The story everyone tells about AI is a story about chips. But ask anyone trying to bring a hyperscale data center online in 2026, and they’ll tell you the chips aren’t the problem. The problem is the building. The site. The crew. The 4,000 to 5,000 workers needed on a single campus, and the simple fact that there aren’t enough of them anywhere.

That’s the bet behind SoftBank’s newest move. The Japanese conglomerate is creating a standalone company called Roze AI, designed to deploy autonomous robots into data center construction sites and speed up the buildout of AI infrastructure. According to TechCrunch, Masayoshi Son is eyeing a U.S. IPO at a $100 billion valuation, possibly within the back half of this year. If that sounds like a lot of zeroes for a company that hasn’t built anything yet, some of SoftBank’s own executives reportedly think so too.

In short: Roze AI is a vertical integration play. SoftBank already owns the chips (through Arm), the model layer (through its OpenAI stake), and the capital (through Vision Fund). Roze would give it ownership of the physical layer, the actual concrete, steel, and copper that AI runs on. If it works, SoftBank becomes indispensable to every hyperscaler. If it doesn’t, it joins WeWork and Katerra in the Vision Fund’s hall of expensive lessons.

Why Roze AI exists, and why now

The bottleneck moved from silicon to site

For three years, the AI infrastructure conversation has been about compute. Whether there are enough H100s. Whether TSMC can keep up. Whether power availability in northern Virginia will hold. Roze AI exists because the answer is shifting toward something more old-fashioned: people on jobsites, with hard hats and tools.

According to the Information Technology and Innovation Foundation, the U.S. construction industry was short roughly 439,000 workers as of late 2025, with the deficit concentrated in the trades data centers depend on most: electricians, pipe layers, HVAC techs. Add the 400+ hyperscale data centers in development, and you get backlogs of nearly a year on construction firms taking the work. SoftBank’s pitch is that you can’t recruit your way out of this. The math doesn’t work.


The labor bottleneck behind every AI announcement

Why every gigawatt of AI compute needs a small city to build it

Worth pausing on the scale. DataBank’s VP of Construction recently noted that peak crew sizes on hyperscale builds, which used to top out around 750 workers, are now hitting 4,000 to 5,000 on sites like the Red Oak campus in Dallas. That’s a small city deployed for a single building. The Stargate consortium SoftBank announced last year alongside OpenAI and Oracle has two buildings live in Abilene, Texas, with another six due online by mid-2026, and IEEE Spectrum reports hyperscalers want construction pace ramped from 1 GW per year to 1 GW per quarter, eventually 1 GW per month. Without automation, that’s not a stretch. It’s impossible.

“The real constraint on global tech growth isn’t solely related to a shortage of microchips, energy, or capital; it is the severe scarcity of the specialized talent required to build it.”

That quote, from Randstad CEO Sander van’t Noordende speaking to CNBC, is the Roze investment thesis compressed into a sentence.


SoftBank’s vertical AI stack, with one piece missing

The Son strategy moves down the stack, again

Roze isn’t a standalone bet. It’s the missing layer in a portfolio Masayoshi Son has been quietly assembling. Arm Holdings gives SoftBank exposure to the chip layer. The OpenAI position covers the model layer. Stargate provides the build pipeline. What was missing was the actual delivery mechanism, the entity that puts those data centers in the ground.

The asset base SoftBank is reportedly bundling into Roze tells you how serious the play is. CNBC reported that the new entity could include energy holdings, land, infrastructure investments, and crucially, ABB Robotics, the Swiss industrial automation business SoftBank agreed to buy last year for $5.4 billion. Add Ampere Computing ($6.5 billion) and DigitalBridge ($3 billion), and Roze starts to look less like a startup and more like a pre-built conglomerate-within-a-conglomerate.


The $100 billion question

Ambitious, or detached from gravity

A $100 billion IPO target for a company that hasn’t shipped a product places Roze in extremely rare company. The valuation thesis depends entirely on a story: data center construction is a trillion-dollar bottleneck, and SoftBank, through Roze, owns the platform to break it.

The skepticism inside SoftBank is informed. The Vision Fund’s track record is studded with valuation misfires: WeWork, the modular construction startup Katerra (which collapsed in 2021), and Zume, the AI pizza-delivery robotics company that absorbed hundreds of millions before shutting down in 2023. According to the Financial Times, several SoftBank executives view both the valuation and the second-half-2026 timeline as overly aggressive, citing Middle East tensions and a crowded U.S. IPO calendar. Roze is reportedly planning an analyst day at a Texas data center facility in July. That event, more than the spreadsheet, will probably determine whether the $100 billion figure holds.


