Is DeepSeek unique? What China’s generative AI breakthroughs mean for equity investors

A trio of Chinese generative artificial intelligence (gen AI) breakthroughs in late January have shaken the perception of US dominance. For equity investors, they present clear signs that a growth cycle driven not by hardware build-out, but accelerating gen AI adoption, could be taking shape.

In our view, the potential upside for internet platforms and Edge AI hardware providers could offset or even eclipse an expected cyclical downturn in the graphic processing unit (GPU) and compute power sector going into 2026. We explain how this is shaping our convictions on the gen AI theme below.

China’s gen AI prowess

Three distinct events in late January thrust China’s impressive gen AI capabilities, achieved in the face of US export restrictions on advanced semiconductors, into the international spotlight:

• On 20 January 2025, Chinese startup DeepSeek released its latest large language model (LLM), DeepSeek-R1. An open-source model allowing for free commercial and academic use, it can perform the same text-based tasks as other Western models but at significantly lower development and usage costs
• Nine days later, Alibaba¹ released a new version of its LLM, Qwen 2.5-max, which it claims surpasses DeepSeek-v3, OpenAI’s GPT-4o and Meta’s Llama-3.1-305B models. ByteDance (TikTok’s owner) also released an update to its flagship gen AI model, claiming it outperformed OpenAI’s o1 in AIME, a benchmark test for AI models to understand and respond to complex instructions
• On the same day, The Spring Festival Gala – China’s most-watched television show –  showcased 16 H1 robots joining a human troupe for a Chinese folk dance performance. These general-purpose humanoid robots –  developed by Unitree, a company backed by Alibaba Cloud¹ – were processing high-frequency spatial data in real-time, allowing them to recognise where they were and coordinate their movements.

The release of DeepSeek-R1 model captured the most attention. The open-source model, issued under an MIT license, uses a mixture-of-experts (MoE) approach with 37 billion active parameters out of 671 bn total parameters. The key development relative to previous versions is its comparable performance versus other gen AI models, including ChatGPT4, across multiple reasoning tests (see Figure 1) and at much better pricing.  The DeepSeek-R1 model was apparently trained on a limited number of Nvidia¹ chips, meaning its costs are 90% lower than those of ChatGPT owner OpenAI’s models in both training and inference (see Figure 2).

FIG 1.  DeepSeek-R1 vs other reasoning models²

FIG 2.  Pricing comparison for inference: DeepSeek-R1 vs OpenAI models³

Market response

Unsurprisingly, this raised questions about Silicon Valley tech firms’ plans to spend billions on building their gen AI infrastructure. Yet, in our view, there are two major concerns for the market. First, is there a risk of lower demand for advanced compute hardware, including semiconductors? And second, what should investors expect from China’s ‘sudden’ advancement in gen AI?

In our view, the initial sharp market sell-off can be attributed to several factors:

• Tech-sector equity market positioning was already very crowded, particularly after the recent Stargate announcement
• There is always cyclicality for tech hardware spending as the investment cycle matures, as evidenced by the sell-off last July. The question is: when will cyclicality really kick in? Needless to say, investors are becoming more demanding and anxious, amid 2026 demand projections (tech stocks typically trade on expectations one year ahead) and uncertain prospects for significant gen AI monetisation  
• There is also a lower possibility that Nvidia¹ will beat earnings estimates and raise guidance at its results announcement at the end of February, due to its production yield issue. Historically, upbeat guidance from the firm has served to restore investor confidence in the global tech supply chain, but its next announcements may just extend the period of ‘no good news’.
   

What should investors expect from China’s ‘sudden’ advancement in gen AI?

