The open-model ecosystem is not disappearing. But two multibillion-dollar acquisitions show that the neutral infrastructure developers rely on to find, download, compare and run AI models has become some of the most strategically valuable territory in technology. History suggests builders should pay attention — and build escape hatches now.
Nvidia confirmed Thursday that it has agreed to acquire Hugging Face for $12.93 billion, giving the dominant maker of AI accelerators ownership of the platform that has become perhaps the most important distribution point for open and open-weight AI models.
The numbers help explain the attraction. Nvidia says more than 18 million developers, researchers and creators use Hugging Face, which now hosts more than 3 million models, 500,000 datasets and 1 million applications. More than 200,000 companies use the service to discover, evaluate, customize and deploy AI.
The deal comes barely two weeks after another major piece of model infrastructure found a buyer. On August 19, Stripe announced an agreement to acquire OpenRouter, the model marketplace and API gateway that gives developers a common interface for hundreds of AI models. The companies did not disclose financial terms, although Reuters reported the transaction at slightly more than $8 billion, while Axios separately reported a price above $8 billion. OpenRouter says it now handles more than 10 trillion tokens per day across more than 400 models for a community exceeding 10 million developers and companies.
That means buyers have committed roughly $21 billion in a matter of weeks to two companies whose primary value comes not from owning the most powerful AI model, but from sitting between model makers and model users.
For independent developers and enterprises that deliberately embraced open models to avoid dependency on OpenAI, Anthropic, Google or another single AI lab, that should be a signal.
It doesn't mean open source or open weights AI is about to go proprietary and closed — instead, it's that the battle over AI lock-in is moving up a layer.
The model is open. The infrastructure around it may not be.
It is worth being precise about what developers commonly call "open-source AI."
Hugging Face hosts everything from conventionally open-source software to downloadable model weights released under permissive licenses, custom research licenses, models requiring users to accept additional terms and completely gated repositories.
Under the Open Source Initiative's Open Source AI Definition, truly open-source AI requires more than downloadable weights. Users need broad rights to use, study, modify and share the system, along with access to parameters, relevant code and sufficient information about the training data and process to meaningfully modify the system. Many popular models commonly described as open source therefore fit more accurately into the broader open-weight category.
Hugging Face itself reflects that diversity. Model cards can specify standard or custom licenses, and model creators can place repositories behind access gates and require users to submit information before downloading their files.
That distinction becomes particularly important after the Nvidia acquisition.
The model weights are one layer. Hugging Face increasingly represents several others: storage, discovery, versioning, documentation, model cards, download statistics, community reputation, datasets, applications, enterprise collaboration, authentication and, increasingly, inference and deployment.
Many model repositories themselves remain portable. Hugging Face's documentation explicitly supports cloning repositories through Git and downloading model files locally.
But replicating a model file is much easier than replicating an ecosystem.
That is the central issue for AI builders.
Duane O’Brien, executive director of the Open Source Initiative, told VentureBeat that this distinction between open technology and the platforms surrounding it is precisely what developers should watch as the acquisitions unfold. "These platforms would not exist without Open Source software and open collaboration," O’Brien said. "Hugging Face and OpenRouter became central to the AI ecosystem because they clearly met the needs of developers, who value openness, platform neutrality and interoperability."
His point places the burden less on the identity of either buyer than on what Nvidia and Stripe actually do after taking control. O’Brien said developers have historically demonstrated that they will move when commercial platforms undermine those characteristics: "History shows that when a platform pushes Open Source developers through proprietary workflows, they find or build more open alternatives."
Nvidia is making an unusually explicit neutrality promise
Nvidia appears acutely aware of the potential conflict.
In announcing the acquisition, CEO Jensen Huang said Hugging Face would continue supporting models from across the ecosystem, as well as multiple clouds and accelerator platforms.
Most significantly, Huang wrote: "NVIDIA compute will not be required to build on or deploy through Hugging Face."
He also committed to preserving developers' ability to select their own models, frameworks, cloud providers, inference services and computing platforms.
Those are unusually specific assurances for an acquisition announcement — and for good reason.
Nithya Ruff, chair of the board of directors of the Linux Foundation, told VentureBeat that such assurances should be judged continuously rather than as one-time commitments. "While these acquisitions unlock scale and resources, the long-term health, security, and competitiveness of the AI ecosystem ultimately depend on open source infrastructure, transparency, and developer choice," Ruff said. "Neutrality is a discipline a company must choose time and again, not a promise it makes once."
