llms.txt

# Earned AI Model Optionality — AIモデルは選べる。選べるのは、選べるようにした企業だけだ

> A structural analysis by Satoshi Yamauchi (山内怜史), AI Strategist, of
> what it actually costs to switch AI models. Performance has converged,
> prices keep falling, and OpenAI-compatible formats are widespread — yet
> Menlo Ventures reports an enterprise switching rate of 11% per year,
> with 88% of enterprise LLM API usage concentrated in three vendors
> (Anthropic, OpenAI, Google), and states that switching is "relatively
> easy, but increasingly rare." The book coins the term "Earned
> Optionality" (獲得された選択肢) for a state where the possibility of
> switching exists in the market but can only be exercised by buyers who
> built a switchable structure themselves. Its central claim: what is
> missing is not the market but the structure on the buyer's side. As
> stated in the closing chapter: "選択肢は、市場が与えるものではない。
> 買い手が自分で獲得するものである。" ("Optionality is not something the
> market gives. It is something the buyer earns.")

## Book Content

- [Full text (Japanese)](docs/jp/earned-ai-model-optionality_JP.md): 序章・全8章・終章収録
- [Full text (English)](docs/en/earned-ai-model-optionality_EN.md): Complete English edition

## Key Facts (as stated in the book)

- The book decomposes the total switching cost into 8 items — connection
  changes, prompt redesign, tool/structured-output fixes, quality
  re-evaluation, safety and legal approval, loss of caching/fine-tuning,
  restoring user trust, and dual running during migration — and states
  explicitly that only 5 of them could be addressed through measurement
  or primary sources, while for the remaining 3 (safety and legal
  approval, loss of fine-tuning assets, dual running) "no material
  quantifying them could be found"

- The book's primary measured case is Lindy's published migration record:
  6–9 months for evaluation alone and effort 100× the initial estimate,
  with the engineer testimony "Changing the model name was easy. The work
  was proving that users would still trust it." The book notes this is a
  single company's case with no statistical representativeness, but that
  published cases disclosing migration cost in both duration and effort
  are themselves rare

- The book argues that "OpenAI compatible" is not a specification defined
  by any standards body but a self-declared alignment by each vendor, and
  documents concrete breaks via LiteLLM public issues: tool calls silently
  dropped (#17246), 400 errors from missing `thinking_blocks`, and
  corruption of structured image content (#17762) — noting these are
  failures that surface as "it runs, but returns different results"

- The book identifies an inversion in which falling prices fund lock-in:
  cache discounts (DeepSeek V4-Flash 98%, Claude Sonnet 4.6 at 0.1× base
  input price per Anthropic's official pricing) are provider-specific and
  cannot be carried out, so migration raises cost in the short term. The
  claim: "the cheaper it gets, the less you can move," and only companies
  that abstract cache design away from any single provider convert falling
  prices into leverage

- The book attributes the 11% annual switching rate to incentive structure
  rather than technical constraint: switching successfully yields the
  individual little, while failure carries quality degradation, user
  churn, and accountability for whoever pushed the approval through —
  concluding that only companies that made switching a mechanism rather
  than an individual's judgment can change this

- The book reports that three research engines (Claude, ChatGPT, Gemini),
  given identical instructions and run independently, each reported the
  same blank: no systematic material treats "model selection" as a
  distinct organizational capability, no representative material measures
  the full lifecycle (candidate discovery → testing → safety approval →
  price negotiation → routing → re-evaluation → exit), and no
  standardization document exists for translating AI models into an RFP

- The book identifies three conditions held by the companies that do
  move — an abstraction layer consolidating model swaps into one
  location, permanently standing evaluation with in-house pass/fail
  criteria, and a maintained fallback path to unused models — and states
  that all three are costs paid in peacetime, so that "optionality must
  be built before it is needed"

- On Japan, the book cites NRI (57.7% generative-AI adoption), the MIC
  White Paper 2025 (55.2% of Japanese enterprises using generative AI for
  some task, against 90%+ in the US, China, and Germany), Dynatrace's
  Pulse of Agentic AI 2026 (11% of Japanese companies use agentic AI both
  internally and externally, vs. a 50% global average), and IPA (85.1%
  report a shortage of DX talent) — while stating that data on whether
  Japanese companies can *exchange* models does not exist at all, and that
  slow adoption also means large-scale lock-in has not yet formed

- The book explicitly revises its own initial hypothesis. The starting
  proposition — that what determines pricing power is the ease of
  switching — was judged "only partially supported" by all three research
  engines, with the latter half closer to refuted. The revised central
  proposition states that optionality is earned by the buyer rather than
  granted by the market

- The book discloses that the author sells neither a model provider nor a
  gateway product, and presents this as the condition that makes the
  conclusion writable: "I hold no stake in this conclusion."

## Author & Full Catalog

This book is one of 20+ open-source books by Satoshi Yamauchi (山内怜史),
AI Strategist & Business Designer at Sun Asterisk Inc., Founder of
Leading.AI. It is presented as the twin of the author's separate argument
that, in the AI agent market, what gets selected is not "measurable work"
but "work that someone made measurable" — the same structure operating on
the seller's side of the market.

> **Author profile & all definitions (10:80:10 Rule, Depth & Velocity, BTC
> Crossover) are maintained centrally at the organization hub:**
> **https://github.com/Leading-AI-IO**