Analysis
Beneath this year's gigawatt-scale AI infrastructure headlines -- Crusoe's reported $3 billion raise, National Grid's $1.75 billion investment in Joulent's Texas gas-fired power project -- a smaller, less-discussed layer of funding is going into the unglamorous physical mechanics that keep that infrastructure running. Oxmiq, a chip-design startup, raised $35 million to collapse GPU, CPU and TPU functionality into a single unified architecture, aiming to simplify the increasingly fragmented AI hardware stack.
Omen AI raised $31 million for a narrower but equally physical problem: sensors that monitor datacenter coolant systems for early signs of machine failure, letting operators catch cooling problems before they take AI clusters offline. Neither round comes close to the billion-dollar scale of the infrastructure megadeals, but both represent the same underlying thesis playing out at a smaller scale -- that durable value in AI infrastructure increasingly sits in the physical layer, not just in compute contracts or GPU access.
This 'picks and shovels beneath the picks and shovels' pattern matters because it's harder to replicate than a compute lease: a proprietary chip architecture or a coolant-monitoring dataset built over years of deployments creates a moat that a well-funded competitor can't simply out-capitalize the way they might outbid for GPU allocation.
For infrastructure-focused early-stage investors squeezed out of the billion-dollar rounds dominating headlines, this tier -- $25 million to $50 million rounds solving specific physical reliability or design problems within the AI stack -- represents one of the few remaining entry points that doesn't require competing directly with sovereign wealth funds and hyperscaler balance sheets.
What to watch: whether either Oxmiq or Omen AI lands a marquee hyperscaler or neocloud customer that validates the thesis at scale, and whether more capital rotates into this smaller, more technical tier of AI-infrastructure investing as the biggest rounds get harder to access.
