Giving Away a Model That Cost Hundreds of Millions to Train — What’s the Play?

On January 20, 2025, a Chinese company called DeepSeek put an AI reasoning model on Hugging Face. MIT license—download it, modify it, use it to make money, no licensing fee required.

The model was called R1. It matched OpenAI’s o1 across multiple benchmarks. o1 was OpenAI’s strongest reasoning model at the time—closed-source, priced per token, backed by Microsoft’s tens of billions in investment and the densest GPU clusters on Earth. Sam Altman himself said GPT-4’s training cost exceeded $100 million. DeepSeek said R1’s training cost about $5.6 million.

Three days later, January 27, a Monday. Nvidia plunged at the open, wiping out $589 billion in market cap in a single day—the largest single-day loss for any company in U.S. stock market history. The entire AI sector followed, nearly $1 trillion in value evaporating. Silicon Valley was in panic. Marc Andreessen called R1 “one of the most impressive breakthroughs” he’d ever seen. Trump called it a “wake-up call.”

Two completely different moves on the same track. OpenAI spends hundreds of millions training a model, locks it up, charges per call, with 2025 annual revenue exceeding $20 billion. DeepSeek spends a few million training a model, gives it away for free, with API prices at one-seventeenth of OpenAI’s. One quietly rakes in money; the other gives it away. Who’s smarter?

Intuition says the one making money is smarter. Spending millions to train something and then giving it away— isn’t that charity? But if you stretch the timeline a bit, you’ll find that this “giving away” move has recurred over the past four decades, and every time it appeared, someone built an empire on it. Linux has been free for thirty years and now runs most of the world’s servers. Android has been free for seventeen years and captured 71% of the global phone market. Red Hat built a $34 billion market cap selling technical support for free Linux.

Giving things away is a way to make money. It’s just that the path to monetization is longer, more circuitous, and depends on one prerequisite: what you give away must be good enough that people are willing to build an ecosystem around it. This prerequisite is harder than it looks.

Free for Forty Years — How Did Linux Not Starve?

To understand the open-source vs. closed-source debate in AI models, you need to look at a longer time window. This question didn’t originate in the AI era. For forty years, the software industry has been answering the same question: how do you survive giving things away for free?

In 1991, Linus Torvalds wrote an operating system kernel in his dorm room at the University of Helsinki and posted it to a Usenet newsgroup. He said it was just a hobby project, nothing big or professional. Thirty-plus years later, Linux runs most of the world’s servers, all Android phones, and the top ten supercomputers globally. How did Linux survive? Not on Linus alone. Over 15,000 developers worldwide have contributed to the Linux kernel. Red Hat, IBM, Google, and Intel all contribute people and money. Linux isn’t maintained by one person—it’s maintained by a globally distributed coalition of companies.

But not all open-source projects are this lucky. Academic research shows that a large number of GitHub projects are abandoned shortly after creation. A project’s survival depends heavily on the continued participation of core developers. Once core developers leave, a project is either taken over by the community or dies outright. Closed-source software sustains teams through profit; open-source relies on community enthusiasm and corporate sponsorship. The two maintenance logics are completely different. Linux has thrived for thirty years because it became important enough that companies are willing to pay to keep it alive. Most open-source projects don’t have this luck.

If Linux is free, how did someone make $34 billion from it? Red Hat’s answer: software is free, service costs money. It took the community Fedora Linux, made it stable, certified it, ran security tests, and turned it into Red Hat Enterprise Linux, sold by subscription. You’re not paying a software license fee—you’re paying for technical support, security patches, and certification. Red Hat went public in 1999, the first publicly traded company to make money from open source. In 2018, IBM acquired it for $34 billion. But Red Hat’s story has an easily overlooked sequel. In 2023, it announced it would no longer post RHEL source code on public websites, only in its customer portal for paying users. Before this, a group in the community (CentOS, Rocky Linux, AlmaLinux) specifically took RHEL’s public source code, recompiled it, and made free RHEL-compatible versions—essentially using code Red Hat spent heavily to maintain to create a free alternative that competed with it. Red Hat closed this loophole. Technically, it didn’t violate the GPL (the GPL requires providing source code to users, and paying users could still get it), but non-users could no longer access it. The community considered this undermining the spirit of the GPL. Red Hat’s response was blunt: we’ve invested enormous resources maintaining RHEL, and we don’t want others repackaging it to make money. Even the most successful open-source company constantly adjusts the boundary between “open source” and “profitability.”

