Intelligence Track
4 June 2026
The Brief
Today's signals split cleanly into two registers: supply-side stress in Indian aviation and a maturation signal in AI tooling economics.
On aviation, the Air India and IndiGo capacity cuts are not a short-term scheduling adjustment — ATF above ₹1 lakh per kilolitre with a weak rupee is a structural constraint, and the cuts through August will produce a fare environment that is less forgiving for price-sensitive OTA traffic. Cleartrip's funnel — search, pricing display, alternate-date suggestions — was built for a more liquid inventory environment; that assumption needs pressure-testing now. On the tooling side, three articles converge on a single theme: AI costs are real, and the teams that manage them deliberately will outperform those that don't. Figma's stickiness numbers confirm that usage-based AI pricing is here to stay in design tooling; the tokenmaxxing piece makes the same case for LLM infrastructure; and the design systems article reveals a subtler cost — implicit gaps in component contracts that AI fills with generic patterns, quietly degrading product identity. The thread connecting all three: the gap between disciplined and undisciplined AI adoption is now measurable in both spend and output quality, and it will show up in product differentiation faster than most teams expect.
HeyDesigner · 4 Jun 2026
Design systems that treat components as agent-readable APIs still fail when AI encounters undefined states — unspecified props, unnamed spacing tokens, undesigned edge cases — and fills those gaps by defaulting to the statistical average of the internet rather than the product's intent. The author argues this is not drift (conflicting definitions) but absence: decisions that were never made explicit and therefore cannot be retrieved.
Coined against the backdrop of Jensen Huang's token-consumption challenge and enterprise AI overspend reports — including an Uber case that burned through its annual AI budget in four months — the piece diagnoses 'tokenmaxxing' as defaulting to frontier models for tasks that cheaper models handle adequately. Three fixes are proposed: deliberate model routing by task complexity, context window hygiene (stripping irrelevant history), and output format discipline to avoid verbose responses that inflate billing.
Why it matters
For product teams running AI features in production — search personalisation, itinerary generation, fare explanations — unmanaged model selection and context passing can produce 5–10x cost overruns relative to a disciplined routing architecture, directly affecting unit economics of AI-powered features.
Also in AI & Design
A provocation on how creative work is shifting from fixed visual outputs to system-defined, input-driven outcomes — where brand identity lives in the rules and tokens that generate interfaces, not in the artifacts themselves.