Ruff nevertheless offered a relatively optimistic reading of the buyers involved. Having collaborated with both Nvidia and Stripe on open-source initiatives, she said she expects them to continue choosing open source as a means of driving AI innovation. Her response highlights the dual nature of both transactions: large corporate owners can provide infrastructure and investment that independent platforms could struggle to match, even as their ownership creates new incentives developers need to watch.
Nvidia is not merely another enterprise software company acquiring a developer community. It makes the hardware on which a huge percentage of modern AI workloads run, develops CUDA and other software surrounding that hardware, builds its own Nemotron AI models and has increasingly invested in the open-model ecosystem.
In March, Nvidia launched the Nemotron Coalition, bringing together organizations including Mistral AI, Black Forest Labs, LangChain, Perplexity and others to develop open frontier models using Nvidia infrastructure. Nvidia says it has also published more than 500 models and more than 250 datasets on Hugging Face.
Buying Hugging Face therefore gives Nvidia something even more valuable than another model family: proximity to the place where developers decide which models matter.
Reuters noted that ownership could give Nvidia greater insight into the models and technologies developers are collaborating on, testing and sharing.
But Nvidia's potential information advantage goes beyond knowing which models developers want. Hugging Face increasingly has visibility into how at least some of those models are actually being run — a strategically important view as Nvidia faces competition from AMD, cloud providers developing their own accelerators and custom-chip companies working with major AI customers. Reuters has reported that major Nvidia customers including Meta, Microsoft and OpenAI are developing chips partly to reduce their dependence on Nvidia.
Broadcom expects AI-chip revenue to reach roughly $115 billion in fiscal 2027 as cloud companies increasingly turn to custom silicon, while Google has struck a major custom-chip agreement with Marvell. Amazon continues to develop Trainium and Inferentia, Google has its TPU family, and AMD supplies its Instinct accelerators to AI customers. Hugging Face already supports non-Nvidia hardware including AMD GPUs and AWS accelerators, while its inference marketplace connects developers to Cerebras and Groq. Reuters reported Broadcom's expected growth amid that custom-silicon shift.
That point requires an important qualification. Nvidia would not automatically learn what processor runs every model downloaded from Hugging Face. A developer can download weights and run them entirely outside Hugging Face's infrastructure, including offline, leaving the service with little or no visibility into the eventual deployment.
However, substantial portions of Hugging Face's commercial stack already process hardware-level information. When a customer creates an Inference Endpoint, the configuration identifies the model repository, cloud vendor, region, accelerator type, instance type and size. Hugging Face's documentation shows that its API exposes hardware fields including accelerator architecture, GPU memory and number of accelerators, while endpoint analytics measure CPU, memory and GPU utilization.
Hugging Face also collects narrower hardware signals elsewhere. Its TRL training library's usage telemetry can report whether a workload uses CUDA, CPU, Apple's MPS or another accelerator type and, when available, the specific GPU model. Hugging Face says that telemetry does not include model identifiers or other user-provided information. Separately, users can tell the Hub what CPUs, GPUs or Apple Silicon hardware they own so it can recommend compatible models, while keeping that information private rather than placing it on a public profile.
Its Inference Providers product offers another view of AI consumption: account dashboards break usage down by model and inference provider. Hugging Face says routed requests do not store request or response bodies and debugging logs remain for up to 30 days. This should not be confused with Nvidia suddenly obtaining the content of every enterprise prompt passing through the service.
The strategic value instead lies in aggregated market intelligence. A platform sitting between millions of developers, thousands of models, numerous inference companies and competing hardware architectures could potentially detect shifts in demand earlier than conventional market research: which architectures are growing, which models perform efficiently on competing accelerators, which inference providers are gaining users and whether particular workloads are moving away from GPUs.
Hugging Face's current systems do not necessarily form one universal database capable of answering all those questions. Privacy rules, enterprise contracts and technical separation between services may restrict how information can be combined or used. Yet the acquisition creates an information advantage that did not exist when Hugging Face was independent.
That matters because Nvidia has both the resources and incentive to respond. If it sees a competing architecture gaining traction for a category of open models, it could investigate why and answer through software optimization, faster product development, pricing, cloud credits, partnerships or customer incentives. It could also identify technologies as investment, licensing or acquisition targets.
