BRICKS & BYTES BULLETIN
INTELLIGENCE FOR CONSTRUCTION LEADERS, brought to you by Document Crunch
THIS WEEK
Autonomy Money Hits $2 Billion in Five Weeks. Permission Might Be the Real Constraint.
Five companies raised $2 billion in five weeks betting AI needs to learn the physical world. Plus: July housing starts collapsed, and data center moratoriums are making permission the new bottleneck.
THE EXECUTIVE BRIEFING
THIS WEEK’S KEY TAKEAWAYS
Key Takeaway 1:
Five companies building autonomous heavy machinery raised close to $2 billion in the past few months, but they compete through five distinct models, retrofit kits, autonomous contractors, single-task robots, mining fleets, teleoperations, each putting buyers in a different relationship with the tech.
Key Takeaway 2:
Travis Kalanick's Atoms, Yann LeCun's Advanced Machine Intelligence and Fei-Fei Li's World Labs have each raised roughly a billion dollars on one bet, that AI needs to understand the physical world, and construction sites are becoming the training ground.
Key Takeaway 3:
US housing starts fell 12.4 percent in July while data center moratoriums spread from Fort Worth to New York State. Permission has become construction's real bottleneck this cycle, ahead of technology or capital.
“Site risk is now political risk, and it's mappable.”
7 THINGS WORTH YOUR ATTENTION
ON THE RADAR THIS WEEK
ONS 2026 opens Monday in Stavanger, Norway's biggest energy forum, spanning offshore, hydrogen and infrastructure. (More)
Nigeria Mega Project Expo opens Tuesday in Lagos, West Africa's largest construction and infrastructure show. (More)
IWF Atlanta opens Tuesday, North America's largest woodworking technology fair covering machinery and automation. (More)
AASHTO's Agency Administration Leadership Conference begins Wednesday in Boston for state DOT leaders. (More)
Advancing Construction Technology starts Wednesday in Chicago, covering BIM, VDC and AI adoption. (More)
ExpoCAMACOL opens Wednesday in Medellin, Latin America's largest construction fair with 500-plus exhibitors. (More)
PWX 2026 opens Sunday in Houston, North America's biggest public works conference and expo. (More)
POWERED BY:
FULL EXECUTIVE BRIEFING
Autonomy Money Hits $2 Billion in Five Weeks. Permission Might Be the Real Constraint.
Five companies building autonomous heavy machinery raised close to two billion dollars between them over the last five weeks, and the pace picked up as the month went on.
It started with TerraFirma in mid-July, a $115 million round from Kleiner Perkins for autonomous earthmoving robots. A week later, Travis Kalanick, the Uber cofounder who now runs the industrial AI company Atoms, closed a $1.7 billion round covering mining trucks and food robotics under one roof.
By the end of July, Komatsu had signed a manufacturing partnership with a startup called AIM Intelligent Machines. The run culminated on August 17, when two more deals landed the same day: Bedrock Robotics said its excavators are now digging without anyone in the cab, and Swiss firm Gravis Robotics closed $200 million from SoftBank.
$2 billion raised across five separate companies since mid-July Five distinct business models packed inside that one number Zero independently audited performance figures among them so far
Investors are betting autonomous construction equipment has already crossed from lab demo to real business, and they are paying premiums to grab a seat before the field closes. Five companies is not the same thing as one market, though, and each of these bets puts a buyer in a different relationship with the technology.

Image: Gravis Robotics
Different Models, Different Relationships
Retrofit platform: Bedrock, Gravis and AIM bolt cameras, sensors and a computer onto machines contractors already own, so a fleet gets upgraded the way a phone gets a software update. Gravis says its kit runs across Caterpillar, JCB, Volvo and John Deere equipment, and the pitch comes down to buying nothing new while keeping what you already have.
Autonomous contractor: TerraFirma does not sell software to earthworks contractors. It wants to win the earthworks contract itself, delivered with its own remotely supervised machines, which puts TerraFirma in competition with the contractors it might otherwise have sold to.
Single-task specialist: Built Robotics concentrates on driving piles for solar farms and reports fifty thousand operating hours across forty deployments, while smaller players like Bluelight Machines and Crewline focus on compaction rollers, since rolling a surface flat is repetitive work nobody wants.
Mining at scale: Caterpillar reports close to 700 autonomous trucks working worldwide, hauling more than 11 billion tonnes between them, and China's EACON listed in Hong Kong in July with over two and a half thousand autonomous trucks already deployed. Travis's Atoms sits here too, through its Pronto acquisition, which has hauled more than two million tons at a Texas quarry.
Teleoperations: keeps a human in the loop but moves them off the machine entirely, running equipment from operators sitting a hundred miles away.
