On April 29, 2026, the Financial Times reported that SoftBank plans to create and publicly list a new AI and robotics company called Roze, with a target valuation of roughly $100 billion. The pitch is that robots will build the data centers that house artificial intelligence. The story is bigger than one IPO. It marks the moment the AI industry admitted its biggest bottleneck is not chips, models, or capital. It is concrete, copper, and skilled labor.
What SoftBank Actually Announced
According to the Financial Times via CNBC, SoftBank is preparing a US listing for Roze, an entity targeted to roll up the company's growing portfolio of AI infrastructure assets. The target valuation is approximately $100 billion. SoftBank executives are aiming for a second-half 2026 listing, though the timeline could slip into 2027 due to broader macro uncertainty.
Roze is structured as a vertical bet on the physical layer of the AI buildout. Per TechCrunch reporting, the planned bundle includes:
- ABB Robotics, the industrial robotics business SoftBank agreed to acquire in 2025
- Existing energy and land assets held inside SoftBank's broader portfolio
- Active data center projects, including a large campus already under way in Ohio
The mission, according to multiple Reuters and FT-sourced reports, is to deploy autonomous robots to make data center construction in the US faster and cheaper. Where peers buy GPUs and lease colocation, SoftBank is wiring together robots, real estate, and power so it can deliver finished gigawatts directly to AI buyers.
Why Q1 Earnings Set the Stage for Roze
The Roze story landed in the same week the largest AI infrastructure spenders reported Q1 2026 earnings, and the connection is not coincidental.
According to CNBC reporting on April 30, 2026, 2026 hyperscaler capex guidance was raised across the board after Q1 calls: Alphabet $185 billion, Amazon $200 billion, Meta $135 billion, and Microsoft $190 billion. That is roughly $710 billion in a single year from four companies. Bank of America and Evercore both pushed 2027 estimates above $1 trillion.
Goldman Sachs projects combined hyperscaler capex from 2025 through 2027 will reach $1.15 trillion, more than double the $477 billion deployed across 2022 through 2024. Capital is no longer the limiting reagent. Putting it into the ground is.
Our take: Investors used to evaluate AI infrastructure as a chips and software story. In 2026 the binding constraint has shifted to the physical layer. Land, transformers, switchgear, water, electricians, and time are the rate-limiting inputs. Roze is a financial-market expression of that reality, packaging the unsexy parts of the AI stack into something investors can own at scale.
The Construction Crisis Behind the AI Buildout
The bottleneck is concrete, both literally and figuratively. The Information Technology and Innovation Foundation reports that the US construction industry is short approximately 439,000 workers, with most of the gap in skilled trades like electricians and pipe layers. Around 85 percent of 2026 construction workforce demand is tied to data center projects. A single hyperscale campus can require roughly 4,000 workers during peak construction.
The math does not work without automation. Big Tech can commit hundreds of billions in capex, but you cannot pour foundations or bend conduit with money alone. CNBC has separately documented the AI-driven shortage of skilled trades workers, and Fortune's coverage of the AI data center electrician gap confirms the same pattern from a labor-market angle.
This is the gap Roze is designed to close. Robotics that can lay cable, install HVAC, place rebar, or commission high-voltage equipment turn a labor constraint into an automation problem. ABB Robotics gives SoftBank a credible starting point. Whether it works at hyperscale, on active job sites, in unionized markets, is unproven.
How This Reshapes the AI Vendor Map
For business buyers of AI, the Roze announcement has three operational implications.
Compute timing risk is now a procurement variable. A year ago, AI roadmaps could safely assume cloud regions, GPU instances, and dedicated capacity would be available roughly when needed. In 2026 that is no longer a safe assumption. We saw the pattern clearly in Anthropic's $100 billion AWS compute deal, where the binding contract variable was the gigawatts AWS could stand up over five years, not the price per GPU-hour. The Roze pitch confirms hyperscalers are willing to pay a premium for any incremental capacity. If you have AI workloads scheduled for 2027 launch, the question is no longer just which model to use. It is which region, which substation, and which delivery date.
