Building the AI Native Cloud: Alex Yeh on Power, Pricing, and Financing the GPU Buildout
Our second guest on Where Capital Meets Compute is Alex Yeh, founder and CEO of GMI Cloud. He started the company five years ago in Bitcoin mining, building it into the largest miner in Arkansas at over 100 megawatts of capacity, before pivoting to AI cloud as GPT-3 changed what the market needed. That mining background shapes how he thinks about infrastructure today: what kills hardware, what holds its value, and what it actually takes to secure power at scale.
Not a NeoCloud
Kony opened by framing GMI as a NeoCloud, alongside peers like CoreWeave, Nebius, and Lambda. Alex pushed back immediately.
I wouldn’t consider myself to be a NeoCloud, but an AI native cloud.
The distinction, in his telling, comes down to the stack. Most peers stop at bare metal, racks and power, nothing more. GMI builds the full vertical: bare metal, Kubernetes, APIs, and what Jensen Huang calls the “AI token factory,” plus agentic infrastructure and CPU sandboxes. Built by developers for developers, not just an energy company that bought GPUs.
The moat: reach, supply chain, capital
Asked what stops competitors from replicating GMI, Alex named three things. First, geography: GMI is the only AI native cloud operating at scale across both the US and Asia, while most peers are US- or Europe-centric. Second, supply chain, four decades of family relationships in Taiwan’s semiconductor ecosystem let GMI get GPUs and full data center systems delivered on time, in an industry where delays are common. Third, capital cost: borrowing through Taiwan and Japan relationships lets GMI raise debt at roughly half of what public peers pay.
The same relationships underpin what he calls the Marriott model. GMI can act as full owner-operator, or partner with asset owners, power owners, or bare-metal operators who lack management software, wrapping their infrastructure in GMI’s brand and tooling.
The bottleneck has shifted to power
GMI’s capacity is sold out through Q1 and Q2, and increasingly through Q4 of next year. But the constraint, Alex says, isn’t chips.
It’s not availability of power, but power in deliverable form, transformers, diesel generators, gas turbines, transmission studies, regulatory hurdles.
He’s seen this before, in Bitcoin mining, but at a different scale. Institutional capital like Blackstone, KKR, BlackRock is now competing for the same land and power that used to be contested only among miners.
Why old GPUs don’t lose value
The Bitcoin story trained the market to expect specialized hardware to go obsolete fast. Alex argues GPUs don’t follow that pattern. A100s, launched in 2019, still can’t be booked in volume on AWS or GCP, seven years past a conventional five-year depreciation schedule.
At worst, it’s worth higher.
His explanation is model diversity: as newer, larger models push demand toward the newest chips, they also spawn smaller, distilled models that run efficiently on older hardware. Every generation keeps finding new jobs to do, rather than being retired outright.
What lenders actually look at
Unlike the leveraged Bitcoin miners who got wiped out when prices turned, Alex describes GPU-backed lending today as heavily de-risked. Deals over $100 million are underwritten by private credit funds, PE firms, and banks who run full technical and legal due diligence, and are typically backed by investment-grade offtakers or ring-fenced SPV structures.
It’s not us that’s the risk, it’s on the contract, on the cash flow, and the GPUs itself.
What lenders actually worry about, in his account, isn’t the chip going obsolete. It’s contract terms: can a hyperscaler walk away after a year, what SLAs are promised, and if the offtaker isn’t investment-grade, what is their underlying credit.
Bubble, or not?
With data centers across Asia sold out, Kony raised the obvious comparison to 1999 telecom, which was also booked solid before demand collapsed. Alex’s answer wasn’t about bookings, it was about usage.
Look at the actual traffic. It’s zero for telecom, but our GPUs are running at 98%. Can you actually live your life without AI now?
Kony’s counter-argument ran the other way: most industries and individuals haven’t adopted AI at all yet, which suggests more demand for headroom, not less.
What’s still missing
Asked what would unlock a genuinely open capital market for AI infrastructure, something Jensen Huang has called for, and what Kony is building toward, Alex’s answer was standardization. Chips like Trainium or TPUs can’t easily be borrowed against, because too few operators know how to run them. NVIDIA GPUs, by contrast, are fungible enough that GMI can take over someone else’s servers and run them immediately, which is what makes them financeable at scale.
It’s the same underlying idea Carmen Li described in Episode 1 with Silicon Data’s pricing indices, different corners of the market, same missing piece.
Listen to the full episode of Where Capital Meets Compute for the complete conversation.