Nvidia and the Price of Control

This article is a two-part work written in collaboration with Wroc++ (wrocpp.github.io) and Filip Sajdak. Companion post: CUDA lock-in is real, and it is not in the C++. Reading the other part is strongly recommended, the two are designed to be read together. Working on this article has worsened my outlook on Nvidia, whether sadly or gladly.

A Q&A session in 2012 at Aalto University. A student asks a question: if she could run a technology called NVIDIA Optimus on her new laptop, which allows for switching between GPU and integrated CPU graphics. A question went out there, and a brief silence followed.

The problem was, NVIDIA Optimus was fundamentally broken on Linux, because NVIDIA refused to support it or to document how it worked for open-source developers. Users had to rely on reverse-engineered efforts that were rather clunky.

After a little context, Linus Torvalds followed up with:

"Nvidia has been the single worst company we've ever dealt with. So, Nvidia, fuck you!"

Watch on YouTube

The question is, how come a company can get to be hated so much, while itself heavily relying on the very open-source software in order to run its products?

The truth is, the relationship had been strained long before that moment in 2012. A lot had been building up.

The usual setup for manufacturers to support their hardware is to integrate their drivers directly into the Linux kernel under open-source licenses like the GPL. But what NVIDIA has always done is something else. While everybody contributed and made interactions with their hardware public, as it should be, Nvidia shipped binary blobs. These binary blobs effectively act as black boxes without any way of knowing how or in what way they interact with the kernel.

In addition to that, Nvidia refuses to publish its GPU architecture specifications for open-source software to be written accurately.

And the irony is, they did that at the time while banking heavily on Tegra, an Android processor, where Android heavily relies on Linux.

Nvidia was more than happy to take the Linux kernel for free to power billions of dollars' worth of chips, while giving absolutely nothing in return, and stating outright they would give nothing.

After the statement was made by Linus, NVIDIA responded days later. They didn't issue an apology or change course; instead, they doubled down through a formal PR statement.

We Do Things Our Own Way

Nvidia publicly acknowledged the criticism but flatly defended their choice to use closed-source proprietary code. The argument was that "Nvidia common code" allowed them to deliver a uniform cross-platform experience. That uniformity has a concrete cost at source level: including a single header like cuda_runtime.h drags in a large transitive surface that builds and tooling must carry everywhere (example: header pollution example). Portability in theory becomes stickiness in practice. The blame for Optimus failing to work appropriately on laptops was simply shifted elsewhere. The general message was clear: if you want a driver, write it yourselves, we are not helping you.

Yet... when it came to the very projects that tried to solve these problems such as Nouveau, an open-source driver for Nvidia GPUs, NVIDIA actively obstructed those efforts architecturally. The power management of the cards was locked behind cryptography in the hardware itself. To get full performance, a driver has to tell the card to "reclock" (speed up) under load. In practice, Nvidia locked down the power-state algorithms and required cryptographically signed firmware blobs for reclocking to be possible. The open-source Nouveau drivers were effectively crippled.

With every single passing generation, Nvidia locks down its firmware deeper behind cryptographic bounds - Maxwell, Pascal, and beyond. They use free software, refuse to contribute back, and continuously strengthen the obfuscation of their firmware so nobody can ever reverse engineer it.

Historical Bad Faith

Friction with the Linux community was merely a symptom, not the epitome of the company itself. It simply demonstrated the way they act and how they think of others. It is a systemic playbook, a philosophy and character of how the whole organization operates that we should pay closer attention to.

GameWorks and Tessellation

Nvidia's developer programs for GameWorks and PhysX provided game studios with proprietary graphics libraries. In titles like Batman: Arkham Knight and The Witcher 3, Nvidia implemented extreme, developer-unoptimized hardware tessellation. This degraded performance to a point where competing AMD graphics cards struggled heavily, while the Nvidia architecture wasn't impacted in the same way. Nvidia was accused of anti-competitively optimizing drivers to cripple competition, but nothing came of it. They got away with it.

GeForce Partner Program

In 2018, the GeForce Partner Program (GPP) launched. It required major hardware vendors like ASUS or MSI to tie their elite gaming sub-brands, like ASUS ROG, exclusively to Nvidia hardware. If a partner wanted to sell AMD GPUs under their premium branding, they would have had to spin off their AMD Radeon lines into entirely separate brands with no marketing heritage. Effectively, this barred AMD from top-tier mindshare. The program was canceled after severe backlash and regulatory scrutiny. This wasn't the first, nor the thousandth time Nvidia pulled something like that.

