Date: 2026-03-25
A common intuition suggests that if an AI research breakthrough significantly reduces inference costs or enhances model capabilities, the company that discovered it should remain silent to maintain its exclusive advantage. While this intuition holds true in many industries, it is repeatedly broken in the AI sector. Google DeepMind published over 175 papers at NeurIPS 2025. Meta makes the weights of its Llama series directly available for download. DeepSeek not only released its model weights but also disclosed detailed training engineering specifics like Multi-head Latent Attention. NVIDIA open sourced its entire inference operating system, Dynamo.
These actions might seem like a voluntary surrender of competitive advantage, but the companies making these choices are far from foolish. To understand why they do this, we need to examine the issue at two levels: first, what exactly is being made public; and second, who benefits once it is public.
The most common mistake when discussing the open behavior of AI companies is treating all forms of disclosure as the same thing. In reality, from research papers to product distribution, there are at least six distinct layers, each with a completely different logic for openness.
Papers and Methods represent the layer that diffuses the fastest. An architectural innovation—such as Transformer, knowledge distillation, or Mixture of Experts—will likely be independently discovered or reverse-engineered by peers within weeks or months, even if a company chooses not to publish. The actual gain from keeping papers secret is low, while the cost of not publishing is high: researchers will leave due to a lack of freedom to publish. One reason Yann LeCun left Meta in late 2025 was his perception that the freedom to publish at the FAIR lab was being restricted. His subsequent founding of AMI Labs with a $1.03 billion seed round demonstrates that the loss of top researchers is not just a blow to reputation, but the direct creation of a new competitor.
Open Source Code and Toolchains go a step further than papers. While a paper explains the principles, code allows you to run them directly. Once open sourced, third parties will proactively help with hardware adaptation, framework porting, performance optimization, and secondary encapsulation. This distributes the engineering costs, which would otherwise belong to the publisher, across the entire community. This is a mechanism we will explore in more detail later.
Protocols, SDKs, and Interface Standards are an often underestimated layer. Anthropic donated the Model Context Protocol (MCP) to the Linux Foundation, reaching 97 million monthly downloads, which forced OpenAI and Google to adopt it as well. Whoever defines the default interface first is more likely to become the central node of the subsequent tool ecosystem. Opening a protocol is not an act of generosity; it is a battle for architectural definition.
Model Weights represent the most significant step in open behavior. Once weights are released, they cannot be retracted, and anyone can fine-tune and deploy based on them. Meta (Llama), Alibaba (Qwen), and DeepSeek have chosen this path because their respective business models do not rely on the exclusivity of the weights themselves. In contrast, the weights of OpenAI and Anthropic remain strictly closed-source because differentiated API pricing is their core revenue source. Google is hedging its bets: the smaller Gemma models are open, while the flagship Gemini remains closed.
Training Data and Alignment Data are almost never made public by any company. Training data is a source of sustainable differentiation; accumulating it requires time and capital, and it involves copyright controversies. OpenAI has never disclosed the specific sources of Books1 and Books2, and Anthropic settled a $1.5 billion copyright case in 2025. Alignment training data is even more sensitive, as it involves both commercial value and safety considerations.
Product Distribution and User Behavior Data are the strictly closed bottom lines for all companies. No company will share the network effects and behavioral data from ChatGPT’s 400 million weekly active users, Meta AI’s 500 million users, or the usage data from Google embedding Gemini into search. It can be said that all the open behaviors at the layers above are ultimately aimed at protecting or expanding this closed distribution layer.
Once you understand this hierarchy, it becomes clear that calling a company “open” or “closed” is an oversimplification. A more accurate description is that every company is open at certain layers and closed at others, and this boundary precisely reflects where its profit pool is located.
Returning to the initial question: the fundamental reason companies do not choose to use research secretly is not altruism, but an economic concept: commoditize your complement—turn the layer you don’t make money from into a commodity, thereby concentrating profits in your stronger complementary layer.
Joel Spolsky articulated this mechanism long ago. IBM did not open the PC architecture out of kindness, but because its profits lay in services and mainframes. Google’s open sourcing of Android was not about sentimentality, but because the more smartphones become a commodity, the higher the search advertising revenue. The exact same thing is happening in the AI industry, only with more participants and more complex layers.