In 2008, Google released Android, open source. Phone manufacturers could use it for free. By 2025, Android held 71% of the global phone market. In 2007, Apple released the iPhone with iOS, closed source, licensed to no one. By 2025, iOS held 28% of the global phone market. Who won? Android took 71% of the share, but Apple took 63% of global phone profits. The average iPhone sells for $930; the average Android for $370. iOS users spent $94 billion on the App Store; Android users spent $55 billion on Google Play. Both models are thriving, coexisting for fifteen years, neither eliminating the other.

But Android’s “open source” has a precise design. Android’s core code, AOSP, is open source, but Google Mobile Services (GMS) is closed source, including Play Store, Google Maps, and Gmail. To use Play Store, you must sign agreements, pass compatibility tests, and pre-install a suite of Google apps. The open-source base layer pulled 71% of the world’s phones into the Android ecosystem; the closed-source service layer locked those users into Google’s monetization pipeline. In 2025, Google went further, moving Android development from a public branch to an internal private branch. AOSP’s public code became an empty placeholder.

Huawei phones once used Android. In 2019, Google cut off GMS for Huawei, and Huawei phones overseas lost Play Store and Google Maps overnight. You thought you were using an open-source system, but what you depended on was the closed-source service layer on top of it. That layer can be cut off at any time. HarmonyOS was Huawei’s response to this vulnerability. In 2024, HarmonyOS NEXT completely abandoned the Android foundation and became an independent operating system. By 2025, it held approximately 19% market share in China.

In 1824, British bricklayer Joseph Aspdin patented Portland cement. The patent’s core was a deal: he would disclose the formula in exchange for a period of monopoly. After the patent expired, the formula entered the public domain, and anyone could use it. This system forced inventors to disclose, in exchange for a monopoly period. Once the formula was public, others could improve on it, and the entire industry advanced together. Aspdin’s son William improved the formula but deliberately kept it secret—only for commercial rivals to discover a similar formula through independent research. Secrecy didn’t produce a permanent lead. The essence of the patent system is “temporary monopoly in exchange for eventual disclosure.” The logic of open-source models is analogous: you give away model weights and get back attention, ecosystem, and developer mindshare. Both are transactions, not charity.

Giving Away a Model That Cost Hundreds of Millions — What’s the Play?

That weekend in January 2025, after DeepSeek released R1, over 600 forks appeared on Hugging Face within days. Hugging Face’s team worked through the weekend to reproduce R1’s training method. Meta’s engineers were studying R1’s technical report. The entire global AI community was discussing DeepSeek. This attention doesn’t translate into direct API revenue, but it bought something scarcer: talent and influence. DeepSeek went from a Chinese company almost no one had heard of to the most discussed name in global AI overnight. This brand effect can’t be bought with any amount of ad spend.

Meta understands this logic best. It began open-sourcing the Llama series in 2023; to date, Llama has over 350 million cumulative downloads on Hugging Face and more than 60,000 derivative models. AT&T, DoorDash, Goldman Sachs, Spotify, and Zoom all use it. Developers worldwide build tools, write tutorials, and create applications on Llama—the entire AI open-source ecosystem revolves around Meta. Meta hasn’t collected a cent in API fees, but it has gained infrastructure status in the AI era. Meta’s real business is ads and social platforms, not selling models. As long as the AI ecosystem revolves around it, its ads and social platforms have a moat of AI capability. Llama’s open-source cost is, for Meta, marketing plus R&D spending.

But there’s a critical boundary condition: your model must be good enough. If your open-source model is worse than the closed-source alternatives, what you’re giving away is a mediocre product. No one builds an ecosystem around a mediocre model; no one spends time making derivative models. You give it away, and you get nothing back.

