I founded Seed Club, a new model for early-stage investing built around networks, shared intelligence, and coordinated support. I’m interested in what happens when AI makes context, memory, and coordination more legible, and what that means for how companies and organizations get built.
My agent drafts this site from what I save. Green is me.
Week of July 6, 2026
LLM inference bottleneck is interconnect, not compute
Modern LLM inference leaves roughly 99% of GPU compute idle, with memory bandwidth and inter-chip communication as the real limits. This flips conventional assumptions about where AI infrastructure investment compounds.
- Tensor parallelism trades memory pressure for interconnect latency pressure
- Interconnect optimization matters more than raw FLOP count increases
- The strategic supply chain constraint may be networking, not chips
AI inference bottleneck may be interconnect latency, not HBM · @KSimback Modern LLM inference is not primarily compute-bound · @0xBADB01E As LLMs get sharded across more chips, decode speed may stop being limited by memory bandwidth and start being limited · '@KSimback' AI inference speed is often limited less by raw compute than by memory bandwidth and interconnect latency · jbrukh (@jbrukh) Inference speed is bottlenecked by memory and interconnects · @0xBADB01E Inference speed is bottlenecked by memory and networking · @jbrukh
AI progress constrained by missing data, not just compute
Scaling laws require data and compute to grow proportionally, but large domains have little training coverage. Sample efficiency gains cannot substitute for structurally absent data, making exclusive data pipelines the new strategic bottleneck.
- Vertically integrated data pipelines hold value that broader access erodes
- Missing domain coverage may rate-limit economic automation more than chips
Better sample efficiency alone may not remove the core bottleneck in AI progress, because many important domains are · Will Depue (@willdepue) As model training moves from abundant public internet data toward scarce private and expert data, the durable edge · @hypersoren AI progress may be constrained less by chips than by access to broad, high-quality training data · industriaalist (@industriaalist) AI progress may be constrained by scarce training data · @industriaalist Missing data may limit AI more than sample efficiency · @willdepue AI data boom may reward owners of exclusive pipelines · @hypersoren
Venture consensus systematically filters out outlier conviction bets
Investment committees protect against obvious mistakes but reliably eliminate the non-consensus, high-conviction bets that generate outsized returns. Early-stage investing compounds this because there is no verifiable product, making narrative the only asset under evaluation.
- One overwhelmingly convicted investor outperforms ten merely comfortable partners
- Mega IPOs reveal whether a return reflected skill or one lottery ticket
Venture committees may screen out the best outlier bets · @E_Bruxxx In venture, narrative can fund products long before proof · @imlaurieowen Investment committees reduce obvious mistakes, but they can also systematically kill the highest-conviction · '@E_Bruxxx' In venture-backed markets, especially frontier categories, narrative is not just marketing layered on top of an · '@imlaurieowen' A wave of outsized IPOs can make venture track records look brilliant while making repeatability harder to judge · @Open_LP Mega IPOs could expose luck versus repeatable venture skill · @Open_LP
Enterprise data privacy pushes production AI toward local models
Enterprise usage data is the most valuable input for continual learning, but companies will not authorize frontier labs to train on it. This structural barrier routes production AI toward small, fine-tuned open-weight models rather than frontier APIs.
- Frontier labs keep owning discovery, open source increasingly owns production
- Latency demands in production agents further favor smaller local models
Enterprise data may push AI toward local learning systems · @iamgingertrash If enterprise usage data is the most valuable input for continual learning, large central labs may struggle to capture · '@iamgingertrash' Production agent deployments may push value toward small open-weight models rather than the most capable frontier APIs · sarthakgh (@sarthakgh) Open models may dominate production agent deployments · @sarthakgh