NVIDIA is the purest example of this logic. At GTC 2026, Jensen Huang explicitly shifted the competitive focus from model training to continuous inference. He compared models to recipes on an assembly line; the recipe can belong to anyone, but the factory must be NVIDIA’s. By open sourcing the Dynamo inference operating system and fostering a multipolar model ecosystem (the Nemotron Coalition, which includes France’s Mistral, India’s Sarvam, and independent Silicon Valley teams), NVIDIA ensures one thing through its model-agnostic infrastructure positioning: regardless of which model wins, the hardware layer profits belong to NVIDIA. The more the model layer is commoditized, the greater the demand for inference, and the better the GPU sales. While Dynamo is open source, the optimal implementations of disaggregated serving and KV cache pinning deeply depend on NVLink bandwidth and HBM4 memory hierarchies. What is open is the software interface; what is locked in is the hardware consumption.
Meta’s logic for openness differs from NVIDIA’s, but the underlying reasoning is consistent. Meta’s core profits do not come from model APIs, but from its advertising system, social distribution, and platform control. The open sourcing of Llama is a classic commoditize your complement move: as the base model layer becomes more commoditized, the pricing power of companies like OpenAI and Anthropic, which charge directly for model APIs, weakens. Meanwhile, the supply of AI capabilities for Meta’s own distribution channels (WhatsApp, Instagram, Facebook, Ray-Ban smart glasses) becomes more abundant. The core argument of Mark Zuckerberg’s 2024 open letter, “Open Source AI is the Path Forward,” is that open source is a way to avoid being locked in by a single model provider. There are over 85,000 derivative models of Llama 4 on HuggingFace; the developers of these models are performing the work of fine-tuning, adaptation, tutorials, and secondary encapsulation for Meta.
However, Meta’s open strategy showed signs of shifting in 2026. Upon his departure, LeCun criticized Meta’s internal research for becoming increasingly closed; meanwhile, Meta began developing the closed-source Avocado model. This indicates that an open strategy is not a permanent commitment but a dynamic choice adjusted continuously based on the competitive landscape and profit pools.
The boundaries of openness for Google and DeepMind are different again. They are more willing to open research toolchains (JAX, Flax, Optax), small models (Gemma), and basic science assets (AlphaFold). The common feature of these is their ability to amplify the attractiveness of Google Cloud and the influence of the research ecosystem. Once developers become accustomed to Google’s technology stack during the research phase, it becomes easier to direct workloads to Vertex AI later. However, the flagship Gemini capability remains firmly closed-source, as it is the core of Google Cloud’s differentiation. DeepMind’s high volume of paper output (the highest in the industry at NeurIPS 2025) is a necessity for its talent strategy; its long-term researcher retention rate is among the highest in the AI industry, sustained by the promise of academic freedom and a culture of publication.
At this point, one might ask: even if publishing research is economically rational, why do many companies go further and open source code, toolchains, and even weights?
The answer is that the problems solved by open source and those solved by papers exist at different levels. Papers solve the problem of knowledge expression: letting peers know what you have done. Open source solves the problem of deployment and adoption paths: allowing others to receive, integrate, run, teach, and deploy your work in real-world scenarios. This link from knowledge to adoption can be called shipping friction. It is much greater than research diffusion friction, and open source is the most effective means of reducing it.
Specifically, open source reduces shipping friction through at least five mechanisms.
First, it externalizes adaptation costs. When only a paper is published, all replication and integration must be done by the publisher or by others starting from scratch. Once the code and inference implementation are open sourced, third parties will proactively complete hardware optimization, framework porting, cloud service encapsulation, and enterprise customization. This means the adaptation costs that would originally belong to the publisher are distributed across the entire ecosystem. Qwen is a classic case in this regard: it is already natively supported by vLLM, HuggingFace Transformers, TensorRT-LLM, and LlamaFactory. There is almost no additional technical integration cost for developers using Qwen. Alibaba did not perform the adaptation work for each framework individually; the community did it for them.
Second, it turns protocols and frameworks into default options. Open sourcing a set of interfaces or toolchains is a battle for architectural definition. Whoever defines the default interface is more likely to become the center of the subsequent ecosystem. NVIDIA’s open sourcing of Dynamo might seem like a convenience for users, but it actually ensures that all inference optimization work is conducted around its own hardware stack. Anthropic’s open sourcing of MCP follows the same logic: once MCP becomes the industry standard for Agent-tool interaction, Claude’s central position in the tool ecosystem is secured.
Third, it reduces developer education costs. Developers first form habits and muscle memory on open source models and toolchains, making the subsequent transition to commercial platforms more natural. Once the path between the research and production phases is cleared, market education costs drop significantly. Google’s release of Gemma considers this: the API patterns and tech stack habits developers learn on Gemma can be smoothly migrated to Gemini on Vertex AI in the future.
