Why national security – not market forces – will define the US-China AI race

IN A recent commentary carried in The Business Times, “When the disruptor gets disrupted: How Chinese open-source AI is eating its own industry”, veteran economic affairs columnist Vikram Khanna astutely shows how Harvard professor Clayton Christensen’s theory of “disruptive innovation” is visibly under way.

In short, Khanna said that the US artificial intelligence labs are being overtaken by the Chinese ones, because models from China are cheaper to use and perform almost as well as their Western competition.

Yet disruption theory on its own cannot account for Beijing’s decision to block Meta’s US$2 billion acquisition of Manus AI in April 2026.

This was followed two months later by Washington’s restriction on Anthropic’s most advanced models.

Both were justified on national security grounds and point to a different logic at work.

I propose that the future of AI will be shaped more by security imperatives than by pure market forces, in which the contest no longer happens between business models, but between state systems.

And in security competitions, the rules of market disruption are constrained by factors that Dr Christensen’s framework was never designed to handle.

AI is a dual-use technology

First, the security dimension changes the analysis.

Every major power with a stake in the AI race has formally designated it a dual-use technology, meaning one with both civilian and military applications.

The US has done so through Executive Order 14110, the EU through the military exemption in its AI Act, and China through its Military-Civil Fusion strategy, which legally binds civilian AI firms to share technology with the People’s Liberation Army.

In Eastern Europe, the Ukrainian military runs more than 50,000 video feeds a month through AI systems, allowing target identification across hundreds of kilometres.

Drone units are already using AI-guided loitering munitions that can lock onto targets and complete the final attack run autonomously when enemy jamming cuts radio links.

In the Middle East, the US used AI to screen data so that military leaders can make decisions faster than the conventional command chains.

AI also slashes the cost of offensive cyber operations, such as generating malware, automating the discovery of vulnerabilities and spear phishing at scale.

Anthropic disclosed 16 million unauthorised attempts to extract information from its Claude model, from roughly 24,000 fraudulent accounts linked to overseas labs.

That is a sign that frontier systems are already being used as tools for intelligence collection, not just productivity improvements.

A capitalist vs state-led model

Second, the incentive architecture matters.

Closed AI models – Anthropic’s Claude, OpenAI’s GPT series and Google’s Gemini – are products of a particular incentive architecture, where individuals accrue direct economic gains from their work.

Equity stakes, talent compensation and proprietary revenue streams flow back to those who build.

That structure has attracted significant research talent from around the world to Western AI labs.

Can China’s state-directed approach replicate this at scale?

Yes, China has delivered real breakthroughs before, most notably in high-speed rail and solar panels.

But the model has come under some scrutiny of late.

Let’s take the electric vehicle sector. China poured an estimated US$230 billion into EV subsidies between 2009 and 2023.

By 2024, the country was producing about 70 per cent of the global output. As at August 2025, though, only three Chinese EV makers – BYD, Li Auto and Aito – were profitable.

Meanwhile, both vehicle and battery capacity had been built to well above domestic demand, with EV battery output alone projected at roughly three times what China’s carmakers could absorb in 2025.

Analysts said that dozens of local firms were kept alive in no small part by provincial governments, with some reportedly inflating sales through “zero-mileage” used-car exports.

By October 2025, Beijing’s 15th Five-Year Plan dropped EVs from its list of strategic industries for the first time in a decade.

The risk is that the AI sector follows the same trajectory: an injection of state support leading to oversupply and structural weakness.

That danger is compounded by broader macroeconomic headwinds – the International Monetary Fund assessed China as having “persistent structural constraints on productivity” and the World Bank warns of “slower productivity growth, high debt and an ageing population”.

A state managing those pressures has less fiscal latitude to absorb the losses that sustaining frontier AI development will inevitably require.

On the other hand, Western AI is not without its own vulnerability.

OpenAI posted around US$39 billion in net loss in 2025 and expects tens of billions more in cumulative burn, while Anthropic’s newly projected operating profit rests on thin margins.

And if the investment cycle turns – as it has before in technology – the risk of private capital starvation is real.

That said, both risk scenarios are not equivalent.

A capital crunch would likely force Western labs to consolidate and cut costs – painful but ultimately reversible.

State control, by contrast, obliges Chinese AI labs to operate within a political logic that shapes not just what they can build, but also what they are permitted to say, which data they can train on and whose geopolitical interests their models must serve.

That is a structural constraint on what Chinese AI can become, and it is one that becomes more binding, not less, as the technology grows more strategically significant.

Is open-source truly open?

Third, not all open-weight AI models function under the same logic.

The Chinese breakout AI model DeepSeek is technically open-weight, which means that users are ostensibly allowed to use and examine its code for their own purposes.

However, DeepSeek’s data infrastructure, content filters and governance chain are nothing like what “open source” implies in a Western context.

For example, concealed code in the model was reportedly found to be capable of transmitting user login data to China Mobile, a telecommunications company owned by Beijing.

Governments from Australia to the Netherlands have banned DeepSeek from government devices because, under Chinese national intelligence laws, any data it holds is accessible to state intelligence agencies on-demand.

By contrast, Mistral AI, the Paris-based open-weight model often mentioned alongside DeepSeek as an alternative to US proprietary systems, operates on a different logic.

In November 2025, SAP and Mistral announced a partnership explicitly billed as delivering “sovereign AI” for European regulated industries.

Closer to home, HTX (Singapore’s Home Team Science and Technology Agency) partnered with Mistral AI and Microsoft to build Phoenix, a family of models fine-tuned on Singapore-specific context and deployed on HTX’s own on-premise infrastructure disconnected from unsecured networks – precisely to avoid routing through any foreign cloud.

Khanna is right that creative destruction repeats, and that cheap Chinese open-source AI models represent a genuine competitive challenge to Western AI labs.

But in international politics, the disruption narrative is incomplete if it is framed mainly as an economic story.

AI now sits at the intersection of economic productivity and military capability – and will be shaped, above all, by the security architecture within which those systems operate.

The writer is chairman of the public affairs group at Public Relations and Communications Association Asia Pacific, and a director at Temus, a Singapore AI and digital transformation firm. He is reading war studies at King’s College London.