From isolated pilots
to embedded intelligence.
A three-horizon plan to make artificial intelligence an embedded capability across Keppel's asset management and operating platform, written as I would write it in my first week in the role: where we stand, what I believe, what we build, where the value is, and how we will be measured.
01 / Starting pointWhere Keppel stands on Day 1
Keppel is no longer a conglomerate; it is a global asset manager and operator with three operating divisions and a fast-growing funds platform. The AI mandate has to be read against that shape: the value pools are in assets, decisions and the infrastructure Keppel sells to the AI economy itself.
Up 39% year on year; recurring income S$941 million, up 21%. ROE 18.7%.
Target S$100 billion by end-2026 and S$200 billion by 2030; asset management net profit S$189 million.
Asia Pacific, including 720 MW secured for an AI data centre in Melbourne; KDC SGP 9 AI-ready campus breaks ground mid-2026.
Sakra Cogen, Singapore's first hydrogen-compatible CCGT, in operation; renewables 4.7 GW toward 7 GW by 2030.
Trans-Pacific cable in commercial operation since December 2025; about S$200 million O&M fees per fibre pair over 25 years.
Energy savings from the AI-powered cooling solution launched in 2025; the Operations Nerve Centre at Changi runs ML-driven monitoring of cooling and energy assets.
Record recurring income; decarbonisation solutions EBITDA S$130 million, up 32%; S$7.1 billion of long-term contracts.
Annual cost-savings target by end-2026 (S$98 million achieved 2023–2025); M1 on a 90-day AI-supported efficiency plan.
AI is present, but as point solutions
Cooling optimisation, the ONC, M1's digital platform and pockets of analytics exist. What is missing is a shared data and semantic foundation, a common engineering practice and a path that moves a good pilot into every relevant asset. The job is to turn islands into a system.
The value is in assets and decisions
Keppel's economics are recurring income from operating assets and fee income from managing capital. AI that improves dispatch, cooling, maintenance, capacity and deal decisions moves those lines directly. Generic productivity tools matter, but they are not where the S$ is.
Keppel sells the AI economy its infrastructure
Over 1 GW of data centre powerbank, hydrogen-ready generation, subsea capacity and AI-driven cooling mean AI is also a product line. Every capability we build for our own assets should be designed to be packaged for tenants, partners and LPs.
02 / PrinciplesFive operating beliefs
These come from building an industry large model and a 100-agent platform at a listed group, from founding a hardware company, and from a year of shipping agentic systems for manufacturers. They decide what we fund and what we stop.
A business capability, not an IT project
Every initiative has a business owner with a P&L, a value hypothesis in S$ and a named decision it improves. The AI office builds platforms and squads; the divisions own outcomes.
Ontology before agents
A Keppel Asset Ontology is the executable contract between data teams and decision makers: what a plant, a chiller, a rack, a lease, a fund and a commitment mean, with constraints the machine checks before any recommendation.
Deterministic core, generative shell
Optimisers, physics and control produce the plan; language models understand, explain and orchestrate. Anything that changes a commitment, a dispatch or a bid stops in front of a person. Autonomy is graded L0–L3 and earned.
Production is the metric
No proof of concept without a path to production, a data owner and an evaluation set. "No eval, no deploy." Value is tracked in the division's P&L, not in a slide.
Governance as an accelerator
Responsible-AI foundations are built once, as reusable controls, evaluation pipelines and model-risk tiers, so that each new use case inherits compliance instead of renegotiating it.
03 / Target architectureThe Keppel AI Stack
Five layers, one ontology, and governance cutting across all of them. Each division plugs its assets into the same platform; each product, internal or external, is assembled from the same engines and agents.
Copilots for Keppel teams; Result-as-a-Service for assets in Keppel-managed funds; external offerings packaged with the operating businesses.
Agents that decompose tasks, call engines and tools, cite evidence and stop at human checkpoints. Autonomy graded L0 (advise) to L3 (act within guardrails).
Forecasting and optimisation for dispatch, cooling and capacity; computer vision for inspection and site safety; digital twins; document and knowledge models for deals, contracts and regulation.
Keppel Asset Ontology across power, cooling, data centres, buildings, funds and contracts; a unified data platform over SCADA, BMS, IoT, ERP and fund administration; lineage, quality gates and decision-readiness checks.
Multi-model by design: frontier APIs, open-weight models and Singapore-hosted options chosen per task; ML infrastructure on Keppel's own data centres; MLOps, evaluation harnesses, secrets and OT/IT segregation.
