The thesis

The AI buildout will be won by whoever owns the chokepoints.

HBM stacks, CoWoS packaging, EUV optics, gigawatts of power, optical interconnect, liquid cooling. Compute Bottlenecks maps the physical supply chain behind every GPU cluster — and the listed companies that own each choke. Own the strait, not the tanker.

The market stares at the obvious names. The binding constraints sit further upstream, where a handful of companies own each physical chokepoint — and most of them are not household names. That is exactly the point.

We walk the supply chain layer by layer. For each link we ask the same three questions: who physically owns the choke, what it would take for someone else to credibly compete, and what the next catalyst is that forces the market to reprice.

The answers land on a living map of 60 listed companies across 10 chokepoints — each with a conviction tier, a bull and bear case, a catalyst path, and the source trail behind every mapping. It is a research map, not a portfolio. Own the strait, not the tanker.

The chokepoints

  1. IHBM & memory bandwidth

    Every accelerator is bandwidth-starved; high-bandwidth memory comes from three companies and is sold out years ahead.

  2. IIAdvanced packaging & CoWoS

    The GPU shortage has repeatedly been a packaging shortage: 2.5D/3D interposer and hybrid-bonding capacity, not wafer starts, sets accelerator output.

  3. IIILithography & EUV optics

    Leading-edge logic and DRAM both route through one EUV toolmaker — and the single-source optics and laser chain inside it.

  4. IVLeading-edge foundry

    Every frontier accelerator is fabbed by one company, mostly in one geography; the alternatives are years behind.

  5. VPower, grid & generation

    Datacenter interconnect queues and transformer lead times are measured in years; gigawatts are the new wafer starts.

  6. VILiquid cooling & thermal

    Rack densities past 100 kW make direct-liquid cooling mandatory — and the cold-plate, CDU and rear-door supply chain is thin.

  7. VIIOptical interconnect & networking

    Scaling clusters beyond a rack is an optics problem: 800G/1.6T transceivers, co-packaged optics, and the lasers and DSPs inside them.

  8. VIIISubstrates, resist & wafer materials

    ABF substrates, advanced photoresists, and 300mm silicon are quiet few-supplier markets sitting under every chip.

  9. IXAccelerators & the capex layer

    The visible layer: accelerator designers and the hyperscaler capex programs that every other choke feeds.

  10. XPrivate giants & exposure vehicles

    OpenAI, Anthropic, xAI — the AI cap tables without tickers, and the listed and quasi-listed paths into them.