Baidu is a living example. It was one of the first companies to build Chinese large language models—ERNIE Bot was released in 2023 and remained closed source. But its performance was inferior to later open-source models like DeepSeek and Qwen. By January 2025, ByteDance’s Doubao had 78.6 million monthly active users, DeepSeek had 33.7 million, and ERNIE Bot had only 13 million. In technical evaluations, DeepSeek-V2.5 outperformed ERNIE 4.5 on most metrics at one-seventh the price. Baidu neither earned closed-source monopoly profits nor captured open-source ecosystem dividends. In February 2025, it announced it would open-source ERNIE 4.5 in June. Not because it had an epiphany, but because it was backed into a corner.

This is the core contradiction of open source. Give away a good product, and you get an ecosystem. Give away a mediocre product, and you get silence. To win with closed source, your model must be the best. To win with open source, your model must also be the best—only the monetization path differs.

Five Companies, Five Choices — None of Them Wrong

Line up the most important companies in the AI industry, and each one’s open-source/closed-source choice is different, but each choice fits its commercial position like a glove.

When OpenAI was founded in 2015, the “Open” in its name was serious. It was a non-profit organization, and even GPT-2’s full weights were released in 2019. The turning point came in 2019, when it created a capped-profit subsidiary and accepted a $1 billion investment from Microsoft. The reason was straightforward: training frontier models cost tens of millions of dollars, and donations couldn’t sustain it. GPT-3 was released in 2020, no longer open source, switching to API pricing. Since then, GPT-4 and GPT-5 have all been closed source. The logic of this shift isn’t that OpenAI turned bad—it’s that its business model changed. GPT-3’s capability reached a tipping point: the model itself could be sold directly as a product. The revenue potential of API pricing exceeded the ecosystem value of open source. By 2025, OpenAI’s annual revenue exceeded $20 billion, with 900 million weekly active users. Microsoft has invested over $13 billion cumulatively. All this money is predicated on OpenAI’s closed-source monopoly position.

Meta took the opposite path from OpenAI. Open-sourced Llama 2 in 2023, Llama 3.1 in 2024, Llama 4 in 2025—open source all the way. The reason isn’t complicated. Meta’s core business is ads and social platforms; it doesn’t make money selling models. If the AI ecosystem were monopolized by OpenAI, Meta would become a downstream company passively accepting someone else’s infrastructure in the AI era. Open-sourcing Llama is Meta’s means of preventing its own marginalization. Meta’s Chief AI Scientist Yann LeCun is the most prominent advocate for open-source AI. He has repeatedly stated publicly that if AI is controlled by only a few companies, culture and democracy are in trouble. He criticizes OpenAI and Anthropic for using safety narratives to justify closed source, calling it regulatory capture. But Meta’s open source has its limits too. In 2025, Zuckerberg hinted that future superintelligence models might not be fully open-sourced. Models that can build ecosystems are released; truly cutting-edge ones might be kept for internal use.

Google is the most mature practitioner of selective open source. Android’s AOSP is open source; GMS is closed source. Open source captures market share; closed source locks in users and developers. On the AI side, Gemini is closed source, served through API and Google Cloud. In 2025, it open-sourced the Gemini CLI coding tool, while the core model remained closed. It opens the tool layer that can build ecosystems and closes the model layer that can directly generate revenue. Google wants everything, with a precise dividing line between the two determined by whether it can be directly monetized.

Anthropic’s Claude series has never been open-sourced. Its stated reason is safety, but the commercial logic is API revenue and enterprise customers. The AI market in 2026 has a reality: the gap between Claude, GPT, and Gemini on benchmarks has narrowed to the first decimal place. Raw capability is no longer a moat. Anthropic sells trust. It demonstrates reliability by showing “things we refuse to do,” attracting high-premium clients in regulated industries like finance, healthcare, and law. Claude Code grew from a research preview to a billion-dollar product in six months. Closed source isn’t a stopgap for Anthropic—it’s a core business strategy. Open-source it, and the safety selling point is discounted.

The Chinese camp chose open source not because they don’t want to make money, but because they can’t win on the closed-source track. OpenAI has Microsoft’s tens of billions in investment; Anthropic has Amazon and Google’s billions. On the closed-source track, you compete on capital density and compute scale. U.S. export controls on high-end chips mean Chinese companies can’t even buy H100s—DeepSeek had to train on downgraded H800s. Open source is a different track. Instead of competing on API revenue, they compete on ecosystem influence. By early 2026, Qwen had over 1 billion cumulative downloads on Hugging Face, surpassing Meta’s Llama as the most-downloaded open-source model family globally. Chinese developers’ download share on Hugging Face reached 17.1%, surpassing the U.S. at 15.7% for the first time. On OpenRouter, the largest neutral LLM routing platform, Chinese open-source models accounted for approximately 61% of token consumption. Four of the top five most-used models were Chinese. Meta’s Llama had dropped to less than 1%.

