Citi Summit: Scaling Robotics Is a Decade-Long Marathon
Miles Bennett
Citi's annual robotics summit concluded that physical AI is moving from proof-of-concept to commercial deployment, but data scarcity, battery life, and high deployment costs make scaling a ten-year build — not a chatbot-style breakout.
Why is data the biggest bottleneck?
Attendees stressed repeatedly: even if the industry collects tens of millions of hours of real-world data by 2026, that volume amounts to mere "basis points" — not "percentage points" — of what high-performance robotics actually needs. This means → the data gap is not small; it is orders of magnitude wide.
In plain terms = large language models train on the vast text already on the internet. Robots must collect data scene by scene in the physical world, starting nearly from scratch each time the task changes.
This reflects a fundamental split between physical and digital AI: value sits not in the model itself but in proprietary data + purpose-built hardware + safety certification.
Why do purpose-built machines make money faster than humanoids?
Citi analyst Heath Terry noted that near-term returns are driven by specialised autonomous mobile robots from companies like Locus Robotics and Dexterity, not general-purpose humanoids.
The fastest-commercialising companies share a playbook: enter through a high-pain-point labour problem, skip the quest for general capability; adopt a Robotics-as-a-Service (RaaS) model to lower the customer's upfront barrier; and prioritise safety and reliability over model complexity.
In plain terms = humanoid robots attract headlines and investment excitement, but near-term profits come from machines that do one thing and do it reliably.
Where has $20 billion been deployed?
Over the past two years, physical AI has attracted roughly $20 billion in investment across warehousing, logistics, trucking, construction, aviation, and defence.
The core end-markets are logistics, warehousing, and auto manufacturing — all characterised by high-frequency, highly repetitive tasks. Last week BMW disclosed that upgraded humanoid robots are already working on the production line at its Spartanburg, South Carolina plant.
This means → capital is flowing down the path of highest certainty — the more repetitive and labour-short the task, the sooner automation lands.
How does RaaS unlock the mid-market?
High upfront capital expenditure has long been the biggest barrier for small and mid-sized companies. RaaS converts a one-time purchase into pay-per-use operating expenditure, sharply lowering the adoption threshold.
Terry highlighted Symbotic's "warehouse-as-a-service" offering (GreenBox/Exol), arguing it can extend warehouse automation to firms that previously balked at the cost.
In plain terms = buying a robot used to be like buying a house. RaaS turns it into renting — no down payment, just a monthly bill.
What is Citi's endgame call?
Terry's conclusion: physical AI is a decade-scale build, not a chatbot-style surge.
Three forces are accelerating enterprise demand — labour shortages, manufacturing reshoring, and a favourable regulatory environment — but data scarcity, talent bottlenecks, battery life, and deployment costs remain major friction points.
This means → Citi believes long-term value will concentrate in companies that master the data flywheel, solve real-world deployment problems, and meet the highest safety standards — and the speed of data accumulation will be the single most important differentiator in this race.
Content is for reference only, not financial advice.