GMI Cloud raises $668m for AI capacity, two-thirds of it as debt

GMI Cloud has raised $668 million, two-thirds of it debt: $223 million of Series B equity led by ARCHIV and a $445 million credit facility led by CTBC, with Nvidia, DSC Investment, Trend Micro, KB Investment, Kyobo Life and KT Corporation named among the other participants.

That participant list bears reading twice, because it puts a chip designer, a security software company that also appears on GMI's customer roster, two Korean financial institutions and a Korean telecom into the same capital structure as a bank-led facility. The announcement does not break out individual commitments, so how the equity and the credit divide among those names is not public.

GMI says the proceeds fund capacity expansion in the US, Taiwan and the wider APAC region, development of its inference services, and hiring. Yeh framed the raise around customer demand: "Our customers are scaling faster than ever, and they need infrastructure that keeps pace. AI is driving a new renaissance, and reliable compute is its foundation. Our goal is to build that foundation across continents, with an ecosystem of products on top of it."

Founded in 2023, GMI operates data centers in the US, Taiwan, Singapore, Thailand and Japan; it secured an $82 million Series A in November 2024, launched a $12 billion sovereign AI infrastructure initiative in Japan in March, and is working with Magna AI on what the pair describe as a global network of sovereign AI factories. Set against that Series A, the $223 million equity tranche is roughly 2.7 times the earlier round; the credit facility has no comparable predecessor, though reports earlier this year said GMI was hunting larger debt.

The clearest window on where the money goes is Taoyuan, where GMI announced a $500 million data center in November 2025 aiming to deploy around 7,000 Nvidia GB300 GPUs across 96 racks at roughly 16MW. GMI will house those racks inside a Vantage Data Centers facility, placing the company on the tenant side of a wholesale lease rather than the owner's side of the shell — the arrangement a deployer picks when the constraint is time rather than real estate.

Two ratios follow from those numbers. Seven thousand accelerators across 96 racks translates to roughly 72 per rack, the density of rack-scale systems rather than a shelf of individual servers; 16MW across the same rack count is about 167 kilowatts apiece, below the 230-kilowatt rack class Siemens Grid Software has been telling utilities to plan around but still deep into the range where the substation and the switchgear, not the lease, set the delivery date.

The site also carries a financing history: the raise lands a few months after GMI was reportedly seeking up to $635 million in loans backed by its customer contracts, financing that reporting tied specifically to the Taoyuan AI factory project. Whether the CTBC-led facility is that transaction is not disclosed, and at $445 million it is the smaller number; whether this credit is secured by offtake, and which offtake, is the detail that determines what the liability actually is.

GMI's customer list is why the security question matters: Fireworks, Higgsfield, Nous Research, OpenRouter, Reflection, Cartesia, Trend Micro and Utopai Studios, mostly AI-native companies, several of them venture-funded model developers and inference providers whose own revenue is young, and one of them simultaneously an investor in the raise. A loan secured by contracts with that roster is a bet on the counterparties' continued financing as much as on GMI's operations. That is why tenant commitments keep showing up at the center of AI capacity finance: the customer's promise carries more weight than the developer's balance sheet. Nebius's 50MW, 12-year capacity deal with AIB Data Centers will be funded by tenant prepayments alongside project debt and equity at a site that already holds a 65MW electric service agreement.

Nvidia's presence among the participants is the other thread worth pulling: in the SB Energy case, the argument was that the chipmaker's guarantee, more than its equity, is what would make a $500 billion Ohio campus financeable, and here the announcement lists Nvidia as a participant without saying what it contributed or whether its money sits in the equity or the credit facility. For a company whose Taoyuan deployment is built on Nvidia's GB300 parts, that distinction is not academic.

In the broader market, assets anchored by hyperscaler or frontier-lab leases get priced like utilities, with long contracts and debt that banks compete to lead, while everything else assembles capital from strategics who want the compute and lenders who want the security, and pays a spread for the privilege. At $223 million of equity against a $500 million Taoyuan build and a Japan initiative whose capital commitment is not publicly quantified, what GMI raises next — and in what form — is the number to watch.

This was a priced event in the sense that matters for comparison: unlike the unpriced announcements of September, GMI put a number on the total. It did not put terms on the credit facility, or say what secures it; when those details surface, they will say more about what an AI-native customer base is worth to a lender than anything in the $668 million headline.