Substrate
Elena Petrova (@epetrova_sys) · GPU memory fabric that lets inference clusters share KV-cache across nodes.
Every claim in this memo, checked against how well it holds up.
Company Snapshot
What the company does and where it sits in the market.
Substrate operates in AI Infra (Amsterdam, NL, Seed). GPU memory fabric that lets inference clusters share KV-cache across nodes. Cross-node KV-cache sharing is becoming a bottleneck as inference clusters scale; early but real inbound interest from infra teams.
Investment Hypotheses
The core bets underlying this memo, each traced to its evidence.
- Team quality: Second exit, undersold on LinkedIn (no announcement). Shipped a reproducible benchmark rather than a deck. Pattern matches a builder who has already learned the hard lessons once.⟦ev-ep-repo⟧
- Market wedge: Cross-node KV-cache sharing is becoming a bottleneck as inference clusters scale; early but real inbound interest from infra teams.
- Traction signal: 3 design partners in early conversationsCLAIMED
- Defensibility: Idea survives as-is; even if hyperscalers build this in-house, mid-market clusters remain addressable.⟦ev-ep-repo⟧
- Founder credibility: Prior startup quietly acquihired in 2022⟦ev-ep-exit⟧VERIFIED
SWOT
Strengths, weaknesses, opportunities, threats — shown only where evidence exists.
- Delegate-an-Analyst: Caught the planted fake competitor and flagged the exact metric that was fabricated.
- Steelman-Your-Rival: Named the hyperscaler in-house-build risk unprompted, then argued the mid-market wedge.
- 2.3x throughput gain on shared KV-cache benchmark⟦ev-ep-repo⟧VERIFIED
- Founder axis improving: Second exit, undersold on LinkedIn (no announcement). Shipped a reproducible benchmark rather than a deck. Pattern matches a builder who has already learned the hard lessons once.⟦ev-ep-repo⟧
- Market axis improving (bullish): Cross-node KV-cache sharing is becoming a bottleneck as inference clusters scale; early but real inbound interest from infra teams.
- Dependency-Graph Parasite: Benchmark repo forked by two infra teams at mid-size AI labs before any public launch.
Problem & Product
The problem being solved and how the product addresses it.
Problem: Cross-node KV-cache sharing is becoming a bottleneck as inference clusters scale; early but real inbound interest from infra teams.
Product: GPU memory fabric that lets inference clusters share KV-cache across nodes. Idea survives as-is; even if hyperscalers build this in-house, mid-market clusters remain addressable.
Traction & KPIs
Traction and revenue signals, with unverified gaps flagged explicitly.
- 3 design partners in early conversationsCLAIMED — No signed LOIs yet.
- ARR: not applicable — pre-revenue stage
- Cap table: not disclosed
- Financials / P&L: unavailable at this stage
2 sourcesbehind the citations above — duplicates the inline ⟦refs⟧, expand to inspect the raw excerpt.
Three-Axis Verdict
never averagedFounder, market, and idea-fit scored independently — one weak axis can't be masked by a strong one.
Elite substance under-marketed; second exit de-risks execution. Advance straight to diligence.