There is no evidence Nvidia intends to use Hugging Face data to suppress a hardware rival, manipulate rankings or disadvantage models optimized for other chips. Observing market trends and responding with better products or prices is normal competition. The structural concern is that one hardware competitor would own a platform capable of supplying unusually early signals about how developers are evaluating that market.
Nvidia's broader strategy makes the issue less theoretical. The company licensed technology from inference-chip startup Groq in a transaction valued at roughly $17 billion and acquired SchedMD, the company behind Slurm, the open-source workload manager widely used in high-performance computing. Reuters reported that the SchedMD acquisition prompted concern among some supercomputing specialists that software spanning competing hardware could eventually favor Nvidia's GPUs and networking products. Nvidia said Slurm would remain open source and vendor-neutral.
Nvidia has also experimented with financing and revenue-sharing arrangements intended to make Nvidia-powered AI infrastructure easier for cloud companies to deploy. Reuters reported in August that Nvidia paused parts of one such program after concerns about the influence the arrangements could give it over participating providers. Nvidia said the broader program remained in place and was evolving.
But the potential conflict-of-interest now exists. A change doesn't have to look like an Nvidia-only Hugging Face to matter. The more plausible form of platform leverage would be subtler: default deployment options, featured models, preferred inference services, optimized runtimes, benchmarking integrations, recommended hardware configurations or simply a faster path from a Hugging Face repository into Nvidia's own software and cloud ecosystem.
Those kinds of defaults can shape developer behavior without anybody removing a single model.
OpenRouter creates a parallel test for model routing
OpenRouter presents a different version of the same problem. It's not an open-source project in the traditional sense. Its value lies in giving developers an abstraction layer across both proprietary and open models.
Instead of rewriting an application every time a better model appears, a developer can use OpenRouter to reach models from multiple providers through a common interface and route workloads according to performance, availability and price.
That makes OpenRouter's perceived neutrality central to its product. The company addressed exactly that concern in announcing the Stripe deal, saying it would continue with the "same mission, same name, same product, same roadmap" and that its routing decisions would continue to prioritize users rather than any model or provider.
Stripe creates fewer obvious conflicts than Nvidia because it does not operate a leading foundation model or GPU platform. But the acquisition still illustrates the same larger economic trend. The abstraction layers between AI developers and AI models have become strategic assets.
Model labs want distribution. Cloud companies want workloads. Chipmakers want applications optimized for their hardware. Payment platforms want to meter and monetize AI consumption. The company controlling the gateway sits in an enviable position between all of them.
David DeSanto, CEO of Anaconda, argued in comments provided to VentureBeat that both transactions validate how commercially important open-weight AI has become. "These acquisitions highlight the impact and reach of open weights AI models as both platforms have helped show the power of what is now available without being tied to a single frontier lab," DeSanto said.
But DeSanto also pointed directly at the question that hangs over both deals: whether platforms originally valued for broad developer choice begin reflecting the strategic priorities of their new owners. "Ultimately time will tell if these acquisitions shift from AI on your own terms to AI through the lens of the acquirer," he said.
Mazin Gilbert, executive director of the Linux Foundation's Agentic AI Foundation, said the same openness question increasingly extends beyond the model weights themselves. "Open weights make advanced AI accessible to anyone without requiring them to train a model from scratch – whether that's small startups, university research teams, or even hospitals," Gilbert told VentureBeat. "As AI shifts toward agents, that same principle of openness must extend beyond models to the protocols and infrastructure agents depend on, so innovation isn’t constrained by the decisions or boundaries of any one platform."
That distinction may prove particularly important for OpenRouter. If developers increasingly build applications around model gateways, agent protocols and other abstraction layers precisely so that individual models remain replaceable, control of those abstraction layers becomes a new form of leverage. Openness at the model layer cannot by itself guarantee openness in the system that selects, authenticates, bills for and routes requests to those models.
The closest historical analogy may be Microsoft and GitHub
The most obvious precedent for Nvidia and Hugging Face may not actually be Linux. It may be Microsoft's acquisition of GitHub.
Microsoft agreed to buy GitHub for $7.5 billion in 2018 amid widespread questions about whether a platform hosting enormous quantities of open-source software could remain neutral after being absorbed by one of the world's largest software vendors.
Microsoft CEO Satya Nadella made promises remarkably similar to the ones Nvidia is making now. In its announcement of the acquisition, Microsoft said GitHub would remain open, continue supporting developers regardless of programming language, tool, operating system or cloud, and operate independently.
