Wednesday’s AI Agenda: Capital, Work and Autonomy

NVIDIA will publish its second-quarter fiscal 2027 results on Wednesday, followed by a 2pm Pacific call. The prior quarter set a demanding baseline: $81.6 billion in revenue and $75.2 billion from Data Center. ¹² The headline will move markets. For operating leaders, the mix matters more: demand concentration, supply, product transitions, customer financing and management’s forward view.

The IEA estimates that capital expenditure by five technology companies exceeded $400 billion in 2025 and may rise another 75% in 2026. It also warns that not every announced data-centre project will be built. ³ A strong result would support continued infrastructure intensity; it would not prove that your organisation is earning value. Use it to revisit supplier concentration, capacity assumptions and contract flexibility. Keep the internal business case tied to accepted work, customer outcomes and operating cost.

In Bangkok, the ILO and China’s Ministry of Human Resources and Social Security will bring ASEAN policymakers, employers and workers together around employment-centred AI governance. The programme covers public employment services, vocational education, enterprise development, regulation and social dialogue. ⁴

Watch for practical mechanisms, not only ambition: who funds skills, how smaller firms participate, how informal work is treated and what employers and workers will measure together. ASEAN economies differ in infrastructure, labour formality and institutional capacity, so no single regional conclusion should become a global workforce forecast. The transferable question is stronger: does your workforce plan include credible training, role redesign and worker voice before productivity targets are booked? The ILO–Singapore partnership shows that this institutional work is intended to continue beyond one event. ⁵

The AI Daily Brief will run a public lab on agentic loops for knowledge workers: systems that continue checking and refining work until a finish line is reached. ⁶ That matters because autonomy is moving from specialist teams into ordinary functions. OpenAI’s own enterprise analysis also reports a shift from assistance toward execution inside its customer base, although that vendor sample is not a global productivity measure. ⁷

Treat the lab as capability discovery, not rollout evidence. Any real loop needs a named outcome, allowed tools and data, a cost ceiling, an escalation path and a person who can stop it. If those cannot be stated plainly, the process is not ready for delegated execution.

Two new version-one preprints point in the same direction underneath the public agenda. One derives reusable task models from observed computer work; another routes work between models according to expected value and the cost of further analysis. ⁸⁹ Neither is production proof.

For leaders, the direction is enough to watch: AI systems may increasingly learn how work is performed and choose how much intelligence to spend on each step. Test that promise on one repeated workflow. Require traceability, permission controls and a cost per accepted outcome before buying at scale.

Every market faces some version of the same three questions: how much infrastructure is justified, how work should change and how much authority software should receive. The answers vary. Energy access, trade controls and supplier concentration shape capital. Labour institutions, informality and training capacity shape workforce transition. Privacy, employment and AI rules shape which data and decisions an agent may touch.

European transparency duties already affect some AI interactions and synthetic content, with different transition dates for existing systems. ¹⁰ Other jurisdictions rely more heavily on existing privacy, consumer, employment and sector rules.

Ask the CFO to record whether the infrastructure outlook changes a supplier or capacity assumption. Ask the people leader to identify one missing workforce mechanism. Ask the operating owner to nominate one workflow that is ready—or not ready—for a bounded loop.

Then change only what the signal supports. A market result can reprice exposure. A policy dialogue can sharpen workforce design. A demonstration can justify a controlled test. None should be allowed to approve the whole AI strategy by association.