{"slug": "faced-with-less-compute-and-fewer-tokens-chinese-ai-labs-are-tightening-the-gap", "title": "Faced with less compute and fewer tokens, Chinese AI labs are tightening the gap with the U.S. by just being more efficient", "summary": "The FBI, the National Security Agency, and the Cybersecurity and Infrastructure Security Agency alleged Tuesday that DeepSeek, Moonshot, and four other Chinese AI companies extracted \"capabilities worth billions\" since 2024 by buying bulk subscriptions to U.S. rivals and training on their outputs, a method the agencies said let DeepSeek understate its $5.6 million training cost. China's foreign affairs ministry called the accusations \"groundless,\" saying its AI development \"is a result of high-level scientific and technological self-reliance.\" Analysts told Fortune that Chinese labs also closed the gap by cutting the computational complexity of the attention mechanism by an order of magnitude, letting models such as GLM 5.2 and Kimi 2.6 and 2.7 handle around 75% of engineering tasks at a fifth of the cost of U.S. models, while the U.S. holds 74% of the world's compute.", "body_md": "The race between the world’s two biggest economies to dominate AI recently escalated with new accusations and raised the question of how China came to rival the U.S.’s AI capabilities.\n\nSix Chinese AI companies, U.S. officials alleged earlier this week, found a shortcut to closing the gap with American AI labs by buying bulk subscriptions to their American rivals and then training on the outputs.\n\nThe FBI, the National Security Agency, and the Cybersecurity and Infrastructure Security Agency [said Tuesday](https://www.cisa.gov/news-events/cybersecurity-advisories/aa26-251a) that DeepSeek, Moonshot, and four others extracted “capabilities worth billions” in this way since 2024, a method the agencies claimed let DeepSeek understate its [$5.6 million](https://arxiv.org/pdf/2412.19437) training cost.\n\nChina’s foreign affairs ministry said the country’s AI development “is a result of high-level scientific and technological self-reliance” and called the accusations “groundless.”\n\nThe allegations are one potential explanation for why Chinese AI models are estimated to be neck-and-neck with their U.S. counterparts, with a [report from Stanford](https://hai.stanford.edu/assets/files/ai_index_report_2026.pdf) putting Anthropic’s top model ahead of DeepSeek’s by just 2.7% earlier this year. \n\nBut analysts say there’s a different advantage Chinese labs developed to compete: squeezing out more value from limited resources.\n\n## **More value for fewer tokens**\n\nThere’s a technique Chinese labs perfected that stems from “attention,” the mechanism Google researchers introduced in a [landmark 2017 paper](https://arxiv.org/pdf/1706.03762) that underlies every large language model.\n\nAttention is what lets a model look at each piece of text in relation to all the others and decide which connections matter most. That helps it understand context, but it gets computationally more expensive the longer the context window gets.\n\nBrendan Burke, the semiconductors and supply chain analyst at tech research firm Futurum Group, told *Fortune* that Chinese labs figured out a shortcut to make the attention mechanism cheaper and more efficient. \n\n“Chinese labs found algorithms that reduce the complexity of those calculations by an order of magnitude, and then achieve better results because they’re able to summarize the most relevant tokens,” he said.\n\nNecessity was the mother of invention, as the U.S. [restricted China’s access to Nvidia’s best chips](https://www.cnbc.com/2026/08/19/china-ai-nvidia-chips-us-export-controls.html), pushing it toward domestic [alternatives like Huawei](https://fortune.com/asia/2025/06/10/china-us-tech-controls-open-source-huawei-founder-ren-zhengfei/), which limited China’s access to the highest-performing compute available in the U.S. Meanwhile, the U.S. has 74% of the world’s compute, according to a [White House report](https://www.whitehouse.gov/wp-content/uploads/2025/03/Artificial-Intelligence-and-the-Great-Divergence.pdf), and it also has [hyperscalers pouring billions](https://fortune.com/2026/04/30/big-tech-hyperscalers-will-spend-700-billion-on-ai-infrastructure-this-year-with-no-clear-end-in-sight-eye-on-ai/) into generating more through data centers. \n\n“Because they had less compute to work with, they found that computationally efficient method instead of just throwing more compute at an inefficient technique, as U.S. labs initially did,” Burke said about China.\n\nBy contrast, U.S. frontier labs can be “token hogs” because their AI systems are designed to be exploratory and to test base models, he explained.\n\nThe trade-off between the two can be quantified. Ameya Kanitkar, cofounder of the AI measurement platform Larridin, told *Fortune* that in the enterprise workflows Larridin tracks, Chinese models like GLM 5.2 and Kimi 2.6 and 2.7 handle around 75% of engineering tasks “reasonably well” at a fifth of the cost of the U.S. ones.