GitHub did remain a broadly cross-platform developer hub, one that has grown dramatically under Microsoft. Microsoft said in July that GitHub had reached 225 million users, with more than 90% of Fortune 500 companies using it. GitHub Copilot had reached 50 million users, Copilot revenue had increased more than 60% quarter over quarter, and one out of every three pull requests on GitHub involved an AI agent.
That record offers a legitimate reason for optimism about Nvidia and Hugging Face: a large corporate entity can invest heavily in infrastructure while preserving interoperability because interoperability is what keeps the platform valuable.
Microsoft didn't need to make GitHub exclusive to its own products to turn it into a strategic part of its developer business. GitHub became the distribution point for Microsoft's AI assistant Copilot, which expanded from code completion into a development platform spanning editors, terminals, repositories and autonomous agents. Microsoft's Agent HQ strategy also supports outside agents and models from OpenAI, Anthropic, Google, Cognition and xAI.
The strategy has a clear parallel to Nvidia's opportunity with Hugging Face: do not necessarily force developers to use your product; own the place where they choose among everybody's products.
Cursor's rise makes that distinction clearer. The AI coding company became one of GitHub Copilot's strongest challengers by moving the assistant deeper into the development environment. Reuters reported in 2025 that Cursor had reached roughly $100 million in annual revenue and a $10 billion valuation. SpaceX agreed in June 2026 to acquire Cursor maker Anysphere in a $60 billion all-stock transaction, and Cursor said in August that the deal had closed.
Days after announcing the closing, Cursor began rolling out Origin, its own Git repository hosting service. Git itself remains a free and open-source distributed version-control system, allowing repositories to move among GitHub, GitLab, Bitbucket and other hosts. Origin operates at the commercial layer above Git: it can host repositories, browse and search code, manage pull requests and run Cursor's cloud agents. It can mirror GitHub repositories, but Cursor says Origin becomes the source of truth for repositories created directly there, taking GitHub out of the path.
That gives Cursor and SpaceX a route around a dependency that previously persisted even when developers chose Cursor over Copilot. Cursor's agents commonly still worked through GitHub repositories and permissions covering code, pull requests, issues, continuous integration and workflows, as described in Cursor's GitHub integration documentation.
The competition is consequently moving from which assistant writes a developer's code to which company owns the environment containing the code, agents, models, pull requests, deployment workflows and organizational context.
Microsoft's ownership did not destroy GitHub's openness. But it positioned Microsoft exceptionally well when AI transformed software development because it already owned one of the most important places where developers stored their work. Hugging Face occupies an analogous position one layer further down the AI stack.
GitHub's new competition also underscores the value of real portability. GitHub cannot own Git, and Cursor can build Origin because Git's architecture and licensing allow repositories to move. The equivalent question for AI builders is not simply whether Hugging Face remains open, but whether the models, metadata and workflows accumulated there remain portable enough for a credible alternative to emerge if developers decide they need one.
Microsoft then extended its reach further into software distribution when GitHub acquired npm in 2020, bringing the package registry at the center of the JavaScript ecosystem under the same corporate umbrella. GitHub pledged that npm's public registry would remain free.
That history demonstrates why an acquisition need not destroy an open ecosystem. A wealthy owner can finance infrastructure, security, enterprise tooling and global scale that would otherwise be difficult for an independent company to provide — that's the hope for Nvidia's acquisition of Hugging Face.
GitHub also demonstrates something else: once developer infrastructure becomes indispensable, its strategic value compounds. Code hosting leads naturally into package distribution. Model hosting can similarly lead into evaluation, deployment, inference and compute.
Red Hat shows the difference between open code and open governance
Linux offers an even deeper lesson: IBM completed its approximately $34 billion acquisition of Red Hat in 2019, while explicitly promising to preserve Red Hat's independence, neutrality and commitment to open source.
The acquisition did not make Linux proprietary and couldn't have, on a technical level. Linux kernel development remains highly distributed. Kernel documentation describes a distributed maintainer structure spanning large numbers of subsystems and repositories, with technical governance extending well beyond any single company. The Linux Foundation provides a neutral institutional layer around that ecosystem, such that corporation can employ important Linux developers and commercialize Linux aggressively without simply owning Linux.
Mike Milinkovich, executive director of the Eclipse Foundation, pointed to that history as reason not to assume that corporate ownership and open-source development are inherently incompatible. "We are both encouraged and unsurprised by Nvidia’s decision to maintain HuggingFace as an open source platform," Milinkovich told VentureBeat. "History has shown that prolific, impactful technologies will almost always move towards an open source model, and AI is proving no different."
