The Open-Weight Paradox: Why Chinese AI Labs Are Losing Revenue to Third-Party Platforms

A Distribution Problem, Not an Adoption Problem

China's latest generation of large language models has achieved something rare in AI: genuine global traction without the marketing budgets or cloud infrastructure of Western incumbents. DeepSeek, Moonshot AI, and a handful of other mainland developers have released model weights - the parameter configurations that encode an AI system's capabilities - under permissive licences over the past two years. Developers from Jakarta to São Paulo now download these weights, fine-tune them for local applications, and deploy them in production.

Yet a study released this month reveals a striking asymmetry. Whilst adoption has surged, the commercial returns are accruing primarily to third-party platforms that host, serve, or wrap these models in APIs, rather than to the labs that trained them. The finding underscores a structural tension in the open-weight strategy: release accelerates reach but dilutes capture.

At Opentechwire, we've tracked this trade-off across multiple funding rounds in the region. The question is no longer whether open weights can compete with closed APIs - DeepSeek's inference benchmarks and Moonshot's context-window performance have settled that. The question is whether the labs building these models can extract enough value to sustain the next training run, or whether they are effectively subsidising a new layer of intermediaries.

How Open Weights Became the Default in China

The open-weight wave in China did not emerge from altruism. It arose from a combination of regulatory pressure, competitive necessity, and pragmatic calculation. Beijing's 2023 guidelines on generative AI required model providers to register with authorities and assume liability for outputs - a compliance burden that made closed, pay-per-token APIs less attractive for smaller developers who could not afford legal overhead or content-moderation teams.

Releasing weights instead shifted that burden downstream. A startup in Shenzhen or Bangalore that downloads a set of weights and runs inference on its own hardware becomes responsible for what the model generates. For the originating lab, this structure reduces regulatory exposure whilst enabling wide distribution.

Competitive dynamics reinforced the shift. With Alibaba, Tencent, and Baidu each fielding proprietary models and cloud platforms, smaller labs faced a choice: compete head-on for API revenue against entrenched ecosystems, or pursue volume and mindshare through open release. DeepSeek, which spun out of quantitative trading firm High-Flyer, chose the latter. Its 67-billion-parameter model, released in early 2025, was optimised for low-latency inference and required less VRAM than comparable Western models - a design decision that made it particularly appealing to developers in markets with constrained GPU access.

Moonshot AI, backed by Alibaba and Sequoia China, followed a similar path with its Kimi series, emphasising long-context performance - up to 200,000 tokens in some configurations. Both labs made weights available via Hugging Face and domestic mirrors, with minimal restrictions beyond attribution.

Where the Revenue Actually Lands

The study - conducted by a Beijing-based AI policy institute and drawing on transaction data from cloud providers, API aggregators, and developer surveys across twelve markets - traced revenue flows from model deployment back to originating entities. The results were unambiguous.

Approximately 68 per cent of revenue generated by applications built on Chinese open-weight models accrues to third-party providers. These include cloud platforms that offer managed inference endpoints, API aggregators that bundle multiple models behind a single interface, and specialist firms that fine-tune and resell weights for vertical markets such as healthcare, logistics, or customer service.

Roughly 21 per cent stays with the end-application developers themselves - software-as-a-service firms, chatbot providers, and enterprise AI teams that integrate the models directly. Only 11 per cent flows back to the originating labs, primarily through optional support contracts, enterprise licensing arrangements, or donations.

The concentration is even starker at the top of the value chain. A handful of API platforms, several of them based in Singapore and the United States, account for more than half of the third-party revenue. These platforms typically charge developers between USD 0.30 and USD 1.20 per million tokens, depending on model size and latency guarantees. The originating lab, meanwhile, receives nothing per token - only the reputational benefit of widespread use and the optionality of converting some users into paying customers later.

The Structural Costs of Open Release

Training a frontier language model costs between USD 5 million and USD 50 million, depending on parameter count, data quality, and the efficiency of the training stack. For Chinese labs, those costs are partially offset by lower electricity prices and access to domestic GPU supply chains, but they remain substantial. DeepSeek's 67B model reportedly required around 18,000 GPU-hours on a cluster of Nvidia A100 equivalents; Moonshot's long-context variant demanded even more, due to the quadratic scaling of attention mechanisms.