Most of the operating hours and tonnage figures quoted this week are self-reported, with little independent verification so far. Which model a contractor buys into shapes the relationship they end up in. A retrofit kit makes autonomy a capability you own. A robot sold as a service makes it a subcontract. TerraFirma's model makes it something closer to a competitor at the bid table.
Why the Money Is Really Chasing the Physical World
If you trace the autonomy money back far enough, you’ll find it landing on a research argument that started well before any funding round. Yann LeCun, the Turing Award winner who spent twelve years at Meta, left the company and raised $1.03 billion for a startup called Advanced Machine Intelligence, built on the argument that language models understand nothing about how the physical world actually behaves. Fei-Fei Li raised roughly a billion dollars for the same mission at World Labs, and Autodesk followed with its own $200 million bet, investing directly in World Labs to bring its models into design software.
Travis has now put a name on the commercial version of that thesis, treating digital and physical computing as parallel systems, where a processor changes bits, storage holds them, and a network moves them, while manufacturing changes atoms, real estate holds them, and transport moves them.
"Multiple trillion-dollar industries rapidly transforming over the next decade, powering an engine of prosperity unparalleled in human history."
Uber, according to him, was never a taxi company. It was the network layer for moving people and things, and Atoms is building the same layer for food, mining, and the wider built environment.
A retrofitted excavator on a job site does more than dig. It also maps the ground while it works, sensing how dirt behaves when wet, how a loaded bucket swings, what a person walking nearby is likely to do next. Alain Waha, Buro Happold's CTO, told us on this show that current AI does not yet understand the built world and that whoever solves it moves AI from reading about buildings to reasoning about them.
We saw a version of that argument in person at Bedrock's Austin site, where what struck Kevin Peterson's team most was the constant watching, the machine tracking everything around it continuously, exactly the capability LeCun and Fei-Fei are trying to build in a lab.
Understanding travels in a way a survey never does, since a map only captures one site on one day, while a system that genuinely understands terrain and physics can drop into a dozer or a haul truck on ground it has never seen. Every hour a machine works, it logs situations no one can buy off a shelf, the same dynamic that decided the self-driving car race, where logged miles beat clever code as the determining factor.
Where the Actual Work Is Going
Two headlines from the week point at the same underlying question, and neither has anything to do with autonomy. On Tuesday the Census Bureau reported that US housing starts fell 12.4 percent in July, to an annualized rate of 1.24 million homes, well below the 1.35 million economists expected and 13.5 percent under a year earlier. Single-family starts dropped to 808,000, the weakest pace since November 2022, according to HousingWire's July numbers.
"Builders remain cautious as elevated mortgage rates, rising construction costs, and economic uncertainty continue to limit demand."
The permit landscape is even more alarming, rising five percent to 1.44 million. A permit is an option that keeps the door open cheaply, while breaking ground brings the materials orders, subcontractors and carrying costs of real commitment. Builders are buying the option and declining the commitment, telling you, in the plainest language a builder speaks, that they do not trust demand at today's prices and mortgage rates. HousingWire's Logan HousingWire's Logan Mohtashami pushes that argument further, suggesting that at these prices America may not have a housing shortage at all, in which case soft starts are simply the market finding its own level.
The backlash against data centers stopped being an online argument this week and became a series of actual votes:
Fort Worth's city council voted unanimously to begin a 90-day pause on new construction, though it will not take effect until roughly February.
Louisville's Metro Council voted 24 to 1 to pause new data center approvals for six months.
New York became the first state to freeze environmental permits statewide for any data center of 50 megawatts or more.
Virginia's local councils are assembling what amounts to a moratorium one jurisdiction at a time.
Demand for that capacity remains close to bottomless, and the money is still queuing up, but permission has become the input everyone underestimated, and it is now being decided council by council. Anyone pricing site risk should run it against the two-speed market carrying US construction and add a jurisdiction column to the model.
Every story this week points to the same shift in what counts as valuable. Capital used to chase the software that managed a job. It is now chasing the systems that can actually sense one, and the businesses building those systems already treat the data as the product, with the digging or the hauling as the thing that pays the bills while the training set builds. Kalanick's framing of atoms and bits is the cleanest version of that argument on offer this month.
The housing numbers and the moratorium votes are the other half of the same story, and they move at a completely different speed. A funding round closes in weeks. A permit needs a person who was elected by people living next to the site, and that person is under no obligation to move at the pace of venture capital.
Site selection used to be an engineering problem: power, water, access roads, soil. It is now also a political one, decided council meeting by council meeting, and the firms treating it that way will spend next quarter building it. The ones still pricing it as an engineering problem will have a tough time in explaining the delay.