Fourth, it lowers the barrier to enterprise adoption. Many organizations—banks, medical institutions, government departments—cannot send data to third-party APIs due to data compliance requirements. Open weights allow them to deploy locally, directly eliminating a key barrier to adoption. This is why the penetration of Alibaba and Meta in enterprise scenarios does not rely solely on API call volume; locally deployed Qwen and Llama have become default choices in many scenarios, though this usage is not reflected in API statistics.
Fifth, it benefits complementary assets. If your core profits lie in the cloud, hardware, advertising, or distribution rather than the model itself, then making models and tools cheaper and easier to ship amplifies the value of your core assets. This is the specific manifestation of commoditize your complement at the shipping level.
Taken together, these five mechanisms mean that open source essentially outsources the GTM, integration, evangelism, and education costs—which the publisher would otherwise have to bear—to the entire ecosystem. This is not charity; it is the socialization of costs.
With these mechanisms understood, the strategic differences between companies become easy to explain. The root of the difference lies in the fact that each company’s profit pool is located in a different place; therefore, the layers they want to commoditize and the layers they want to protect differ.
NVIDIA’s profit pool is 100% in the hardware layer. Its open strategy is the most aggressive among all players: open sourcing the inference OS, releasing Nemotron model weights and training recipes, and embracing third-party Agent frameworks. Simultaneously, it maintains pricing power at the hardware layer through control points like NVLink binding and Privacy Routers. Its optimal world is one where the model layer is completely commoditized, inference demand grows infinitely, and all optimizations are tailored to its own hardware.
Meta’s profit pool is in advertising and social distribution. It releases Llama weights to offset its disadvantage in closed-source model capabilities compared to OpenAI and Anthropic, to ensure the open source ecosystem is built around its distribution channels, and to avoid dependence on any single model provider. However, the controversy over Llama 4 benchmarks, LeCun’s departure, and the launch of the closed-source Avocado project suggest that Meta is reassessing the costs and benefits of openness. For Meta, it remains unclear whether its $600 billion investment in AI infrastructure over three years can be converted into a sustainable competitive advantage.
OpenAI and Anthropic’s profit pools are in differentiated API pricing and subscription revenue. They choose to keep flagship model weights closed-source and selectively publish papers and safety research. OpenAI still publishes selectively for three reasons: to maintain academic reputation for recruiting researchers; to guide the industry’s direction through technical papers (the RLHF paradigm was defined by OpenAI); and to maintain the public image of welcoming competition that Sam Altman has long cultivated. Anthropic’s strategy is even more interesting: the core models are closed-source, but it is active in open sourcing at the protocol layer (MCP) and high-profile in publishing safety research (Constitutional AI, interpretability). Its safety brand is a significant source of its 40% enterprise market share, and its 6:1 talent inflow-to-outflow ratio (the highest in the AI industry) is partly due to the academic influence of its interpretability team. Anthropic’s openness is not at the model layer, but at the brand and standards layer.
Google and DeepMind’s profit pools are in search advertising and cloud services. Gemma is open to maintain a presence in the developer community, preventing the ecosystem from being entirely dominated by Llama and Qwen, while using small models to cultivate developer familiarity with Google’s AI tech stack. The Gemini flagship remains closed because it is a core asset for Google Cloud’s differentiated competition. The AlphaFold model is different again: opening basic science tools and public science assets while keeping high-value commercial applications built upon them closed-source. Google and DeepMind do not open the layers closest to their profit pools, but rather the layers that can amplify the influence of their cloud platform, research ecosystem, and industry standards.
Chinese Open Source Players (DeepSeek, Qwen, MiniMax, Moonshot) have the most aggressive open strategies: model weights are completely open, and API pricing is 80% to 95% cheaper than American closed-source models (Kimi K2.5’s input token is priced at approximately $0.60/M, while Claude Opus 4.6 is $5/M). They also actively integrate into global inference frameworks to gain native support. This strategy achieves a triple objective: gaining global technical influence through adoption (Qwen has over 700 million cumulative downloads on HuggingFace, with derivative models accounting for over 40% of new language models); attracting price-sensitive global developers with low-cost APIs (four of the Top 5 models by call volume on OpenRouter in February 2026 were from China, accounting for 85.7%); and bypassing training-phase restrictions through the diffusion of the deployment layer. This last point involves a key asymmetry in US-China competition that requires separate elaboration.
In March 2026, the U.S.-China Economic and Security Review Commission (USCC) released a research report titled “Two Loops: How China’s Open AI Strategy Reinforces Its Industrial Dominance.” It presented a judgment worth noting for Chinese AI practitioners: the U.S. has officially begun to evaluate the distribution capabilities of open source models, API pricing, framework compatibility, and industrial deployment as an independent dimension of competition.