Without meaningful per-token revenue, labs must rely on alternative monetisation paths. Some, like Zhipu AI, have pivoted towards enterprise contracts, offering fine-tuning services and on-premise deployments for state-owned enterprises and financial institutions. Others are exploring hybrid models: open weights for smaller configurations, closed APIs for the largest or most capable versions.

But these strategies introduce their own trade-offs. Enterprise sales cycles are long, and the total addressable market within China is constrained by the fact that many large organisations prefer to build in-house or partner with Alibaba Cloud, Tencent Cloud, or Huawei. Hybrid release, meanwhile, risks fragmenting the developer community and ceding the open tier entirely to competitors.

The study's authors note that the revenue gap is widening. As third-party platforms mature and build sticky relationships with developers - offering not just inference but also monitoring, fine-tuning pipelines, and compliance tooling - they become harder to displace. The originating lab, by contrast, remains a largely invisible infrastructure provider, credited in README files but absent from invoices.

A Familiar Pattern from Other Platforms

The dynamic is not unique to AI. Open-source software has long grappled with the tension between broad adoption and concentrated value capture. Linux powers the majority of cloud servers, yet the Linux Foundation's annual budget is a rounding error compared to the revenue of AWS, Azure, and Google Cloud. MySQL and PostgreSQL underpin millions of applications, but it is managed database providers - not the core maintainers - that generate most of the associated revenue.

What distinguishes AI is the capital intensity. A database can be maintained by a small team on a modest budget. A frontier language model requires continuous investment in compute, data, and talent. If the labs releasing open weights cannot capture enough value to fund the next generation, the strategy becomes unsustainable - unless they are willing to accept a role as research organisations rather than commercial entities.

Some Chinese labs appear to be making that calculation explicitly. DeepSeek has positioned itself as a research-driven project, with revenue expectations subordinated to the goal of demonstrating that high-performance models can be built outside the Western AI oligopoly. For founders with backgrounds in quantitative finance, where model performance matters more than market share, this framing is internally coherent.

But it is unclear whether venture investors share that view. Moonshot AI raised USD 300 million across two rounds in 2024 and 2025, at a valuation reportedly exceeding USD 2 billion. Those investors will eventually expect returns, and reputational wins do not pay dividends.

Policy and the Shape of the Market

Beijing's posture towards open-weight models remains ambivalent. On one hand, regulators welcome the diffusion of Chinese AI capabilities as a counterweight to US dominance in closed, API-first models. On the other, they are wary of losing control over how those capabilities are used, particularly in sensitive domains or by actors outside mainland jurisdiction.

The registration framework introduced in 2023 does not explicitly discourage open release, but it does impose ongoing compliance obligations on labs that make weights available. Labs must monitor "known uses" and report misuse - a requirement that becomes nearly impossible to enforce once weights are mirrored across dozens of platforms and jurisdictions.

Some policy researchers have floated the idea of a tiered licensing regime, in which weights released under permissive licences would be subject to lighter compliance burdens, whilst those released under commercial licences would trigger registration and liability. The proposal has not advanced beyond working-group discussions, but it reflects growing recognition that the current framework was designed for closed APIs and does not map cleanly onto open weights.

In the meantime, the market is solving the problem through vertical integration. Several large Chinese cloud providers are acquiring or investing in open-weight labs, not to close the weights but to bundle them into managed services and capture the downstream revenue. Alibaba's investment in Moonshot AI is one example; Tencent's partnership with Zhipu AI is another.

These arrangements offer labs a path to sustainability, but they also risk replicating the centralisation dynamics that open release was meant to circumvent. If the only viable business model is to become a subsidiary or strategic partner of a hyperscale cloud provider, the open-weight ecosystem becomes an extension of the cloud oligopoly rather than an alternative to it.

