In the Global Open Source War, How Many AI Players Will Be Exposed?

Once open source loses the support of a competitive position, it can be withdrawn faster than it was expanded.

The global large-model world is in turmoil.

English text is being prepared.

Among the world's leading large-model vendors, DeepSeek thus became the first company to publicly disclose a production-environment native inference framework. Previously, inference scheduling, operator optimization, and Token alignment logic had long been regarded by MaaS service providers as core trade secrets for maintaining the premium of official APIs and widening the cost gap with community deployments.

This open-sourcing is not a simple technical gift. In the accompanying technical document "DeepSeek V4.1 Flash Large-Scale Inference Deployment Practice," DeepSeek disclosed the operating baseline of its own online clusters: relying on the DeepInfer scheduling system, a single cluster carries about 155 trillion Token requests per day, which translates to peak throughput approaching 1.794 billion Tokens per second; at off-peak API pricing, the revenue corresponding to a single compute card reaches $38 per hour.

The system natively supports two hardware routes, NVIDIA SM100 GPUs and Huawei Ascend 950 NPUs. The accompanying deepseek-recipe is responsible for unifying the dialogue message format, prompt encoding, streaming output parsing, tool calling, and JSON Schema structured returns, so that the calling experience after private deployment by cloud vendors can remain consistent with DeepSeek's official API.

As the news spread, the global open-source track had already entered a high-density release cycle. Just two days after DeepSeek's open-sourcing, iFlytek, another major domestic large-model company, announced the official release of Spark X2.6, which for the first time achieved breakthroughs in code and agent capabilities based on fully domestic compute, while also announcing continued open-sourcing. Six months earlier, it had launched AstronClaw, the country's first enterprise-grade open-source, self-hosted AI skill management platform.

The pace overseas is equally urgent: Meta's Llama series continues to iterate weight versions, Google continues to supplement its inference optimization suite after the Gemma series of lightweight models, and NVIDIA led an $800 million investment to support Reflection AI in launching the Beam large model while completing the acquisition of Hugging Face.

It is worth mentioning that Jensen Huang led 25 tech companies in publishing the open letter "Open Weights and American AI Leadership," advocating maintaining the survival of the open-weight ecosystem and opposing one-size-fits-all regulatory bans. Microsoft, Meta, and Mistral were all present on the joint signatory list, while Anthropic and Google's closed-source team still insist on the closed-source path—the rift within Silicon Valley between the open-source and closed-source routes has been publicly placed on the table.

An unavoidable question then emerges: why, within just two years, have global large-model companies collectively turned to open source? If open source means model weights, inference tools, and underlying code are open to the outside, meaning anyone can download, fine-tune, deploy, and modify them, is the moat of traditional API sales disappearing?

01

To understand this AI open-source wave, the most suitable reference sample is not today's large models, but two classic contests already written into the history of the software industry: Microsoft's decades-long strategy of repeatedly wavering between closed source and open source, and Android's business experiment of rapidly capturing the smartphone market through an open-source strategy.

The feud between Microsoft and open source is the best historical negative for understanding this competition. From the late twentieth century to the early twenty-first century, Microsoft regarded Linux and the GPL as a fundamental threat to the Windows licensing model. Then-CEO Steve Ballmer once called Linux a "cancer." But after Satya Nadella took over as CEO in 2014, the wind reversed.

In June 2018, Microsoft acquired GitHub for $7.5 billion. Nadella's wording in the official statement was: "Microsoft is a developer-first company, and by joining GitHub, we strengthen our commitment to developer freedom, openness, and innovation." He also said the line that was later repeatedly quoted: "Microsoft is all-in on open source." The reason Microsoft embraced open source as if "returning to its original form" was not that open source suddenly became morally correct, but that its profit center shifted from Windows licensing to the Azure cloud service, and the cloud runs a large number of Linux and open-source workloads. Without supporting open source, Azure would be merely a "Windows Only" cloud, which would have no chance of winning in the cloud war.

The Android system that Google built in the mobile internet era had already validated the power of open-source expansion even earlier. After the iPhone was released in 2007, Google had almost no foothold in mobile operating systems. If Android had been closed-source, no phone manufacturer would have had any reason to abandon the system it was already using to bet on a newcomer. Google's choice was to open-source Android's underlying layer, attracting Samsung, HTC, Motorola, and other manufacturers to join in collectively with zero licensing cost, using the approach of "uniting all manufacturers" to counter iOS.