The core insight of this report lies in a policy blind spot. The design logic of U.S. chip export controls is built on a premise: restricting China’s access to the most advanced chips will restrict its ability to train frontier models. This logic is effective in the training phase. However, the global diffusion of Chinese models occurs in the deployment phase. Deployment typically uses small models that have been distilled or quantized, which have a much lower dependence on frontier chips than pre-training. Export controls have almost no leverage over this path.
To put it more bluntly: the U.S. can control training, but it cannot control deployment.
This gives the open strategy an additional dimension in US-China competition that goes beyond commercial logic. By aggressively opening weights, offering low-cost APIs, and ensuring framework compatibility, Chinese companies have established a competitive path in the global developer ecosystem that is completely different from closed-source frontier models. Derivative models of Qwen already account for over 40% of new language model derivatives on HuggingFace, surpassing Meta Llama’s approximately 15%. When Qwen becomes the default choice for global developers, China’s model architecture, API design, and inference optimization solutions effectively become part of the industry standard. The USCC calls this “alternative pathways to AI leadership”: even if not the strongest in frontier capabilities, technical definition power can be gained through adoption and standard penetration.
This creates a strategic dilemma for American companies. If they continue to remain closed to protect the gap in frontier capabilities, global developers will accelerate their shift toward Chinese open source alternatives. If they choose to open up to compete for the developer ecosystem, it will accelerate model commoditization and compress their own profit margins on closed-source APIs. NVIDIA’s response is the clearest in this dilemma: using a model-agnostic open source software stack and a multipolar model ecosystem to ensure that regardless of whether American or Chinese models win, the hardware layer profits belong to NVIDIA.
It is important to avoid simplifying this competition into a single-dimension race. The U.S. still maintains a clear lead in frontier training and capital density. China has advantages in open source distribution, low-cost APIs, framework compatibility, and the speed of industrial deployment (67% of Chinese industrial enterprises have deployed AI in production, compared to only 34% of similar U.S. enterprises). However, progress in adoption does not equate to overall leadership, just as a lead in benchmarks does not equate to industrial dominance. The real competitive landscape is multidimensional, with the U.S. and China’s respective advantages concentrated in different dimensions.
Even if you are not making strategic decisions at a large company, understanding why AI companies publish and open source research directly affects your daily judgments as an AI practitioner.
First, it affects how you evaluate the commercial implications of a paper. A strong paper does not mean the publisher has a strong moat. If the publisher’s profit pool is not in the model layer, publishing the paper may precisely be an attempt to accelerate model commoditization to benefit its own complementary assets. When you see a new breakthrough in training efficiency being published, it is worth asking: is this company giving up an advantage, or is it targetedly compressing its competitor’s space for differentiation?
Second, it affects how you understand a company’s true advantages. Often, a company’s strongest assets are not in the model weights but in places you won’t see in papers: the scale of computing power, the pace of engineering execution, the depth of customer relationships, control over distribution channels, and the quality of the data flywheel. Anthropic’s 40% enterprise market share was not won through Claude’s benchmark scores; it was built on its safety brand, enterprise toolchains, and compliance narrative.
Third, it affects your career judgments. As model capabilities themselves are rapidly commoditizing (the equivalent inference cost of GPT-4 has dropped about 50 times in three years), the position of pure model research capability in the value chain is changing. System architecture design, Agent orchestration, quality evaluation, product integration, enterprise deployment, and standard setting—the weight of these capabilities in the profit chain is rising. This does not mean research is no longer important, but rather that the full-link capability from research to shipping has become scarcer.
Fourth, it provides you with a framework for judgment. When you see a company announce the open sourcing or closing of a technology, you can ask three questions: At which layer is this company’s profit pool? Is the layer being opened the same as the profit pool layer or a complementary layer? After opening, how much will shipping friction be reduced, and who will benefit? If a complementary layer is opened while the profit pool is elsewhere, the action is likely a rational competitive strategy. If the layer where the profit pool itself resides is opened, then either the moat of that layer is no longer defensible, or the company is making a larger strategic bet.
The final judgment is also the most macro: these open strategies are not static. Meta shifted from an open source pioneer to the closed-source Avocado; Anthropic shifted from a closed-source model company to open sourcing MCP; DeepSeek shifted from an academic lab to a contender for the global developer ecosystem. A company’s open boundaries adjust continuously with the competitive landscape and profit pools. Labeling any company as permanently “open” or “closed” will likely be outdated within a few months. What is truly useful is not remembering who opened what, but understanding the underlying interest mechanisms and continuously tracking the movement of the boundaries.
Primary Reports and Official Statements
Industry Analysis
Adoption and Competition Data
Talent Flow