MiniMax wants to find the next "10x"

With the surge of ClaudeCode, AI is shifting from a chat tool to an agent. As models begin to truly carry out tasks on behalf of humans, token consumption will grow exponentially.

Whoever can truly bring AI into the production process will gain the most stable and sustained token consumption. After the collective spike in China’s AI concept stocks at the beginning of the year, followed by a quick pullback, domestic large-model companies have started looking for new growth narratives again.

After catching the vibe coding trend and the lobster craze and getting a taste of success, domestic model players like MiniMax are rushing to expand their network of contacts in search of the next gold mine.

On May 11, MiniMax released a new collaboration plan called “10xTeam.”

In addition to the vertical areas already bound—industrial software, game engines, chip design, finance, and financial services—this time MiniMax is mainly publicly inviting experts from economics, life sciences, and materials chemistry, fields that are more globally oriented and more likely to be deeply integrated with large models, to co-create together. At the same time, it also launched the “10xTeam Researcher” role on recruitment platforms.

The ambition behind it is obvious: they want to replicate the “10x efficiency leap” that has appeared in the programming field into more industries.

This will be a win-win. Through this approach, MiniMax can enhance the capabilities of the general intelligence foundation, while also pushing the models’ penetration into deeper, more in-depth scenarios across additional industries.

In fact, “general large models + co-building with industry experts” has already become a consensus among leading companies.

Anthropic has long been recruiting academic and industry researchers; its EconomicIndex further incorporates the impact of models on economic activities across different industries into its evaluation scope. OpenAI has launched HealthBench for healthcare and treats legal and financial scenarios as key optimization directions for the GPT series. GoogleDeepMind has long flown the flag of “breakthroughs in the scientific field”: AlphaFold (structural biology), GNoME (materials science), and others have shown that collaboration between top domain experts and foundational research teams can produce “domain-level leaps.”

By the end of 2025, Baidu also floated a similar “Wenxin Mentor” plan—inviting experts from both industry and academia to guide large models in knowledge transfer, quality assessment, and professional calibration.

Over the past year, the programming field has become the earliest phenomenon-level scenario where large models show “10x efficiency.” Tools such as Cursor and ClaudeCode have effectively reshaped the software development process, and competition among related infrastructure has largely been completed.

After ClaudeCode went viral, the entire AI industry quickly reached a consensus: AI’s most important capability is no longer simply “answering questions,” but “completing tasks.” Once AI enters real production systems, it becomes a must-have.

Programmers need to call it every day; enterprises need to run it every day; team collaboration must be continuously integrated; and reasoning chains will keep growing. Model usage shifts from occasional needs to ongoing consumption, and token revenue naturally begins to increase exponentially.

But such certainty has also attracted players to carve up the pie. Eighteen months ago, AI programming was dominated by Copilot alone. Now, overseas, Cursor, Windsurf, Cline, Claude Code, and Aider are fighting fiercely; domestically, DeepSeek TUI, Kimi Code, MiniMax-M2.5, ByteDance’s Trae, Tongyi Lingma, Wenxin Kuai Ma, Zhipu’s CodeGeeX, Alibaba’s Qoder, and others are all competing for market share.

When programming dividends enter a bottleneck period, “what will be the next 10x field?” will become the question every company needs to answer.

MiniMax’s answer is: to push model capabilities into fields with high professional knowledge density, complex workflows, and where standardized playbooks have not yet formed.

This is precisely something that cannot be solved by optimizing behind closed doors within model teams alone. It requires top industry experts to step in to define problems, co-build evaluation and workflow systems, and then have the models drive industry transformation in return.

Industry knowledge naturally has very high barriers.

Chip design involves complex verification processes; industrial software has massive engineering systems; finance has its own risk-control logic and regulatory framework; and life sciences are full of implicit experimental experience and specialized knowledge structures. None of these naturally exist in openly available internet corpora.

For a truly usable industry agent, the challenge isn’t primarily model inference capability—it is whether it understands the industry’s workflows.

This has caused large-model companies to increasingly resemble a hybrid of research institutions, industry organizations, and consulting firms. MiniMax’s “10xTeam,” to some extent, is also the first time Chinese large-model vendors have clearly brought this “scientific collaboration model” to the front stage.

In MiniMax’s view, this is more like an industry research partner mechanism. The model team is responsible for foundational capabilities; industry experts define problems, build workflows, and establish evaluation systems; and then the agent is deployed into actual production scenarios.

Because when AI’s goal shifts from “answering questions” to “completing tasks,” the importance of industry experts will be rapidly amplified.

Looking back, in the internet era, the most important talent was product managers, because they defined user needs. In the agent era, the truly important people may become those who understand industry processes best.

Programming is just the first industry being reconstructed by agents. What all large-model companies are truly trying to find now is the next scenario that can generate massive token consumption while truly creating industry value.

Over the past year, the valuation growth speed of the large-model industry has begun to make more and more people think of the internet bubble around 2000.

Recently, economist Ma Guangyuan pointed out that upstream infrastructure like computing power, optical modules, and hardware does indeed have orders, revenue, and profitability, because the world is frantically stockpiling computing power. But midstream large models and downstream applications such as humanoid robots, general AI, and ToC/ToB deployment scenarios still remain at the concept and storytelling stage so far, with no large-scale commercialization, no sustained profitability, and no real demand explosion—yet all these future expectations have already been fully priced into current valuations.

The whole industry is actually well aware that if AI cannot truly enter industries, and cannot help companies continuously improve efficiency and make money, then this capital game is unlikely to be sustainable in the long run. Only when AI truly starts working for enterprises, participates in production processes, and helps industries make money can the entire industrial chain operate properly.

That is also why today’s global leading AI companies are rushing into deeper waters of industry.

Anthropic no longer emphasizes only model capabilities, but instead focuses on how Claude can enter enterprise workstreams. OpenAI continues to strengthen vertical scenarios such as healthcare, legal, and finance. GoogleDeepMind has long treated “scientific breakthroughs” as an important strategic direction.

Because everyone knows that AI must genuinely start helping industries make money, improve efficiency, and reduce costs for the industry narrative to keep moving forward. Otherwise, the bubble will eventually be burst.

And once the bubble bursts, it will not only affect a few model companies. From GPUs to cloud providers, from data centers to AI startups, from the primary market to the secondary market, the entire AI upstream and downstream could experience a severe winter.

So today, all large-model companies are essentially racing against time to prove one thing: AI is not just a concept—it is real productivity. And MiniMax’s “10xTeam,” in essence, is also a strategic positioning for the industry in this context.

It aims to bind industry experts in advance, embedding model capabilities into complex industrial processes such as chip design, industrial software, financial analysis, and life sciences, and then gradually forming its own data barriers, workflow barriers, and commercialization barriers.

Because when AI’s goal shifts from “answering questions” to “completing tasks,” industry knowledge becomes a new scarce resource. Programming is just the first industry reconstructed by agents. What the entire AI industry now truly wants to prove is: could the next one be the entire industrial world?

Risk Disclaimer and Exemption of Liability

        There are risks in the market; investing requires caution. This article does not constitute personal investment advice, nor does it take into account the special investment objectives, financial conditions, or needs of any particular user. Users should consider whether any opinions, viewpoints, or conclusions in this article are appropriate for their specific circumstances. Investments made based on this are at your own risk.