The AI Spring Festival and the Watchtower

A choice the Global South can no longer defer

In the span of a single week around the 2026 Chinese New Year, three world-class AI models dropped in rapid succession: Zhipu’s GLM-5, a 744-billion-parameter behemoth released under the MIT open-source license; Alibaba’s Qwen3.5, supporting 201 languages under Apache 2.0; and ByteDance’s Doubao 2.0, designed for the emerging era of AI agents. All three emphasized agentic capabilities -- the ability to plan, reason through multi-step tasks, and use external tools -- signaling that the competition had moved well beyond chatbot fluency. Chinese media called it the “AI Spring Festival.”

The public responded in kind. Over the nine-day holiday, Chinese consumers generated more than ten billion AI interactions across major platforms. ByteDance’s Doubao alone handled 1.9 billion exchanges on New Year’s Eve, peaking at 63.3 billion tokens per minute -- more than ten times the rate reported for OpenAI’s infrastructure. Some 130 million people used AI to assist their holiday shopping for the first time; nearly 400,000 of them were over the age of 60.

Just one year earlier, during the 2025 Spring Festival, a single company called DeepSeek had open-sourced its R1 reasoning model and sent shockwaves around the world. What had happened in twelve months to turn one company’s breakthrough into an entire ecosystem’s eruption?

From DeepSeek to Ecosystem

The numbers tell a striking story of global diffusion. On OpenRouter, a platform where developers route queries to different AI models, Chinese open-source models went from a 1.2% share of weekly token volume in late 2024 to nearly 30% by the second half of 2025. DeepSeek alone processed 14.37 trillion tokens over a twelve-month span, more than any other single model provider on the platform. On Hugging Face, the world’s largest repository of AI models, Qwen derivatives now account for over 40% of all new language model fine-tunes -- overtaking Meta’s Llama, which has fallen to about 15%. Martin Casado, a general partner at the venture capital firm Andreessen Horowitz, estimated that roughly 80% of startups using open-source stacks for their demos are now running on Chinese models. The shift reflected something deeper than cost savings: these models were genuinely competitive. Meanwhile, the market itself had consolidated fast: a JPMorgan report noted that the number of funded model developers in China had shrunk from over 200 to fewer than 10, but the survivors emerged stronger -- Zhipu and MiniMax both completed IPOs on the Hong Kong exchange, becoming some of the first pure-play AI companies to go public anywhere in the world.

The domestic adoption story was equally dramatic. China’s generative AI user base doubled in six months, from 257 million in December 2024 to 515 million by mid-2025 -- 2.6 times the size of America’s AI user population. During the Spring Festival itself, Alibaba’s Qwen chatbot saw its daily active users surge from 7 million to 58 million. Across all platforms, companies spent over 8 billion yuan (roughly $1.1 billion) on AI red envelope promotions -- a strategy reminiscent of how WeChat popularized mobile payments through Spring Festival red envelopes in 2014, except this time the adoption curve was steeper. AI was no longer a tool for programmers and tech workers; it had become a household utility.

The key catalyst behind this mass adoption was straightforward: every major Chinese AI application is free for consumers -- compared to $20 per month for Claude Pro or ChatGPT Plus -- and their API prices run 80 to 97% cheaper than Western equivalents. The models powering these apps are released under MIT or Apache 2.0 licenses, meaning any government, company, or developer in any country can download them, deploy them locally, and modify them without restriction.

Then, right at the peak of this celebration, Anthropic published a blog post titled “Detecting and Preventing Distillation Attacks.” The company intended to demonstrate its security capabilities. What it actually revealed went far beyond distillation.

Under the Watchtower

Anthropic’s blog accused three Chinese AI labs -- DeepSeek, Moonshot AI, and MiniMax -- of using approximately 24,000 fraudulent accounts to generate over 16 million exchanges with Claude, extracting its capabilities through a technique called distillation. Distillation -- training a less capable model on the outputs of a stronger one -- is a widely used and legitimate training method. Frontier labs routinely distill their own models to create smaller, cheaper versions for customers. Anthropic itself acknowledged this. But the blog’s most revealing passages were not about the accused. They were about the accuser.

Consider what Anthropic openly described doing. The company attributed each campaign “with high confidence through IP address correlation, request metadata, infrastructure indicators, and in some cases corroboration from industry partners who observed the same actors and behaviors on their platforms.” In plain language: Anthropic records the IP addresses, request patterns, and hardware fingerprints of all users, and shares this information with other AI companies for cross-referencing. For DeepSeek, they disclosed that “by examining request metadata, we were able to trace these accounts to specific researchers at the lab.” For Moonshot, they “matched the public profiles of senior Moonshot staff.” For MiniMax, they “confirmed timings against their public product roadmap” -- meaning Anthropic’s data collection was detailed enough to infer a competitor’s unreleased product schedule. The blog further described how the company detected MiniMax’s campaign “while it was still active -- before MiniMax released the model it was training,” giving Anthropic what it called “unprecedented visibility into the life cycle of distillation attacks.” Read together, these disclosures paint a picture of a surveillance apparatus that extends far beyond catching bad actors: it is an intelligence operation capable of mapping the identities, affiliations, and strategic plans of anyone who uses the service.

