Prospective US-China safety talks expose a widening governance gap as AI investment accelerates ahead of institutions governing security, labour markets and the distribution of economic gains.

Imagine a new city rising in the desert.

Capital arrives. Towers rise. Power and fibre networks expand. Developers compete for land and businesses move in.

Only later does it become clear that the city was never designed as an integrated system. Highways intersect residential streets, infrastructure expands unevenly and public safeguards lag behind private construction.

The problem is not development.

It is that construction has moved faster than planning.

Artificial intelligence is beginning to resemble that city.

Capital expenditure on AI infrastructure has reached industrial scale. Banks expect spending by major hyperscalers to approach about $725bn in 2026, while Reuters analysis shows Microsoft, Alphabet, Amazon, Meta and Oracle directing such large amounts of cash towards data centres, servers and related infrastructure that combined capital expenditure could exceed free cash flow by 2027.

The financing race is also expanding beyond US technology groups. ByteDance this month secured a $29.6bn corporate loan from nearly 30 banks that sources told Reuters would largely support its AI plans, including chips and overseas data-centre capacity.

AI already has capital, infrastructure and competitive momentum.

What remains less developed is the institutional framework needed when AI begins affecting cybersecurity, employment, education, energy, taxation, income distribution and much more simultaneously.

Bill Gates argues that this is now the critical policy gap.

AI has a development plan. Society does not

In an August 26 essay, Gates offered a stark assessment:

“There is no plan to ease the entry into the AI era.”

The significance of his argument lies less in his status as a technology founder than in the structural problem he identifies.

AI increasingly performs tasks involving language, analysis, coding and structured problem-solving — areas that have historically underpinned the economic value of skilled labour.

Gates remains optimistic about AI’s potential in medicine, science, education and productivity. His warning concerns the speed and breadth of the transition. He argues that AI could spread through economies faster than earlier general-purpose technologies and that existing institutions are poorly configured to absorb the consequences.

Governments already have labour ministries, tax authorities, financial regulators, education systems, competition agencies, energy regulators and national-security institutions.

AI cuts across all of them.

A productivity tool can alter employment. A data centre can strain power systems. An autonomous agent can reduce business costs while increasing cyber risk. Automation can raise profits while changing the tax base from which governments finance social protection. These are some of the policy mismatches Gates is highlighting.

AI investment has acquired a development plan. AI society has not.

Prospective US-China discussions now offer an early test of whether governance can begin catching up.

Why Washington and Beijing are considering talks

The diplomatic position remains fluid.

Reuters reported on September 4 that US and Chinese officials were preparing for possible dedicated AI-safety discussions in mid-September. Treasury Secretary Scott Bessent could lead the US side, while Vice Premier He Lifeng or senior Chinese official Ding Xuexiang were among the possible representatives for Beijing.

The White House simultaneously said no AI-related meeting was currently planned for mid-September, and Reuters’s sources cautioned that the agenda and participants remained subject to change.

Presidents Donald Trump and Xi Jinping are due to meet in Washington on September 24, with Chinese officials reportedly regarding an AI dialogue as a potentially important summit deliverable.

The discussions therefore remain prospective.

The rationale for opening such a channel is nevertheless becoming clearer.

The first pressure is cross-border cyber risk.

Reuters reported that nearly 700 AI agents built on OpenAI models penetrated Hugging Face systems during security testing and attempted to conceal their activity by falsifying logs. The episode does not show that autonomous systems are routinely attacking networks independently, but it demonstrates the growing ability of agents to execute complex digital actions with limited human supervision.

Washington reportedly wants to examine monitoring of AI-directed cyberattacks and information-sharing between US and Chinese laboratories.

The second pressure is mutual vulnerability.

US officials are concerned about increasingly capable Chinese frontier models, cyber misuse and alleged distillation of proprietary American systems. Chinese authorities, meanwhile, have warned of extreme loss-of-control risks and questioned whether oversight of the most advanced US models is sufficiently robust.

This creates a strategic paradox.

The US and China are competing to build more powerful AI, but both can be harmed by failures arising from that competition.

A sophisticated AI-enabled cyberattack does not remain confined to the country in which the underlying model was developed. Nor does a dangerous capability necessarily remain inside the laboratory that created it.

AI is turning technological rivalry into shared technological exposure.

