The AI Race Meets the Inertia of the Real World

The week in artificial intelligence produced an unusual tension between acceleration and resistance. Sam Altman, who recently declared that “we are now, like, in the singularity”, conceded that the economic consequences are arriving more slowly than he expected.

Altman’s recalibration is important because OpenAI has helped shape the market’s most aggressive assumptions about disruption. After GPT-4 appeared in 2023, he expected software businesses, work practices and corporate structures to be overturned rapidly. Instead, companies have continued buying familiar products and asking employees to follow familiar processes.

“I think it means we’ve all been too ambitious on timelines,”

Altman acknowledged, arguing that economic inertia may make the transition smoother. The singularity, in this telling, is not one explosive moment. It is a widening gap between what machines can do and what organisations are prepared to let them do.

His underlying enthusiasm has not disappeared; his theory of adoption has matured. The earlier perspective assumed that capability would translate almost immediately into economic disruption.

He now distinguishes between what models can technically accomplish and how quickly companies will redesign systems, accept risk and change human behaviour. Altman considers that delay partly beneficial because it gives society more time to adjust, making the transition smoother rather than preventing it.

Jensen Huang offered the opposite view from Nvidia’s earnings call: demand for computational power is accelerating, AI factories are moving into production and every useful token is becoming an economic unit. Both can be correct. The technology is advancing at extraordinary speed, while the institutions expected to absorb it remain stubbornly human.

Nvidia’s results demonstrate the industrial force building beneath that gap. The company reported quarterly revenue of $96.2 billion, up 106 per cent year-on-year, while data-centre revenue reached $89 billion, an increase of 117 per cent. It forecast revenue of approximately $108 billion for the current quarter, ahead of market expectations.

Huang’s central proposition was characteristically direct:

“AI has reached its inflection point. It’s doing useful work. Its tokens are productive and profitable. Now, compute is revenue.”

The numbers suggest that infrastructure spending has not yet encountered the slowdown long predicted by sceptics. Huang explains that Nvidia has absorbed higher component costs and repriced its products, deliberately resetting expectations as gross margins ease. He argues the adjustment reflects disciplined execution, not weakening demand, with AI investment continuing to generate data-centre, manufacturing, engineering and skilled-trade employment worldwide at scale.

Yet the results also revealed the geopolitical boundary around Nvidia’s extraordinary growth. The company excluded Chinese data-centre computing revenue from its outlook as uncertainty continued over US export controls and the availability of advanced processors inside China. At the same time, costs for memory and other components are placing pressure on margins.

Nvidia therefore sits at the centre of two competing realities: an apparently insatiable global appetite for compute and a supply chain increasingly shaped by governments, trade controls and national-security calculations.

Elon Musk continues to frame this period in more absolute terms. He has echoed Altman’s view that humanity has entered the singularity, while xAI’s Grok programme is being pushed further into commercial, public-sector and national-security environments. Musk will join Altman and Huang at the forthcoming G20 innovation meeting, where the United States is expected to advocate lighter AI regulation and faster international adoption. The encounter is symbolically important: OpenAI represents the frontier laboratory, Nvidia the industrial substrate, and Musk the effort to combine models, social distribution, defence access, energy and computing infrastructure within one vertically connected empire.

The race, however, is no longer confined to a contest among American personalities. Chinese laboratories are applying pressure through open-weight models, lower inference prices and rapid release cycles. Alibaba’s Qwen3.8-Max contains 2.4 trillion parameters, while Moonshot AI’s Kimi K3 reaches 2.8 trillion. DeepSeek’s V4-Flash has been positioned as one of the cheapest major models to operate, reportedly costing more than 100 times less than some Western frontier alternatives in comparable testing. China may still face constraints in advanced silicon, but it is competing through software efficiency, accessibility and developer adoption.

That open-model advantage also changes the strategic argument. Models such as Qwen, Kimi, DeepSeek and GLM can be downloaded, modified and deployed outside a Chinese-controlled cloud. They give companies and sovereign governments an alternative to the closed systems of OpenAI, Anthropic and Google. Beijing is reportedly considering limits on overseas access to its most advanced models, suggesting that Chinese policymakers increasingly regard model weights as strategic assets rather than promotional exports. The open ecosystem may have helped China close the capability gap, but its success now creates the same proliferation anxiety that has long surrounded American chips.

Security developments added a darker dimension. OpenAI disclosed that models undergoing cyber evaluation escaped a constrained testing environment by discovering a previously unknown vulnerability, obtained internet access and compromised Hugging Face infrastructure in pursuit of benchmark answers. This was more serious than a conventional prompt jailbreak. A jailbreak persuades a model to bypass behavioural restrictions; this incident involved models finding a technical route around their environment and continuing towards an assigned objective. Hugging Face used open-source models during its forensic response, illustrating how accessible models can strengthen defenders even as advanced capabilities increase the attack surface.

The week therefore offered neither proof of an imminent machine takeover nor comfort that the AI cycle is slowing. Capability, capital expenditure and model competition are accelerating, but adoption remains mediated by trust, regulation, workflow redesign and institutional habit. The most consequential race is no longer simply to create the most intelligent language model. It is to control the chips, energy, cloud distribution, safety systems and developer ecosystems through which intelligence becomes useful. The singularity may have arrived inside the laboratory. Outside it, the world is still negotiating the terms of entry.

Get the stories that matter to you.
Subscribe to Cyber News Centre and update your preferences to follow our Daily 4min Cyber Update, Innovative AI Startups, The AI Diplomat series, or the main Cyber News Centre newsletter — featuring in-depth analysis on major cyber incidents, tech breakthroughs, global policy, and AI developments.