Milinkovich said Eclipse is simultaneously seeing "unprecedented interest in Sovereign AI," which he described as deeply tied to open-source solutions. That suggests another force pushing against complete platform capture: governments and enterprises increasingly want enough control over their AI infrastructure to deploy models under their own legal, geographic and technical constraints rather than leaving every layer to one provider.
But the Red Hat story also provides one of the strongest warnings for today's AI developers.
In December 2020, the CentOS Project announced that it would shift resources from CentOS Linux, which functioned as a downstream rebuild of Red Hat Enterprise Linux, toward CentOS Stream, which sits upstream of RHEL. CentOS Linux 8 consequently reached end of life at the end of 2021 rather than continuing on its previously expected trajectory.
The change occurred after IBM bought Red Hat, although that chronology by itself does not establish that IBM ordered or caused the decision. What matters for builders is what happened next: the software was open, so the ecosystem could respond.
Alternatives emerged. AlmaLinux formed around the demand for a stable RHEL-compatible distribution, while other projects pursued similar goals.
Open source worked exactly as intended. But users still had to migrate. That is an important distinction: the right to fork is not the same as frictionless portability. Applications have dependencies. Businesses certify software against particular distributions. Operations teams build deployment systems around specific infrastructure. Communities accumulate documentation and expertise. Open licensing provides an escape hatch but doesn't make the escape hatch free.
Java may be an even better warning
Java provides another useful case study because its history separates an open technology from its commercial distribution.
Sun Microsystems began releasing Java under the GPL in 2006. Oracle completed its acquisition of Sun in January 2010, obtaining Java among several enormously important software assets. Oracle itself described Java at the time as one of the central strategic technologies it was acquiring.
Java didn't suddenly become closed. OpenJDK remains available under GPLv2 with the Classpath Exception, and alternative production Java distributions subsequently flourished. Eclipse Temurin provides tested OpenJDK runtimes through the Eclipse Foundation.
Oracle's own Java distribution and commercial licensing changed. Oracle says its Java license changed for releases beginning April 16, 2019, making some uses that had previously been available without payment subject to different terms. Oracle subsequently introduced its Java SE Universal Subscription, whose published pricing uses a company's employee count rather than merely counting the individual employees running Java.
That didn' make Java itself universally paid or proprietary. OpenJDK remained open. It did, however, force enterprises to understand exactly which Java they were running, under what license and from which supplier. AI builders should recognize the analogy. "Open model" does not answer every infrastructure question.
Which weights? Which license? Which revision? Hosted where? Served through whose API? Optimized by whose runtime? Authenticated through whose account system? Subject to whose acceptable-use rules? Those details become super important if and when ownership changes.
Oracle's other Sun acquisitions show what happens when communities lose confidence
Two other Sun-era technologies provide an even sharper precedent.
After Oracle obtained MySQL through its acquisition of Sun, MySQL founder Michael "Monty" Widenius created MariaDB because of concerns about Oracle's stewardship. MariaDB remains an open-source database project today.
Members of the OpenOffice.org community followed a similar path after Oracle acquired Sun. They formed the independent Document Foundation and created LibreOffice, a fork of OpenOffice.org. Oracle eventually donated OpenOffice.org's code and trademarks to the Apache Software Foundation.
Again, open source provided resilience. The technology survived even when its institutional home changed. But every fork also duplicated effort, fractured communities and required developers to choose a new center of gravity. That's why governance can matter as much as licensing.
JavaScript shows the value of putting governance somewhere neutral
JavaScript offers yet another variation.
Oracle currently owns the JavaScript trademark, an asset inherited through Sun, even though Oracle does not control the technical evolution of the programming language developers generally mean when they say JavaScript. Ecma International's TC39 maintains the vendor-neutral ECMAScript standard, with participation from companies and developers across the industry.
That separation has allowed JavaScript to evolve independently of the company holding its trademark. Even the trademark itself remains contested: Deno filed a petition in 2024 seeking cancellation of Oracle's JavaScript registration, and the U.S. Trademark Trial and Appeal Board listed the cancellation proceeding as pending and suspended as of August 28, 2026.
The broader lesson is more important than that dispute: ppen ecosystems are most resistant to corporate capture when the code, standards and governance do not all live under the same roof.
