Google’s $15 billion Malaysia plan puts power and water at the center of AI investment

Original analysis by Open Market Notes. Information cutoff: November 5, 2026, 00:00 UTC. Investment commitments, company estimates and our analysis are distinguished below.

From building facilities to coordinating resources

Google announced on November 4 that it plans to invest $15 billion in Malaysia during 2027 and 2028. This is a forward investment program, not spending already completed. Its announcement brings data centers, skills, water and additional clean energy into one plan; Reuters independently reported the amount and timetable.

The structure of the commitment matters as much as its size. AI infrastructure is often presented as competition among chips, racks and models. Once expansion reaches a particular community, the questions change: which transmission line provides power, where water comes from in a dry period, who maintains public assets, and what employment remains when construction ends.

Our reading is that the plan illustrates a more complete investment model. A technology company must help strengthen the local conditions that allow its equipment to operate, not merely purchase the equipment. This does not mean resource conflicts have been resolved. It means evaluation should shift from capital announced to infrastructure that capital actually makes available.

A gigawatt is not a promise of continuous supply

Google plans to add more than one gigawatt of clean-energy capacity to Malaysia’s grid. Capacity measures potential power at a moment in time. Energy depends on actual output over time. Neither a capacity target nor an annual procurement total demonstrates that a data center receives clean electricity in every hour.

This is a physical distinction, not an allegation about the company. Solar generation and computing demand may occur at different times. Even if yearly purchased generation equals yearly consumption, nights and cloudy periods still require storage, dispatchable resources or other grid support. Reliability depends on matching hourly supply and demand, not simply balancing annual figures.

The Malaysian Energy Commission’s CRESS guidelines provide a route for corporate consumers to buy renewable power from developers through the grid. Contractual and access arrangements connect buyers, generators and the network. They cannot replace substations, transmission lines or dispatch capability. Signing a contract begins the supply chain; the regional benefit depends on generation connecting on time and delivering when it is needed.

Cost allocation matters too. Funding a connection can cover direct project costs without automatically settling the wider treatment of reserves, congestion and long-term maintenance. The appropriate question is not whether a data center can operate without a grid. It is whether the relationship between new demand and additional system costs is transparent.

Low PUE is not the same as low water use or emissions

Google states that its campuses in tropical settings such as Malaysia can achieve power usage effectiveness of 1.12–1.14. That is a company performance statement, not independent verification of full-year measurements at every proposed site.

The US Department of Energy’s design guide defines PUE as total facility energy divided by IT-equipment energy. It describes overhead from cooling, power distribution and other supporting systems. It does not measure energy per useful AI task, water consumption or carbon emissions.

Consider a simple illustrative example. With 100 units of IT energy and PUE of 1.12, total facility energy is 112 units. Double the equipment load at the same PUE and total energy doubles. Efficiency reduces overhead per unit of IT demand without proving that aggregate demand falls.

Two facilities with identical PUE can also use electricity with different carbon intensity and employ different water and cooling arrangements. Efficiency, absolute electricity use, water intensity and emissions should therefore be evaluated together. One attractive ratio cannot resolve every environmental trade-off.

Water benefits must be tested by watershed and season

A separate Google announcement commits $160 million to shared water infrastructure in Negeri Sembilan: an off-river reservoir and treatment plant, designed to supply 65 million liters daily, with ownership passing to a state entity upon completion. These are design capacity and ownership arrangements, not additional water already reaching residents.

Treatment capacity, available raw water and delivered water are separate constraints. A finished plant cannot fully translate its rating into service if distribution links are missing or dry-season supplies are inadequate. Conversely, coordinated improvements in storage, treatment and distribution can benefit the public even when a data center consumes part of the output.

The right questions concern place, season, quality and beneficiaries. Upstream restoration, rainwater harvesting and additional potable-water treatment can all help, but they perform different functions. Their volumes should not simply be added together as interchangeable offsets for consumption anywhere and at any time.

Ownership also needs an operating counterpart. Who funds maintenance, monitors water quality and decides allocations during drought affects whether a capital project delivers lasting public value. Investment is the starting point; dependable service is the outcome.

Employment should be judged by capabilities that remain

Infrastructure initially creates construction demand and then moves into a smaller, longer-lived operating phase. Google’s employment estimates distinguish these stages and include related supplier activity. Construction-period jobs supported annually are not automatically permanent new positions, and not all supported jobs are direct Google employees.

For Malaysia, transferable skills may be more important than the peak headcount. Electrical engineering, cooling maintenance, water treatment and commissioning expertise can serve multiple customers. Training tied to one temporary build may lose value after completion. Local suppliers able to win repeat work can retain benefits beyond the original project.

Investment is not equivalent to domestic value added either. Imported servers, local construction contracts, maintenance services and wages affect the economy differently. Development outcomes should be judged partly by the knowledge and business retained locally, rather than counting the full cost of globally sourced equipment as local income.

Three schedules must align

The first schedule covers computing facilities: construction, equipment, testing and service. The second covers resources: grid connections, generation, treatment and distribution. The third covers public benefits: training becoming employment, supplier capability developing, and communities receiving more reliable services.

Misalignment changes the economics. Racks arriving before power can leave expensive assets idle. Facilities operating before public infrastructure is ready can impose pressure before benefits arrive. Training delivered too early or too late can miss actual vacancies. These are execution risks to manage, not events this article claims have occurred.

Verification should focus on concrete evidence: actual spending versus announced commitments, commissioning and connection progress, site-level water and efficiency data, public service from shared assets, and sustained employment following training. Those observations will establish whether computing capacity and the resources supporting it are improving together.

Google’s plan shows that the next phase of AI expansion increasingly resembles an engineering program spanning several systems. Resource coordination may become a deployment advantage for companies. For host countries, success means more than additional buildings: new computing should develop alongside more dependable public infrastructure. The $15 billion commitment is an important beginning. Its final test remains in power, water and everyday local life.