Nvidia Just Funded a $668M Bet on Taiwan's AI Supply Chain

GMI Cloud $668M FAQ

What is GMI Cloud? An AI-native GPU cloud headquartered in Mountain View, California, built inside Taiwan's server supply chain. Founder and CEO Alex Yeh — a former venture capitalist with a materials-science degree from Johns Hopkins — pivoted the company from Bitcoin-mining data centers (Arkansas and Texas) to AI and GPU cloud in 2023. Its products: GPU Compute (on-demand and reserved Nvidia H100 and H200 nodes wired with InfiniBand), a Cluster Engine for GPU scheduling built by engineers hired from Google X, Alibaba Cloud and Supermicro, and a usage-billed Inference Engine.

How much did GMI Cloud raise, and how is it structured? $668 million total: $223 million in Series B equity led by ARCHIV — a San Francisco firm specializing in AI and robotics — with Nvidia participating alongside DSC Investment, Trend Micro, KB Investment, Kyobo Life Insurance and KT Corporation; plus a $445 million credit facility led by Taiwan's CTBC. Equity makes headlines; the debt builds data centers.

What is the Taiwan AI Factory? GMI Cloud's $500 million build with Nvidia, announced in November 2025: 7,000 Blackwell Ultra GPUs across 96 high-density GB300 NVL72 racks, 16 megawatts of power, rated at up to two million tokens per second. Anchor users include Wistron, Chunghwa Telecom and Trend Micro — one of the largest AI infrastructure commitments in Asia outside the hyperscalers.

What is the $12 billion Japan project? A sovereign-AI initiative in Kagoshima announced at Nvidia's GTC in March 2026: a $12 billion AI factory targeting a full gigawatt of power, developed with Wistron and a local joint venture called Kai Shin Digital Infrastructure, backed by the Kagoshima prefectural government and Satsumasendai City. Designed specifically for physical AI — robotics, autonomous vehicles, industrial automation.

Eighteen months ago, $668 million for a GPU cloud nobody had heard of would have been a typo. This morning it was a press release.

GMI Cloud — Mountain View headquarters, Taiwanese supply chain, founder who used to run Bitcoin mines — announced a $668 million raise: $223 million in Series B equity led by a San Francisco firm called ARCHIV that most of Silicon Valley couldn't pick out of a lineup, plus a $445 million credit facility from Taiwan's CTBC. Nvidia participated in the equity. That's the sentence that matters. Nvidia doesn't write Series B checks for fun. Nvidia has been seeding the AI developer stack all month, including its Hugging Face deal

Here's what I keep coming back to. The GPU cloud business has a pitch so boring it barely qualifies as a pitch: buy chips, rent them out, repeat. The reality is less boring, and it's where the entire game is played — getting the chips on time. GMI Cloud's wager, and apparently ARCHIV's too, is that living inside Taiwan's server supply chain is a better moat than anything CoreWeave or Nebius can pour concrete for in Virginia. Contracted revenue up more than 9x in nine months suggests someone is buying it.

The structure tells you what the money is actually for

Equity rounds make headlines. Credit facilities build data centers. Of the $668 million, only $223 million is equity — the Series B, led by ARCHIV, which the release describes, with admirable vagueness, as "a new investment firm based in San Francisco that specializes in AI and robotics." The other $445 million is debt. That split is the whole story in miniature: this is a construction company that happens to rent out GPUs.

GPU clusters are a capex business. You borrow against the revenue the hardware will generate, you deploy fast, you pay it back. CoreWeave practically wrote the playbook — we covered its own wall of convertibles and debt facilities weeks ago — and Nebius has been raising GPU rental prices because demand keeps outrunning supply. GMI Cloud is running the same playbook. Except its banker is in Taipei.

