GPUs as Collateral, Round Two: GMI Cloud Raises $668 Million With Nvidia at the Table

The most interesting number in AI infrastructure this week is not a parameter count. It is $668 million — the amount that GPU cloud provider GMI Cloud has just raised, according to an exclusive from The Information — because of how the money is structured. The round combines fresh equity with Nvidia among the investors and roughly $445 million in credit led by Taiwan’s CTBC Bank, extending an experiment that is quietly rewriting the rules of data-center finance: lending against GPUs and the contracts written on them, rather than against real estate or corporate balance sheets.

What was announced

Per The Information’s briefing, published September 30, GMI Cloud raised $668 million from Nvidia and others. The company is described as one of seven firms that build facilities entirely based on Nvidia’s reference stack — the tightly integrated compute, networking, and software platform that Nvidia certifies through its Nvidia Cloud Partner (NCP) program. Reporting syndicated through Reuters and picked up across financial media identifies the credit component as approximately $445 million led by CTBC, the Taiwanese bank that has anchored GMI’s previous GPU-backed facilities. By subtraction, the equity slice of the package is on the order of $223 million — a major step up for a company whose entire recorded fundraising history stood near $93 million before this year’s debt deals.

Neither the full investor list nor a valuation has been disclosed. But Nvidia’s direct participation matters beyond the dollars. When the company that makes the chips also holds equity in the company that operates them, the chipmaker’s incentives align with capacity actually getting deployed, filled, and renewed — a flywheel Nvidia has replicated across the “neocloud” landscape as it battles to keep its architecture the default for AI workloads.

The company behind the number

GMI Cloud is a San Francisco-headquartered, AI-native GPU cloud provider. Its profile rose in May 2025 when it was named a Reference Platform Nvidia Cloud Partner — a specialization Nvidia reserves for clouds meeting its highest bar for performance, security, and full-stack operation. Its earlier funding was modest: an $82 million Series A announced in late 2024, led by Headline Asia with participation from Banpu Next and Wistron, split between $15 million of equity and $67 million of debt. Wistron’s presence was a tell — the Taiwanese ODM’s involvement foreshadowed GMI’s pivot toward island-based infrastructure.

That pivot became concrete in November 2025, when GMI unveiled plans for a $500 million “AI Factory” in Taiwan: a 16-megawatt facility designed to house roughly 7,000 Blackwell Ultra GPUs arranged across 96 GB300 NVL72 racks, targeted to come online in March 2026. The project was pitched as sovereign, high-density compute for Asian enterprises — bringing frontier-class training and inference capacity onto Taiwanese soil rather than routing it through US or Singapore regions. In July 2026, GMI announced a further $500 million commitment to expand infrastructure for frontier AI customers, explicitly structured around selling Nvidia-powered cloud services.

The real story: GPUs as loan collateral

The $668 million headline is best read as the third act of a financial innovation GMI has been pioneering since mid-2026.

In July, Bloomberg reported that GMI was seeking a NT$20.45 billion (about $635 million) multi-tranche loan backed not by land or buildings but by its customer contracts for GPU capacity — the signed commitments of AI companies to rent compute. International Financing Review later detailed the structure: a five-year term loan of NT$14.05 billion (about $438 million) with CTBC Bank as mandated lead arranger, a NT$6.4 billion twelve-month bridge, and a NT$150 million five-year revolving facility, all wrapped around total server capital expenditure of roughly $546 million. It was the first GPU syndicated loan in APAC, priced at Taipei interbank rates plus 175 basis points — roughly 3.4 percent at the time, strikingly cheap for a startup-friendly facility.

Demand surprised even the arrangers. By early September, GMI had drawn $947 million in total loan commitments, per Tech in Asia — oversubscribed to the point that the deal landed in the 97th percentile of all-time debt raises of its type, according to Dealroom’s tracking. Malaysia’s CTBC-coordinated syndicate found strong appetite among Asian banks hungry for AI-linked yield with tangible collateral attached.

Then, in mid-September, Reuters partners reported GMI was lining up a separate $300 million facility to buy chips for a data center in Thailand — with yields in the low teens and, notably, Tencent reportedly set to consume the capacity.

Today’s $668 million package — equity plus a further $445 million CTBC-led credit line — extends that playbook. GMI has effectively demonstrated a repeatable template: sign hyperscale-class customers to GPU contracts, borrow against those contracts at investment-grade-adjacent rates, buy Nvidia silicon, deploy it in jurisdiction-advantaged locations, and repeat. Nvidia’s equity check at the center of the new round adds a supplier’s balance-sheet vote of confidence to a structure that already had banks convinced.

Why it matters

It validates a new class of AI debt. Goldman Sachs recently tallied over $500 billion in AI-linked debt issuance — a figure this blog covered yesterday. GMI’s deals show the debt is no longer confined to giant balance sheets like OpenAI’s or Meta’s. Mid-sized neoclouds can now tap syndicated bank markets by collateralizing GPUs and their revenue contracts, at margins once reserved for utility infrastructure. That broadens who can build AI capacity — and who is exposed if demand disappoints.

It hardens Nvidia’s ecosystem lock-in. GMI is one of only seven firms building entirely on Nvidia’s reference architecture. Every financing round Nvidia co-invests in deepens that alignment: the cloud partner gets cheaper capital and chip allocation priority; Nvidia gets guaranteed absorption of its latest Blackwell-class systems. Rival accelerator makers must contend not just with CUDA’s software gravity but with a vendor that now participates directly in its customers’ capex.

It export-frames AI sovereignty for Asia. Taiwan’s first AI Factory and the planned Thailand site show GPU clouds being sited for data-residency, latency, and geopolitical reasons — sovereign compute for enterprises and (per reports) Chinese hyperscalers reaching around regional constraints. The geography of AI inference is being decided by where these financed facilities land.

It prices the risk. A 175-basis-point margin over Taibor for GPU-collateralized lending is a statement: Asian credit markets currently view contracted AI compute as roughly as safe as infrastructure lending gets. If AI demand stays tight, that view is vindicated; if it softens, the collateral — chips that depreciate fast and resale poorly in a downturn — will be stress-tested in ways real estate never was.

The open questions

No valuation was reported for the equity portion, which invites speculation about how neoclouds are being priced after CoreWeave’s public-market volatility. The $445 million credit line’s pricing, tenor, and covenants remain undisclosed. And GMI’s total commitments — $947 million in loans, $300 million sought for Thailand, $668 million announced today — now stack into the low billions against a company that was a $93 million startup two years ago. The leverage is the point of the model, and also its risk.

What is no longer in question is that “GPU-backed finance” has graduated from novelty to template. GMI Cloud just ran the playbook again, with the chipmaker itself now holding a seat at its cap table.