From the canonical brief: commercially viable as a sovereignty-led offer for a narrow set of enterprises — not as a broad cost-saving play, and not differentiated from Palantir on architecture alone.
Pricing shape: layered, not per-token — platform subscription, infrastructure at visible cost-plus, bundled public-model allowance with pass-through overage, security/monitoring where sovereignty drives the sale, short paid POC. Pike already follows this shape.
Assumptions that drive the conclusion (by weight): (1) share of spend that moves local — easiest work is cheapest publicly; (2) fully loaded local cost (staff + GPU utilisation); (3) price direction — frontier mid/small cuts vs rising GPU rental; (4) hard processing boundary; (5) who operates the environment and under which jurisdiction.
Enterprises buy public AI mainly as seat-priced copilots and metered model APIs, and they overwhelmingly choose closed frontier models. Sensitive data does leak into AI tools, but the main exposure runs through unmanaged personal accounts, not sanctioned enterprise contracts.
Menlo Ventures (2025): ~$37B enterprise genAI spend; ~$12.5B foundation APIs; ~$7.2B copilots; 76% of use cases bought rather than built. Anthropic / OpenAI / Google hold ~88% of enterprise LLM API usage; open-source share fell from 19% to 11%.
Netskope (2026): heavy sensitive-data incidents into AI apps; ~47% of genAI users still on personal unmanaged accounts. Cost predictability is also shifting as seats turn into meters for agentic use.
Narrow, repetitive, data-sensitive work can move local with little quality loss. Frontier coding, long-horizon agents and novel reasoning cannot — and that is where spend is concentrating.
Epoch AI: best open-weight models trail closed frontier by ~4 months / 8 capability points (since Jan 2026). Strongest open models need large GPU clusters; workstation-class models score lower. Hard research-math tests still show frontier ahead.
Routing literature (e.g. LMSYS RouteLLM) shows large savings when easy tasks leave the frontier — but that is share of calls, not share of spend. The easiest work is also the cheapest publicly, so spend displacement is smaller than token displacement.
Public-model token use can fall while total inference rises (embeddings, routers, guards, retries, eval, agents). Demand also expands as unit cost falls. The defensible claim is slower growth in paid external usage, not lower overall AI spend.
On the brief’s illustrative fully loaded local costs (~$0.6M to ~$2.0M/yr depending on utilisation and staff), localisation saves money only above roughly $1M–$13M/yr of in-scope public-model spend — a few hundred very large users, not the broad market. Central case threshold ~$1.7M/yr at a blended ~$4/M tokens.
Keep three ledgers apart: customer price (Pike $15.275M proposal layers), Cosmic cost (blocked / fully loaded), customer value. Pike’s $1.4M AI allocation is customer price, not Cosmic provider COGS. BAIT is device/dashboard pricing. Assurant is draft.
Attractive when the customer meets at least two conditions, and one is usually a hard constraint rather than cost:
| Condition | Test |
|---|---|
| Sustained volume | In-scope public spend above ~$1.7M/yr (central), steady not bursty |
| High utilisation | Local GPUs busy most of the day (idle capacity kills the case) |
| Local model good enough | Passes the customer’s own eval set |
| Hard boundary | Law/contract/OT/jurisdiction that residency-only does not solve |
| Operable environment | Clear who runs it and under which jurisdiction |
On cost alone it clears only for the largest users. Public stays cheaper for small/bursty loads and for hard frontier work.
Do not price a localised LLM primarily per token — that invites a line-by-line fight with public prices that fall every quarter. Prefer a layered structure that prices the governed platform and passes model/infra through transparently — which is what Pike already does.
| Shape | Cosmic reference | Fit |
|---|---|---|
| Platform subscription | Pike $12M; Assurant $12M draft | Strong — prices the governed environment |
| Bundled AI allowance + overage | Pike $1.4M AI allocation (customer price) | Strong — avoid pure token retail |
| Infrastructure cost-plus | Pike $75K host line | Needed — brief flags $75K as likely too low for local inference compute |
| Security / monitoring | Pike $1.8M | Fits when sovereignty drives the sale |
| Per-device / active user | Assurant diagnostics; BAIT $0.14/user | Device/dashboard — not LLM token economics |
| Paid POC | BAIT 30-day | Right discovery path; users/execution unconfirmed |
Palantir already sells local-plus-public inside a governed platform. Models are a swappable input: provider-hosted, Palantir-hosted open models, and customer self-hosted GPUs as first-class AIP models — pitched for sovereignty, air-gap and cost control, with no extra platform fee for registered models.
Value sits in integration, Ontology, governance and operational apps. Cosmic is not differentiated by “local model” alone. Any edge must be proven on device-native intelligence, ingestion/compression at source, and volumetric data — none of which are evidenced yet.
Do not sell Palantir as a thin ChatGPT wrapper.
“Not used for training” is now standard — it cannot be Cosmic’s argument. Sovereignty turns on where data is processed, under which law, who holds the keys, and who can access it. Cosmic must answer those questions about itself too.
OpenAI Business/Enterprise: no training by default, ZDR for qualifying API customers, customer-managed keys, storage residency in multiple regions. Bedrock: does not store/log prompts for training; provider models in AWS-owned accounts. For many US/EU buyers that is enough. Cosmic wins where processing location / air-gap / OT boundaries still fail under residency-only offers.
Until these are evidenced, the proposition is a credible hypothesis, not a proven product: