Venture Compute Moonshots
David@wdavidturner
· 1
−18.8%
since Sep 30
The idea
I believe the real multiplier in AI sits with the small builders of compute, not the megacap buyers of it, so this take skips them entirely. The core is neocloud names living and dying on GPU demand, alongside memory and challenger silicon, backed by power builders and the plumbing of interconnects and cooling that every data center needs. I'm fine with a few positions going near zero if one or two multiply, and I hold about a tenth in cash while treating this as a multi-year bet, not a quick trade.
What it owns
- Young GPU-cloud companies whose entire business is renting AI compute, the purest single-story bets in the take. · CoreWeaveA young company that lives or dies on renting GPUs to AI workloads, the purest single-story compute builder., Nebius GroupA challenger GPU cloud whose whole business is selling raw AI compute to customers that megacaps cannot absorb., Applied DigitalIt converts sites into AI datacenters and leases them out, a small builder that matters only while the compute buildout runs., IREN LimitedA bitcoin miner pivoting its power and sites into AI hosting, an option that multiplies if the pivot lands real revenue., Cipher MiningIt owns power sites with land and grid connections that AI builders need and cannot get quickly.29%
- Compute runs on electrons, and the companies building and selling generation capacity get paid whichever chip wins. · GE VernovaIt sells the gas turbines that new datacenters queue up years for, the cleanest way to own the fight for electrons., VistraIt owns the plants and sells power straight to compute buyers, the electrons side of the bet in its most direct form., Talen EnergyIt already sells reactor output directly to AI buyers and the bet is more of those contracts at better prices., OkloSmall reactors purpose-built for datacenters, years from revenue, priced on whether it gets licensed and built., Bloom EnergyIt sells fuel cells that give datacenters power now, while grid connections take years to complete.24%
- Challenger chips and the memory that every AI server needs more of, a direct route into the hardware itself. · Micron TechnologyEvery GPU cluster needs high-bandwidth memory, and Micron is one of three companies on earth that makes it., Cerebras SystemsA single-product company betting wafer-scale chips can win real inference workloads from larger competitors., SanDiskTraining data and model checkpoints need fast flash at scale, demand that keeps pulling storage upward.20%
- Interconnects, optics and cooling are the bottlenecks between chips and power, small companies with venture-sized upside. · Astera LabsIt makes the chips that connect GPUs to everything else inside a cluster, plumbing that scales with accelerator shipments., Vertiv HoldingsIt sells the cooling and power gear inside every AI datacenter, so its backlog is a direct read on the buildout itself., Credo TechnologyIt makes the high-speed cables linking servers inside AI clusters, small plumbing that multiplies with cluster size., CoherentIt supplies the optical parts that move data between racks, demand that rises as clusters get bigger and farther apart.17%
- Cash10%