The recent sell-off in the AI sector has also affected Neocloud shares, such as $IREN (-9,38 %) —coinciding with the Kimi K3 announcement from China and the debate over whether the billions in investments by U.S. AI labs are still sustainable. A closer look, however, reveals that the trigger for the sell-off affects IREN’s business model significantly less than that of the model providers—it could even bolster it.
The background:
IREN rents out computing power instead of building models and diversifies its customer base across three levels—Microsoft as the anchor client ($9.7 billion, five years, 20% upfront payment; handover of the first Horizon 1 phase scheduled for mid-July according to the company’s plans), Nvidia as a cloud customer ($3.4 billion, plus a partnership for up to 5 GW, including an option to purchase 30 million shares at $70), and, looking ahead, enterprise and government customers with their own dedicated GPU clusters.
Why the trigger for the sell-off barely affects IREN’s business model
The Kimi-K3 shock targets the training side of the AI economy: the one-time, multi-billion training costs that U.S. labs must recoup through premium pricing. When China creates nearly equivalent models at a lower cost and Nvidia distributes its top-of-the-line models—such as Nemotron 3—for free, this very business model comes under pressure.
IREN’s business, however, increasingly relies on the other side: inference, i.e., the ongoing operation of the models. According to Deloitte, this will already account for about two-thirds of total AI computing power by 2026 (2023: one-third), and over the lifetime of a model, 80–90% of computing costs are attributable to operation, not training.
The key point here is that a free model requires exactly as many GPUs/computing power to run as a paid one. The more interchangeable and affordable the models are, the more they will be used—and the more valuable the scarce computing power required to run them tends to become. Added to this is data sovereignty: Companies are likely to increasingly want to avoid having sensitive data run through third-party model providers—especially when it comes to models from China, or even $AAPL (+4,07 %) compared to OpenAI.
Freely available models on self-rented, dedicated clusters solve exactly this problem, and IREN provides the infrastructure for it.
Opportunity: If the open-source trend takes hold, margins would shift from the model level to the infrastructure level. Furthermore, if direct enterprise sales are successful, this would mitigate the concentration risk associated with Microsoft—but it would come at the cost of sales force development and the financial security provided by long-term, prepaid hyperscaler contracts.
Risk – Second-Round Effect: If U.S. labs were to stretch out their training investments, this would eventually also affect demand for training clusters like the ones IREN builds for Microsoft. The inference wave cushions this impact but does not automatically replace it on a one-to-one basis.
Overall, the current decline in IREN’s stock price appears to be more of a sector-wide correction than a reassessment of its own business model: Doubts center on the training economics of model providers, while IREN’s value hinges on growing, model-independent inference demand.
Sources:
Deloitte via Computerworld – Inference Share in 2026: https://www.computerworld.com/article/4114579ces-2026-ai-compute-sees-a-shift-from-training-to-inference.html
Not investment advice.
