00Opening read
The next scarce input in AI may be time.
Anthropic expects to commit at least $518 billion to infrastructure over the next decade. Data-center developers are buying smaller gas turbines because waiting for the grid can take too long. Meta is reserving compute in Southeast Asia. AMD is spending $8.2 billion to bring spatial intelligence and physical-AI research directly into the company.
These are not simply bets on more capacity. They are attempts to secure capacity earlier.
That distinction matters as the physical AI economy grows more complicated. A data center that receives power in two years can be worth more than a cheaper one waiting six. A compute contract signed today can protect a model company from tomorrow's shortage. A chip architecture informed by emerging robotics workloads can arrive one generation sooner.
The buildout is starting to pay for time.
01What moved
Anthropic makes compute a decade-long commitment
Anthropic says it expects to spend at least $518 billion on AI infrastructure with six partners over the next decade.
The striking part is not only the number. Roughly 80% of those commitments are non-cancelable or require payment regardless of usage, according to a confidential IPO prospectus reviewed by Reuters.
That changes the character of compute procurement.
Cloud capacity is starting to resemble industrial offtake: reserve the scarce input now, then carry the risk that demand, technology and utilization develop as expected.
Data centers build around the grid queue
U.S. data-center developers are increasingly turning to smaller gas turbines and other behind-the-meter generation.
Enverus expects 29.6 GW of this gas capacity to be added by 2030, with data centers accounting for most of the demand.
The reason is time. Large combined-cycle plants can take years to arrive, while grid interconnections can be equally slow.
Smaller systems can cost more over their lifetime. Developers are buying them anyway.
The price premium is increasingly a schedule premium.
AMD buys into physical AI
AMD agreed to acquire Fei-Fei Li's World Labs for $8.2 billion.
World Labs is working on spatial intelligence: systems that understand three-dimensional environments and physical interactions. Those capabilities can feed robotics, simulation and design.
For AMD, this is more than another AI acquisition.
It puts a chip company closer to the workloads that may shape future hardware. Physical AI is moving upstream into semiconductor strategy.
Southeast Asia becomes contracted compute territory
Meta has signed agreements with Australia's Firmus for GPU capacity at data centers in Southeast Asia.
Firmus will provide contracted capacity and expansion options built around Nvidia infrastructure.
This is how a new compute geography starts to harden. Large customers do not have to wait for a regional market to mature. Their contracts can help create it.
Public markets start asking harder questions
RoboTechnik raised about $660 million in Hong Kong, then closed nearly 5% below its offer price on its first trading day.
The company has real exposure to AI infrastructure through equipment used in silicon-photonics assembly and testing. Its revenue also grew sharply in the first half.
That was not enough to guarantee a strong debut.
Meanwhile, French electrical-infrastructure supplier Legrand raised its medium-term targets on data-center demand and its shares jumped.
AI exposure still matters. Execution and valuation matter more than they did.
02The throughline
The physical AI economy is becoming a critical-path economy.
Companies are no longer optimizing only for the cheapest chip, megawatt or building. They are asking when that resource will actually be available.
That changes behavior.
Compute gets reserved years in advance. Power moves behind the meter. Hyperscalers spread capacity across geographies. Chip companies move closer to the workloads that will define future architectures.
RCW's deployment chain now needs a clock beside every gate:
announced, financed, permitted, accepted, powered, connected, qualified, controlled, productive.
The next advantage may come from spending less time between those words.
03What to watch
- October 1 for South Korea's official trade data and another read on AI-memory demand.
- October 1 as IROS closes in Pittsburgh, with embodied-AI research moving toward commercial deployment.
- Mid-October for DayOne's expected public IPO filing and a closer look at energized versus planned capacity.
- Through 2028 for new U.S. turbine manufacturing and whether it shortens the power bottleneck.
04Source trail
Reuters, Anthropic infrastructure commitments
https://www.reuters.com/business/anthropics-518-billion-ai-buildout-hinges-largely-deals-that-cannot-be-canceled-2026-09-29/
Reuters, behind-the-meter turbines
https://www.reuters.com/business/energy/dash-small-gas-turbines-set-impact-data-center-costs--reeii-2026-09-29/
Reuters, AMD / World Labs
https://www.reuters.com/technology/amd-acquire-fei-fei-lis-world-labs-82-billion-deal-2026-09-28/
Reuters, Meta / Firmus
https://www.reuters.com/world/asia-pacific/meta-taps-australias-firmus-ai-computing-capacity-southeast-asia-2026-09-28/
Reuters, South Korean exports
https://www.reuters.com/world/asia-pacific/south-korean-exports-seen-rising-16th-month-solid-ai-chip-demand-2026-09-29/
Reuters, RoboTechnik Hong Kong debut
https://www.reuters.com/world/asia-pacific/chinas-robotechnik-slip-hong-kong-trading-debut-2026-09-29/
Reuters syndication, Legrand
https://www.streetinsider.com/Reuters/Legrand%2Braises%2Bmid-term%2Btargets%2Bon%2Bdata%2Bcentre%2Bgrowth/27112291.html