Khurram Badar / Archive / Writings / Ant Group's Cost-Cutting AI Techniques with Chinese Chips

Ant Group's Cost-Cutting AI Techniques with Chinese Chips

Article · 2025-03-24 · 710 words · Khurram Badar

In the competitive world of artificial intelligence, a silent revolution is taking place in Hangzhou. Ant Group is accomplishing what many think

The AI Revolution: How Ant's Domestic Strategy Changes the Game While Nvidia Remains the Unstoppable Titan of AI Power

In the competitive world of artificial intelligence, a silent revolution is taking place in Hangzhou. Ant Group is accomplishing what many think impossible: developing AI training techniques using Chinese-made semiconductors that slash costs by a remarkable 20%. Who knew that something named after a tiny insect could cause such a giant stir in the tech world?

The breakthrough comes through an ingenious approach. By harnessing domestic chips from Alibaba and Huawei, Ant implements the Mixture of Experts (MoE) machine learning technique, achieving results that rival those produced by Nvidia's coveted H800 processors. It turns out David might have a shot against Goliath after all—as long as David has access to some pretty sophisticated silicon.

While Ant hasn't abandoned Nvidia entirely, a significant shift is underway. For its latest models, the company increasingly favors alternatives from AMD and Chinese manufacturers. This strategic pivot echoes a broader trend in the industry following DeepSeek's demonstration that powerful AI models can be trained for a fraction of the billions being poured into development by major Western tech companies. Who needs a gold-plated GPU when a silver one works just fine?

The research paper Ant publishes this month tells a compelling story. Its models—named Ling-Plus and Ling-Lite—reportedly outperform Meta's offerings on certain benchmarks. If the claims prove accurate, these platforms mark another leap forward for Chinese artificial intelligence by dramatically reducing the cost of inferencing and bolstering AI service capabilities. Meta might need to do more than just change its name to keep up.

MoE models have become the darling of the AI world for good reason. By dividing complex tasks into smaller data sets—much like having specialized experts tackle different aspects of a project—they achieve remarkable efficiency. It's the digital equivalent of "many hands make light work," though unfortunately, these digital hands still can't help you move your furniture.

Ant's innovation changes this equation. Traditional methods require approximately 6.35 million yuan ($880,000) to train 1 trillion tokens using premium hardware. Ant's optimized approach brings this down to 5.1 million yuan using more modest specifications. That's a savings of 1.25 million yuan—enough to buy a small island, or at least a really nice virtual one in the metaverse.

The real-world applications are already taking shape. After acquiring the Chinese healthcare platform Haodf.com, Ant deploys an AI Doctor Assistant to support 290,000 medical professionals with tasks like managing patient records. Other innovations include an AI "life assistant" app called Zhixiaobao and a financial advisory service named Maxiaocai. Finally, an AI that can both diagnose your illness and tell you how to afford the treatment.

Perhaps most impressive are the performance metrics. Ling-Lite outperforms Meta's Llama model on English-language understanding benchmarks, while both Ling models surpass DeepSeek's equivalents for Chinese-language tasks. Who knew an Ant could outrun a Llama? Nature documentaries never prepared us for this.

Industry analysts note that this development signals China's accelerating innovation in AI and growing self-sufficiency, developing computationally efficient models that circumvent export controls on advanced foreign chips. It's like being told you can't have the fancy kitchen knife set, so you invent a better way to chop vegetables with what you already have.

This approach stands in stark contrast to Nvidia's strategy. While Ant focuses on efficiency and cost reduction, Nvidia continues building increasingly powerful GPUs with more processing cores, transistors, and memory capacity, betting that computational demand will continue to grow regardless of model efficiency improvements. It's the classic "bigger is better" philosophy—which works great until someone invents a smarter approach and leaves you holding a very expensive paperweight.

As the AI landscape evolves, one thing becomes clear: the race isn't just about raw processing power anymore—it's about finding smarter, more efficient paths to intelligence. And in that race, Ant's domestic strategy is changing the rules of the game. In the world of AI, it seems the mighty ant might just move the rubber tree plant after all.

← Procrastination: A MemoirAnt Group's Cost-Cutting AI Techniques with Chinese Chips →
Two years of working thought, indexed.
Ask me to present it in your conference room — WhatsApp +971 55 623 9111
Book Session →