or lower-binned chip remarked or firmware-flashed to report itself as a higher-tier SKU, sold at a price that reflects the claimed tier rather than the actual one. A less common but more damaging pattern is a card built from harvested or recycled components, assembled to pass a basic power-on test but without the reliability margin of a genuinely new unit.

Grey-market hardware is a different problem. It is frequently genuine, unmodified product that has been diverted outside its intended authorized distribution channel — moved across regions to bypass allocation controls, resold outside an authorized partner agreement, or routed around an export restriction. A grey-market card can be entirely authentic silicon while still arriving with no transferable manufacturer warranty, no verifiable chain of custody back to an authorized distributor, and no assurance that its firmware has not been altered somewhere in that unofficial path.

Both problems can pass a glance. Packaging can be counterfeited alongside the product itself, and a card that boots and installs drivers normally gives no indication, on its own, that anything is wrong.

Why a Successful Boot Is Not Verification

The instinct to treat "it works" as sufficient proof is understandable and wrong. A relabeled GPU is not a non-functional fake; it is a real chip that genuinely does less than its label claims, or that lacks the warranty and support pedigree the price implies. It will boot. It will install a standard driver package. It will run a light workload without triggering an obvious error. The misrepresentation is in the specification and the provenance, not in whether the silicon functions at all — which is exactly why intake verification has to check specification and provenance directly, rather than inferring them from the fact that the card powers on.

What a Defensible Intake Process Actually Checks

None of the individual checks below is exotic. The point is running them as a standard sequence at intake, before hardware is racked and powered, rather than treating any single one as sufficient on its own.

  • Serial number pedigree against the source, not just the sticker. Reading a serial number off a label confirms only that a label exists. Checking that serial number against the vendor's or an authorized distributor's own records, where that lookup is available, confirms whether the unit is recognized as a legitimately distributed part at all.
  • Firmware-reported identity cross-checked against the physical label. Vendor diagnostic utilities report device ID, subsystem ID and VBIOS version from the card's own firmware. Comparing that reported identity against what the physical label and packaging claim catches the case where the two disagree — though a sophisticated relabel can also falsify the firmware report, which is why this check is necessary but not sufficient on its own.
  • Sustained power and thermal behavior under a standardized load. This is the check least dependent on anything the hardware's own firmware is willing to report. A card's actual power draw and thermal output under sustained load are a function of its physical die and memory configuration, not of what it claims to be. A lower-tier chip relabeled as a higher-tier part will often diverge from the vendor's published power or thermal envelope for the claimed SKU under sustained load, even when every reported identifier passes.
  • Documented chain of custody back to an authorized source. A purchase order and invoice from a named reseller is not the same as a documented path back to the manufacturer or an authorized distributor. Grey-market hardware frequently arrives with paperwork that is genuine but incomplete — it proves a transaction occurred, not that the transaction traces back to an authorized channel.
  • Physical condition consistent with the claimed state. PCB revision markings, connector wear, and solder joint condition should be consistent with whether the unit is represented as new, refurbished, or used. A card sold as new with connector wear patterns consistent with prior service life is a mismatch worth investigating on its own, independent of any electronic test.

Where This Fits Relative to Other Safeguards

Intake verification is not a substitute for the controls that sit on either side of it. Export control compliance during procurement and broader supply chain resilience planning address where hardware is sourced from and whether that sourcing is compliant and reliable at the business level; intake verification addresses whether the specific physical unit that arrived matches what was ordered. Cluster acceptance testing happens after intake and focuses on whether a fully assembled cluster performs to its aggregate specification under production-like conditions — a useful backstop, but one that can mask a single relabeled card's deficit inside cluster-wide averages rather than isolating it to one unit. Physical access control at the facility governs who can reach hardware once it is inside the building, which is a related but distinct problem from verifying what that hardware actually is when it first arrives.

The same discipline has a mirror image at the other end of a GPU's service life. Lifecycle management from deployment through retirement and decommissioning and IT asset disposition both depend on the same kind of verifiable identity record this article describes at intake — a card whose provenance was never properly established going in is equally difficult to properly account for, resell, or responsibly retire coming out.

A Practical Intake Checklist

  • Record the serial number and, where a vendor or distributor lookup is available, confirm the unit is recognized as legitimately distributed before it is unboxed further.
  • Run a vendor diagnostic utility and compare the firmware-reported device ID, subsystem ID and VBIOS version against the physical label and packaging.
  • Run a standardized sustained-load test and compare power draw and thermal behavior against the vendor's published specification for the claimed SKU.
  • Request documentation that traces the unit back to an authorized distributor or the manufacturer directly, not only a reseller invoice.
  • Inspect physical condition — PCB revision, connector wear, solder quality — against the condition the hardware is represented as being in.
  • Log the outcome of every check against the unit's serial number so the record exists before the card is racked, not reconstructed afterward if a problem surfaces.

Frequently Asked Questions

Can a counterfeit GPU actually boot and run normally?

Yes. A relabeled or remarked GPU is usually a real, functioning chip of a different specification than the label claims, not a non-functional fake, so it can install drivers, boot, and run workloads without triggering an obvious failure. The deception is in the identity and the binning tier, not in whether the silicon works at all, which is exactly why a boot test and a visual check are not sufficient verification on their own.

What is the difference between a counterfeit GPU and a grey-market GPU?

A counterfeit GPU has falsified identity: a relabeled die, a remarked serial number, or firmware reporting a specification the hardware does not actually meet. A grey-market GPU is typically genuine hardware that has been diverted outside its intended authorized distribution channel, often to bypass regional allocation, export controls, or pricing agreements. Grey-market hardware can be functionally authentic while still carrying no manufacturer warranty, no verifiable purchase trail back to an authorized source, and no guarantee that its firmware has not been altered in transit.

Why can't nvidia-smi or GPU-Z alone confirm a card is genuine?

Diagnostic utilities read values reported by the GPU's own firmware, including device ID, subsystem ID and VBIOS version. If that firmware has been altered to misreport the card's identity, a diagnostic tool will faithfully display the falsified values rather than detect the falsification. These tools are useful for cross-checking the firmware-reported identity against the physical label and against sustained-load behavior, but a clean readout by itself is evidence of what the firmware claims, not proof of what the silicon actually is.

Does sustained power and thermal testing really catch a relabeled GPU?

It is one of the more reliable checks available at intake, because a card's physical power draw and thermal output under a sustained, standardized load are tied to its actual die and memory configuration, not to what its firmware claims. A lower-tier chip relabeled as a higher-tier part will often show a power draw, thermal profile, or sustained clock behavior under load that does not match the vendor's published specification for the claimed SKU, even when every reported identifier looks correct.

Who is responsible for verifying GPU authenticity in a colocation arrangement?

This depends on the hosting model. In a fully managed arrangement where the operator procures the hardware directly from authorized distribution, the operator carries that verification responsibility as part of its own procurement chain. In a bring-your-own-hardware colocation arrangement, the customer typically owns procurement, but a colocation operator's intake process is still the practical last checkpoint before that hardware is racked, powered and placed on a shared power and cooling infrastructure, which is why many facilities run a verification step regardless of who purchased the equipment.

Hardware You Can Verify, Infrastructure You Can Trust

Rax Data & Energy applies intake verification to GPU hardware entering our AI compute and colocation facilities. Tell us about your deployment and we'll walk through how intake, acceptance testing and ongoing lifecycle management fit together.

Contact Us AI Compute