The Ghost Clouds: Inside Southeast Asia's Speculative GPU-Rental Boom

JAKARTA — The pitch deck that crossed dozens of investor inboxes last autumn was twelve slides long and breathtaking in its confidence. A company called Meridian Compute, registered in Singapore with operations promised in three countries, would secure 12,000 high-end accelerators, wrap them in a developer-friendly cloud, and rent them to the AI startups of Southeast Asia at rates 30 percent below the global hyperscalers. It raised a reported 90 million dollars in months.

There was one wrinkle, which the deck did not dwell on. When BriefAsia began tracing Meridian's supply chain this spring, the company could account for fewer than 1,500 accelerators physically installed and powered in the region. The rest existed as allocation letters, forward contracts, reseller commitments and, in several cases, chips that had been bought and quietly sold on to someone else before they ever reached a rack.

Meridian is not alone, and this is not principally a story about one company. Over four months, BriefAsia reviewed corporate filings in three jurisdictions, contracts and term sheets shared by eleven sources, and interviewed founders, financiers, datacentre operators and two former employees. What emerges is a portrait of a regional 'AI cloud' boom in which a meaningful share of the businesses are not really compute providers at all. They are arbitrageurs — trading on the gap between who can get scarce chips and who is desperate to use them.

The findings matter beyond the fortunes of a few startups. Governments across the region are leaning on exactly these companies to deliver the 'sovereign compute' they have promised voters. If the clouds are partly hollow, so is some of the strategy built on top of them.

How the arbitrage works

The mechanics are simpler than the rhetoric. Advanced accelerators are allocated, not freely sold. Suppliers and their channel partners decide who gets how many, and when. An allocation — the right to buy a quantity of chips at list price on a delivery schedule — is itself a valuable asset in a shortage, because the secondary market price runs well above list.

A company that secures an allocation has three choices. It can build a cloud and rent the chips out over years, the hard and capital-intensive path. It can flip the allocation or the delivered hardware to a buyer further up the queue, booking an immediate margin. Or it can do both — build a modest showcase cluster for credibility, and quietly monetise the rest of its allocation through resale. The third path, several sources said, is where much of the regional 'cloud' boom actually lives.

Half the AI clouds I diligenced last year were allocation traders wearing a cloud costume. The cluster in the photos was real. It just wasn't the business, said a Singapore-based infrastructure investor who passed on three such deals.

None of this is necessarily illegal. Reselling hardware is a legitimate business, and arbitrage in a shortage is as old as commerce. The problem is disclosure. Investors and government partners were told they were funding durable compute capacity. In several cases reviewed by BriefAsia, what they were funding was a position in a fast-moving commodity trade — with the risk profile of a trade, not an infrastructure asset.

The Meridian paper trail

Meridian's structure, reconstructed from filings and contracts, illustrates the pattern. The Singapore holding company sat above operating entities registered in Indonesia and a third market. Allocation letters from two channel partners, dated last September, committed a combined headline figure of roughly 12,000 accelerators across staggered deliveries running into 2027.

An allocation letter is not a delivered chip. It is a place in line, contingent on payment, export approvals and the supplier's discretion. By treating those letters as inventory, Meridian could present investors with a fleet that existed mostly as future rights. The 90 million dollars raised was, in effect, working capital to convert paper rights into either real machines or profitable resales — and the incentives pulled toward resale.

Two former employees described an internal tension that resolved the same way every quarter. The cloud business lost money on day one; depreciation and power and engineers are expensive, and paying customers were few. The resale desk made money immediately. When a delivery of several hundred accelerators landed in March, one source said, a majority were sold to an intermediary serving a customer outside the region within weeks, never racked locally.

Meridian, through a representative, disputed the characterisation. It said it operated 'a hybrid model that includes hardware trading alongside cloud services', that all transactions complied with export controls, and that its installed base was 'growing in line with deliveries and power availability'. It declined to provide a current accelerator count or utilisation figures.

Why the region was such fertile ground

Three conditions made Southeast Asia unusually hospitable to the ghost-cloud model. The first was genuine, acute demand. Local AI startups, priced out of foreign clouds and frozen out of allocation queues, would sign almost any term sheet for guaranteed access. That demand was real, which made the supply stories credible.

The second was the sovereign-compute narrative. Governments wanted national champions in AI infrastructure, and were willing to lend their endorsement, their land and sometimes their subsidies. A founder who could stand next to a minister at a launch acquired a legitimacy that no diligence could easily puncture.

