The “CoreWeave of Korea” may just be the latest canary in the AI coal mine, though US hyperscalers are still mining and breathing easy.
Shares of South Korean “neocloud” data center builder Elice plummeted more than 30% following its debut on the KOSDAQ exchange on Tuesday, another flashing red sign that investors are skeptical of AI-adjacent businesses squeezed between steep financing costs and slowing model revenue growth. On the other hand, seemingly nothing can stop hyperscalers’ data center buildout, except maybe access to energy.
Not a Neocloud in the Sky
Elice’s nosedive adds to neocloud woes and marks the perfect case study for others tiptoeing into the public market. Nvidia-backed Australian neocloud company Firmus aborted its IPO plans earlier this month amid skepticism over its $30 billion valuation (the abrupt cancellation has since drawn the attention of local regulators). Meanwhile, the similarly Nvidia-backed English neocloud company Nscale is now facing investor pushback to its own planned IPO valuation of $30 billion, Bloomberg reported on Tuesday.
The widespread neocloud wobble is hardly a surprise; providers of surplus compute have long been viewed by investors as the most vulnerable players in the AI ecosystem. “Everyone in the AI race is running for a last-man standing position,” John Lovelock, chief forecaster at Gartner, told CNBC earlier this week. “Pretty much every market that Gartner is tracking in our AI race is likely to collapse down to a duopoly.” The research firm had previously projected that 80% of independent neoclouds will be forced to cease operations by the end of the decade, thanks to an expected roughly $5 trillion gap between AI buildout expenses and AI revenue. The big league players, meanwhile, are still full steam ahead. New non-China data center capacity is expected to reach 51 gigawatts in 2027, according to new Semianalysis data, with three-quarters of projects already under construction. That’s roughly double the capacity expected to be added this year, with most coming from hyperscalers, who are more constrained by physical realities than compute crunches these days. Consider, for instance, Amazon:
- The e-commerce-turned-cloud-computing giant now has billions of dollars worth of Nvidia GPUs stored in intermediary warehouses as power supply constraints and other physical bottlenecks limit its ability to expand its data center footprint, sources told Business Insider in a story published Monday.
- In related news, Amazon on Monday announced a deal with industrial 3M that could see it spending as much as $6 billion on server connectors and other data center products.
Liftoff: Still, at least one company is finding success as an AI compute provider. In its second earnings report ever as a publicly traded company, SpaceX on Tuesday said revenue from its AI unit had increased ninefold year over year in the quarter to $7.6 billion, which came mostly from renting out compute capacity to Anthropic, Google and others. That helped the company, worth some $2 trillion, score its first ever quarterly profit of $4.4 billion.