What this means for the AEC industry

Pragmatist’s read on what to actually watch

Strip away the IPO drama and Roze AI is the loudest signal yet that data center construction is the segment of AEC that big capital has decided to industrialize first. For contractors, that’s both opportunity and warning. Opportunity, because hyperscaler demand isn’t going anywhere. Warning, because if SoftBank pulls this off, it changes the shape of who you compete against on the next mission-critical bid.

Two things to watch over the next twelve months. First, ABB integration: how SoftBank fuses ABB’s hardware with construction-specific automation will tell you whether Roze is a real operating company or a holding structure with a story. Second, Stargate flow. SoftBank’s role in the Stargate consortium gives Roze first-look access to gigawatts of build pipeline already capitalized and waiting on labor. Even a meaningful slice of that routing through Roze gives the company implied multi-year backlog before its IPO.


How Roze stacks up against the field

Robotics-meets-construction is a crowded category

Roze isn’t entering an empty room. Several construction robotics companies have spent years on automation for site work, layout, masonry, earthmoving and prefab. Here’s a quick lay of the land.

Company Focus Stage / Backing Approach
Roze AIData center construction at hyperscalePre-launch SoftBank spinout, eyeing $100B IPOBundled assets including ABB Robotics, energy, land, infrastructure
Bedrock RoboticsHeavy machinery autonomy$350M+ raised (Series B Feb 2026), ex-Waymo teamRetrofit autonomy onto existing earthmoving fleets
Dusty RoboticsLayout printing$69M raised across Seed to Series BSells robots to GCs and trades for floor layout
MonumentalBricklayingAmsterdam-based, scaling in EuropeOutcome-as-a-service: sells finished walls, not robots
Tempo WorksTrades and finishing workRevenue-positive, expandingOaaS model, robots deployed by own crews on live sites

The pattern across the rest of the field is small, focused, often outcome-as-a-service. Roze is proposing the opposite: vertical and scaled from day one, financed by the largest tech investor on earth. Whether that beats the focused-and-fast playbook is the experiment everyone in AEC robotics will be watching.


FAQ

What is Roze AI?
Roze AI is a new company SoftBank is spinning out, designed to automate the construction of data centers in the U.S. by deploying autonomous robots. The plan is to bundle existing SoftBank assets, including ABB Robotics, energy holdings, land and infrastructure investments, into the new entity. (Source)
How much is SoftBank trying to value Roze at?
According to the Financial Times, SoftBank executives are targeting a valuation of around $100 billion for a U.S. IPO, with some pushing for the listing to happen in the second half of 2026. The valuation and timeline are reportedly fluid, and some inside SoftBank consider both ambitious. (Source)
Why is data center construction such a bottleneck right now?
The U.S. construction industry was short roughly 439,000 workers as of late 2025, concentrated in skilled trades like electricians and pipe layers, exactly the trades data centers need most. Over 400 hyperscale data centers are in development, creating backlogs of nearly a year for firms taking on the work. (Source)
How does Roze fit with SoftBank’s other AI bets?
Roze completes a vertical stack across the AI economy. SoftBank already has exposure to chips through Arm Holdings, models through its OpenAI position, and a build pipeline through the $500 billion Stargate consortium with OpenAI and Oracle. Roze would give it ownership of the physical delivery layer, the actual construction of the data centers themselves. (Source)
What is ABB Robotics, and why does it matter to Roze?
ABB Robotics is the robotics division of Swiss industrial giant ABB, and one of the world’s largest suppliers of automation and machine systems. SoftBank agreed to acquire it last year for $5.4 billion. The plan is to integrate ABB’s hardware with AI capabilities and use it as the technical foundation of Roze’s data center construction automation. (Source)
Why are some SoftBank executives skeptical of the plan?
According to the Financial Times, internal concerns center on the $100 billion valuation, the aggressive 2026 IPO timeline, and broader uncertainty including Middle East tensions and a crowded U.S. IPO market. SoftBank’s track record also includes high-profile valuation misfires like WeWork, Katerra and Zume, all of which weigh on the conversation around Roze. (Source)
What does Roze mean for construction tech competitors?
Roze enters a category that already includes well-funded specialists like Bedrock Robotics, Dusty Robotics, Monumental and Tempo Works. Most of these companies have built credibility through narrow focus and outcome-as-a-service models. Roze is proposing the opposite, a vertically integrated, hyperscale-from-day-one approach backed by SoftBank’s capital. The next 12 months will test which playbook actually works on live sites. (Source)