The blog then announced that Anthropic is developing “countermeasures” -- “Product, API and model-level safeguards designed to reduce the efficacy of model outputs for illicit distillation.” The word “illicit” is defined by Anthropic alone, with no legal proceeding, no judicial review, and no external oversight. This means that any paying customer whom Anthropic unilaterally deems suspicious could receive degraded outputs without knowing it.

The legal foundation for these measures is shaky at best. As researchers at the University of Montreal’s Cyberjustice Laboratory have argued, distillation primarily transfers knowledge rather than model parameters, and under current law it is not clearly illegal. The Laboratory’s analysis made a pointed observation: “Large AI development companies cannot reasonably claim they should be allowed to use others’ works for training without permission or compensation, while simultaneously prohibiting others from using their model outputs for similar purposes.” Elon Musk, hardly a sympathizer of Chinese tech, was even blunter. “Anthropic is guilty of stealing training data at massive scale and has had to pay multi-billion dollar settlements for their theft,” he wrote on X. “This is just a fact.” (Anthropic settled a $1.5 billion copyright dispute over books used to train its models last year.)

But perhaps the most consequential lines in the blog were the political ones. Anthropic declared that it “has consistently supported export controls to help maintain America’s lead in AI,” and described the accused labs as “foreign labs, including those subject to the control of the Chinese Communist Party.” CNN, Fortune, and The Verge all covered the story, but none focused on the surveillance infrastructure that the blog inadvertently exposed. For Global South countries watching from the sidelines, these two sentences turned a technical dispute into a geopolitical declaration: the company providing your AI services is not a neutral vendor. It is an active participant in a great-power technology contest, and it is watching.

What This Means for the Global South

Place these two pictures side by side and the decision framework becomes remarkably clear. On one path, you use closed-source AI from a company that records your metadata, shares it with “industry partners and relevant authorities,” can degrade your service at will, and explicitly supports export controls designed to maintain one country’s technological advantage. On the other path, you download an open-source model under an MIT or Apache 2.0 license, deploy it on your own servers, and keep your data within your borders. No one monitors your queries. No one poisons your outputs. No one can cut off your access because of a shift in geopolitical winds.

The performance gap between these two paths has nearly closed. GLM-5 scores 77.8 on SWE-bench Verified, a widely used coding benchmark -- less than four percentage points behind Claude Opus 4.6’s 80.9. It also achieved what VentureBeat described as a record-low hallucination rate among open-source models. Its API costs roughly 80% less. DeepSeek’s V3.2 is 97% cheaper. And the consumer-facing applications built on these models are entirely free. For a government ministry in Lagos, a startup in Jakarta, or a research lab in Sao Paulo, the math is hard to argue with: comparable capability, a fraction of the cost, and full control over your data.

The Tony Blair Institute’s recent report on open-source AI and middle powers captured the strategic logic precisely: “The paradox is that what is important for middle powers is not to build models, but to have the capacity to build models.” Chinese open-source models now account for 50-60% of new open-model adoption globally, overtaking American models at 25-30%. The infrastructure to run them is already being built: India’s national AI mission has deployed over 34,000 publicly funded GPUs; Mexico’s Coatlicue supercomputer has reached 314 petaflops of capacity. On z.ai, Zhipu’s global platform, users from India, Brazil, Japan, and the United Kingdom are already among the most active outside the US and China. As CSIS has noted, developing countries face a crossroads: Western AI systems that can be “tortuous to deploy” versus Chinese alternatives that are more cost-effective but carry their own implications. Open-source models offer a third way -- one that sidesteps both dependencies. This is not a theoretical possibility. The migration is underway.

I have experienced this shift firsthand. As someone who has written about agentic AI workflows -- where an AI model acts as an autonomous programmer, planning and executing multi-step coding tasks -- I have begun moving my daily work from Claude to GLM-5. It plugs directly into the same agentic tooling I was already using. The experience is, for practical purposes, indistinguishable. GLM-5 handles complex multi-step programming tasks, reasons through debugging chains, and manages project-level context with the competence I need. Demand for the model has been strong enough that z.ai had to temporarily restrict new subscriptions to its GLM coding plan. This is not an endorsement based on benchmarks alone -- it is a report from daily use, and a migration path that anyone can replicate.

The Next Twelve Months

Open source has not just changed China’s AI landscape. It is redrawing the global map of AI capability. When the world’s largest AI user base -- 515 million people and counting -- commits to an open path, the catalytic effect compounds: more users generate more feedback, more developers build more applications, more fine-tuned models serve more languages and use cases, and the entire ecosystem accelerates in a virtuous cycle that closed systems cannot replicate. A year ago, open-source AI was a niche movement led by Meta’s Llama and a handful of academic projects. Today, four of the five most-used model providers on OpenRouter are Chinese AI startups.

If no one predicted the leap from the DeepSeek moment to the AI Spring Festival in just twelve months, there is little reason to assume the next twelve will be any less surprising. The network effects of open ecosystems reward early adopters disproportionately. For countries in the Global South weighing their options, the window to join is open now -- and the cost of waiting grows with each passing quarter.

Between the jubilant noise of the AI Spring Festival and the quiet gaze of the watchtower, the choice is becoming hard to ignore. One path invites you to the celebration. The other asks you to be watched.