While hyperscalers built clouds for the internet, SoftBank is building one for the agent economy, and it's betting 10 gigawatts that compute will be the new oil.
The Summary
- SoftBank Group and its telecom unit will launch AI cloud services in the US next fiscal year, targeting the enterprise AI compute market dominated by CoreWeave and Nebius
- The play: leverage a growing pipeline of data centers at 10-gigawatt scale to compete in the rental compute market where demand is outstripping supply
- This positions SoftBank as infrastructure for Web4: not just funding AI companies, but powering the agents they're building
The Signal
SoftBank is making a category shift. For years, the company placed bets on AI companies through Vision Fund checks. Now it's moving down the stack to become the compute layer those companies run on. This isn't just vertical integration. It's a recognition that in the agent economy, compute access is the chokepoint.
The 10-gigawatt number tells the story. For context, a typical data center runs on 20-50 megawatts. SoftBank is planning for 200-500x that scale. That's infrastructure for a world where every company runs dozens of agents, where training runs happen continuously, where inference costs drop low enough that AI becomes ambient.
"SoftBank is betting that enterprise AI demand will grow faster than AWS, Azure, and Google can expand capacity."
The competitive set is telling: CoreWeave and Nebius, not the hyperscalers. These are the pure-play GPU cloud providers that emerged because the big three couldn't move fast enough. CoreWeave went from zero to $19 billion valuation in three years by doing one thing well: renting H100s to companies that couldn't get capacity from the majors. Nebius spun out of Yandex specifically to chase this market. SoftBank sees the same opening.
The timing matters. Next fiscal year means operations start April 2027 at the latest. That's when the current GPU shortage should be easing, when Blackwell and whatever comes after are in volume production, when the first wave of companies have figured out which AI workloads are worth running and which were science projects. SoftBank isn't rushing to catch the hype cycle. It's positioning for the deployment phase.
Key advantages SoftBank brings:
- Capital to build at hyperscale while smaller providers scramble for funding
- Telecom infrastructure from the SoftBank unit for low-latency connections
- Existing relationships with enterprise customers through Vision Fund portfolio companies
The business model shift is the real story. SoftBank built its reputation on big, risky equity bets. WeWork, Uber, Arm. Some worked, many didn't. Compute rental is the opposite: predictable revenue, long-term contracts, margins that compress over time but never collapse. You're selling something customers absolutely need and can't easily replace.
The Implication
Watch for SoftBank to bundle this. The play isn't just "rent us your GPUs." It's "we'll fund your Series B and give you priority compute access." That's a flywheel the pure infrastructure players can't match. If you're building agents or running serious AI workloads, you now have a fourth option beyond AWS/Azure/GCP or the GPU specialists. One backed by deeper pockets than almost anyone and a track record of thinking bigger than the market thinks is possible.
For enterprises, this is validation that the agent economy is real enough for infrastructure plays at 10-gigawatt scale. Plan accordingly.