Open Agent Economy
Clay Norris@claynorris
−10.4%
since Oct 1
The idea
I believe AI keeps advancing fast but stays open and competitive, so no single model maker wins the economics. Instead value flows to what agents actually consume: chips, networking, inference clouds, and the power behind it all, with a small stake in the crypto rails agents might use to transact with each other. Compute and power are the core because they get paid across fast scenarios, while the crypto sleeve is the swing-for-it piece. I hold a bit of cash and avoid concentrating in any one name, expecting this to play out over a few years rather than a decade.
What it owns
- In an open ecosystem, chipmakers who sell to all competitors get paid regardless of which models win. · NVIDIAMost agents at scale run inference on NVIDIA chips, so expanding agent deployment drives GPU demand as long as NVIDIA remains the default choice and supply stays tight., Advanced Micro DevicesOpen-model buyers seek a second source beyond NVIDIA, making AMD credible if open inference keeps growing and wins real deployments., Taiwan SemiconductorEvery serious AI chip gets manufactured at TSMC, so it collects a toll in every fast scenario as long as advanced chip demand compounds., BroadcomHyperscalers building custom AI chips rely on Broadcom for silicon and networking that ties clusters together, benefiting as competition drives custom silicon orders., Micron TechnologyFast agents need high-bandwidth memory next to every accelerator, so Micron wins if HBM stays supply-constrained and priced accordingly., Marvell TechnologyIt builds custom AI chips and connectivity for clouds avoiding NVIDIA's full price, getting paid as competition pushes buyers toward custom silicon.37%
- Swarms of agents run on rented GPUs, so cloud providers serving open models capture the usage surge. · CoreWeaveIt rents GPUs for inference and nothing else, so its bookings show exploding inference spend first if agent deployment accelerates., Meta PlatformsIt pushes hardest to keep frontier models open, which is the foundation this take rests on, and must keep shipping competitive open weights., Nebius GroupIt hosts open models served as cheap inference for builders, working if open models stay near the frontier and developers choose neutral hosts., OracleIt rents data center capacity without betting on any one model, so the neutral landlord gets paid by all competitors as long as its AI backlog converts to revenue.20%
- Agents communicating at high volume saturate the interconnects and optics between chips first. · Arista NetworksAgents talking to agents means data centers move vastly more traffic, and it sells the switches carrying it if AI networking spend keeps growing., Astera LabsIt makes the connective tissue between chips inside AI servers, a bottleneck that tightens as clusters grow and stays designed into next-generation systems., Credo TechnologyIts high-speed cables link racks of accelerators together, an essential piece every inference buildout needs if data center expansion keeps accelerating., CoherentInside big data centers, signals move as light and it sells the optics needed, growing with inference clusters if they keep expanding.16%
- Fast advancement requires massive electricity, and power generation is the slowest infrastructure to build. · Constellation EnergyIts nuclear fleet powers the steady electricity data centers sign long-term deals for, working as long as those contracts keep coming at good prices., VistraIt sells power at market prices, so if agent-scale inference pushes electricity demand up faster than supply, it captures the price rise directly., GE VernovaIt sells turbines and gear to build generation and grid equipment for new data center loads, getting paid in every fast scenario as power buildouts continue.11%
- If agents get wallets and budgets, exchanges and stablecoin infrastructure collect transaction tolls. · CoinbaseIf agents start paying each other, it is the most likely exchange and wallet layer they route through, needing only machine payments to reach scale., Circle Internet GroupIf agent-to-agent payments arrive, stablecoins are obvious, and its regulated USDC is what agents would settle in if machine-driven payments materialize., Galaxy DigitalIt touches trading, custody, and data centers, working if crypto rails become the plumbing agents pay through and it wins on both the crypto and compute sides.9%
- Cash7%