What DeepSeek’s breakthrough means

Despite the ability of DeepSeek-R1 to perform strongly without top chips, we think it is too early to conclude that a structural downshift in compute power required by newer gen AI models will take place, given that:

• Deepseek-R1 is a fine-tuned model based on existing open source LLMs (mainly from Meta, and partially from OpenAI), so the cost cannot be easily comparable. Additionally, there remains a lack of clarity on the actual cost base, as there is debate around the extent of GPUs used and the inclusion of overheads and engineering costs
• We believe that the Scaling Law remains true for gen AI, even with DeepSeek’s breakthrough in reasoning – i.e., that models with more chips or compute power will perform better and achieve more
• With more efficient training and inferencing, gen AI is likely to reach the Jevons paradox sooner. In this scenario, where the increasing efficiency of a technology is accompanied by greater demand, this could result in much faster adoption and eventually require more compute power. Such adoption could provide an answer to many of gen AI’s monetisation and adoption-curve worries, in our view
• AI monetisation worries are not new. Sequoia Capital’s doubts about AI monetisation was a key driver for the AI sell-off in July 2024 but it subsequently adopted a more neutral tone in a couple of months later.

In our view, DeepSeek’s breakthrough is not revolutionary but has evolved from several key concepts, chiefly the MoE approach. In the tech sector, cost per function is always declining, usually at 15-25% per year ⁴, and this can be even faster in the early days of the adoption curve. We believe this is true for gen AI and, therefore, we expect to see the release of more efficient models over the next one-to-two years given how many start-ups focused on this technology have been formed.

Read also: Asian markets: long-term tailwinds fuelling growth, consumption and tech

Our outlook for tech hardware and internet platform firms

We have long seen the rise of gen AI as an attractive investment theme. Using the dotcom bubble as an example, we believe the current stage of the gen AI cycle is closer to the year 1996 than the boom’s 1999 peak. If we are right, a very long adoption curve remains ahead of us. Consequently, the head-turning DeepSeek and Alibaba releases lead us to the following views:

• We believe DeepSeek is not unique and major Asian internet or cloud firms with internal gen AI programs will likely have equivalent technology or models sooner or later, if not already. With continuous development, the commoditisation of LLMs could accelerate and the focus will gradually shift from LLM owners to companies who can benefit most from efficiency gains and monetisation
• This will likely bring advancements in artificial general intelligence (AGI) – autonomous systems that surpass humans in economically valuable tasks – given the lower cost availability and increasing superiority of gen AI models. Key downstream beneficiaries would likely be internet and cloud platforms that could benefit from continuous efficiency gains
• As DeepSeek demonstrates, Edge AI, where a small model can be stored on a smartphone or PC to perform generic AI functions, is likely to undergo faster adoption, in our view
• However, with regards to the high-end GPU segment, we acknowledge that a period of cyclicality could extend through to 2026. For many key AI products (many agentic AI products are currently being announced that perform human-like tasks or act as virtual assistants) monetisation remains a key area to watch. As a result, we favour beneficiaries like Edge AI and custom ASIC (AI chips designed by cloud players for better efficiency instead of generic chips) as we progress through 2025 and into 2026.

We provide an overview of our views on Asia tech equities in Figure 3.

FIG 3. A snapshot of the tech views informing our Asia High Conviction strategy⁵

Read also: Asian markets: find what other investors don’t see

Our focus: shifting from gen AI infrastructure to adopters and Edge AI

Leading semiconductor firms, especially Nvidia¹, and upstream manufacturers have faced the brunt of the market impact resulting from the release of DeepSeek-R1, as these areas could have more to lose from a scenario in which there is a downgrade of growth expectations in advanced GPU demand and compute power requirements. Although these fears are slightly overblown, in our view, we are also cognisant of the cyclicality in this sector going into 2026, and the potential for less-than-expected growth. However, at a macro level, we expect the advancement of China’s various LLMs to drive further and faster the adoption of AGI, which could benefit device manufacturers in the Edge AI segment as well as internet platform companies in Asia.

As such, in our strategy we have shifted towards gen AI adopters and Edge AI hardware providers instead of pure beneficiaries of demand for AI infrastructure. We believe gen AI adoption, including the use of AGI, is a structural phenomenon and we are at the beginning of this three-to-five-year trend. As a result, we are adjusting our positioning, and are currently focused on prominent Chinese internet platforms.