SoftBank Unit, SC ZEUS and Robust HPC Sign ASEAN AI Data Centre MoU

Front panels of rack-mounted servers with status indicators lit, illustrating a report on an ASEAN AI data centre partnership. Photo by heladodementa on Pixabay

ASEAN TECH · 17 Aug 2026 —

SINGAPORE, 17 AUG 2026 — A SoftBank subsidiary, a Singapore data centre developer and a Malaysian high-performance computing firm have signed a memorandum of understanding to build AI data centres in Malaysia, Thailand, Indonesia and Vietnam. No investment figure, capacity target, timeline or site was disclosed.

That combination of four named markets and no numbers makes the announcement worth a careful read. The division of labour and the technical specifications for the buildings are more informative than the headline itself.

What was announced

1 AugustDate the memorandum took effect

4 marketsMalaysia, Thailand, Indonesia and Vietnam first

800 VDCPower architecture specified, with liquid cooling

Not disclosedInvestment, capacity, timeline, sites, financial terms

SB Telecom Singapore, a subsidiary of SoftBank Corp, takes data hall fit-out, integration of information and communications infrastructure, localised deployment support and connectivity. SC ZEUS Data Centers, headquartered in Singapore, sources the land and designs, builds and operates the core facility. Robust HPC Group of Malaysia procures, installs and manages the GPU clusters.

Together the parties say they will develop AI factory solutions and explore GPU-as-a-service platforms and what they call AI token factory commercial models. Hisao Inoue, managing director of SB Telecom Singapore, described the alliance as a step in accelerating digital transformation; Joe Gooi, chief executive of SC ZEUS Data Centers, framed it around AI changing what digital infrastructure has to do.

How to read a memorandum of understanding

A memorandum of understanding is an agreement to work together. It is generally not binding on either party to build anything, spend anything or meet any date, and the region has accumulated a considerable number of them over the past two years.

This does not mean the announcement should be dismissed, but it does require precision about what has and has not happened. Nothing here commits capital. No land has been identified in public. No customer has been named. A memorandum is the paperwork that lets three companies begin doing the expensive work of finding out whether a project is viable, and most of that work happens before anything is announced again.

The contrast with recent announcements is instructive. When CoreWeave said it would build in Indonesia it named a country, a capacity of 360 megawatts and a year. That is still a plan rather than a building, but it is a plan with numbers attached that can be checked later. This one cannot be checked later, because it includes no targets to check against.

The division of labour is the disclosure

Read the three roles together and the commercial model is legible.

SC ZEUS supplies the shell, which is the capital-intensive, slow, permission-heavy part: land, power, construction, operations. Robust HPC supplies the compute, which is the fast-depreciating part that has to be bought in a constrained market and refreshed on a hardware cycle. SB Telecom supplies connectivity and fit-out, which is the part that determines whether the capacity is usable by anyone outside the building.

Splitting a data centre business along those three lines matches each major cost to a company that can bear it. No single partner has to carry the entire risk or fund a shell, a GPU fleet and a network from one balance sheet.

The choice of a Malaysian company for GPU procurement is notable. That partner will operate inside one of the industry's most closely watched supply routes, given the scrutiny on advanced accelerators moving through Malaysia. Nothing in the announcement suggests anything irregular, and none is alleged here. It does mean that the procurement leg of this partnership operates in a compliance environment that did not exist two years ago, and that is a real operational cost.

The technical specification is the other disclosure

The mention of liquid cooling, 800-volt direct current distribution and low-latency networking is not marketing fluff. These specifications describe a building for rack densities that conventional data centres cannot easily support.

Air cooling runs out of headroom as power per rack climbs, which is why liquid cooling has moved from exotic to expected in AI facilities. Direct current distribution at 800 volts reduces conversion losses and copper weight at those densities, and it is the direction the industry has been moving for next-generation accelerator racks.

Most existing capacity in Southeast Asia was not built for either. A specification like this is therefore a statement that the partners intend new-build facilities designed for current-generation AI hardware, rather than retrofitting or leasing existing halls. That is a more concrete claim than anything in the headline, and one that would be expensive to walk back.

Four markets, four different problems

Naming Malaysia, Thailand, Indonesia and Vietnam ahead of Singapore is itself a decision.

Singapore has the demand and the connectivity and comparatively little land and power headroom. The four named markets have more room and each brings a different constraint: Malaysia has attracted the most AI-specific capacity and faces the accompanying scrutiny on power and water; Thailand's capacity is growing quickly from a small base; Indonesia has scale and a grid that has to be planned around; Vietnam has manufacturing demand and a regulatory environment for data that is still settling.

A partnership announcing all four at once has not yet made the choice that matters, which is where to start. Watching which market gets the first site will say more about the economics than the memorandum does.