The Inference Arbitrage Window

One underappreciated factor in the revenue split is the falling cost of inference. Two years ago, running a 67-billion-parameter model required a multi-GPU server and careful optimisation. Today, quantisation techniques, speculative decoding, and improved kernels have brought inference costs down by an order of magnitude. A developer can now run a quantised version of DeepSeek's model on a single Nvidia L4 or even a high-end consumer GPU.

This commoditisation erodes the pricing power of managed inference providers, but it also reduces the addressable revenue pool. If inference becomes cheap enough, developers will run models locally rather than pay for hosted endpoints, and the entire API layer - both third-party and first-party - shrinks.

For open-weight labs, this trajectory is a double-edged blade. On one hand, it validates the technical efficiency of their models and expands the potential user base. On the other, it accelerates the shift from a service economy to a software distribution model, in which the primary value is in the weights themselves rather than in serving them.

Some labs are betting that this shift will eventually work in their favour. If weights become the product, then brand, performance benchmarks, and trust in the training process become the differentiators - and those are assets the originating lab controls. But that outcome depends on developers caring who trained the model, and the evidence so far is mixed. In surveys, developers cite performance, licensing terms, and ease of integration as their top criteria; the identity of the lab ranks lower.

What the Labs Are Doing Now

Faced with these headwinds, Chinese open-weight labs are experimenting with several responses. DeepSeek has begun offering premium support contracts to enterprise users, with SLAs around model updates, fine-tuning assistance, and compliance documentation. Moonshot AI is exploring a dual-track model: open weights for the base Kimi model, closed API access for versions fine-tuned on proprietary data or optimised for specific tasks.

Others are leaning into the research and talent pipeline. By releasing high-quality models, labs attract top PhD students, postdocs, and engineers who want to work on the frontier. This strategy treats the open-weight release as a recruiting tool rather than a revenue stream - a bet that the long-term value lies in the team rather than the model.

A few labs are also lobbying for public or quasi-public funding models, arguing that open-weight AI is a form of digital infrastructure and should be supported the way governments support roads, ports, or broadband. Whether Beijing will embrace this framing - and what strings it might attach - remains an open question.

The Global Context

The Chinese open-weight surge is unfolding against a backdrop of tightening US export controls on advanced GPUs, which constrain the training capacity of mainland labs but also make efficient, open-weight models more strategically valuable. If Chinese developers can achieve comparable performance with smaller, more efficient architectures, they reduce their dependence on cutting-edge Nvidia hardware - and make their models more accessible to developers in other GPU-constrained markets.

This dynamic has caught the attention of policymakers in Washington, who worry that open-weight releases could undermine the effectiveness of export controls by enabling adversaries to build capable systems without access to restricted chips. The debate is ongoing, but it has already influenced the licensing terms of some US-based open-weight projects, which now include use restrictions targeting certain countries or applications.

For Chinese labs, the geopolitical dimension adds another layer of complexity. Open release is both a competitive strategy and a soft-power tool, demonstrating that advanced AI can emerge from outside the US-led ecosystem. But it also invites scrutiny and potential countermeasures, particularly if the models are perceived as dual-use technologies.

Capture or Concede

The central question is whether the current revenue split is a temporary inefficiency that will correct as labs build direct relationships with developers, or a structural feature of open-weight economics that labs must accept and adapt to.

If the former, the prescription is to invest in developer relations, build proprietary tooling around the weights, and create switching costs that make it harder for third-party providers to commoditise the offering. If the latter, the prescription is to treat open weights as a loss leader - a way to build brand, recruit talent, and position for acquisition or partnership - rather than as a standalone business.

The study offers no definitive answer, but it does provide a benchmark. At 11 per cent revenue capture, Chinese open-weight labs are operating far below the threshold needed to self-fund the next generation of models. Unless that number improves, or unless external funding - venture, strategic, or public - continues to flow, the open-weight wave will crest and break, leaving behind a handful of well-funded survivors and a long tail of archived repositories.

For now, the models keep shipping, the downloads keep climbing, and the revenue keeps flowing - just not to the people who trained the parameters.