The biggest Hugging Face risk is therefore not that Nvidia deletes the models
For AI builders, a dramatic lockout scenario is probably the wrong thing to focus on.
The more realistic questions involve incremental dependence. 10 key questions:
Will Hugging Face's discovery mechanisms remain neutral among model vendors?
Will models targeting AMD, Intel, Google TPUs or emerging accelerators receive the same integration effort as models targeting Nvidia hardware?
Will Nvidia-hosted inference become the easiest default?
Will Hugging Face's evaluations, recommendations and model rankings remain transparent?
Will enterprises remain able to export their private models, datasets, metadata and deployment configurations easily?
Will APIs and storage remain interoperable?
Could pricing change?
Could important services become bundled with Nvidia products?
Could independent inference companies receive the same treatment as Nvidia's infrastructure?
And what happens to the extraordinarily valuable behavioral signals created when millions of developers search for, evaluate, download and deploy models through the same platform?
None of these outcomes has been announced. Nvidia's public commitments point in the opposite direction.
But those are now reasonable architectural questions rather than philosophical ones.
Pierre Baqué, founder and CEO of Neural Concept, offered a more optimistic assessment of Nvidia itself, calling the chipmaker "a natural new home for Hugging Face" and saying the acquisition makes Nvidia more credible as a provider of open models and technologies.
But his practical recommendation to enterprises ends up in much the same place as the risk analysis: avoid designing applications around the assumption that any one provider will remain indispensable.
What AI builders should do now
There is little reason for developers to stop using Hugging Face or OpenRouter today. Both companies are explicitly promising continuity, and Nvidia in particular appears to understand that destroying Hugging Face's neutrality would also destroy much of what makes the company worth nearly $13 billion.
The more rational response is to reduce accidental dependency.
"More than ever, model agnosticism should be a design principle, not an afterthought," Baqué told VentureBeat. "The winners won’t necessarily be the companies with permanent access to the most powerful model. They’ll be the companies that can switch models without having to rebuild their AI systems."
For production systems, developers and enterprise AI teams should consider:
Mirror critical model artifacts. Where licenses permit it, maintain internal copies of production model weights, tokenizer files, configuration files and other required artifacts rather than depending entirely on a live Hugging Face repository.
Pin revisions. Treat model versions like software dependencies. Record exact repository revisions or checksums so a future upstream update cannot silently change production behavior.
Archive licenses and model cards. Hugging Face supports standard and custom model licenses. Enterprises should retain the applicable license and accompanying metadata for the exact model revision they deploy rather than relying on a URL remaining unchanged forever.
Separate artifact storage from runtime inference. An application that downloads a model from Hugging Face but can run it independently has much more negotiating leverage than one whose production path requires a Hugging Face-hosted service.
Maintain a direct-provider fallback to model gateways. OpenRouter can dramatically simplify multi-model deployments, but mission-critical applications should know how they would reach essential providers if their aggregation layer changed pricing, policies or behavior.
Test portability before you need it. A disaster-recovery plan that says a model can theoretically run elsewhere is different from actually deploying it on another provider or accelerator and measuring whether it works.
Vinay Thakker, co-founder and CTO of KloudStax, a Google Cloud Premier Partner, made a similar argument in unsolicited comments provided to VentureBeat. "If you're building something that only runs on one company's chips, ask yourself what happens the day that company raises the price or a deal like this changes who's in charge of a piece of your stack," Thakker said. "That's not a reason to panic, it's a reason to build differently from the start."
That advice applies equally to models, chips, clouds and gateways. A production architecture that can genuinely move between providers turns an ownership change or pricing shift into an operational problem; an architecture in which the identity of one vendor is embedded throughout the application can turn the same change into a rewrite.
This isn't preparation for an Nvidia betrayal — it's good infrastructure discipline.
Open weights AI is becoming big business — and that changes the meaning of "open"
For much of the software industry's history, open source was framed as an alternative to commercial software.
AI is producing something different: Open models themselves can be free to download while enormous businesses emerge around finding them, storing them, evaluating them, routing them, optimizing them and running them.
Nvidia doesn't need to close Hugging Face to benefit from owning that layer. In fact, at present, it's more valuable to keep it as the de-facto source for open source AI weights, applications, and services.
Stripe doesn't need to own an AI model to profit from OpenRouter becoming one of the standard ways businesses consume models.
The infrastructure surrounding open AI may ultimately be more commercially valuable than many of the individual models flowing through it.