Across infrastructure, tooling, and capital deployment, the week's sharpest voices are all circling the same unresolved tension: the layer that creates durable value in AI isn't the model itself but the workflow, context, and proprietary data wrapped around it — yet the market keeps pricing the model layer as if it's the moat.
Bridges across posts
Stratechery's read on Apple — that Siri's failure was broken search indexing, not bad models — is the infrastructure-first thesis, and it rhymes precisely with Upstarts Media's Squint piece: Squint's edge isn't its 2B-parameter model but the tribal knowledge and SOP layer on top. Both voices independently land on 'the app/workflow layer is the moat,' which is the single most important signal for a seed-stage investor evaluating AI infrastructure bets.
Not Boring's Return on Tokens framework — value created minus token costs, with most work better served by deterministic code — directly challenges the tokenmaxxing culture Stratechery diagnoses in Anthropic's Fable 5 guardrails piece, where Anthropic deliberately makes its model token-intensive to extract revenue from heavy compute users. The disagreement matters for founders: are you building toward ROT optimization or are you accidentally building a product whose unit economics worsen as adoption scales?
Dwarkesh Patel's data on frontier models training on 10s-100s of trillions of tokens versus humans' ~200 million lifetime tokens frames a founder quality signal that Lenny's Newsletter surfaces obliquely: the 'Proven, Better, New' framework works precisely because human intuition is sample-efficient (right direction 95% of the time) while AI is not. A seed-stage founder who understands this asymmetry — leaning on human judgment for category selection, AI for execution throughput — is operating with a structural advantage that most AI-native teams are currently ignoring.
Across the authors
Converging Not Boring and Stratechery independently arrive at the same conclusion this week: the AI infrastructure race is producing enormous waste and misaligned incentives. Not Boring names it 'tokenmaxxing' driven by perverse board and lab incentives; Stratechery shows Anthropic actively degrading model performance for potential competitors while publishing safety reports — both see the current infrastructure layer as captured by incumbent incentives rather than value creation.
Where they split Stratechery and Upstarts Media split on where the AI infrastructure moat actually lives. Stratechery's Apple piece argues the moat is platform re-architecture and on-device inference — deep, hard-to-replicate OS-level work. Upstarts Media's Squint piece argues the moat is the application and workflow layer sitting on top of commodity models. This split matters acutely for seed-stage allocation: are you backing infrastructure depth or application specificity?
The outlier Danny Crichton at Lux Capital is the outlier — while every other voice this week is focused on AI capability and tooling, Crichton is tracking gray-zone warfare, weaponized economic data, and hybrid tactics as the macro risk frame. For a seed-stage VC tracking AI infrastructure, Crichton is the only voice asking what happens to these bets when the geopolitical environment actively degrades the data and trust layers that AI infrastructure depends on.
Worth carrying forward
- Before evaluating any AI infrastructure startup, apply Not Boring's ROT test: can you actually measure value of output minus token costs, and does the unit economics improve or worsen at scale? If the founder can't answer this, it's a red flag regardless of benchmark performance.
- The Squint benchmark result (78% accuracy vs. 53% for Claude Code on industrial multi-document questions) is a concrete template for how to pressure-test 'we beat GPT-4' claims — always ask: on what domain-specific dataset, with what proprietary context layer, and does 78% actually clear the bar for production use?