SoftBank Group is planning to rent AI computing capacity directly to US companies, according to The Information. The move transforms Masayoshi Son’s massive AI infrastructure investments — including the $500 billion Stargate project — from a capital expenditure into a commercial revenue stream. SoftBank isn’t just buying GPUs anymore. It’s becoming an AI compute landlord.
🔍 THE BOTTOM LINE
SoftBank is pivoting from AI infrastructure investor to AI infrastructure operator. The plan: rent GPU capacity from its Stargate data centers to US enterprises that need AI compute but can’t justify buying their own GPU clusters. It’s the classic cloud business model — but focused exclusively on AI workloads, and backed by the largest private data center buildout in history.
From Investor to Operator
SoftBank’s AI infrastructure strategy has evolved through three phases. Phase one was equity investment: buying stakes in Nvidia, Anthropic, and OpenAI. Phase two was the Stargate project — a $500 billion joint venture with OpenAI and Oracle to build AI data centers across the US. Phase three, now emerging, is commercialization: turning those data centers into a rentable service.
The Information reports that SoftBank is targeting US companies specifically, not just its existing portfolio companies. This positions it in direct competition with AWS, Google Cloud, and Microsoft Azure — but with a narrow focus on AI training and inference rather than general-purpose cloud computing.
The economics are compelling. An H100 GPU costs roughly $30,000-40,000. A cluster of 100,000 GPUs — what a frontier AI model requires — represents $3-4 billion in hardware alone. Most enterprises can’t justify that capital expenditure. SoftBank’s bet is that renting access to pre-built GPU clusters, managed and maintained by Stargate’s infrastructure team, will be cheaper than building in-house capacity.
The Stargate Connection
The Stargate project, announced in January 2025, is the physical backbone of this strategy. The joint venture between SoftBank, OpenAI, Oracle, and MGX plans to build 10 data centers across the US, each housing up to 500,000 GPUs. The first facility in Texas is already operational.
SoftBank’s existing compute arrangements with companies like SpaceX and Anthropic have been bilateral deals — bespoke infrastructure built for a single tenant. The new plan is different: multi-tenant AI compute, available to any US company willing to pay.
This is a significant shift. It means SoftBank is no longer just a capital partner for AI labs; it’s becoming an infrastructure provider in its own right. The competitive implications for the hyperscaler cloud market are real. AWS and Azure charge premium margins for GPU instances. A SoftBank-backed alternative with Stargate’s scale could pressure pricing across the entire AI cloud market.
Why US Companies Specifically?
The US focus isn’t accidental. American enterprises face the most acute AI compute shortage — demand from US companies for GPU capacity vastly exceeds domestic supply, and export controls have complicated access to alternatives. SoftBank is also navigating Japan’s own sovereign AI ambitions, but the revenue opportunity is in the US market where willingness to pay is highest.
There’s a geopolitical dimension too. The US government has been actively encouraging domestic AI infrastructure investment through the Stargate partnership. SoftBank renting compute to US companies aligns with that policy goal — keeping AI compute within US borders rather than offshoring it.
The Risk: Margin Compression and Competition
SoftBank’s plan faces real obstacles. AWS, Google Cloud, and Microsoft Azure are not passive competitors. They have existing enterprise relationships, mature billing and management platforms, and bundled services that SoftBank would need to replicate. A pure GPU rental business is also structurally lower-margin than a full cloud platform — it’s closer to a commodity play.
SoftBank’s recent financial pressures add another layer of risk. The company has committed enormous capital to AI infrastructure. If GPU prices decline — which they tend to as new generations arrive — the value of those investments erodes. Renting compute is how SoftBank recovers its capital expenditure, but the window to do so profitably may be narrowing as the hyperscalers expand their own GPU fleets.
NZ Angle
For New Zealand, SoftBank’s move highlights a structural gap. NZ companies needing AI compute currently rely on AWS Sydney or direct contracts with US cloud providers. A SoftBank GPU rental service doesn’t change that — the compute stays in the US. But if the model succeeds, it validates the idea of AI compute as a rentable utility, which could eventually support regional providers. The NZ Super Fund’s sovereign AI infrastructure push is betting on exactly this: that AI compute becomes something nations and institutions rent rather than own.
❓ FAQ
How is this different from AWS or Azure? SoftBank’s plan is narrower: AI-specific compute rental, not general-purpose cloud. AWS offers everything from databases to lambda functions. SoftBank would focus on GPU clusters for training and inference, potentially at lower cost by avoiding the full cloud platform overhead.
What’s the Stargate project’s current status? The first data center in Texas is operational. The full 10-site buildout is planned over several years. SoftBank’s compute rental plan would leverage these facilities as they come online.
Would this compete with OpenAI, SoftBank’s own partner? Potentially, yes. OpenAI is both a Stargate partner and a potential customer of the compute rental service. The relationship is complex — SoftBank is simultaneously an investor, infrastructure provider, and now potentially a competitor to OpenAI’s own cloud ambitions.
Is GPU rental a sustainable business? That depends on GPU pricing trends. If next-generation chips (Blackwell, Rubin) make current H100 fleets obsolete faster than expected, the residual value of Stargate’s hardware drops. SoftBank needs to earn back its investment before that depreciation curve accelerates.
🔍 THE BOTTOM LINE
SoftBank is making the most aggressive infrastructure bet in AI history, and now it needs that bet to generate revenue. Renting GPU capacity to US enterprises is the obvious commercialization path — but it puts SoftBank in direct competition with the three companies that already dominate cloud computing. The Stargate data centers give SoftBank physical scale that no startup could match. Whether that translates into a sustainable compute business, or just a very expensive way to buy AI exposure, depends on execution over the next 18 months.