But this story began to turn at the end of 2025. Alibaba started releasing closed-source models in September. Qwen3-Max was the first Qwen flagship without open-source weights. In April 2026, Qwen3.6-Max was fully closed source, API only. The strategy became: small models open source to build ecosystem, large models closed source to make money. Baidu followed a similar path. Open source can buy traffic, but traffic doesn’t directly equal money. Investors and the market are pushing these companies to prove they can be profitable.

Releasing a Model Isn’t the Same as Releasing an Ecosystem

Back to the original question: if OpenAI and Anthropic open-sourced their models, could they directly crush China’s lead in the open-source space?

Intuitively, it seems yes. OpenAI’s models are the most capable—if a GPT-5-level model were open-sourced on Hugging Face, wouldn’t all developers flock to it? But reality isn’t that simple.

First, open source isn’t about releasing weights; it’s about building an ecosystem. Uploading model weights to Hugging Face takes minutes, but building an ecosystem around a model takes months or even years. By early 2026, Qwen had over 113,000 derivative models, and approximately 40% of new LLM derivative models on Hugging Face were based on Qwen. Developers worldwide have invested significant time writing adaptation code, doing fine-tuning, and building toolchains on Qwen. These investments are sunk costs. Release a GPT-5-level open-source model, and developers won’t migrate immediately, because migration means discarding existing investments and starting over. And whether the new model will be continuously updated is unknown. Once Linux built its ecosystem, Windows couldn’t steal the server market no matter how it tried. Ecosystems have first-mover advantages; migration has costs.

Second, open source cannibalizes your own revenue. OpenAI’s API revenue in Q1 2026 was approximately $2.8 to $3.2 billion annually. If the flagship model were open-sourced, many users would self-deploy and stop paying API fees. OpenAI’s total revenue in 2025 was approximately $13 billion, but by Q1 2026 it had annualized to over $20 billion, with a gross margin of about 33% and inference costs of $8.4 billion. It doesn’t expect to reach cash-flow break-even until 2029-2030. Under this financial structure, open-sourcing the flagship model means opening a hole in the most critical revenue source. Anthropic’s situation is even more noteworthy. By May 2026, its annualized revenue had reached $47 billion, exceeding OpenAI’s, and it was projected to achieve its first profitable quarter in Q2 2026—single-quarter revenue of $10.9 billion with operating profit of approximately $560 million. About 80% of Anthropic’s revenue comes from enterprise API calls, with over 1,000 customers spending more than $1 million annually. Under this financial structure, open-sourcing the flagship model means directly cutting off the most critical revenue source. And Anthropic’s brand is indeed built on the safety narrative—closed source is a core component of that narrative. Open-source it, and this selling point disappears.

Third, whether the gap is large enough is a key variable. The reality of 2025 to 2026 is that the gap isn’t that large. DeepSeek R1 matched OpenAI’s o1 on reasoning tasks. Qwen 3.5 was comparable to GPT-4o and Claude Sonnet on multiple benchmarks. When the gap isn’t large enough, developers have no incentive to migrate. Your model is 10% better, but my ecosystem is already built on another model, and migration costs far exceed a 10% performance gain. Only when the gap is so large that not migrating means being eliminated will the ecosystem shift rapidly.

No Optimal Solution, Only Positional Solutions

Back to the DeepSeek scenario from the beginning. A model trained for $5.6 million given away for free, causing those who spent tens of billions to collectively bleed. It’s not a paradox; it’s a precise game-theoretic strategy. Open source trades model weights for attention, ecosystem, and developer mindshare. Closed source trades model capability directly for money. Each strategy has its conditions for success and its possibilities for failure. What determines which one you should choose is your commercial position.

If you want to judge whether an AI company—or your own company—should go open source or closed source, you can start with four questions.