Wayland: GBM vs EGLStreams

When the Linux ecosystem started transitioning from X11 to the modern Wayland display server protocol, the community developed Generic Buffer Management (GBM) as the unified standard. Nvidia simply refused to support it for years, instead pushing its own proprietary API called EGLStreams. This stalled Linux desktop progression for Nvidia users, forcing developers to code separate display pathways until Nvidia finally relented and added GBM support in 2022.

Review Samples and Hardware Unboxed

In late 2020, Nvidia temporarily blacklisted the YouTube channel Hardware Unboxed from receiving review samples. The reason? The channel didn't focus enough on ray tracing performance and chose to review hardware on their own terms. Following widespread industry outrage and backlash from fellow reviewers, Nvidia reversed the ban within a week.

GTX 970 Memory Split

In 2014, Nvidia released the GTX 970, advertised with 4 gigabytes of VRAM. However, it was later discovered that the memory was split: 3.5GB operated at full speed, while the final 0.5GB shared a severely crippled memory controller running at roughly one-seventh of the speed. When games exceeded 3.5GB of VRAM usage, players experienced severe stuttering and performance drops. NVIDIA eventually settled a class-action lawsuit, paying out partial refunds to affected users.

Bundling and Mellanox

As the artificial intelligence boom accelerated, antitrust regulators worldwide started investigating Nvidia's business practices. Allegations involved Nvidia leveraging its near-monopoly on AI training GPUs to force customers into bundled purchases, such as requiring buyers to purchase Mellanox networking hardware alongside H100 and B200 chips, while deprioritizing and penalizing clients considering silicon from AMD or custom hyperscaler chips. They went full-on with it regardless, favoring full racks that integrated Mellanox. It demonstrates they never stopped doing what they always did.

The ARM Bid

In 2020-2022, Nvidia attempted to buy ARM for $40 billion. The deal immediately triggered global alarm from tech giants like Google, Microsoft, and Qualcomm, as well as regulatory bodies like the FTC, UK CMA, and EU. Everyone knew Nvidia was going to impose the same practices and ruin ARM's open licensing model, that's what happens to hardware or software Nvidia touches. The deal was eventually abandoned due to insurmountable regulatory hurdles.

Money-Gating Hardware: Feature Gating and Limiters

Nvidia artificially disables capabilities on hardware via software drivers and firmware, despite the physical silicon fully supporting these features:

• vGPU: disabled on consumer GPUs and treated as an expensive license on enterprise GPUs.
• P2P (Peer-to-Peer DMA): deliberately disabled on consumer cards so direct interoperation between them isn't possible.
• ECC memory: artificially disabled on consumer GPUs despite the silicon supporting it.

During the massive cryptocurrency boom, the LHR (Lite Hash Rate) anti-mining limitation was easily undone via an official Nvidia driver that was leaked "accidentally."

Fake Open-Source Kernel Modules

In May 2022, Nvidia made headlines by announcing they would open-source their GPU kernel modules (open-gpu-kernel-modules). It was praised widely at first, but then the realization came.

Nvidia did not open-source the actual driver logic. They split the stack, moving the critical logic, memory management, scheduling, and power management, into a black-box GPU System Processor (GSP) running on an internal RISC-V co-processor inside the GPU. The same pattern repeats one level down: CUDA exposes PTX as supposedly portable assembly, while the actual SASS that executes stays hardware-specific and driver-controlled. Compare the two for the same kernel (PTX next to SASS example): what you write is not what runs, and Nvidia reserves the final lowering for itself.

The released open-source kernel modules were just a heavily abstracted wrapper communicating with that black box. They did it just to say they were open-source friendly and had open-sourced their drivers. Massive PR and goodwill stunt.

Nouveau

Nouveau was the reverse-engineering effort by the Linux community to build a true open-source Nvidia driver. Nvidia consistently undermined these efforts to prevent a viable open-source alternative from maturing.

As Nvidia introduced more complex architectures, they cryptographically locked the GPU firmware. They refused to provide open-source developers with signing keys, effectively locking custom firmware out and forcing everyone into their proprietary ecosystem.

Why? Because control is everything to them.

The Leather Jacket PR Veil: "General Purpose Utility"

On an episode of the Dwarkesh Podcast, Jensen Huang addressed software ecosystems like CUDA and the proprietary hardware stack that entraps clients. In deflecting that argument, Nvidia relies on framing itself as a liberator, an "accelerator of the general-purpose computing revolution." The narrative is that tightly coupled, proprietary stacks are a necessary evil to drag the world out of legacy CPU-bound computing.

