Target verticals research plan

Objective (from Keli)

Assess whether there is a viable enterprise market for Cosmic Wire’s localized LLM model — local work for recurring or sensitive ops, with selected tasks sent to public frontier models when needed. This is a viability analysis, not a final GTM plan, vertical ranking, or binding price recommendation. Companion verdict page: cosmic-localized-llm-viability.pages.dev.

Cosmic pricing examples — how to use them

These show monetization shapes (platform, consumption, infrastructure, security, devices, active users, outcomes). They are not Cosmic’s underlying provider-token cost.

Analytical guardrails

• Separate a drop in paid public-model usage from a drop in total tokens / inference (retrieval, retries, agents can raise total work).
• Keep three ledgers apart: customer price, Cosmic cost, and customer economic value.
• Do not present hypothetical routing / offload % as measured Cosmic performance — label scenarios only.
• Do not assume every sensitive workload can run locally.
• “Not used for training” ≠ full sovereignty. Also do not imply public providers necessarily train on enterprise data.
• Treat Pike as a proposal, Assurant as a draft, BAIT users / execution as unconfirmed unless new evidence arrives.
• Relativity Engine needs a Cosmic definition. Do not equate it with RAG, ontology, or entity resolution without confirmation.

Questions this research answers

1. Market need — How do enterprises use public LLMs on ops / sensitive work? What do they buy, how are they charged, and what pain creates demand for local or hybrid?
2. Displacement — Which work can reasonably leave paid public models? Factor complexity, quality, latency, sensitivity, context, multimodality, tools, reliability. Offload % = labeled scenarios only.
3. Net economics — Could paid public tokens fall while total inference rises? Compare avoided public charges with fully loaded local cost (compute, host, storage, orchestration, security, staffing, redundancy, support).
4. When it works — Usage, local quality, utilization, and control needs that make localization attractive — and when public models stay cheaper or better.
5. Pricing shapes — Public seats / tokens vs platform, usage, per-device / active user, hosting, security, paid POC, outcome-linked, implementation fees (map to Pike / Assurant / BAIT).
6. Palantir — Value from integration, ontology, governance, orchestration, apps, deploy flexibility (incl. BYOM / self-hosted). Where Cosmic is comparable vs differentiated (device / volumetric / local inference). Never “thin wrapper.”
7. Sovereignty — Broader than no-training: location, jurisdiction, keys, prompts, outputs, embeddings, retrieval, logs, fine-tunes, retention, backups, support access, subprocessors, deletion, audit. Acknowledge enterprise privacy products (ZDR, residency, private networking).
8. Preliminary conclusion — Viable or not, which assumptions drive that, and what Cosmic evidence would validate it.

Markets

From the Target Verticals deck (attached to Keli’s note). Mark’s four to start; everything else stays on the map until someone reopens it.

Shared framework

Every market note uses the same eight headings so Tue/Thu packs are easy to compare:

1. Buyer & use cases — who pays, and which day-to-day jobs chew through AI
2. Data gravity — what can stay private vs what needs a big public model
3. Token economics baseline — sticker price vs what enterprises actually pay (list prices are often soft)
4. Localized LLM delta — rough share that could stay local, and what still needs the public model
5. Sovereignty / compliance — why “keep it on site” matters in this market
6. Cosmic wedge — how Cosmic makes the customer’s own ops data useful (no invented product claims)
7. Enhance vs replace — tools already in the plant that Cosmic should sit next to (e.g. Samsara-class)
8. Open questions for Cosmic — only Jonathan / product can answer these

Mark every number so nobody confuses a guess with a fact:

Verified from the vendor or a filing · Observed from a public source · Inferred our scenario · Blocked waiting on Cosmic

Sequence — deep → thin

1. Data centers first Mark meeting

Three kinds of buyer, not one market — money can show up at more than one step:

What we owe Mark: one solid data-center brief, plus a simple cost table (small / medium / large daily usage — all on the public model vs mostly local). Label the table as scenarios, not Cosmic results.

2. Critical infrastructure & utilities

Grid and plant ops — spotting trouble early, keeping the lights on. Overlaps with data-center power. Use public CenterPoint / Chain Reaction stories as context about the market, not as Cosmic claims. CIP-style rules (critical infrastructure protection) are why “keep it on site” often matters here.

3. Manufacturing

Machines, sensors, uptime, throughput. OT = the factory floor systems; IT = the office systems. Lots of small, repeated AI jobs that can stay local; send only the hard questions to a big public model.

4. Telecom

Network ops (NOC = the room that watches the network), devices, constant signals. If we can find real usage numbers, this is where “do most of it locally” can look cheapest.

Research method

Done for the month

Keli’s bar: preliminary and worth iterating — not final. By the third week of November we’re done when:

• All four markets have been touched, and data centers was Mark-ready by Tue 10 Nov
• The cost story is clearly labeled as scenarios — no fake precision
• We say why keeping data close matters, and that Cosmic sits next to existing tools rather than ripping them out
• Cosmic gaps are listed so each Tue/Thu check-in is a refine loop, not a stall

What we’re waiting on — keep moving anyway

November calendar through third week (checkpoints Tue + Thu)