The equity cap table is the more interesting document. Alongside ARCHIV and Nvidia: DSC Investment, KB Investment, Kyobo Life Insurance, KT Corporation — Korean money, heavily — plus Taiwan's Trend Micro, which is both an investor and a named customer. Korean financial press noticed. Chosun Biz, the only genuinely independent newsroom to report the round today, framed it as Korean capital following Nvidia into the GPU value chain. The logic, laid out plainly: US hyperscalers need low-latency capacity in Asia; Asian enterprises often can't move data offshore and need production-grade AI infrastructure close to home. GMI Cloud wants to sell to both sides of that sentence. Crusoe just killed its own $125B boom running a version of this debt playbook

The numbers that need a footnote

Let's be honest about the release, because the release isn't being honest with you. The sub-headline claims contracted ARR "tops $600 million." That figure appears nowhere except the company's own wire copy. We looked. So treat it as a company claim, not a confirmed fact.

The growth ratesare better grounded. Contracted ARR up more than 9x since the end of 2025. Live production ARR up more than 4.5x. An inference platform processing roughly 4 trillion tokens a week. The ARR multiples were echoed by Chosun Biz's independent reporting; the token figure is the company's own, so keep a hand on your wallet there. No valuation was disclosed anywhere we could find — and when a company raises $668 million and won't talk valuation, that's a data point too.

One more footnote, because the sources can't agree: the company's founding year. TechCrunch's Series A coverage and Korean company profiles point to different years in the early 2020s. We say "early 2020s" and move on. It's a $668 million company; the founding year is trivia.

A cloud built for the supply chain, not the servers

GMI Cloud's product story is three products wearing one name: GPU Compute (on-demand and reserved Nvidia H100 and H200 nodes wired with InfiniBand), a Cluster Engine for GPU scheduling built by engineers the company pulled from Google X, Alibaba Cloud and Supermicro, and an Inference Engine that bills by usage. None of that is exotic. The exotic part is the address.

"In AI infrastructure, a delivery date is a promise. Customers plan launches, hiring, and revenue around it. Our place in Taiwan's supply chain is how we keep that promise, cluster after cluster globally," Alex Yeh, the company's founder and CEO, said in the announcement. It reads like marketing — until you remember that every neocloud CEO in America has spent the last two years apologizing for missed deployment dates. Taiwan builds essentially all of the world's AI servers. Being close to the factories in Taipei beats being close to the customers in Texas, at least for the part of the business that involves hardware actually arriving.

The proof point is the Taiwan AI Factory, announced in November 2025: a $500 million build with Nvidia, 7,000 Blackwell Ultra GPUs across 96 high-density GB300 NVL72 racks, 16 megawatts of power, rated at up to two million tokens per second. Anchor users include Wistron, Chunghwa Telecom and Trend Micro. It's one of the largest AI infrastructure commitments in Asia outside the hyperscalers. Assembled by a company that, two years earlier, most people in the industry couldn't have named.

The customer roster in today's release runs through the usual AI-infrastructure middle class: Fireworks, Higgsfield, Nous Research, OpenRouter, Reflection, Cartesia, Utopai Studios. No independent outlet has confirmed the list, so read it as the company's version. But one named customer put a quote on the record, and it's the most revealing line in the whole release: "Capacity that arrives late is capacity we can't use," said Chenyu Zhao, co-founder of Fireworks. "GMI Cloud has been one of our strongest and most reliable providers across Nvidia GB200 and GB300 NVL72 systems." When your testimonials are about showing up on time rather than about technology, that tells you what the differentiator really is. It's logistics. It's always logistics.

The $12 billion moonshot in Kagoshima

If the Taiwan AI Factory is the proof of concept, the Japan project is the moonshot. In March 2026, during Nvidia's GTC, GMI Cloud announced a sovereign AI initiative in Japan: a $12 billion AI factory in Kagoshima targeting a full gigawatt of power, developed with Wistron and a local joint venture called Kai Shin Digital Infrastructure, backed by the Kagoshima prefectural government and Satsumasendai City. Construction is slated to start late this year. It's designed specifically for physical AI — robotics, autonomous vehicles, industrial automation — which would make it the first gigawatt-scale facility purpose-built for that workload instead of chatbots.

"Japan has built some of the world's most sophisticated industrial and manufacturing systems. The next frontier is ensuring those systems are powered by AI that Japan owns, controls, and can trust," Yeh said at the time, in remarks carried by Data Center Dynamics. The public-private structure — prefectural government involvement, domestic capital, a foreign AI cloud as operator — is the sovereign-AI model in practice. Japan's version just happens to be the biggest one with an actual groundbreaking date.

