Updated every Monday · 07:00 SGT
AI capital is in, but returns are not.
Boards are funding the AI P&L. Few CEOs can yet show both cost takeout and new revenue on the same page.
Frontier Watch
Lab and platform moves this fortnight. TCO, vendor mix, residual risk.
Key points
- US closed models and China open-weight families moved in the same fortnight — treat them as one vendor-mix problem.
- Switching cost and residual risk are the CEO questions; origin of the weights is a residency note, not a loyalty test.
- A two-vendor, two-stack posture remains the default hedge for an APAC group with a China P&L.
For the CEO
- Review whether the board pack still splits ‘US labs’ and ‘China labs’ into different tabs.
- Consider one residual-risk line that covers both stacks: identity, evals, kill-switch, residency.
4 min
Audience · Board · ELT
McKinsey — State of AI trust 20265 min listen · 9 min read
Key points
- Companies plan to lift AI spend to about 1.7% of revenue in 2026; 94% will keep investing even without immediate returns (BCG).
- Trailblazer CEOs (~15%) report systematic productivity and decision-quality gains; Followers still fund pilots that miss the operating statement.
- 72% of CEOs now say they own the AI decision — twice last year’s share.
For the CEO
- The board question has moved from whether to fund, to how capital is ring-fenced and when the dual P&L will show both cost and revenue.
- Residual risk on agentic workflows belongs next to cyber — not under an innovation update.
9 min
Audience · Board · ELT
BCG AI Radar 2026Written executive summary
5 min listen · 9 min read
For Board Preparation
Capital is no longer the constraint. BCG’s AI Radar 2026 finds companies planning to lift AI spend from about 0.8% to 1.7% of revenue, with 94% intending to keep investing even if returns are not immediate (BCG). The board question has shifted from whether to fund AI to how the capital is ring-fenced, how residual risk is carried, and when the dual P&L will show both cost and revenue. Trailblazer CEOs — about 15% — are already reporting productivity and decision-quality gains where they applied AI systematically; Followers are still funding pilots that never hit the operating statement (BCG). Risk sits next to capital: agentic systems now touch payments, customer files, and production workflows, so the control plane is a board risk topic, not an IT footnote (McKinsey).
For ELT Attention
Ownership of ROI is still split across digital, data, and the line. The ELT conversation this week is who carries the production path from pilot to P&L, and which two workflows are allowed to consume the new capital. McKinsey’s 2026 trust survey finds only about 30% of organisations at higher maturity across strategy, governance, and agentic controls — the rest are still experimenting with agents they cannot yet supervise (McKinsey).
Must-know AI headlines
EU AI Act high-risk duties are now a buyer and insurer expectation, not a Brussels-only file.
Procurement and D&O language will start to assume documented residual risk, even for groups whose primary market is Asia Pacific.
(EU)US frontier-model testing (AISI) is becoming a reference standard for enterprise model due diligence.
Model cards and red-team evidence will show up in vendor questionnaires the way SOC 2 did a decade ago.
(NIST)
Audio briefing
5 min listen
Three numbers
1.7%
Planned AI spend as a share of revenue in 2026, up from 0.8%
BCG
72%
CEOs who say they are now the main decision maker on AI
BCG
15%
Trailblazer CEOs reporting systematic productivity and decision gains
BCG
Weekly triage
Board prep
ELT attention
Delegate
From the Library
Deep dive · Same theme · ~25 min
PwC Global CEO Survey — value, workforce, confidence
How peer chief executives describe AI value versus what they can show.
Must-know AI headlines
- 1
EU AI Act high-risk duties are now written into buyer and insurer questionnaires, including for groups whose primary market is Asia Pacific.
A Singapore- or Tokyo-headquartered group selling into Europe — or raising from European LPs — will be asked for residual-risk evidence as a commercial term, not a legal footnote.
EU AI Act (Regulation 2024/1689) - 2
US AISI frontier-model testing is becoming the reference language for enterprise model due diligence.
Vendor packs that cannot point to independent evaluation will start to fail procurement the way unaudited cloud once did.
NIST AISI - 3
Singapore’s Model AI Governance Framework remains the practical regional reference for boards that need something implementable, not only principled.
For APAC groups, IMDA / PDPC language is the shortest path to a board-readable control narrative that counterparties recognise.
IMDA / PDPC Model Framework - 4
Data and model access remain uneven across Japan, ASEAN, and China — the same foundation model is not a single regional product.
A group operating across those jurisdictions needs a vendor-mix and residency story, not a single-stack decision.
METI AI policy (Japan)
Watch list
OECD AI Principles continue to be the common diplomatic language behind national rules.
Useful when the board asks ‘which standard are we aligning to’ across more than one jurisdiction.
OECD AI PrinciplesPwC’s Global CEO Survey still shows a wide gap between stated AI confidence and reported value.
A useful peer check against internal optimism on the dual P&L.
PwC Global CEO Survey
Asia Pacific
Singapore
Governance as the regional reference
Singapore’s Model AI Governance Framework and IMDA’s GenAI extension remain the most board-usable pack in the region: roles, data, testing, and human oversight in language a risk committee can read. Groups using Singapore as the control location for regional AI can treat this as the default narrative, then delta Japan, ASEAN, and China.
PDPC Model AI Governance FrameworkJapan
Promotion-first statute, still a binding culture
Japan’s AI policy stance is promotional rather than a Brussels-style prohibition list, but METI guidance and customer expectation still constrain training data, transparency, and public-sector procurement. For a Tokyo listing or a Japan P&L, ‘we follow Singapore’ is not a complete answer.
METI AI policyASEAN
Access is not one market
ASEAN is not a single data or model regime. Cross-border inference, localisation, and sector licences (finance, health, telecoms) still fragment what a regional platform can do. The CEO question is which markets share a stack, and which require a sovereign or local-partner path.
OECD AI Principles (regional reference)China
Model access is a strategic constraint
China’s generative-AI and algorithm rules, plus export and interconnection limits, mean the China P&L often cannot sit on the same foundation-model contract as the rest of Asia Pacific. Treat China model access as a portfolio decision with its own residual-risk note.
Cyberspace Administration of China — Generative AI measuresArchive edition · Return to this week