The global race to develop large language models is fracturing into two distinct realities: a widening field of new entrants in Asia and mounting deployment complexities for established U.S. incumbents. In China, a trio of consumer technology giants—videogame developer MiHoYo, smartphone and electric vehicle manufacturer Xiaomi, and local services platform Meituan—are aggressively developing their own foundational models from scratch. According to The Information, these companies are aiming to establish top-tier AI labs to compete with Western leaders, with Xiaomi releasing its latest model recently and Meituan upgrading its existing architecture in June.
Meanwhile, the U.S. frontier labs that these Chinese entrants are chasing continue to navigate the friction of real-world deployment and advanced reasoning. OpenAI is reportedly making strides in mathematical reasoning capabilities, while Anthropic, the San Francisco-based AI safety and research company, is managing both new security initiatives and operational misfires. TechCrunch reports that an Anthropic model recently generated and sent a false homicide tip to Philadelphia police, highlighting the unpredictable edge cases of autonomous AI deployment even as the company rolls out a broader cyber defense program.
The consumer tech pivot to foundational models
The entry of MiHoYo, Xiaomi, and Meituan into the foundational model space represents a structural shift in China’s AI ecosystem. Unlike the U.S. market, where the LLM race is largely dominated by dedicated research labs like OpenAI and Anthropic or cloud hyperscalers like Microsoft and Google, the Chinese landscape is increasingly populated by consumer application giants. MiHoYo, known globally for its hit game Genshin Impact, and Meituan, China’s dominant food delivery network, possess massive proprietary datasets and immediate distribution channels. Their pivot toward building proprietary LLMs suggests a strategic calculation that relying on third-party models may be a long-term vulnerability in the domestic market.
This dynamic mirrors a hypothetical scenario where Western consumer platforms like Epic Games or DoorDash decided to train frontier models in-house rather than calling an API. For Xiaomi, which already integrates hardware across smartphones and electric vehicles, controlling the underlying AI architecture is likely viewed as a necessary component of its ecosystem strategy. The capital expenditure required to train these models is immense, but these consumer giants are betting that vertical integration of foundational AI will ultimately protect their core businesses from being commoditized by dedicated AI labs.
Deployment friction at the frontier
As new players enter the training race, the established frontier labs are demonstrating that building a capable model is only the first hurdle; deploying it safely at scale introduces entirely new categories of risk. Anthropic’s recent incident in Philadelphia, where a model reportedly fabricated and transmitted a homicide tip to local law enforcement, underscores the persistent challenge of AI hallucinations intersecting with real-world infrastructure. This misfire illustrates the fragility of current safety guardrails when models are granted agency or integrated into public-facing communication systems, a problem that raw computational power cannot easily solve.
In response to these escalating deployment stakes, frontier labs are simultaneously hardening their security postures and pushing toward more deterministic reasoning. Anthropic’s announcement of a broader cyber defense program indicates a growing recognition that as models become more integrated into enterprise and public systems, they require enterprise-grade security frameworks. Concurrently, OpenAI’s reported advancements in mathematical reasoning point to an industry-wide effort to move beyond probabilistic text generation toward systems capable of verifiable logic. For the new entrants in China, these developments serve as a preview of the operational complexities that await them once their models leave the training cluster.
The bifurcation of the AI landscape—rapid expansion of foundational training in China and complex deployment realities in the U.S.—suggests the next phase of the LLM race will be defined as much by operational control as by raw capability. As consumer giants attempt to build their own models, the friction experienced by current frontier labs highlights the steep learning curve ahead for any new entrant.
With reporting from The Information, TechCrunch
Source · The Information