\n\n“[Frontier](https://fortune.com/company/frontier-group-holdings/) U.S. models still have an advantage on the most complex tasks, but Chinese open-weight models are becoming more than capable enough for the majority of everyday enterprise engineering work,” Kanitkar said.  \n\nThe cost difference might end up mattering as AI spending [consumes a bigger share](https://fortune.com/2026/09/02/companies-getting-the-most-from-ai-rethinking-how-work-gets-done-cfo/) of corporate budgets, with [20% of business leaders](https://fortune.com/2026/09/02/companies-getting-the-most-from-ai-rethinking-how-work-gets-done-cfo/) surveyed by McKinsey saying AI-related costs like buying tokens is constraining their use of it. \n\n## **U.S. enterprises warming up to Chinese models**\n\nCost-efficiency isn’t the only appeal. DeepSeek’s R1 reasoning model was [made available for download](https://fortune.com/2025/02/04/deepseek-open-source-ai-cybersecurity/) through platforms like Hugging Face, allowing companies to run and adapt versions of the model themselves. This helps companies by giving them an open-source model to work with and fine-tune to meet their needs without depending on a closed model, with the option to run it through a U.S.-based cloud provider [like Amazon Web Services](https://aws.amazon.com/blogs/aws/deepseek-v3-1-now-available-in-amazon-bedrock/). \n\nThis flexibility helped U.S. businesses that might have been wary of sending data to a China-based company warm up to DeepSeek, and this applies to other Chinese models as well. Hugging Face reported Chinese open-source models [accounted for 41%](https://huggingface.co/blog/huggingface/state-of-os-hf-spring-2026) of total downloads last year, a larger share than U.S. ones.  \n\n[DoorDash](https://fortune.com/company/doordash/) CEO Andy Fang [said](https://x.com/andyfang/status/2074252174226493584) using Moonshot AI’s Kimi was “cheaper” and “better quality” without degrading the quality of code. AI coding startup Cursor [also used Kimi](https://fortune.com/2026/07/17/businesses-turn-to-cheaper-chinese-ai-models/) to help build its Composer 2 coding agent. [Airbnb](https://fortune.com/company/airbnb/) and [Siemens](https://fortune.com/company/siemens/) are also experimenting with [Alibaba](https://fortune.com/company/alibaba-group-holding/) and DeepSeek models, with AirbnB CEO Brian Chesky calling Qwen “[fast and cheap](https://www.bloomberg.com/news/articles/2025-10-21/airbnb-ceo-brian-chesky-says-chatgpt-integration-not-ready-for-airbnb-app?sref=LqVYNnVJ).”\n\nChinese models are even starting to replace their American counterparts in more specialized work. Thomson Reuters said it built an in-house model called Thomson-1 by [adapting Alibaba’s open-source Qwen](https://fortune.com/2026/08/27/chinese-open-source-ai-is-starting-to-win-over-u-s-businesses/) model to handle document-review work previously done by Claude.\n\nData suggests the enterprise [shift](https://fortune.com/2026/08/27/chinese-open-source-ai-is-starting-to-win-over-u-s-businesses/) is becoming more visible. [Ramp’s AI index](https://ramp.com/data/ai-index-august-2026) showed that the share of businesses paying for platforms with access to open-source and Chinese-developed models rose to 6.1% of total AI-spending businesses in July from 4.5% in January. \n\nBut companies experimenting with Chinese models doesn’t necessarily mean they’re leaving behind U.S. ones, which are [still months ahead in performance](https://www.csis.org/analysis/what-know-about-chinese-ai-models). Mike Finley, chief technology officer of enterprise AI analytics firm AnswerRocket, told *Fortune* the output of U.S. AI companies still serves as the “existence proof” for Chinese labs to innovate off of. \n\n“The work they do would simply not be possible without the frontier labs blazing the trail,” Finley said.\n\n**Exclusive:** In a new sit-down interview with\n\n*Fortune*, OpenAI CEO\n\n**Sam Altman** explains safety standards are \"not at a place\" to push AI capabilities much further and warns AI beyond human control is \"absolutely\" possible.\n\n**Watch or listen here.**", "url": "https://wpnews.pro/news/faced-with-less-compute-and-fewer-tokens-chinese-ai-labs-are-tightening-the-gap", "canonical_source": "https://fortune.com/2026/09/13/china-us-ai-models-efficient/", "published_at": "2026-09-13 10:00:00+00:00", "updated_at": "2026-09-13 10:34:25.343009+00:00", "lang": "en", "topics": ["artificial-intelligence", "large-language-models", "ai-policy", "ai-chips", "ai-research"], "entities": ["DeepSeek", "Moonshot", "FBI", "National Security Agency", "Cybersecurity and Infrastructure Security Agency", "Anthropic", "Huawei", "Nvidia"], "alternates": {"html": "https://wpnews.pro/news/faced-with-less-compute-and-fewer-tokens-chinese-ai-labs-are-tightening-the-gap", "markdown": "https://wpnews.pro/news/faced-with-less-compute-and-fewer-tokens-chinese-ai-labs-are-tightening-the-gap.md", "text": "https://wpnews.pro/news/faced-with-less-compute-and-fewer-tokens-chinese-ai-labs-are-tightening-the-gap.txt", "jsonld": "https://wpnews.pro/news/faced-with-less-compute-and-fewer-tokens-chinese-ai-labs-are-tightening-the-gap.jsonld"}}