CalcLLM v2 — Cost Intelligence Platform

Concept Exploration • 2026-05-22 • Slug: cost-intelligence

Summary

An LLM cost intelligence platform that connects to provider billing/usage APIs, pulls usage data, and models the true cost by surfacing hidden multipliers — retry overhead, cache inefficiency, context window waste, embedding recomputation, and provider-specific pricing quirks. Helps teams understand why their bill is higher than raw token pricing implies.

Clean-slate rebuild of an existing v1.

Problem Hypothesis

Confirmed LLM bills are opaque. Sticker pricing ($/1M tokens) dramatically understates actual spend due to compounding hidden multipliers that teams can't see until the invoice arrives.

Hidden Multipliers

Retry storms — provider errors and rate-limits silently multiply costs
Cache miss rates — prompts reprocessed when they could be cached
Context window waste — oversized prompts pay for tokens that don't contribute
Embedding recomputation — unchanged content re-embedded unnecessarily
Provider quirks — billing granularity, rounding, minimum charges

Beneficiary Hypothesis

Hypothesis Engineering leads and AI/ML teams shipping LLM-powered products.

Org Size	Consumption Pattern	Confidence
Solo / small (1-10)	Direct dashboard use. Builder = bill payer.	High
Mid (10-50)	Platform/AI lead owns costs. Needs attribution + alerts.	Medium
Large (50+)	Leadership wants distilled reports, chargebacks. Open question.	Low

Value Wedge

Core Differentiator

Hidden multiplier modeling. Competitors show token counts; CalcLLM shows why your bill is higher than your token count implies by detecting and quantifying cost amplifiers.

Potential Core Differentiator #2

Agentic velocity cost modeling. Help leadership understand the true cost of product velocity driven by agentic coding tools (Cursor, Claude Code, Codex, etc.). Surface cost-per-feature or cost-per-sprint from AI-assisted development, and help teams decide which tasks justify agentic coding spend vs. hand-coding.

This reframes LLM cost from a pure infrastructure line item into a strategic product velocity decision.

ICP-dependent Resonates most with engineering leadership at orgs heavily adopting agentic workflows. Relevance depends on the ICP ultimately selected.

Secondary: read-only provider sync (no proxy), bill analysis, flexible team sharing (dashboard + alerts).

Market Evidence (May 2026)

Recent events validate both the core problem and the agentic velocity differentiator.

Uber — Confirmed (The Information)

Detail	Finding
Budget burn	Full-year AI budget exhausted in ~4 months after Claude Code rollout to ~5,000 engineers (Dec 2025)
Adoption curve	32% → 84% adoption by March 2026
Per-engineer cost	$500–$2,000/month; CTO spent $1,200 in 2 hours personally
AI code share	70% of committed code from AI tools; 11% of live backend updates AI-generated
CTO quote	"I'm back to the drawing board because the budget I thought I would need is blown away already."
Compounding factor	Internal leaderboards gamified adoption, accelerating uncontrolled spend

Microsoft — Substantiated (The Verge, Windows Central)

Detail	Finding
Scope	Experiences & Devices division (Windows, M365, Teams) — thousands of employees
Budget burn	Full annual AI budget burned within months via token-based billing
Response	License cancellation deadline June 30, 2026 (fiscal year end)
Internal framing	"Toolchain unification" around GitHub Copilot CLI — but cost is the clear driver
Strategic tension	Paying Anthropic while trying to sell Copilot widely noted as contradictory
Official statement	None. Reporting based on internal sources from credible outlets

Implications for CalcLLM

The problem is real, urgent, and happening now at the largest engineering orgs
Budget planning for agentic coding is fundamentally broken — CTOs are "back to the drawing board"
The agentic velocity differentiator is not speculative; Uber's CTO is publicly asking for exactly this capability
Token-based billing creates unpredictable, compounding costs that flat seat-licensing masks — the hidden multiplier problem in action
Orgs need tools to answer: "What is our AI-assisted velocity actually costing, and which tasks should we hand-code instead?"

Key Decisions

Decision	Resolution	Source
Provider integration model	Read-only API pull, not a proxy	from interview
"Billing" scope	Analyze customer LLM bills, not SaaS monetization	from interview
V1 relationship	Clean slate, no code carry-over	from interview
Team sharing model	Both dashboard + alerts, varies by org size	from interview

Constraints

Solo developer, ship-fast mindset
Read-only provider API integration only
Must support multiple providers (OpenAI, Anthropic, Google minimum)
No monetization model defined yet

Non-Goals

Not an LLM proxy or gateway
Not a prompt engineering or model evaluation tool
Not a provider switching recommendation engine

Riskiest Unknowns

#	Unknown	Impact
1	Data granularity gap: Can provider APIs expose retry counts, cache rates, context utilization? Or does this need request-level telemetry?	Blocks core value
2	Org ownership: Who owns the LLM cost problem — AI lead, platform team, finance, eng mgmt?	Shapes GTM
3	Standalone viability: Dedicated cost tool vs. feature inside existing observability?	Shapes product
4	Monetization: If CalcLLM analyzes bills but doesn't charge, what's the business model?	Deferred

Next Steps

Primary

/bootstrap-repo — Set up project foundation (README, package structure, project metadata) before research begins.

Then

/pack install business-discovery — Enable ICP, competitive analysis, and GTM research lane.

Then

/icp — Research the ideal customer profile using the beneficiary hypotheses above.

Deliverables: research/concept-brief-cost-intelligence.md • research/concept-brief-cost-intelligence-interview.md
ICP Readiness: Ready with caveats (test data granularity assumption first)