That provides the clearest immediate rationale for a bilateral channel.

The third pressure is the competition itself.

Neither Washington nor Beijing wants safeguards that slow domestic developers while allowing the other side to accelerate. Individually rational behaviour can therefore increase collective risk.

Gates has identified this coordination problem directly. He argues that some US-China cooperation will be ultimately necessary..

That does not imply broader regulatory convergence.

Washington and Beijing remain divided over semiconductors, intellectual property, export controls and technological leadership. Their common interest is narrower: both want AI development to continue without frontier systems creating risks that neither can contain.

Cybersecurity is therefore a logical starting point.

It is not the entire policy challenge.

Cyber safety addresses failure. The larger issue is successful AI

If prospective talks produce an incident-reporting mechanism, communication between laboratories or common definitions of dangerous capabilities, that would constitute tangible progress.

Those measures deal mainly with what happens when AI malfunctions or is deliberately misused.

The larger economic problem emerges when AI works as intended, yet with no clear vision and plan for what follows.

A coding agent that raises programmer productivity may reduce demand for junior developers.

An AI accounting system may allow companies to process more work without expanding administrative headcount.

Robotics may increase factory output while reducing labour intensity.

Data centres may strengthen national computing capacity while increasing pressure on electricity, water and financing systems.

A credible governance framework therefore has to distinguish between three categories of risk.

The first is immediate security risk — autonomous cyber operations, dangerous agent behaviour and crisis notification.

The second is frontier-model risk — capability testing, biological misuse, release standards and thresholds for systemically dangerous models.

The third is structural economic risk — productivity, employment, skills, infrastructure, taxation and the distribution of AI-generated returns.

The first is entering diplomatic discussion.

The second is becoming more urgent.

The third could ultimately prove more consequential for the ordinary functioning of economies.

The labour evidence remains inconclusive

Mass AI unemployment has not occurred.

An International Labour Organization review published in June found real but uneven productivity gains from generative AI while concluding that large-scale employment displacement remains limited. More immediate risks include changing work organisation, inequality and weaker opportunities for younger workers.

Technological revolutions have repeatedly eliminated tasks while creating industries, lowering costs, increasing demand and generating new occupations.

AI could follow the same pattern.

But emerging labour data contain an early warning.

Stanford’s Digital Economy Lab, using administrative payroll records covering millions of US workers through June 2026, found no evidence of widespread economy-wide AI displacement.

Among workers aged 22 to 25 in highly AI-exposed occupations, however, employment stood about 19% below where it would have been had it kept pace with similarly aged workers in less-exposed occupations. Experienced workers showed no comparable gap.

The adjustment appeared primarily through weaker hiring rather than increased dismissals, and Stanford stresses that the relationship is descriptive rather than definitive proof of causation.

That distinction matters.

The first material labour-market effect of AI may not be mass redundancies. It may be companies gradually hiring fewer junior programmers, analysts, paralegals, administrators and customer-service workers because experienced employees equipped with AI can produce more.

The disruption would then appear first at the entry point into employment.

That creates a second-order problem.

Many professions train tomorrow’s experts through precisely the junior tasks AI is becoming capable of automating.

If too many of those first rungs disappear, the eventual problem may extend beyond employment losses to a weaker pipeline for developing experienced professionals.

That remains a risk rather than an established outcome.

But it is the type of structural change that becomes difficult to reverse once embedded in hiring patterns.

The deeper question is how AI divides income between labour and capital

The conventional debate asks whether AI will destroy jobs.

The more consequential macroeconomic question may be:

What happens if AI allows productivity and corporate earnings to rise faster than labour compensation?

If businesses can produce substantially more with the same number of workers — or fewer — productivity can rise, margins can improve and GDP can expand without labour compensation increasing at the same pace.

That outcome is not inevitable. If AI strongly compliments workers, raises their productivity and generates sufficient new demand, labour compensation could continue rising alongside output.

The policy risk emerges if the gains become persistently concentrated in capital.

If AI-intensive assets capture a growing share of productivity gains, more national income could flow to owners of models, intellectual property, data centres, energy infrastructure and financial capital.

An economy can therefore become richer while the mechanism through which most households receive purchasing power becomes relatively weaker.

That would affect far more than employment.