This is where the $445 million credit facility starts to look small rather than large. Twelve billion dollars of sovereign ambition does not get funded by one Series B. What today's raise buys is the balance sheet to keep winning the intermediate steps — the US, Taiwan and Southeast Asia capacity the release says the money will fund — while Kagoshima assembles its own, much larger financing.

From Bitcoin mines to GB300 racks

The origin story would get laughed out of a pitch meeting if it weren't already working. GMI Cloud started in the early 2020s as a Bitcoin-mining data-center operator, running facilities in Arkansas and Texas. In 2023, Yeh pivoted the entire company to AI and GPU cloud — swapping crypto's boom-bust for the AI buildout's bottomless demand. We checked the paper trail: TechCrunch covered the $82 million Series A in October 2024 (a $15 million equity slice plus $67 million in debt, led by Headline Asia with Thailand's Banpu Next and Taiwan's Wistron participating), when the company had about 35 staff in Asia, 18 in the US, and data centers in Taiwan, Thailand and Malaysia.

Yeh is an odd founder for this business, and I mean that as a compliment. He's a former venture capitalist — Infinity Ventures Crypto, HOF Capital — with a materials-science degree from Johns Hopkins, not a systems engineer. It shows in how the company talks: less about FLOPS, more about financing structures and supply chains. It shows in the cap table. The company that once raised from Thai energy firms and Taiwanese electronics manufacturers now counts Nvidia and a San Francisco AI fund as backers. The pivot, at least on paper, is complete.

Our read: the thesis is unromantic, which is why it might work

Here's my honest assessment, and I'll put it bluntly: GMI Cloud's pitch is refreshingly free of magic. No software breakthrough claimed. No secret sauce in the orchestration layer. No "AI-native" fairy dust. The claim is that it can get GPU clusters delivered and running on schedule because it lives inside the supply chain that manufactures them — and that everything else in this market is a commodity.

That's a defensible thesis. It's also, notably, a thesis that doesn't require you to believe a single unverifiable product claim. The Taiwan AI Factory exists. The Kagoshima project has a prefectural government attached to it. Nvidia's cloud-partner roster is short, and GMI Cloud is on it. You don't need to take anyone's word for the interesting parts.

The shoddy parts are the parts the company chose not to let anyone verify. No valuation. A $600 million contracted-ARR figure that lives only in the company's own press release. A customer list no independent reporter has confirmed. Growth multiples that are real and impressive but measured from an undisclosed base — 9x of what, exactly? None of this is disqualifying; every growth-stage company shades its numbers. But the right posture toward this announcement is impressed, not convinced.

The tell, as always, is who else wrote a check. Nvidia does not participate in Series B rounds for fun — its cloud-partner ecosystem is the distribution channel for every GPU it can't sell directly, and backing a Taiwan-connected neocloud hedges against the possibility that the American neoclouds can't build fast enough. The Korean insurers and KT aren't making a technology bet either. They're making a geography bet: that the AI buildout's center of gravity is shifting east. If they're right, $668 million will look like the moment GMI Cloud stopped being a curiosity and started being infrastructure.

Meanwhile the custom-silicon side of the buildout keeps moving — SemiFive just landed $52 million to build custom AI chips in the US — a useful reminder that everyone in this market is racing the same clock. Get capacity online before the demand curve does something embarrassing to your forecast.

Sources: PR Newswire (GMI Cloud announcement, Oct 1, 2026); Chosun Biz (Oct 1, 2026); Data Center Dynamics (Japan sovereign AI initiative, March 2026); TechCrunch (Series A coverage, Oct 2024); YesPress company and founder profiles.

By Daniel Rao, TechPulse USA — covering the companies, models, and money moves shaping artificial intelligence. We read the fine print so you don't have to.

File photo: a modern data-center aisle. GMI Cloud is spending its $668M raise on GPU capacity across the US, Taiwan and Southeast Asia. Photo: PiDatacenters / Wikimedia Commons (CC BY-SA 4.0).