Policy, model-risk tiers, human oversight, data protection, IP, transparency and resilience, implemented as reusable controls and evaluation pipelines rather than per-project reviews.
Group platform at the base, division models in the middle, scenario applications at the edge. Data flows back into the ontology and models; models land in the next asset. Systematic AI is the moat; point solutions are not.
- Run our own compute where it pays. Keppel builds AI-ready data centres; our ML platform should live in them and become a reference customer.
- One ontology, many models. Model vendors will change every year; the semantic layer and evaluation sets are the durable assets.
- Evidence fields are mandatory. Every agent recommendation carries its sources, assumptions and confidence, or it is not shown.
- OT stays deterministic. Plants, chillers and switchgear take set-points from validated optimisers and controllers, never from a language model directly.
04 / Value mapWhere AI creates measurable value
Twenty-two opportunity areas mapped by value at stake and readiness to reach production, based on public information and the first-100-day assessment that will replace these estimates with Keppel's own numbers.
Infrastructurelargest earnings share
- Dispatch, bidding and fuel optimisation across Sakra, Merlimau and BESS1–3% margin · indicative
- Predictive maintenance for turbines, boilers and chillers15–30% maint. cost
- AI cooling scaled across all district cooling and EaaS sites, then sold as a serviceup to 20% energy
- Waste-to-energy combustion and availability optimisationavailability +1–2 pts
ConnectivityAI-ready campus
- Data centre PUE, cooling and capacity planning; incident predictionPUE −0.05 to −0.1
- Floating data centre design and operations twinde-risk first-of-kind
- Bifrost capacity sales analytics and network operationsfibre-pair pricing
- M1 customer-service and network-operations agents supporting the 90-day plancost to serve −20–30%
Real EstateRE-as-a-Service
- Building energy and tenant-experience optimisation8–15% energy
- Sustainable urban renewal: AI audit to retrofit scope and business caseweeks, not months
- Leasing and asset-plan analyticsfaster decisions
Asset ManagementFUM S$95bn → 200bn
- Investment copilot: screening, data-room Q&A, IC memo drafting with evidenceDD cycle −30–40%
- Portfolio and asset-performance monitoring across fundssame platform as ops
- LP reporting, ESG and climate data automationdays to hours
05 / Lighthouse portfolioSix initiatives for year one
Chosen for value, readiness and what they teach the platform. Each has a division owner, a value hypothesis, a production path and a 12-month milestone. The bar shows the share of the initiative that is foundation work versus new capability.
Asset Intelligence for AI-ready data centres
Cooling and PUE optimisation, predictive maintenance and capacity planning across the Genting Lane campus first, then the portfolio. Designed so tenants running AI workloads see it as a service.
- Value
- PUE −0.05 to −0.1; unplanned downtime −30%; capacity released for AI racks
- 12 months
- Two campuses in production; RaaS offer defined for REIT assets
Energy Operations Agent
A deterministic dispatch and bidding optimiser for Sakra, Merlimau and storage, wrapped by an agent that explains, runs what-ifs and prepares decisions for traders and plant managers. Plant health models feed the same loop.
- Value
- Spark-spread capture +1–3%; maintenance cost −15%
- 12 months
- Live in the control room at L1 (advise), L2 (propose) for storage
Investment Intelligence Copilot
Screening, data-room question answering, investment-committee memo drafting with mandatory evidence, and portfolio monitoring that reads the same asset data the operators use.
- Value
- Due-diligence cycle −30–40%; more deals screened per analyst
- 12 months
- Used in every IC paper of two flagship funds; LP reporting automated
Keppel Asset Ontology & data platform
The semantic layer for power, cooling, data centres, buildings, funds and contracts, with source-system mappings, constraint axioms and a decision-readiness gate. The contract every lighthouse is built on.
- Value
- Second use case at a fraction of the first; data disputes settled once
- 12 months
- Ontology v1 covering three divisions; readiness gate in front of every agent
AI Cooling and EaaS productisation
Package the Operations Nerve Centre and the award-winning AI cooling into a repeatable Cooling-as-a-Service product, following the Chennai model, with an AI performance guarantee and a partner delivery kit.
- Value
- New recurring revenue; higher margin on decarbonisation EBITDA (S$130M, +32%)
- 12 months
- Productised offer; three new sites under contract outside Singapore
Responsible AI & model-risk foundation
Policy, model-risk tiers, evaluation pipelines, human-oversight rules and an AI inventory, aligned with Singapore's Model AI Governance Framework, AI Verify and ISO/IEC 42001. Enables the other five to move fast.
