China’s GenAI debut realigning the tech investment universe

The launch of China’s home-grown large language model (LLM) DeepSeek earlier this year carries “significant implications” for investors, shifting the centre of gravity for the technology’s future development beyond the US.

Anton du Plooy, global technology sector analyst at Anglo-South African asset management firm Ninety One, says the shift will have “far-reaching” consequences, not only for the future diversification of the technology, but also for investors who will likely need to realign from their US-centric allocations.

“AI is no longer a US-centric phenomenon— and accelerated adoption creates opportunities far beyond traditional infrastructure plays in this next phase of AI-driven transformation,” he wrote in a recent analysis.

What is more, the US Government’s efforts to contain China and protect local companies’ AI innovation edge appear to have backfired, leading to an acceleration of China’s home-grown AI capability.

DeepSeek a Hangzhou–based AI lab affiliated with the Chinese quant hedge fund High-Flyer, was launched to significant fanfare (and palpable shock for many in the West) in January this year, being among the first non-US developed LLMs capable of delivering performance on par with US offerings, including OpenAI’s GPT-4 and Google’s Gemini.

The more streamlined (or indeed less infrastructure-intensive) DeepSeek LLM appears also to have been developed “at a fraction of the cost” of the US-built models.

“The immediate response to DeepSeek’s announcement was seismic,” du Plooy wrote.  

“NVIDIA, the dominant force in AI computing hardware, saw its stock plummet, wiping out over $500 billion in market capitalisation. The reason? DeepSeek reportedly achieved comparable AI model performance while circumventing US export restrictions on advanced NVIDIA Graphics Processing Unit (GPU) chips.”

However, du Plooy adds, this initial “panic” from the West has since moderated to some degree, with evidence showing the model’s foundations were more derivative (using “pre-existing frontier models via distillation techniques” – a method of LLMs ‘learning’ from preexisting models) than novel.

DeepSeek born of a ‘failure’ of US containment

For du Plooy, China’s expanding home-grown LLM development capability underscores a broader failure of US containment policies – an attempt to restrict Chinese firms’ access to advanced semiconductor technologies that power AI systems.

Instead, the previous administration’s ‘small yard, high fence’ approach appears to have spurred home-grown innovation and reduced Chinese developers’ hitherto dependence on US technologies.

“By employing H800 chips – procured before the latest sanctions – the company sidestepped trade limitations while proving that AI innovation in China isn’t wholly dependent on NVIDIA’s most powerful hardware,” du Plooy wrote.

He cited the resurgence of Chinese tech powerhouse Huawei following US sanctions (preventing US chipmakers from selling older-generation semiconductors to the company), which is now building chips in-house.

“Initially crippled by US blacklisting, Huawei rebounded through indigenous semiconductor innovation, proprietary software, and diversification into new sectors like electric vehicles.

“DeepSeek’s success follows a similar trajectory, reinforcing China’s broader strategic goal of AI leadership by 2030,” he said.

Implications for investors

For investors, China’s emergence as an AI innovator is rapidly reshaping the playing field.

While its development may not have been as cheap as first suggested – with initial estimates of just under US$6 million likely significantly less than the actual development cost – the overall cost of developing an LLM is falling rapidly.

“The falling cost of AI model development and deployment suggests the opportunity set is diversifying across industries, expanding the range of companies positioned to benefit,” du Plooy wrote.

Curiously, despite DeepSeek proving out the potential for significantly less resource-intensive LLMs, capital expenditure on AI infrastructure – the hardware necessary to power these LLMs – by both China and the US has only increased in the preceding months.

According to du Plooy, this aligns with Jevons Paradox, the economic principle that suggests that resource efficiency can in fact lead to an increase in the consumption of that resource.

He warns that the temptation of policymakers in Washington to further restrict chip exports could lead to a further bifurcation of AI capabilities between China and the West, and ultimately forcing China to accelerate its domestic semiconductor development and rethink its reliance on Western AI ecosystems.

Chinese developers, unlike most private-sector dependent Western operations, also have the advantage of direct state backing, with AI development “deeply embedded in state-driven industrial policies, such as the ‘New Generation AI Development Plan’ and ‘Made in China 2025’, Du Plooy noted.

As Du Plooy notes, AI is no longer a US-centric phenomenon— and accelerated adoption creates opportunities far beyond traditional infrastructure plays in this next phase of AI-driven transformation.