EIP · Pulse 1 of 6 · Agentic Demand Family · daily composite
Enterprise Infrastructure Pulse
How much demand-pull is agentic AI putting on the enterprise hardware stack right now? GPU rental tightness (from CCST), Taiwan AI server ODM monthly revenue (from TAISP), Cloudflare's own Workers AI model distribution as agent-runtime proxy. The Pulse that reads the same hardware data as the Build-Out Family but interprets it from the demand side: are agents actually consuming the supply that's coming online?
Module breakdown
| Module | Weight | Score | Source |
|---|---|---|---|
| M1 · GPU/CPU inference rental | 25% | 53 | Service-Binding → CCST |
| M2 · Cloud AI platform pricing | 20% | 50 (neutral) | pending — Bedrock/Vertex/Azure + Helicone |
| M3 · AI server ODM monthly | 15% | 90 | Service-Binding → TAISP |
| M4 · Networking + storage + KV-cache | 15% | 50 (neutral) | pending — AVGO/MRVL/ANET Earnings-NLP |
| M5 · Power/cooling AI rack | 10% | 50 (neutral) | pending — VRT/ETN Earnings-NLP |
| M6 · Cloudflare/agent runtime | 15% | 81.2 | CF Radar AI Inference + GitHub |
Build status: MVP1 first iteration. Three modules live via Service-Bindings (M1/M3) and direct Radar API (M6). Substance modules (M2/M4/M5) at neutral 50 placeholder until Cloud-AI pricing scraper + Earnings-NLP pipeline ship. Service-Binding pattern = same Worker→Worker pattern used by content-agent / cron-sentinel / tape; production-grade and verified Phase 1.
Multi-Basket Equity Gaps v3 · value-capture decomposition
Four sub-baskets across the AI infrastructure stack — Accelerator Silicon (NVDA/AMD/AVGO/MRVL), AI Server System Build (DELL/HPE), Network + Power (ANET/VRT), and Cloud Tollbooth (ORCL). When system-build runs away from silicon, or when network/power lags both, the rotation is telling you which constraint is being repriced. AI-server-build outpacing accelerator silicon often signals memory/HBM tightness easing while integration capacity tightens.
| Sub-Basket | Vertical | Weight | 21d Basket Return | Module Score | Tickers |
|---|---|---|---|---|---|
| Accelerator + Custom Silicon | GPUs, TPU-ASICs, custom AI accelerators | 50% | +18.65% | 100.0 | NVDA · AMD · AVGO · MRVL |
| AI Server System Build | OEM rack + system integrators | 15% | +37.08% | 100.0 | DELL · HPE |
| Network + Power/Cooling | Switch fabric + datacenter power-thermal | 20% | +0.14% | 50.7 | ANET · VRT |
| Cloud Tollbooth (Hyperscale-Adjacent) | Hyperscale-adjacent compute + DB tollbooth | 15% | +11.42% | 100.0 | ORCL |
Methodology
EIP is Pulse 1 of 6 in the Closelook Agentic Demand Family. Its conceptual twist: same hardware data as the Build-Out Family (CCST · TAISP · MHP · STP), but read as Demand-Pull instead of Supply-Tightness. When CCST shows "rental rates rising" the Build-Out Family asks "is supply running out?" — EIP asks "is the demand strong enough to consume incremental supply?" Same data point, different lens.
Service-Bindings power M1 (CCST) and M3 (TAISP) — Worker-to-Worker fetches at the Cloudflare edge, no duplicate scraping. M6 reads Cloudflare's own Radar AI Inference API which exposes Workers AI model distribution as a proxy for serverless agent-runtime adoption. Daily composite at 05:00 UTC, after all upstream workers + Lake finish their cycles.
Pulse reading is a diary entry on our process, not investment advice.