Multi-cloud is becoming a hedge against physical scarcity. When OpenAI ended Microsoft exclusivity and added AWS Bedrock on April 30, the simple read was vendor diversification. The Roze story adds a complementary reading. Each cloud has different physical capacity, in different regions, on different timelines. Spreading workloads across clouds is no longer just a vendor risk control. It is a way to access the next available rack of accelerators when your preferred provider is queue-blocked.
Cost projections need a physical layer assumption. AI cost models that hold compute pricing flat through 2027 are likely wrong in both directions. If Roze and similar bets work, AI compute pricing softens as capacity unlocks faster than expected. If they fail, scarcity persists and pricing for premium GPU access stays elevated. Either path has a margin impact for AI-native products. Stress test both scenarios in your 2027 plan rather than carrying one assumption.
Why Roze Is Not a Sure Thing
Skepticism is warranted. Robots that lay rebar in a controlled factory setting are very different from robots that work alongside human electricians on a half-built data hall in rural Ohio. SoftBank's track record on ambitious technology bets is mixed. The Vision Fund era produced both winners and high-profile failures, and Masayoshi Son's appetite for vertical industry control has not always converted into operating success.
Two specific risks stand out:
-
Robotics maturity. ABB Robotics is a proven leader in factory and industrial automation. Construction robotics is a much earlier market. Most successful deployments to date are point solutions like rebar tying or bricklaying robots, not general-purpose autonomous workers. Closing the gap from controlled environments to active job sites will take time and capital that may not show ROI before the IPO clock runs out.
-
Regulatory and labor friction. Construction is heavily unionized in many US markets and subject to local building codes, OSHA standards, and inspection regimes. Even working robots will face friction from organized labor and permitting officials, and SoftBank will need to manage those relationships in every jurisdiction it builds.
The valuation is also a narrative anchor more than a measurement. A $100 billion price for an entity that has not yet proven the operating model is consistent with how SoftBank has historically priced strategic bets. Buyers of the eventual stock should treat the IPO valuation as an assumption about the future, not a snapshot of present cash flows.
What Businesses Should Actually Do Now
You do not need to invest in Roze to act on the underlying signal. Three concrete steps fit most AI roadmaps in 2026.
-
Add a compute capacity column to your AI plan. For each significant AI workload on your 12 to 24 month roadmap, document the cloud, region, and instance family it depends on. Confirm with your account team that capacity will be available on the dates you need. This was once a routine procurement detail. In 2026 it is a planning input that can move launch dates by quarters.
-
Avoid single-region, single-vendor lock-in for new AI workloads. Even if your incumbent cloud is the obvious choice, stage workloads so they can shift if a region runs short. Containerized model serving, portable data pipelines, and multi-region training checkpoints are now infrastructure hygiene, not advanced practice. We covered the broader implications in the AI vendor landscape shakeup.
-
Stress test your AI cost model against two compute pricing scenarios. One in which robotic and prefab construction adds 20 percent more capacity by 2028 and prices soften. One in which the buildout stays labor-constrained and premium GPU pricing rises through 2028. Decisions about pricing, hiring, and product investment look very different in those two worlds. Make the assumption explicit so it can be challenged on the merits, not assumed away.
Key Takeaways
- SoftBank plans to spin out and list Roze, an AI and robotics company, at a $100 billion target valuation, per FT reporting on April 29, 2026.
- Roze will bundle ABB Robotics, energy and land assets, and SoftBank's existing data center projects, with the goal of using robots to build AI infrastructure faster.
- 2026 hyperscaler capex from Alphabet, Amazon, Meta, and Microsoft now totals roughly $710 billion, and 2027 capex is projected above $1 trillion per CNBC reporting on Q1 2026 earnings.
- The US construction industry is short roughly 439,000 workers per ITIF, with about 85 percent of 2026 construction workforce demand tied to data center projects.
- The strategic implication for AI buyers is that physical compute capacity is now a procurement variable, not a given. Plan, contract, and budget accordingly.
The businesses that move early on AI infrastructure planning will have a meaningful advantage. If you want to be one of them, let's start with a conversation.