The third was the power gap. Building a real cluster requires megawatts that, in much of the region, simply are not available on the timeline investors wanted. A founder who could not get power had a ready-made excuse for why the chips were not yet racked — and a ready-made cover for the chips that were never going to be.

When you cannot tell the difference between 'my power substation is delayed' and 'I sold the chips,' you have a market that rewards the wrong people, said a datacentre operator in Johor who hosts capacity for several of these firms.

Where the chips actually went

Tracing physical hardware is the hardest part of an investigation like this, because chips move through intermediaries designed to obscure exactly that. But the broad flows are visible in the gap between announced fleets and installed capacity. BriefAsia compiled announced accelerator figures from nine regional 'AI cloud' companies and compared them against the power those firms had actually contracted.

The mismatch was stark. The nine companies had announced fleets that, fully racked, would require power well beyond what any of them had secured — in aggregate, several times the contracted megawattage. Either the chips were not where the companies implied, or they could not be switched on. Both, the evidence suggests, are true in different cases.

Some of the missing capacity is genuinely delayed: real chips, in transit or awaiting power, that will eventually run. But sources across the supply chain were consistent that a substantial slice has been resold — into other regions, to hyperscalers topping up, to buyers willing to pay the shortage premium. The region's 'AI clouds' have, in part, functioned as a conduit moving scarce silicon to wherever it commands the highest price, which is often not Southeast Asia at all.

The customers caught in between

The people who pay for this are the startups that believed the pitch. BriefAsia spoke with four founders who signed capacity contracts with regional clouds and then could not get the compute they were promised. One, a Hanoi-based drug-discovery startup, restructured its entire research plan around an allocation that arrived months late and at a fraction of the contracted size.

'We raised our seed round on the assumption we had compute locked,' the founder said. 'When it did not show up, we were not a startup with a delayed vendor. We were a startup with no product and a burn rate.' The company survived by renting, at a premium, from a foreign cloud — exactly the dependency the local provider was supposed to end.

This is the quiet damage. Every founder burned by a ghost cloud becomes a harder sell for the legitimate operators who are genuinely building durable capacity. Arbitrage poisons the well for infrastructure, because both wear the same costume and only time tells them apart.

The legitimate builders pushing back

Not every regional AI cloud is hollow, and the serious operators are increasingly vocal about the distinction. A Singapore-based provider that has actually racked and powered several thousand accelerators has begun publishing utilisation figures and offering on-site audits to large customers — a transparency play aimed squarely at separating itself from the arbitrageurs.

Its founder argued that the shakeout is coming regardless. As the chip shortage eases with new HBM supply and fresh accelerator generations, the arbitrage spread narrows. The trade that funded the ghost clouds gets less profitable, and the companies that never built anything real are left holding depreciating commitments with no operating business underneath.

'The shortage created these companies and the end of the shortage will kill them,' he said. 'The question is how much investor and government money goes down with them, and how many real startups they strand on the way out.'

What the regulators see

Regulators are starting to notice, though their tools are blunt. Securities authorities in at least one jurisdiction have begun asking AI-infrastructure startups to substantiate 'installed' versus 'committed' capacity claims in fundraising materials. Export-control officials, separately, are scrutinising the resale flows for compliance, since chips bound for one approved destination cannot simply be redirected to another.

But the core problem is not obviously a violation of any single rule. It is a disclosure and incentive failure that sits in the seams between securities law, export control and ordinary commercial discretion. A company can truthfully say it has 'access to' 12,000 accelerators while installing 1,500 and trading the rest, and no single regulator owns the gap between those numbers.

That is what makes the ghost clouds a policy problem rather than merely a cautionary tale for venture capitalists. The region's compute ambitions are being delegated, in part, to a class of company whose actual business is harder to verify than anyone admits.

The reckoning ahead

The boom is not over, but its premise is eroding. New memory supply from Korea is loosening the accelerator shortage that made the arbitrage so lucrative. Several regional clouds that raised aggressively in the past two years now face the same test at once: deliver real, utilised capacity, or be exposed as positions in a trade that is no longer paying.

For the founders who built genuine infrastructure, the shakeout cannot come fast enough. For the investors and governments who funded the costumes, it will arrive as a series of quiet write-downs and unbuilt clusters, dressed up as 'power delays' and 'strategic pivots' for as long as the story holds.

Meridian, for its part, says it is now 'doubling down on owned capacity' and has signed a new power agreement to rack delivered hardware it previously could not switch on. Whether that is a turn toward building something real, or simply the next chapter of a story written to keep the money in, is the question its investors should be asking — and the one this region's compute strategy now hangs on.