Does your core revenue come from the model itself, or from things built on top of the model? If revenue comes directly from the model—API call fees, subscription fees—closed source makes more sense. OpenAI and Anthropic are like this. If revenue comes from the ecosystem above the model—ads, cloud services, hardware, social platforms—open source makes more sense. Meta and Alibaba are like this. For you, the model is infrastructure, not a product. The more open the infrastructure, the larger the ecosystem on top.

Are you in a leading or catching-up position? If you’re leading, closed source is a tool for harvesting profit. OpenAI went closed source in the GPT-3 era because it was leading. If you’re catching up, open source is a tool for changing lanes and overtaking. Chinese companies chose open source because they can’t compete with American capital and compute on the closed-source track.

Is your model good enough? This is the most brutal question. Open-source a good model and you get an ecosystem. Open-source a mediocre model and you get silence. Baidu’s ERNIE was closed source but not as good as DeepSeek’s open-source version, forcing a belated pivot to open source. The product itself being good is the prerequisite; open source and closed source are monetization strategies, not product strategies.

Can you sustain the investment? Open source isn’t a one-time deal. You release a model, and developers expect the next version. If you can’t continuously train and release new models, the ecosystem will be taken by someone else. Meta can continuously open-source Llama because it has money, GPUs, and talent. Linux has thrived for thirty years not because it’s open source, but because people keep investing money and people into it.

After answering these four questions, you’ll find there’s no correct answer, only positional answers. When your position changes, the answer changes too. OpenAI went from open source to closed source because its position changed from a research lab to a product company. Alibaba went from pure open source to selective open source because its position changed from a chaser to a platform with monetization capability. Baidu was forced from closed source to open source because its position changed from China’s AI leader to a chaser caught between DeepSeek and Qwen.

The open-source vs. closed-source game in the AI era may ultimately not be about who won, but about what territory each side occupies. Just as Android and iOS have coexisted for fifteen years, Linux and Windows for thirty. Open source occupies scale and ecosystem; closed source occupies profit and the high-end market. The two compete with and define each other. Without Android’s 71% share, iOS’s 63% profit has no frame of reference. Without Linux being free, Windows’ price wouldn’t be questioned.

That weekend in January 2025, DeepSeek released R1, and Nvidia lost $589 billion. Many asked whether this meant open source had won. It didn’t. It meant there was one more powerful player on the open-source track, a player who proved at extremely low cost that closed source’s moat isn’t as deep as imagined. It forced OpenAI and Anthropic to answer a harder question: when your model is only 10% better but the other guy’s is free, how long can your pricing power last?

The answer to this question will determine the landscape of the AI industry for the next five years.

References

DeepSeek R1 release and Nvidia market cap loss: Reuters
OpenAI training costs and revenue data: Ars Technica, TechSpot, RevenueMemo
OpenAI’s non-profit to for-profit transition: NYTimes, The Guardian, OpenAI
OpenAI gpt-oss open-source model: OpenAI, Built In
OpenAI gpt-oss-safeguard model: OpenAI
GPT-2 “too dangerous” incident: Slate, Sesame Disk
Musk lawsuit against OpenAI: Washington Post, Reuters
OpenAI-Microsoft renegotiation: CNBC, The Verge, VentureBeat
Meta Llama series and downloads: Meta AI Blog, Wikipedia
Zuckerberg hints superintelligence may not be open-sourced: TechCrunch
Meta Muse Spark closed-source release: DailyCrunch, FourfoldAI, StartupHub.ai
Google Android AOSP goes private: Ars Technica, Android Authority
Google Gemma 4 and Apache 2.0: Ars Technica, VentureBeat
Google Gemini pricing and full-stack advantage: Business Insider, TokenRate
Red Hat restricts RHEL source code: Ars Technica
ByteDance Doubao user data: 36Kr, DoNews, 每经网
ByteDance Seed 2.0 and Seed-OSS: ByteDance Seed, GitHub, Hugging Face
Volcano Engine MaaS revenue target: KrAsia
ByteDance 2026 AI priorities: KrAsia
DeepSeek price war and competitive pressure: Decrypt, TheStreet
Qwen downloads and Hugging Face share: Hugging Face
OpenRouter Chinese model share: OpenRouter
Anthropic Claude Code revenue: Reuters
Linux kernel contributor data: Linux Foundation
Portland cement patent history: Wikipedia - Portland cement