- Value
- Approval time per use case measured in days; zero unmanaged models
- 12 months
- Framework adopted by the Board; all production agents tiered and evaluated
06 / RoadmapThree horizons, 36 months
Listen and assess, build foundations while shipping lighthouses, scale across assets and funds, then embed. Quarters from the start date; milestones in red.
Listen, assess, decide
- Asset and technology landscape across all divisions; data audit
- Value map with Keppel's own numbers; six lighthouses funded
- AI Council chartered; hiring plan; two quick wins shipped
Foundations and lighthouses
- Ontology v1, data platform, MLOps, model-risk tiers
- All six lighthouses in production at first sites
- First S$ value tracked in division P&Ls
Scale and productise
- Roll-out across Keppel-managed assets and REIT portfolios
- RaaS for LPs; external Cooling-as-a-Service and DC services
- Partner ecosystem and academic programmes running
Embed
- AI in every operating platform and every investment committee
- AI-native offerings contributing to recurring income
- Group-level value and risk reported like any other line
07 / Operating modelHub, spokes and a stage gate
A small Group AI & Technology Office that builds the platform and lends forward-deployed engineers; division AI squads that own use cases; business leaders who own value. Ideas move through one funnel with clear exits.
- Forward-deployed engineers paired with plant managers, DC engineers and deal teams: the people who turn tacit expertise into ontology and agents.
- Hire for the platform, borrow for the frontier: ML, ontology and data engineering in-house; frontier research through NUS, NTU, A*STAR and AI Singapore, with joint Ph.D. and industry-attachment tracks.
- Partners with skin in the game: hyperscalers already working with Keppel on data centres and cables, model providers, and start-ups evaluated through the same stage gate.
- Keppel as customer zero: our own AI workloads run in Keppel data centres and become the reference for AI-ready DC services.
08 / Responsible AIFast because it is governed
Built with Risk, Cybersecurity, Data and Legal from day one, aligned with Singapore's Model AI Governance Framework (including the 2024 generative-AI edition), AI Verify, the PDPA, ISO/IEC 42001 and the NIST AI RMF, and with the operational realities of plants, networks and funds.
Model-risk tiers and human oversight
Every model and agent is tiered by the decision it touches. Higher tiers require validation, monitoring and a named human approver; the tier decides the autonomy level an agent may reach.
Data protection, IP and provenance
Data classification and residency rules for LP, tenant and customer data; provenance recorded for training and retrieval sources; IP terms settled in every vendor and partner agreement before data moves.
Cyber, OT resilience and evaluation
OT/IT segregation for plants and data centres; adversarial and safety evaluation before deployment; drift and incident monitoring in production; fallback to deterministic control on any failure.
09 / ScorecardHow the role should be measured
Indicative 36-month targets, to be calibrated with the CEO and CFO in the first 100 days. The measure is value delivered and transformation embedded, not models trained.
EBITDA and cost impact tracked in division P&Ls, audited like any other line. Indicative.
Agents and models in production across all divisions, each with an owner, a tier and an evaluation set.
Median time from approved value case to first production site, down from the years typical of enterprise AI.
Share of Keppel-operated capacity (power, cooling, data centres, buildings) covered by the ontology and asset-intelligence services.
Productised AI services (Cooling-as-a-Service, AI-ready DC services) contributing to recurring income.
Every production model in the inventory, tiered and monitored; zero Tier 1 decisions without a human approver.
Platform, ontology, ML and forward-deployed engineers across the Group office and division squads; joint programmes with two universities.
Investment-committee papers for flagship funds prepared with the copilot, with evidence and assumptions machine-checked.
10 / First 100 daysWhat I will do first
Listening before building, but not listening instead of building: two quick wins ship inside the first quarter.
Listen
- Site visits: Sakra, ONC, Genting Lane campus, M1, two funds' deal teams
- 30 conversations with division heads, CIOs, risk and investment leaders
- Inventory of existing models, pilots, vendors and data
Assess
- Data and architecture audit; OT/IT and cyber posture
- Value map rebuilt with Keppel numbers; lighthouse short-list
- Responsible-AI gap analysis against the Singapore framework
Decide
- Roadmap, budget and hiring plan to the CEO and Board
- AI Council chartered; division AI leads named
- Platform choices: compute, models, ontology tooling
Ship
- Quick win 1: AI cooling extended to two more sites with value tracking
- Quick win 2: investment-copilot pilot on one live deal
- Ontology v0 for one asset class; first FDEs embedded
Report
- Board update: baseline, targets, first value, risks
- Public narrative for LPs, tenants and talent
- Horizon 1 starts with all six lighthouses funded