Emre Kazim, co-CEO of Holistic AI, is an expert in AI ethics and governance and holds a Ph.D. in Philosophy from King’s College London.
Each year since 2018, the State of AI Report, put out by AI investor Nathan Benaich and Air Street Capital, acts not just as a scoreboard of model metrics but as a mirror of how technological advances and ecosystem strategies are evolving.
The 2025 edition signals that open-weight models, offered mainly from companies based in China (such as DeepSeek, Alibaba Cloud’s Qwen, Moonshot AI’s Kimi and MiniMax), are closing the gap versus leading AI companies based in the U.S. (such as OpenAI and Anthropic).
As the co-founder of an AI company that relies heavily on open-source software, this is a topic my leadership team regularly evaluates. Based on this experience, openness and cost are likely key factors driving AI adoption preference. Let’s break down how these shifts are reshaping global AI competition and what business leaders can do to stay ahead.
The New Equation: Access × Speed × Scale
Leading open-source models from Chinese-based organizations now rank among the top global models. Cumulative downloads and derivative fine-tunes outpace releases from U.S. companies.
The State of AI report states that Qwen now accounts for about 40% of new fine-tunes on Hugging Face, suggesting a shift in large language model (LLM) leadership. Overall, downloads of open-source models from Chinese firms recently eclipsed those of U.S. firms, marking an inflection point that could signal a widening gap in popularity.
Why? Because, for startups and researchers, open-weight models can lower the barrier to innovation: no billion-dollar compute budgets or exclusive API access required.
When weights, architectures and fine-tuning pipelines are available, niche players can flourish, from vertical-AI firms to localization labs and universities. A closed-API world benefits incumbents; an open world invites experimentation.
It’s not just startups and researchers. Global companies are also adopting Chinese open-weight systems. Airbnb, for instance, recently explained that its AI customer-service agent uses 13 models, including Alibaba’s Qwen line, because they are “faster and cheaper.” When global enterprises place open-weight systems into production, the infrastructure of innovation begins to shift.
Global AI Adoption And Open Models
Regulatory caution has led some analysts to warn that limited access could slow experimentation and innovation. Export controls and bans also incentivize Chinese companies to use domestic technologies over U.S. alternatives, fueling growth and competitiveness.
By prioritizing open source development, companies in China can expand local AI communities and export open innovation abroad. As open-source models proliferate globally, they influence prevailing architectural standards, developer ecosystems and workflows in AI development.
The Governance Challenge: Speed Of Trust
The question is no longer whether open-source models can perform; they can. The bigger question is whether they can compete sustainably on trust, safety and enterprise-grade reliability.
Open need not mean unsafe. It’s time to stop treating “open” and “secure” as mutually exclusive concepts.
My company recently red-teamed open-sourced models from Chinese-based organizations. The results show that these models are narrowing the gap in both safety and performance. Meanwhile, bundling proprietary lock-in under “safety” and “IP protection” risks slowing the very innovation that once fueled Silicon Valley.
For business leaders, the core issue is not geopolitics but operational readiness. Enterprises need AI systems that are transparent, governable and deployable within real-world constraints. Weaker technology may not determine who falls behind in the AI landscape as much as slower openness limiting adoption, experimentation and speed-to-value.
Five Moves To Keep Pace
Against this backdrop, business leaders need to understand how to navigate an AI landscape where openness, cost and accessibility are rapidly reshaping the landscape. Here are a few steps to consider:
1. Adopt a “portfolio approach” to models. Enterprises should diversify across open-weight, proprietary and specialized models. This reduces dependency risk, lowers cost and ensures flexibility as global innovation accelerates.
2. Develop governance-ready licensing and procurement. Create open-source procurement frameworks that mandate transparency in training data, evaluation standards, red-team results and safety mitigations. Incorporate these requirements into RFPs, ensuring every model—regardless of origin—meets your organization’s governance baselines.
3. Enable continuous red-teaming and validation. Support independent testing of all foundation models. Enterprises should treat AI model safety like cybersecurity: ongoing, measurable and independently validated. Think “SOC 2 for LLMs.”
4. Build internal AI benchmarking capabilities. Rather than relying on broad restrictions or assumptions about model origin, organizations benefit from rigorous, consistent benchmarking. By developing lightweight pipelines that measure latency, accuracy, hallucination rate and safety, enterprises can make informed choices and strengthen their competitive position.
5. Invest in global talent and open-source participation. Include global voices in open source and governance forums. Enterprises should contribute to and learn from global communities. Participation unlocks early visibility into emerging innovation and helps shape norms before they solidify.
In 2026 and beyond, AI leadership will hinge on ecosystem participation, not model size. Open, low-cost and high-performing models are gaining rapid global traction, shifting priorities for companies navigating AI adoption.
Business leaders risk ceding advantage if they treat openness as a liability rather than an accelerant. Therefore, the mandate is clear: Build on trusted models, diversify your AI stack, invest in transparency and evaluation and treat openness as a strategic multiplier.
Forbes Technology Council is an invitation-only community for world-class CIOs, CTOs and technology executives. Do I qualify?