When confronted with hyperscalers diversifying away from Nvidia using TPUs and custom ASICs, Huang minimizes them as narrow, single-purpose accelerators. He routinely pivots to state that true general intelligence requires dynamic, flexible computing that only Nvidia can handle. That claim fails at the code level: the same kernel, unmodified, compiles for AMD through SCALE (same kernel on AMD example), which shows the coupling of CUDA source to Nvidia hardware is policy, not technical necessity. See the companion post, CUDA lock-in is real, and it is not in the C++, for the full walkthrough of these examples. This conveniently dismisses competitor chips as rigid tools, despite how effective they have proven to be and how aggressively hyperscalers are shifting away from Nvidia silicon.

Questions of forced network infrastructure or restricted hardware layers are met with a framing of "deep industry collaboration." Rather than acknowledging structural coercion, he projects the persona of a tired pioneer laying the rails for others to ride on. Prioritizing or deprioritizing customers based on their alignment is recast as mere operational triage during a supply chain crunch.

When questioned on exports to China and whether flooding those markets weakens American dominance in AI, or whether Western companies might lose ground and get copied, Huang heatedly stated:

"The premise that, even if we competed in China, that we're going to lose that market anyways... You're not talking to somebody who woke up a loser. That loser attitude, that loser premise makes no sense to me."

Yet after closer analysis, Nvidia didn't just lose ground in China. They were systematically locked out, and Huawei stepped directly into that vacuum.

The Hugging Face Play: Owning the Town Square

While this paper was being written, Nvidia made what may go down as the most predatory software play of the decade: an agreed $12.93 billion acquisition of Hugging Face.

Nvidia and Hugging Face signed a definitive agreement on 2 September 2026, disclosed in an 8-K filed with the U.S. Securities and Exchange Commission and in Nvidia's own announcement. The deal is expected to close in the first half of 2027, subject to customary closing conditions and regulatory approval. It is agreed, not done.

That qualifier matters, and the ARM section above is exactly why. Regulators killed a $40B Nvidia acquisition once before.

Jensen Huang wrapped the agreed buyout in a warm blanket of "philanthropic stewardship," swearing that the platform will remain compute agnostic.

To be fair, Nvidia committed publicly to keeping the platform multi-accelerator. In its announcement Nvidia states Hugging Face "will continue to support multi-cloud and multi-accelerator development and deployment, so builders can use the hardware and infrastructure that best fit their work," with existing integrations spanning AWS Trainium and Inferentia, Google TPUs, Intel Gaudi, and AMD Instinct.

There is no reason to believe that commitment survives contact with Nvidia's track record. This is the company that refused GBM for years to push EGLStreams, that moved driver logic behind the GSP black box while calling the wrapper open-source, that used GameWorks tessellation and GPP branding to punish competitors. A commitment to neutrality from a company whose playbook is control lasts until neutrality becomes inconvenient.

Let's be clear about the game being played: The next war isn't about chips. Silicon is increasingly getting cheaper, more commoditized, and easier to spin up elsewhere.

The real war is about capturing the ecosystem and owning the developer workflows. By moving to buy the definitive library catalog where 18 million developers discover and deploy models, Nvidia isn't just buying a startup, it is trying to buy the front door of the entire global AI revolution.

See the SEC 8-K and Nvidia's announcement for the dated deal terms, closing timeline, and stated multi-accelerator commitment.

Conclusion

Nvidia has demonstrated time and again that it is all about control. They don't want to give up anything. They don't want to create something open enough where others can build alongside them; they would rather the whole world burn than improve together with others.

Seemingly, it stems not from a mindset of abundance, but from a mindset that there is only a fixed pie to go around, and you have to steal it. Dominate, make others lose, win, that is their stance. No other interpretation fits what they do. If they believed the world was abundant, that you could bake a new pie and let everybody enjoy it, they instead chose to say: if we go down, you go down too.

And to go as far as faking open-source contributions to foster goodwill among developers, all while sinking deeper into this rabbit hole of control... It is simply intractable, unimaginable misery.

As some would say: "A strong businessman is not going to allow you to bully him or take his market share." I don't know when the end will come to this mindset, but I certainly hate it.

It leaves us, the developers, the very people Nvidia relies on to sustain its model, to think: should we help the one who wishes in mind to be our master?

Should we align with a company that HAS BEEN consistently aligned with horrible corporate practices that were not accidental, but structural?

For all we care, we might one day wake up in a world where tech is permanently locked in, and the knowledge to create it is guarded, so that there is no way out to a better world that doesn't align with their agenda or politics.