Proposed class action · GPU owners

You bought the whole chip.
You got half.

NVIDIA sells GeForce cards with their AI tensor cores capped to half speed in software. The same silicon, sold as a professional card, runs at full rate. We're organizing the owners who paid for capability they never received.

01You built their empire

CUDA launched in 2006 on the GeForce 8800 — a gamer's card. When AlexNet kicked off the deep-learning revolution in 2012, it trained on two GTX 580s. Gamers, students and researchers bought GeForce cards by the millions and made CUDA the platform every AI lab now depends on. That installed base — financed by consumers — is the moat behind NVIDIA's multi-trillion-dollar valuation.

The platform those buyers built is now used to sell them half of what they paid for, while the same silicon leaves through the back door at many times the price with the locks removed. And now it's pricing them out entirely: the RTX 5090 launched at $1,999 and trades at $6,800–9,400 — up to 4.7× MSRP — while NVIDIA reportedly halted 5090 production to reroute the same GB202 die into enterprise cards. First they capped what you bought. Now they won't sell it to you at all.

02GeForce owners — the cap is in the driver, not the chip

Every RTX 20, 30 and 40 series card ships with FP16/BF16/TF32 tensor throughput — the math nearly all AI workloads use — capped to half speed. An open-source boot-time patch (gpu-unlock) removes the cap. No hardware change. Same card, same day:

RTX 3060 12GB27 → 51 TFLOPS (+87%)
RTX 3080 10GB65 → 113 TFLOPS (+73%)
RTX 4080104 → 193 TFLOPS (+87%)
RTX 4090165 → 329 TFLOPS (+100%)
As shippedCap removed — same hardware
"Why did Nvidia handicap my GPU? Market segmentation, they want businesses to purchase the professional cards without these limitations at significantly higher prices."— gpu-unlock author, project FAQ

Roughly 135 million consumer GPUs across three generations shipped with this cap. It affects AI inference, fine-tuning and image generation — not the gaming performance most reviews measure, which is why it went unnoticed for years.

03Enterprise buyers — the $20,000 is a permission bit

The professional card that runs the same silicon at full rate costs many times more. That price difference isn't manufacturing cost — it's the price of the restriction being off. Consumers pay for a crippled chip; enterprises pay a premium for the identical chip uncrippled. One policy, both sides pay.

04Mining buyers — 8GB sold, 80GB on the package

NVIDIA's CMP 170HX mining card used the same GA100 die as the A100 datacenter GPU. It shipped with 8GB of usable memory — while physically carrying ~80GB of HBM2e, disabled by factory fuses. In 2026 a security researcher showed an $8 voltage-glitching board can rewrite those fuses:

8 GB
sold as
80 GB
physically on the chip
85.9
TFLOPS FP16 unlocked — vs 72.5 on a retail $20,000 A100

Nine of ten tested cards unlocked to the full 80GB. A software-only unlock now exists as well. The lock was never physics — it was policy, burned into fuses.

05They didn't just cap what you bought. They destroyed what they made.

NVIDIA disclosed $550 million in CMP revenue for fiscal 2022. At the 170HX's ~$4,500 launch price, published estimates put the fleet at roughly 12,000–37,000 cards. Do the memory math:

96–296 TB
sold to buyers (8GB per card)
960–2,960 TB
physically manufactured (~80GB per card)
~90%
of the memory produced, fused off on purpose

Up to 2.7 petabytes of HBM2e — produced at enormous energy and material cost, soldered onto packages, then deliberately disabled to protect the $20,000 price of the real A100. That is the memory equivalent of tens of thousands of datacenter GPUs that never needed to be built — destroyed in the factory while the world now faces a memory shortage and DRAM prices have roughly quadrupled.

This isn't yield binning. Nine of ten sampled cards were fully functional at 80GB. And NVIDIA has form on this exact product line: in 2022 the SEC charged the company with misleading investors about how much of its revenue came from cryptomining. The regulator already found deception once, on these very cards.

Unit counts are estimates derived from disclosed revenue and estimated street prices (arXiv:2505.03782); NVIDIA has never published per-model CMP volumes.

06It's not one lock. It's a system.

The tensor cap is not an isolated product decision. Independent researchers have peeled back lock after lock, and every one of them sits exactly on the line where a consumer card could replace an enterprise sale:

One lock is a product decision. Five locks, all removable, all guarding the same boundary — that is a policy. Driver bit, fuse, license: three layers of the same wall, built to keep the people who financed the platform on the cheap side of it.

07This has a precedent — and it settled

In July 2026, John Deere settled with the FTC and five states for $99 million for withholding repair software from owners of its equipment. One product line, one decade. The NVIDIA pattern spans three GPU generations, hundreds of millions of units, and reaches from driver settings down to silicon fuses. Same mechanism. Orders of magnitude larger.

Our draft complaint draws on consumer-protection, warranty and right-to-repair law, and on prior GPU class actions that settled. Registration is free and creates no obligation. If you bought an NVIDIA GPU, you may be part of the class.

Join the class

Register your interest. We'll contact you when the filing is ready — no fees, no obligation, and we never share your data.

You're registered. We'll be in touch.
Something went wrong — please try again.

This is an organizing effort, not a law firm, and registering is not legal representation. Nothing here is legal advice. Claims described are allegations to be tested in court.

Sources