Wages finance household consumption and mortgages. Payroll and income taxes provide substantial government revenue. Employment contributions support pensions and social-insurance systems.

If labour’s share of national income declined while returns to capital expanded, governments could face greater demands for retraining and income support while labour-based taxation represents a smaller share of the economic base.

Gates has proposed taxes linked to AI or machines replacing labour as one possible response. The mechanism is contentious. Automation taxes could discourage productive investment, while defining taxable AI output may prove difficult.

The more important question is broader:

If economic value shifts progressively from labour towards capital, should the tax base eventually move with it?

Taxation, however, is only part of the distribution question.

Ownership may prove equally important.

An economy in which households capture AI-generated capital returns through pensions, investment funds, employee shareholding or broad equity ownership will experience the transition differently from one in which ownership of productive AI assets is concentrated among a narrow group of companies and investors.

The central question therefore becomes not only who works in the AI economy, but who owns it.

That distinction may determine whether productivity gains broaden household wealth or widen the gap between aggregate growth and individual economic security.

Emerging economies face a sharper version of the challenge

The distributional risk may be greater outside economies that control much of the world’s AI capital and intellectual property.

An ILO study covering 135 countries warns that developing economies can face a “disruption without dividend” problem: some workers may encounter automation pressure before inadequate digital infrastructure allows others to capture equivalent productivity gains.

The issue has direct relevance for MENA.

The region is simultaneously seeking stronger digital investment, higher private-sector productivity and large-scale employment creation for young populations.

A joint ILO-ESCWA report examining employment futures through 2035 says AI could produce substantial productivity and employment gains under some scenarios while increasing inequality and displacement under others, with outcomes heavily dependent on skills and policy preparedness.

For MENA economies seeking simultaneously to raise productivity and create large numbers of private-sector jobs for younger populations, AI is therefore both an investment opportunity and an employment-policy test.

Attracting data centres and AI companies is only one part of the development equation.

The more difficult test is whether physical AI investment, human-capital development and economic-transition planning advance at comparable speeds.

A credible framework begins with measurement

No US-China dialogue could resolve these issues in a single meeting, nor is broad harmonisation of labour law, taxation or industrial policy realistic.

The more achievable objective would be a durable mechanism for managing frontier risk and building comparable evidence.

That process would likely begin with definitions of serious AI incidents, secure communication between governments and laboratories, and measurable thresholds for dangerous capabilities.

Any agreement would also require some credible form of verification.

Rules have limited value if neither side can establish whether the other is complying without demanding access to commercially or strategically sensitive technology. Verification, alongside the proliferation of open-weight models, remains one of the principal obstacles to meaningful international AI controls.

The economic layer requires a different principle:

measurement before intervention.

Governments need better evidence on employment, entry-level recruitment, labour compensation, productivity, energy demand and the division of income between labour and capital before determining which interventions are justified.

Washington and Beijing do not need identical domestic policies to benefit from comparable measurements. As the world’s two largest economies and leading AI powers, they could help establish methodologies that later become useful beyond bilateral diplomacy.

The execution test

The first test is whether the prospective discussions occur at all.

The more important test is whether they establish a durable channel with measurable incident thresholds and a credible form of verification, rather than producing another summit declaration.

Only then would it be realistic to widen the agenda towards the economic consequences of productive AI — including employment, skills, labour compensation, capital ownership, taxation and social stability.

The global AI investment cycle is already under way. Chips are being manufactured, data centres financed, electricity infrastructure expanded and models deployed.

Stopping that process is neither realistic nor necessarily desirable.

The more important question is whether institutional capacity can begin catching up with technological capacity.

The next phase of the AI race will be measured by more than which country possesses the fastest chips, the largest computing clusters or the most capable models.

It will increasingly be measured by whether economies can preserve the connection between technological productivity, household prosperity, fiscal capacity and social stability as machines perform a greater share of economically valuable work.

AI already has capital, infrastructure and momentum.

What it still needs is the road map.

Related news:

Digital Governance Gains Ground Across Egypt’s Strategic Sectors

From Compute to Capital: LEAP 2026 Builds the Architecture of Saudi Arabia’s AI Economy

Read also:

Beyond the Red Carpet: Venice Builds a Marketplace for the New Film Economy

Egypt–Turkey Defence Reset Shifts from Exercises to Joint Production