But Google's "open source" was a layered design from the very beginning: the underlying code was opened up to expand the device base, while GMS—Play Store, Gmail, Maps—which actually generated sustained profits, remained closed-source and firmly controlled by Google. Android's victory was the victory of a business architecture in which the open layer pursues scale and the closed layer pursues profit, not a victory of the open-source philosophy alone.

To put it simply, Android's underlying logic is to open the underlying layer and compete for an ecosystem seat; to close off traffic entry points and hold on to monetization channels. Applied to today's large-model industry, DeepSeek's open-sourcing of DeepInfer this time is essentially an upgraded version that replicates this logic: the model weights had already been opened up, and now it is further opening up the intermediate layer of inference scheduling, filling in the most expensive and highest-barrier link in the deployment chain, reducing the cost for cloud vendors and government and enterprise customers to build their own services, while DeepSeek still retains high-value businesses such as natively optimized upper-layer services, enterprise private deployment consulting, and customized fine-tuning. It no longer relies solely on selling Token APIs to earn a margin, but instead relies on ecosystem scale to expand the commercial reach of the entire technology stack.

Putting these two pieces of history together yields a plain but crucial judgment: when enterprises choose open source, it is usually not because open source itself is the goal, but because at a particular competitive position, it is the most effective lever for expanding the ecosystem, sharing costs, and countering incumbent monopolists. When this competitive position changes, the boundaries and sincerity of open source will change accordingly.

Moreover, changes at the geopolitical level have further amplified the strategic value of open source. Domestic AI vendors face an external environment of restricted high-end chip supply, and simply competing on the training compute of super-large clusters does not give them a natural advantage. Open source has become a path for differentiated breakthrough: release models, operators, and inference frameworks to the global community, rely on optimization at the algorithm and scheduling layers to produce competitive models under limited compute conditions, and use the developer community to spontaneously complete adaptation, fine-tuning, and scenario implementation, with millions of developers worldwide jointly building the ecosystem, without enterprises needing to invest huge marketing budgets on their own to educate customers.

As for Nvidia's strong support for open source, it is essentially to prevent the global open-source ecosystem from accelerating its departure from Nvidia's hardware system and damaging its own hardware business. In this game, open source is no longer just a product strategy; it is also a geopolitical bargaining chip for compute vendors and model vendors competing for ecosystem standards.

02

But beneath the boom, a cruel question hangs over all open-source players: open source does not equal victory. Opening weights and underlying code only gets you a ticket to the ecosystem arena. A large number of once high-profile open-source models quickly went silent after the hype faded—as the title says, they were beaten back to their original form amid fierce competition.

In fact, the open-source software field has repeatedly staged this plot: countless open-source projects briefly went viral, then slowly stalled after the core team's resources were exhausted, and were eventually forgotten by the ecosystem. In the large-model track, elimination will only be faster. Model evaluation standards keep updating, inference hardware keeps iterating, and customer requirements for stability, multimodality, and agent capabilities keep rising. A one-time open-sourcing cannot hold a position.

Extrapolating along current industry clues, the future evolution of this global open-source war will not lead to a single winner taking all, but to a landscape of layers, camps, and deep software-hardware binding. Closed source and open source will not eliminate each other, but will coexist for the long term and penetrate each other.

Looking at the current global large-model market, the first layer of the track is competition in underlying inference and deployment infrastructure—the battlefield of inference engines, operator libraries, and codec components. At this layer, compatibility, hardware adaptation, throughput and latency, and cost floors are the core metrics. Its profit model is not directly selling code, but supporting enterprise-level support, cluster deployment consulting, custom operator development, and compute scheduling services.

At this layer, the winner will not be a single vendor. A small number of frameworks that have been validated by trillions of tokens of traffic and are compatible with multiple hardware architectures will form de facto standards. DeepSeek's bet on DeepInfer may be aimed at joining this camp, turning its engineering experience in large-scale inference into an industry standard.

The second layer is competition over foundation model weights, which is also the track that attracts the most public attention. Foundation models will show clear divergence: some models will take the general-purpose all-around route, continuously benchmarking against closed-source flagship capabilities and covering code, mathematics, long text, and multimodality; other models will most likely move toward vertical lightweight development, targeting segmented scenarios such as on-device, industrial, medical, and automotive, pursuing low power consumption, low cost, and high determinism.

Please note that foundation players that can stand firm over the long term must simultaneously satisfy two conditions: sustained and stable compute investment to support iteration, and a large, active developer community that continuously produces fine-tuned versions and application plugins. The domestic large models with relatively high activity are all concentrated at this level, but as for players that can satisfy both conditions at the same time, you might as well count them up.

