Firmus Technologies began in 2019 as a bitcoin mining operation in Tasmania. Last week it announced it would deploy 170,000 Nvidia GPUs across a 360-megawatt data centre campus in Batam, Indonesia, and expects to pull $25bn to $30bn in customer commitments over six years. The trajectory from crypto miner to $5.5bn AI infrastructure firm is sharp. The model is different from what South Korea just unveiled, and that difference is instructive.
South Korea bet on state coordination. The government announced $880bn, choreographed across Samsung, SK Hynix, LG and others, with ministerial support for power, land, workforce and housing. One nation, one signal, clear beneficiaries, ten-year horizon. The model is industrialisation by fiat. It works because South Korea has centralized industrial policy and a stable government-chaebol relationship.
Firmus and Nvidia are doing the opposite. They are using a financing structure that bypasses upfront state capital entirely. Nvidia is not selling chips to Firmus for cash. Instead, Nvidia takes a share of cloud revenue from the campus. Firmus does not own the infrastructure outright. It operates it, sells access to it, and splits the returns with Nvidia. Nvidia keeps hardware revenue plus cloud revenue. Firmus bears operational risk, captures the margin between what it buys and what it sells, and funds everything through a mix of equity raises and debt.
The beauty of this model is that it does not require a government announcement or ministerial coordination. It requires capital markets. Firmus raised $505m in April at $5.5bn post-money from Coatue Management and Nvidia. It closed a $10bn debt facility from Blackstone in February. The company is preparing an ASX IPO. None of that required government permission or a national industrial plan. The private markets provided the capital. The revenue-sharing with Nvidia provided the certainty. The committed customer agreements provide the cash flow.
Why Batam, not Seoul?
Location matters. Batam sits 26 kilometres from Singapore, a 50-minute ferry ride. It is close enough to tap Singapore's financial system, its regulatory infrastructure, its established tech talent, and its reputation. It is far enough away that it avoids Singapore's land scarcity and power costs. Indonesia is permissive on foreign investment and data flows. The island is industrial, not residential. Power is available from Indonesia's grid. Singapore's Green Data Centre Roadmap targets 300 megawatts of new capacity. Batam can absorb American AI companies and Asian startups without the friction of Western regulatory oversight.
Seoul is the opposite play. South Korea is saying: build the entire chain here. Memory chips, packaging, AI data centres, robotics, all domestic. Maintain technological sovereignty. Keep dollars inside. But the cost is centralization. The benefit is control. Seoul will become the hub. Batam is designed to diffuse AI infrastructure across Asia by making compute cheaper and access easier.
Who uses each model?
This matters. Firmus's Australian projects serve hyperscalers: Meta, Amazon Web Services, the mega-cloud operators. They sign long-term contracts. They demand 99.99% uptime. They have credit ratings. They are easy to finance because they are predictable.
The Batam campus is different. It targets AI-native companies, mid-market firms whose entire business is running on accelerated compute. They are startups, research labs, emerging AI companies. They cannot sign ten-year contracts. They don't have investment-grade balance sheets. They need pay-as-you-go access at a price they can afford. That is what Batam delivers. And it is why revenue-sharing works. Nvidia gets paid when the customer gets paid. Everyone's incentive aligns.
Distribution question
South Korea is betting that concentration creates competitiveness. Build everything at home, own the supply chain, control the ecosystem, dominate the market.
Firmus and Nvidia are betting that distribution creates access. Spread compute infrastructure across jurisdictions with low regulatory friction, cheap power, and permissive data rules. Make it cheaper for mid-market companies to train models. Fragment the market rather than consolidate it.
Both strategies could work. South Korea's model scales what worked in memory chips (domestic champions, government backing, supply-chain control. Firmus's model scales what works in cloud) distributed infrastructure, pay-as-you-go pricing, financial engineering instead of industrial policy.
The question for Western policymakers is which they want to compete with. Seoul's consolidation is easier to see. Batam's diffusion is harder to counter.
