La batalla de la IA que se juega al margen de OpenAI

Chinese open-source AI models have demonstrated business demand for alternatives to Silicon Valley's leading artificial intelligence technology. Companies like Mistral, Reflection, and Aleph Alpha are now aiming to make headway with technology developed in the West. In less than a month, several Western firms have moved to compete with OpenAI and Anthropic for AI adoption in enterprise.

French Mistral just unveiled Large 4, while German Aleph Alpha – now merged with Canadian Cohere – launched Kolibri. Meanwhile, US-based Reflection, backed by Nvidia, announced Beam, its first open-source model. Over the past two years, the main contenders to OpenAI, Anthropic, and Google have emerged from China. Systems such as DeepSeek, Qwen (Alibaba), Kimi (Moonshot), and the GLM family (Z.ai) have shown that there is a market for efficient models, even if they are not the most advanced.

This progress, at a time when businesses are closely watching the cost of deploying AI in the era of autonomous agents, has opened a new front in the AI race. Many of these systems are open-source, meaning companies and developers can download their training parameters, adapt them to their needs, and run them on their own systems, unlike closed platforms like ChatGPT and Claude, where the technology remains under the control of its owners.

Companies want to compete in this segment with proposals developed in Europe and the United States, highlighting their Western status and, in the European case, the need to strengthen technological sovereignty. Mistral, Aleph Alpha, and Reflection, as Western companies, are betting on open-source models that allow businesses and public institutions greater control over the technology.

"This market goes beyond mere model performance and focuses on aspects such as governance, transparency, the ability to deploy in controlled environments, and strategic independence," explains Manuel Cid, Capgemini's Director of Insights & Data. According to the consultancy, the opportunity for these players lies in the growing demand from organizations seeking to reduce technological dependencies, maintain control over their data, and meet stringent regulatory requirements.

This comes alongside an evolution from competition for the largest models to the creation of AI platforms tailored to specific business cases. While OpenAI, Anthropic, and Google still lead the technological race, many organizations do not need the most advanced AI for all their tasks and seek more efficient alternatives. According to Vercel data, open-source models account for 56% of tokens (the basic unit of computation processed by a model) consumed in August.

Moreover, Alberto G. Arrieta, responsible for Digital Core at Accenture, notes that there are increasingly more use cases solvable with open-source models, which are "around six months behind" the most advanced ones. Western companies competing in open-source models are catching up with their Chinese counterparts. Mistral's latest system, Large 4, the most advanced in Europe, lags far behind the leaders on the Artificial Analysis index, a reference for IA performance tests.

However, the French company has made a leap with its latest system, which, according to the company, has improved in areas such as programming, finance, or cybersecurity. Mistral uses an architecture that activates only a portion of its parameters in each query to reduce computational costs. Meanwhile, US-based Reflection claims its technology is on par with Z.ai's GLM-5.2, one of the most advanced Chinese models, and can compete in programming and AI agent tasks at a fraction of the cost of some US rivals.

The company is backed by Nvidia, which supports open-source models as they expand the number of organizations capable of deploying AI on their own infrastructure, increasing demand for chips. Financial Times reports that Nvidia is in talks to acquire Reflection. Geopolitically, the reading is clear. US Treasury Secretary Scott Bessent recently warned Congress that the United States needs more homegrown open-source models to reduce its dependence on China and the dominance of big Silicon Valley labs.

In this battle, Meta plays both sides, pursuing closed models like the new Muse Wonder, but also advocating for open-source ones. Its former chief scientist, Yann LeCun, initially supported this approach but has since distanced himself. However, in August, Meta launched Muse Glimmer, an open-source model that can run locally on computers.

CEO Mark Zuckerberg reiterated Meta's support for open-source models as an alternative to the concentration of power in a few labs. The company promised to continue backing this formula. In Europe, the proposal is based on arguments such as technological autonomy, data protection, and the development of domestic AI capabilities.

Cid confirms the "increasing interest" among European organizations to leverage generative AI without losing control over where their data resides, how models evolve, or the strategic dependencies they assume. The question is whether Europe can create its own DeepSeek. "Rather than exactly replicating the US or Chinese model, Europe has the opportunity to lead a proposal for open, reliable, and sovereign AI," Cid argues. China's influence in open-source models is evident.