The world’s supply of computing power is unlikely ever to catch up with AI demand, according to the boss of Nvidia partner Iren, who called the AI data centre build-out “fundamentally different” from past investment cycles.
Dozens of emerging AI data centre developers and Big Tech groups are racing to bring online new facilities packing together racks of thousands of Nvidia and Google AI processors.
Goldman Sachs expects the build-out to double US data centre capacity from 2024 levels by the end of 2027, and to more than triple it by 2030, reaching about 125GW — enough energy to power more than 100 cities the size of San Francisco.
Even as Iren prepares to spend up to $30bn on its AI ambitions over the coming year, Daniel Roberts, co-chief executive and co-founder of Iren, told the FT: “It’s really hard to see how the supply curve ever overtakes [the] demand side, really hard.”
A crypto miner turned AI infrastructure upstart, Iren is one of several Nvidia-backed “neocloud” companies, including Nebius, Nscale and CoreWeave, that are taking on Big Tech’s dominance of the AI cloud computing market. Nasdaq-listed Iren’s customers include Microsoft and Perplexity, as well as Nvidia, which is also a key supplier of chips to Iren.
“This industry is potentially unique, in the sense that every unit of supply you bring online actually feeds a multiple of that in additional demand, rather than solving the existing demand,” Roberts said. The growth of AI agents and faster processing times mean more consumption of computing resources, he explained.
“At risk of saying ‘this time is different’, I think there is something fundamentally different about this” from past commodity and investment cycles, he said.
Roberts added that the physical world was itself a check on oversupply, with builders already “hitting social, political, physical thresholds around the availability of power”. Many communities across the US are mounting opposition to data centre construction in their areas, citing concerns about the impact on local energy and water supplies.
Over the past three years, Roberts has transformed the company he founded as Iris Energy with his brother in Sydney in 2018. The brothers have turned it from a Bitcoin miner that had defaulted on most of its mining-rig loans and drawn a going-concern warning from its auditor into one of Nvidia’s important neocloud partners.
Iren’s market value has risen from a post-crypto crash low of about $60mn in late 2022 to roughly $17bn today; its shares have swung from about $1 to a high of $76.
Its volatile trajectory mirrors that of a number of neoclouds and AI infrastructure developers racing to build the AI industry’s physical backbone. They face huge rewards if the computing boom continues but also large risks, from heavy debt, a concentrated customer base and fast-depreciating chips, should supply outrun demand.
Iren is among the few AI infrastructure upstarts to own the entire stack, from the land and buildings to computing equipment and the software layer on top.
Owning the full stack, and largely forgoing the long-term leases rivals use to secure hyperscaler tenants and cheap debt, lets Iren benefit from rising compute prices. But it has also required greater share issuance, with its total share count now at about 394mn, up from 55mn at its initial public offering.
Roberts compared long-term leases his rivals have signed to sales by another name. “You’ve just given up the greatest, scarcest asset that probably exists in the world at the moment, and sold it to someone else to monetise. We just don’t think that’s very strategic.”
The company has still been able to raise huge sums. In the past 12 months, Iren has secured about $19bn, through issuing convertible notes, chip-backed debt, customer pre-payments and equity. Most of the proceeds are used to buy Nvidia processors and build the data centres to house them.
“It’s incredible the amount of capital that’s starting to mobilise,” said Roberts.
Iren this summer raised $3.65bn in chip-backed financing at 6 per cent, tied to a Microsoft offtake agreement, and last week secured a further $2.4bn from Blue Owl and funds managed by Pimco. That deal priced at 9 per cent, reflecting the weaker credit of its AI-lab customers, among them Cohere and Perplexity, than investment-grade Microsoft.
Both facilities are also secured against Nvidia processors, with the chipmaker’s deepening involvement as supplier, customer and would-be shareholder central to Iren’s transformation.
In May, Nvidia signed a five-year, $3.4bn contract to rent capacity from Iren for its own workloads and took rights to acquire up to $2.1bn of stock, at $70 a share, that vest as Iren takes delivery of up to 600,000 Nvidia chips.
Roberts called the structure “powerful” with a critical supplier tied to “a metric that defines [our] success — ie how big and how many GPUs do we get”. Graphics processing units are the Nvidia chips that have become the workhorses of the AI boom.
He dismissed the circular-financing concerns that analysts have raised about such Nvidia deals, arguing the chipmaker was strategically breaking down barriers to the growth of the AI ecosystem, and, in turn, its own business. “One of the levers they have is their balance sheet.”
Roberts said Iren talks to Nvidia almost daily, including on financial matters. “They continue to evolve how they look to support all aspects of the ecosystem — they continue to provide ideas.”
Iren ended June with $7.6bn of cash on its balance sheet and an equal amount of debt. It has $3.5bn left on a $6bn at-the-market share sale programme and a portfolio of data centres it has yet to mortgage as the group prepares to spend up to $30bn on AI infrastructure in the year to June 2027.
The real ramp, Roberts said in a recent X post, comes next year: “It’s delivery time.”
Copyright The Financial Times Limited 2023
© 2023 The Financial Times Ltd. All rights reserved. Please do not copy and paste FT articles and redistribute by email or post to the web.
