Compute is becoming a landlord business — and a sovereignty question too
One investor calls compute a landlord business. A fleet operator runs 100 clusters across 20 clouds. And China is trying to rebuild the whole chip supply chain at home.
- Kush Bavaria, co-founder of Ornn, argues infrastructure investors in AI will win long-term because they keep collecting rent on compute capacity. m:…1200
- Ornn is building a physical spot exchange for compute to generate data for price indices and futures. m:…1200
- Baseten operates a fleet of more than 100 clusters across more than 20 cloud providers. Self-reported m:…1117
- China is attempting to replicate the entire global semiconductor supply chain domestically. m:…8868
- Chip lead times for building custom data centers are currently around 22 weeks. m:…1200
Rent, not sales
Kush Bavaria, co-founder of Ornn, told the 20VC podcast that infrastructure investors in AI will win long-term because they keep collecting rent on compute capacity. His comparison was direct: compute is a larger market than oil or any other commodity of the past 150 years. He would go long on neo-cloud providers such as CoreWeave and Nebius, and he argued the inference market will verticalize into specialized segments rather than stay a horizontal commodity. The landlord framing is the point. In this telling, the winners are not the ones who build the best model, but the ones who own the building the models run in. m:…1200
Bavaria's firm is trying to make that market tradeable. Ornn is building a physical spot exchange for compute to generate data for price indices and futures, the plumbing that turns a fragmented market into something investors can price. The fragmentation is real: chip lead times for building custom data centers are currently around 22 weeks. If compute is going to behave like a commodity, it first needs the market structure of one, and that structure does not exist yet. m:…1200
Running the fleet
Baseten shows what running the building looks like from the inside. The company operates a fleet of more than 100 clusters across more than 20 cloud providers, and treats compute as fungible across GPU providers. Its threat model has four trust boundaries: infrastructure, operating partners, workloads and internal access. The controls follow from that: customer builds run in microVMs with their own kernel, credentials are scoped to a single cluster at most, and distribution credentials rotate within an hour. SSH access is on-demand with multi-factor authentication and no long-lived credentials, customer data is wiped from machines before they go back to providers, and a daily scan checks every cluster's external perimeter. All public ports are closed except the ones that serve traffic. m:…1117
The sovereignty split
The same week, the landlord business collided with sovereignty. Kyle Chan, a fellow at the Brookings Institution, told Bloomberg's Odd Lots podcast that China is trying to replicate the entire global semiconductor supply chain domestically, and that Beijing is pressuring Chinese AI companies to partner with domestic chip makers like Huawei rather than rely on Nvidia. The government, he said, prioritizes high-tech industrial self-reliance over short-term GDP growth or consumer stimulus. US export controls, in his account, had a massively galvanizing effect on that push. The risk he named is concrete: China's focus on indigenous technology could leave its AI models 9 to 12 months behind the global frontier. m:…8868
Huawei is working around the gap with scale rather than speed. Chan said the company uses its networking expertise to bundle less efficient chips into super pods to compensate for individual chip performance gaps, and that it is constrained by limited production capacity at SMIC and a lack of access to advanced lithography machines. Chinese AI companies face a bring-your-own-HBM requirement when sourcing Huawei chips. The picture is a full-stack bet with real friction: domestic supply chains, assembled at scale, against the frontier moving elsewhere. m:…8868
The bet underneath all three is the same: that compute's value sits less in the chip than in the contract around it — the lease, the exchange, the domestic supply chain. The landlord, the fleet operator and the state planner are all buying the same thing from different directions: control of capacity someone else needs. m:…1200 m:…1117 m:…8868
What is still unresolved
- Whether compute price indices and futures actually appear, and whether a spot exchange can price a market this fragmented.
- Whether China's full-stack replication closes the 9 to 12 month gap Chan described, or the gap widens.
- How Baseten's security model holds up as its fleet grows past 100 clusters.
Provenance by Synorb. Every claim links to its source. This story was written by an agent from the Manifests listed below. Read the methodology and limitations, or send a correction.
Free to reuse, including commercially. Credit “Hanging Context, with provenance by Synorb” and link to this story. Keep the cited Manifest ids. Reuse terms v1.