Taiwan's GMI Cloud stacks equity and bank debt in $668 million AI infrastructure raise

GMI Cloud, a Taiwan-founded GPU cloud provider, closed $668 million in combined financing in the last week of September, splitting the raise into $223 million of Series B equity and a separate $445 million credit facility. The structure, confirmed across company disclosures and three independent trade outlets, is less notable for its size than for who is on both sides of the ledger: Nvidia sits among the equity investors in a company whose entire business is reselling access to Nvidia chips, while the lead lender on the debt piece already has a credit position in the data center landlord hosting GMI's largest facility.

An equity-debt split built for a capital-intensive bet

The equity tranche was led by ARCHIV, a San Francisco investment firm, with participation from Nvidia, DSC Investment, Trend Micro, KB Investment, Kyobo Life, and KT Corporation. No valuation was disclosed, a gap all three outlets reviewed flagged explicitly rather than glossed over. GMI's Series A, by comparison, raised $82 million in November 2024 — meaning the step-up implied by $223 million in fresh equity sixteen months later cannot be assessed without a disclosed post-money figure, something investors evaluating comparable neocloud raises would normally expect as a baseline data point.

The debt tranche, $445 million, was led by Taiwan's CTBC Bank. None of the reporting disclosed the facility's interest rate, maturity, or covenants, and none described collateral terms beyond characterizing it as a credit facility tied to capacity expansion. That is a meaningful omission for a facility of this size: asset-backed lending in the neocloud sector has typically been structured against either hardware or contracted revenue, and without knowing which applies here, the facility's risk profile to CTBC — and by extension to any instruments CTBC later syndicates or securitizes — is not independently verifiable from public reporting.

The same lender, two sides of one campus

GMI's flagship Asian facility is a planned $500 million "AI factory" in Taoyuan, Taiwan, first announced in November 2025, intended to house roughly 7,000 Nvidia GB300 GPUs across 96 racks at about 16 megawatts of power. That facility sits inside a campus operated by Vantage Data Centers. CTBC, now GMI's lead lender, was also one of the banks — alongside Cathay United Bank — that financed Vantage's Taiwan greenfield hyperscale project when it closed in 2024. That earlier financing was reported as what Vantage and its partners called Taiwan's first greenfield hyperscale data center project financing transaction. The practical result: a single Taiwanese bank now carries credit exposure to both the landlord building the shell and the tenant filling it with GPUs on the same AI campus, a concentration that matters for assessing systemic risk in Taiwan's AI buildout even though neither financing discloses enough detail to size that exposure precisely.

Nvidia as supplier and shareholder

Nvidia's participation in GMI's equity round extends a pattern that has drawn scrutiny through 2026: the chipmaker has committed more than $40 billion in equity across AI companies this year, many of which are also customers buying its GPUs, drawing criticism from some analysts that the arrangement amounts to circular financing — demand that Nvidia itself helps fund. Nvidia chief executive Jensen Huang has publicly rejected that characterization as investment scrutiny of the structure has grown. GMI's round does not resolve that debate, but it does extend the pattern to a smaller, privately held neocloud rather than a frontier lab, suggesting the practice has moved beyond the largest, most visible AI companies into the vendor-financing layer underneath them.

Our customers are scaling faster than ever, and they need infrastructure that keeps pace, said Alex Yeh, founder and chief executive of GMI Cloud.

Growth claims outrun audited disclosure

GMI said contracted annual recurring revenue now exceeds $600 million, more than nine times its level at the end of 2025, with live production ARR up more than 4.5 times over the same period and roughly 4 trillion tokens processed weekly on its inference platform. These figures come from the company itself; none of the outlets that reported them cited independent verification, audited financials, or a breakdown of customer concentration. GMI's customer list — which includes Fireworks, Higgsfield, Nous Research, OpenRouter, Reflection, Cartesia, Trend Micro, and Utopai Studios — points to heavy reliance on other AI infrastructure and model companies rather than diversified enterprise demand, which raises the standard neocloud-sector question of how much of that contracted ARR is sourced from counterparties facing their own funding pressure.

Capacity that arrives late is capacity we can't use, said Chenyu Zhao, co-founder of Fireworks, a GMI customer.

GMI's broader expansion plans — a $12 billion sovereign AI infrastructure initiative announced in Japan in March 2026 and a partnership with Magna AI on a network of "sovereign AI factories" — suggest the company is positioning the new capital as a down payment on a far larger international build-out than the disclosed $668 million alone could finance. Earlier reporting had indicated GMI was separately seeking as much as $635 million in loans backed by customer contracts specifically for the Taoyuan project, a figure close enough to the disclosed $445 million credit facility that the two are plausibly the same financing, reported at different stages of negotiation. If so, it indicates the facility is backed by contracted revenue rather than hard assets — a structure that performs well only as long as GMI's own customer contracts hold, tying the credit facility's risk directly to the health of the smaller AI infrastructure and inference firms on GMI's customer list.