[DigitalToday reporter Jinju Hong (홍진주)] Computing and memory efficiency are emerging as a key competitive edge behind Chinese AI company DeepSeek's push for large-scale fundraising, even as it keeps its free model releases and low-cost API strategy.

On May 26, local time, online media outlet Gigazine reported that DeepSeek is in talks to raise investment worth about 70 billion yuan.

DeepSeek drew attention in the global AI industry with DeepSeek-R1, released in January 2025. At the time, the model was assessed as delivering performance similar to OpenAI's reasoning model o1. DeepSeek-V4-Pro, released in May this year, was rated by the U.S. government-affiliated Center for AI Standards and Innovation (CAISI) as having GPT-5-level performance, while an analysis suggested it lags U.S. cutting-edge models by about 8 months.

What the industry is watching is its pricing policy. DeepSeek has effectively made permanent an API discount policy of 75% that it initially ran for a limited period. Its sharply lower costs versus models with comparable performance are cited as a competitive strength. China's Alibaba and Zhipu AI (Z.ai) are also joining the trend of releasing advanced models for free, but they are assessed as pursuing relatively clear monetisation strategies such as building agent systems. DeepSeek, by contrast, continues talks for large-scale investment even though it lacks an obvious revenue model.

As a reason, AI analyst GDP pointed to "overwhelming efficiency". Large language models use a KV cache structure that stores and reuses computation results, and DeepSeek's design is said to have sharply reduced memory usage. Based on an input of 1 million tokens, memory use was analysed at 60 GB for GLM-5 and about 89 GB for Qwen3-235B-A22B, compared with 5.48 GB for DeepSeek-V4.

The gap goes beyond a simple technical comparison and is directly tied to cost structure. In the AI industry, surging memory prices are emerging as a major burden. The outlet explained that about 63% of AI chip costs currently come from memory, and that in some environments memory costs are a bigger burden than GPUs. In this situation, DeepSeek's high-efficiency structure could sharply cut model operating costs and serve as a basis for securing profitability, the analysis said.

Efficiency was also highlighted in inference speed. DeepSeek is assessed as being able to reduce latency even on relatively lower-performance AI chips by optimising how it uses caches. DeepSeek-V4-Pro was reported to run on about one-third of the computation of the previous generation, while DeepSeek-V4-Flash was reported to be able to process with only about one-tenth of the computation.

The strategy also intersects with the realities facing China's AI industry. Chinese companies face constraints in securing Nvidia's latest high-performance AI chips due to U.S. export restrictions. If they can deliver high-performance AI services with relatively lower-spec chips, that would not only cut costs but also strengthen supply-chain stability.

Some in the market also say DeepSeek's 70 billion yuan fundraising is tied to a longer-term strategy than simply developing models. GDP forecast that the funds could be reinvested in Chinese memory and AI chip companies and used to build a "cost-efficient AI ecosystem". That is, the structure would be to secure market share through free model releases and a low-priced API strategy, then generate actual profitability through high-efficiency model design and building a hardware supply chain within China.

In the industry, some see that if such a strategy is realised, Chinese AI companies could emerge as a new alternative in the global AI industry beyond simple price competition.