What we could not establish

Everything commercial. No investment amount, capacity target in megawatts or GPU count, deployment timeline, site location, revenue-sharing arrangement or governance structure was disclosed, and there is no public indication of which market comes first.

Also unestablished is whether an anchor customer exists, if the GPUs are already allocated, how the partnership is financed, if SoftBank Corp's involvement extends beyond its Singapore subsidiary, and whether the memorandum is exclusive. We also could not establish whether the AI token factory model refers to selling inference capacity by token volume, which is what the phrase implies, or to something narrower.

What to watch

A named site would be the first sign of a real project. Land, power allocation and a location are what turn a memorandum into a development, and none are yet in the public record.

The second signal is an anchor tenant. GPU-as-a-service is a business that works when capacity is contracted before it is built and works badly when it is not, and the partners have not said which model they are pursuing.

Finally, watch whether this partnership survives the year without a follow-up announcement. Memoranda that lead to buildings usually generate news within a few quarters — a site, a permit, a financing. Ones that go quiet have generally gone nowhere, and the silence is the result.

Sarah Chew’s Verdict ASEAN

Four named markets and no numbers. No investment figure, no megawatts, no timeline, no site, no financial terms — and I would hold that line firmly, because a memorandum of understanding generally binds nobody to build anything, and this region has collected a great many of them since the AI build-out began. Nothing here commits capital. That said, dismissing it would mean missing the two things that were disclosed, both of which carry more information than the headline. The first is the division of labour. SC ZEUS takes the shell — land, power, construction, operations, which is the slow capital-heavy part. Robust HPC takes the GPUs, which is the fast-depreciating part bought in a constrained market. SB Telecom takes connectivity and fit-out, which decides whether anyone outside the building can use the capacity. That split matches each cost to a party that can carry it, and it means no single balance sheet has to fund all three at once. The second is the specification. Liquid cooling and 800-volt DC distribution are not adjectives; they describe rack densities that conventional design does not serve, and most existing Southeast Asian capacity was not built for either. So the partners are saying they intend new buildings for current-generation AI hardware rather than retrofits or leases. That is the most concrete claim in the announcement and the most expensive to walk back. One note on the Malaysian leg: a Malaysian firm handling GPU procurement operates inside a compliance environment that did not exist two years ago, given the scrutiny on accelerators moving through that route. No irregularity is suggested or alleged. It is simply an operational cost that belongs in anyone's read of the partnership. What I would watch is which of the four markets gets the first site. That single decision will tell you more about the economics than the memorandum does.

Sources & cross-checks

Primary: TNGlobal, 4 August 2026 — for the memorandum taking effect on 1 August 2026 between SB Telecom Singapore, SC ZEUS Data Centers and Robust HPC Group; the target markets of Malaysia, Thailand, Indonesia and Vietnam ahead of the wider region; the division of roles, with SC ZEUS sourcing land and designing, building and operating the core infrastructure, Robust HPC procuring, setting up and managing GPU clusters, and SB Telecom Singapore handling data hall fit-out, ICT integration, localised deployment support and connectivity; the specification of liquid cooling, 800 VDC architecture and low-latency networking; and the plan to develop AI factory solutions and explore GPU-as-a-service and AI token factory models.
Also primary: w.media — corroborating the parties, roles, markets and specification, and for Hisao Inoue, managing director of SB Telecom Singapore, on the alliance accelerating digital transformation, and Joe Gooi, chief executive of SC ZEUS Data Centers, on AI changing digital infrastructure requirements. Both accounts state explicitly that investment amount, capacity, timeline, site locations, financial terms and governance were not disclosed.
Our earlier reporting: The comparison is CoreWeave's Indonesian announcement, which named a country, a capacity and a year.
What could not be established: ⚠ Every commercial term. No investment amount, megawatt or GPU-count target, deployment timeline, site location, revenue-sharing arrangement, governance structure or first-market choice was disclosed. Also unestablished: whether an anchor customer exists, whether the GPUs contemplated are already under allocation, how the partnership is financed, whether SoftBank Corp's involvement extends beyond its Singapore telecom subsidiary, whether the memorandum grants exclusivity between the parties, and whether "AI token factory" refers to selling inference capacity by token volume or to something narrower.
Method and limits: Compiled 17 August 2026. The explanation of what a memorandum of understanding does and does not commit, the reading of the three-way split as a matching of each cost to a party able to bear it, the interpretation of liquid cooling and 800 VDC as a commitment to new-build capacity for current-generation hardware, the account of what distinguishes the four named markets, and the observation about the compliance environment around GPU procurement in Malaysia are RECATOOLS analysis rather than claims by any of the three parties. No irregularity in GPU procurement is suggested and none has been reported.