"Your AI Strategy Is Making Bets. Do You Know Which Ones?" — @DanPupius
Why this is in the vault
Dan Pupius (CTO, The General Partnership; ex-Google, ex-Medium) offers a framework for making the implicit strategic assumptions inside any AI product explicitly auditable — directly applicable to how RDCO positions itself as an AI-deployer and how Sanity Check interrogates AI strategy claims.
The core argument
AI product strategy is always making bets on four axes that shift faster than execution cycles. Most founders don't name them, which means they can't detect when the ground has moved. Pupius argues you should write down your actual bet on each axis and — crucially — pre-specify the signal that would tell you you're wrong.
The four axes:
Token Economics — Scarce vs. Abundant. We're in "fragile token abundance" now (cheap compute enables categories like Cursor). Bet on abundance → your advantage must be proprietary data, domain expertise, or distribution. Bet on scarcity → articulate why cost declines don't destroy you (e.g., volume accumulation in always-on agents, latency-constrained voice AI). Novel insight: cost variance reduction is more defensible than competing on average price — "your bill won't exceed X" beats "we're cheapest."
Model Self-Sufficiency — Needs Scaffolding vs. Handles Natively. The fate of "model wrapper" companies. If your product would be unnecessary if the model had unlimited capabilities, you're building scaffolding (absorption risk). If you're connecting AI to messy external systems (EHR workflows, systems of record), integration survives capability growth. Regulated industries add a structural floor: "the model judged itself compliant" is not acceptable to financial or healthcare regulators — independent oversight of AI decisions is non-negotiable.
Platform Structure — Lock-In vs. Commoditized. Are you building on one provider's ecosystem (OpenAI function-calling, Anthropic computer-use) or staying portable? Lock-in isn't inherently wrong — but you need to know you're doing it. Open-weight models (Llama, DeepSeek, Qwen) remain viable; consolidation to 1-2 dominant providers is years away. Author's advice: have a diversification plan ready even if you're not executing it.
Trust and Governance — Permissive vs. Constrained. The only axis not primarily driven by technology. "Compliance can go viral" via enterprise procurement — Fortune 100 mandates cascade through vendor chains faster than government regulation (SOC2 precedent; Vanta built a business on that wave). A single public AI failure (misdiagnosis at scale, infrastructure disruption) could trigger overnight legislative response.
Practical application: For each axis, answer four questions: (1) what is our implicit bet? (2) what must be true for it to pay off? (3) what signals would tell us we're wrong? (4) how quickly could we adapt? References: Hamilton Helmer's "7 Powers" (moats that survive multiple futures); Andy Grove's "strategic inflection points."
Mapping against Ray Data Co
Strong mapping on multiple dimensions:
Token Economics → RDCO-as-deployer: The cost-variance-reduction insight is directly relevant to how RDCO pitches phData clients. Framing AI deployment value as "bill predictability" rather than "cheapest inference" maps to what enterprise buyers actually care about. This is a Sanity Check article candidate — "stop selling cheap tokens, sell predictable outcomes."
Model Self-Sufficiency → RDCO positioning: RDCO's value as an AI deployer sits in the "integration" quadrant — connecting AI to the messy data infrastructure clients already have (Snowflake stacks, dbt models, domain rules). This is the argument for why the DSA role is durable as models improve: Harvey didn't die when GPT-4 launched, because law firm workflows are still messy.
Trust & Governance → phData DSA role: The independent-model-oversight requirement in regulated industries (Federal Reserve explicitly requires challenge of AI model outputs) is a direct argument for the DSA engagement model. This is the "regulated industries need a human-in-the-loop deployer, not a model vendor" positioning.
Platform Structure → RDCO tool choices: Relevant to current Cloudflare/Anthropic tool stack decisions. Pupius's point that diversification plans should be ready (even if not executed) maps to maintaining portability across Claude/OpenAI/open-weight where it doesn't compromise delivery.
Sanity Check angle: The four-axis framework is a reframeable lens for evaluating any AI company's claims. A Sanity Check piece could apply it to a specific company's stated strategy and surface which bets are implicit (and therefore unexamined). That's original re-frame territory, not derivative coverage.
Related
- [[2026-05-27-joe-schmidt-a16z-yellow-brick-road-app-layer]] — complementary framework on where durable value sits in the AI stack (integration layer vs. model layer)
- [[2026-06-14-satya-nadella-frontier-ecosystem-learning-loop]] — proprietary learning loops as the firm's IP; converges with Pupius's "integration moat" argument
- [[2026-05-08-jaya-gupta-shape-as-moat]] — shape as the moat that survives model capability absorption; same axis as Pupius's Bet 2