China's Hundred Model War and the Race That Reshaped Global AI

Within months of ChatGPT's arrival at the end of 2022, China had answered with more than a hundred large language models of its own, a burst of activity so frantic that observers reached for a martial metaphor to describe it. The phrase that stuck, the Hundred Model War, deliberately echoed a celebrated Communist offensive of 1940, and by the close of 2023 it had been enshrined among the year's official media buzzwords. Read casually, the episode looks like undirected panic, a whole industry sprinting to copy a foreign breakthrough it had failed to anticipate. Read more carefully, it was something closer to a mobilization: a contest whose sheer scale was part of the design, staged so that a small number of durable competitors would emerge from a field everyone understood, from the outset, would have to be thinned. By mid-2025 Chinese institutions accounted for roughly 1,509 of the world's some 3,755 publicly disclosed large models, and what had begun as imitation had resolved, within three years, into a distinctive competitive model built on the levers China actually possessed — efficiency, open distribution, and aggressive pricing — rather than the compute it did not.

The state had prepared the ground years before the war began. The State Council's New Generation Artificial Intelligence Development Plan, issued in 2017, committed the country to world-leading status in AI by 2030, and when generative models arrived Beijing moved with unusual speed to govern them. The Interim Measures for the Management of Generative AI Services, effective from August 2023, made China the first country anywhere with binding rules aimed specifically at the technology. Crucially, those rules regulated public deployment while leaving research and internal development untouched, requiring providers of consumer-facing services to pass a security assessment and file their algorithms with regulators, and to keep their output within the bounds of core socialist values. The effect was to build a lane rather than a wall. By the end of 2025 nearly 1,200 generative AI services had cleared national or local registration. Beijing pointedly declined to pass a single omnibus AI law, shelving a draft in 2025 in favor of incremental measures, and folded artificial intelligence into national statute only gradually. Regulation, in other words, channeled the war rather than suppressing it.

The incumbents rushed in first. Baidu, the earliest of the giants to respond, unveiled its ERNIE Bot in a demonstration that was reportedly pre-recorded rather than live, a stumble that briefly cost the company a tenth of its share price before the product recovered to become one of the first eleven services approved under the new rules. Its later flagship was positioned frankly as a rival to GPT-4, and by 2025 competitive pressure had pushed Baidu to abandon subscriptions and make the assistant free to everyone, a concession that would echo across the market. Alibaba took a different path that proved more consequential. Its Qwen family was released as open weights, downloadable and modifiable by anyone, and by early 2026 those models had been pulled from public repositories more than seven hundred million times, overtaking Meta's Llama to become the most downloaded model family in the world and spawning well over a hundred thousand community-built derivatives. Tencent leaned on the unmatched distribution of WeChat, embedding its Hunyuan-powered assistant directly into the app that mediates daily life for a billion people, while ByteDance turned Doubao into China's single most popular AI application by routing it through Douyin and its sister platforms, reaching well over two hundred million users. Distribution, these firms understood, could matter as much as raw capability.

Huawei occupied a category of its own. Rather than compete for consumers, it built downward through the entire technology stack, from its Ascend processors and software frameworks up through cloud platforms to its enterprise-focused Pangu models. Cut off by American sanctions from the world's most advanced chip fabrication, Huawei shifted production to a domestic foundry and accepted the resulting performance penalty in exchange for something no rival could offer: silicon that remained available to Chinese developers when Western alternatives did not. Its most advanced accelerators deliver perhaps a third of a leading Nvidia chip's throughput, yet their significance lay less in benchmarks than in access. Huawei functioned less as a chatbot competitor than as the domestic infrastructure on which much of the rest of the war was fought, the provider that allowed sanctioned companies to keep training at all.

If the giants supplied the mass of the campaign, a cluster of well-funded startups supplied its restlessness, and one of them supplied its defining shock. DeepSeek, founded and financed entirely by a quantitative hedge fund and led by Liang Wenfeng, took no outside venture capital and pursued fundamental capability over commercial polish. When it released its R1 reasoning model early in 2025, matching the best American systems on mathematics, coding, and reasoning while giving the model away under a permissive license, the market response was seismic. The DeepSeek app vaulted past ChatGPT to the top of the American app store within a week, Nvidia shed roughly six hundred billion dollars of value in a single session — the largest one-day loss for any company in the history of American markets — and commentators reached for the language of a Sputnik moment. Liang was soon summoned to symposia hosted by the premier and then by Xi Jinping himself, a signal that the state now regarded algorithmic ingenuity as a strategic asset in its own right. The episode crystallized the war's central lesson: cleverness could partially substitute for the hardware China had been denied.

That denial was the constant pressure shaping everything. Successive rounds of American export controls, beginning in 2022, first banned the most advanced accelerators, then closed the loopholes opened by deliberately weakened versions, and eventually restricted even the chips Nvidia had engineered specifically to slip beneath the thresholds. The controls imposed real costs, raising the price and complexity of training at scale. Yet their record proved double-edged. They did not stop China from reaching the frontier, and they hardened the country's drive toward self-sufficiency, spurring domestic chip design and retaliatory restrictions on the critical minerals the West needs. By some accounts the constraints even helped produce the efficiency breakthroughs that made a model like R1 possible in the first place. The point was made plain when Washington, in a striking reversal at the end of 2025, approved the sale of a more powerful Nvidia chip to China precisely because officials feared that continued restriction was accelerating Huawei's rise as a homegrown substitute. Restriction and stimulus had become difficult to tell apart.

The consolidation that industry veterans had predicted arrived on schedule. Baidu's chief executive had warned publicly that the proliferation of models represented a vast waste of computing power, and the market soon proved him right. It divided into tiers: a handful of players competing at the global frontier, the established giants defending their positions through ecosystem integration while quietly incorporating rivals' models alongside their own, and a second tier that withdrew from the frontier altogether. The most telling defection came from a startup founded by one of the field's most celebrated figures, which announced it would stop pre-training large models entirely and build instead atop DeepSeek. Underneath the shakeout ran a brutal price war, touched off when DeepSeek and then the giants slashed the cost of their services by staggering margins; average prices for access to Chinese models fell more than ninety percent within a year, dragging the leading Western labs into cuts of their own. The collapse did not shrink the market but enlarged it, as cheaper intelligence summoned far more demand than it destroyed, and even Nvidia recovered to reach a valuation of five trillion dollars.

What the Hundred Model War produced, in the end, was not a replica of the West at a lag but a genuinely different competitive posture, one that leaned into efficiency, open weights, and price expansion because those were the fronts on which China could win. By early 2026 the country counted more than half a billion generative AI users, and Chinese open-source models accounted for roughly a third of all model downloads worldwide, turning a domestic scramble into a force that reshaped the global field. The open question was no longer whether China could build competitive models. It was whether the survivors of the war could sustain the enormous capital demands of frontier development without clear paths to profit, a challenge that, for all but the largest ecosystem players, remained conspicuously unresolved.