The third layer is the binding camp of the hardware-model-inference stack. This is the battlefield Nvidia cares about most, and also the most core opportunity for domestic open-source vendors. The stronger the heterogeneous adaptation capability of open-source models, the faster the ecosystem shortcomings of domestic compute can be made up.

Conversely, the maturity of the compute foundation also determines whether open-source models can be deployed at scale in government and enterprise settings. In the future, when customers choose open-source models, they often do not first look at benchmark scores, but at whether the model can run stably on their existing compute clusters. The power of the hardware ecosystem will continue to reshape models' market share. From this perspective, iFlytek is the only domestic large-model vendor with a domestic compute foundation, and perhaps it will have more say in the future.

The most obvious dividend shows up in ecosystem discourse power. Relying on open-source models, domestic vendors bypass high-end compute barriers, gain channels directly to global developers and government and enterprise customers, build brand awareness in markets that value data autonomy such as Europe, Southeast Asia, and the Middle East, and break the past pattern in which AI capabilities could only be supplied by a small number of closed-source US vendors.

Open source also forces internal engineering capabilities to evolve. Community testing, vulnerability feedback, and secondary development amount to free global quality verification. After inference frameworks and Token components are opened up, developers everywhere will stress-test extreme scenarios on different hardware, helping vendors find defects that internal testing cannot cover. At the same time, open source also creates ecosystem soil for domestic compute. For NPU platforms such as Ascend to grow, what is most needed is a large number of models, operators, and inference frameworks actively adapting to them.

But the challenges absolutely must not be underestimated.

The core problem is the difficulty of closing the commercialization loop. After open sourcing, model weights themselves are no longer a source of revenue. Enterprises must shift to monetization through private deployment, technical support, fine-tuning services, compute leasing, and industry solutions. Such businesses are heavy-delivery, long-cycle projects with gross margins far higher than API calls, but they are harder to scale. Many open-source model teams have decent technology but lack delivery networks aimed at government and enterprise customers. Community popularity is hard to convert into stable cash flow. Once the financing environment tightens, funding for continuous iteration will break, and the project will stall, which is the so-called "being beaten back to their original form."

Governance and compliance risks are an unavoidable topic. Once open-source code and model weights are distributed externally, model output content, data provenance, copyright, distillation derivatives, export controls, and local regulatory rules become complicated. Different countries have vastly different levels of control over open-source weights. Once a model flows out, it is very difficult to control the scenarios of downstream users, and compliance costs continue to rise.

What is easily overlooked is the long-term cost of community governance. An open-source community does not run automatically just because code is released. It requires continuous maintenance of documentation, merging community contributions, maintaining version branches, and coordinating derivative projects. This requires dedicated engineering teams to invest over the long term. Many domestic teams underestimate the manpower consumption of community operations.

Homogeneous involution is likewise a major hurdle. A large number of open-source models have similar metrics on evaluation leaderboards and insufficient differentiation. When customers choose, they easily fall into price comparison, compressing the profit margins of all players. The industry is also prone to vicious price wars, a scenario that has repeatedly appeared in many domestic industries such as photovoltaics and home appliances.

Returning to the initial topic, who will stay at the table?

The global open-source war will not end with "open source defeating closed source" or "closed source eliminating open source." The more likely landscape is that the boundaries of open source will be redrawn: weight open sourcing will become the norm, inference framework open sourcing will become the new battlefield for the few players with genuine engineering depth, while closed-source capabilities will retreat to frontier training that requires sustained large-scale compute investment and highly regulated enterprise service layers.

In this landscape, the one that can "laugh last" will not be the one that shouts open-source slogans the loudest, but the type of participant that can build a structural connection between open source and business. Nvidia's bet is to use capital and platforms to keep open-source developers on its own hardware track; Google uses Gemma to maintain a presence on the open-source side while using Gemini to hold the closed-source frontier. But Meta's exit reminds all participants that once open source loses the support of a competitive position, retreat can be faster than expansion.

As for how many AI projects will be beaten back to their original form in the competition, the answer will certainly not come from any announcement, but from the frequency of commit records in those open-source projects over the coming quarters and years, the quality of questions from enterprise users in the issue sections, and the next round of price restructuring around inference costs.

[This article is for discussion only, not investment advice. Please be aware of investment risks.]

This article comes from the WeChat public account "Daily Capital Talk", author: Baikaishui de Sikao, published by 36Kr with authorization.