Three new cards from three operators. Two themes cut across the scan: who captures the economics in the AI stack, and how to properly delegate to models that can now run for hours without oversight. The third card stands alone. It is a quantified counter to the "Friendslop is saturated" thesis that has circulated in game-dev communities this year.
Theme 1, The ownership layer now determines who keeps the AI economics
Tomasz Tunguz published "Microsoft Resells the Frontier" on July 29, analyzing Q2 2026 earnings across the major cloud providers. The finding is quantified and specific.
Full-stack AI ownership at the model and chip layers produces nearly 2x the revenue growth of the resell-and-rent approach names the mechanism: a cloud provider that owns both its foundation model and its custom silicon eliminates vendor margin at two layers. Google pays neither Nvidia (it runs TPUs) nor an external model lab (it runs Gemini). Azure pays both: OpenAI for the model and Nvidia for the silicon. That double margin cost explains most of the growth gap: Google Cloud at 82% YoY, Azure at 43%, AWS at 28%. Strip out OpenAI's own Azure commitment ($220 billion of the $678 billion contracted backlog) and Azure's organic growth is approximately 25%.
The vulnerability Tunguz names: "every model is substitutable." When a capable open-weight model matches OpenAI benchmark performance, OpenAI has an option to reduce its Azure commitment and the backlog concentration risk materializes. Google Cloud's operating margin expanded from 20.7% to 35.6% in the same period. The retained margin from ownership compounds into reinvestment that a reseller's economics cannot match.
"Google owns its models & chips & spends aggressively on the margin that ownership captures"
- Tomasz Tunguz, Full-stack AI ownership at the model and chip layers produces nearly 2x the revenue growth of the resell-and-rent approach. Full-stack ownership at the model and chip layers captures nearly 2x the cloud revenue growth of the resell-and-rent approach; Google Cloud's 82% YoY versus Azure's 43% reflects retained margin at both layers.
Theme 2, Autonomous frontier models require a different delegation protocol
Dan Shipper published "Taming Opus 5" in Every's Context Window column on July 28, following Every's internal testing of Claude Opus 5 across coding, writing, and knowledge work. The card is a companion to Scaffolding built for a less capable model constrains a stronger one; upgrading models sometimes requires deleting the scaffold, not adapting it from July 24. That card covers when to delete scaffolding entirely; this one covers how to construct the brief before the run.
Front-load the full brief and evaluate the finished artifact; interrupting an autonomous run resets the model's planning state names the operational claim: autonomous frontier models are optimized for single-run execution, not iterative chat. The correct protocol is to write the complete brief in the first prompt and let the model run to completion. Mid-task correction resets the model's planning state. The context narrows from the full-scope original brief to the current exchange, and output quality degrades.
Narration is the trap. Models describe their process while executing. Users who correct based on narration are responding to process commentary, not output quality. The correct review mode is the finished artifact, not the in-progress description.
The pruning corollary:
"Skills preserve assumptions about the model they were written for. Try them freely but treat each new model release as a reason to prune."
This pairs with yesterday's The AI harness now moves coding benchmarks more than the model does: the harness determines performance, and the brief is the first layer of the harness. An over-specified brief written for a weaker model constrains the stronger one. Both cards name the same upgrade-moment risk from different angles.
- Dan Shipper, Front-load the full brief and evaluate the finished artifact; interrupting an autonomous run resets the model's planning state. Autonomous frontier models need a complete front-loaded brief and an uninterrupted run; mid-task correction resets planning state and produces worse output than evaluating the finished artifact.
Standalone, Friendslop demand compounds; the cycle, not the ceiling, is the design constraint
Chris Zukowski published "Is Friendslop saturated?" on July 30 in response to concern in the game-dev community that the co-op party game genre has peaked. The data says the opposite.
Co-op party game demand resets every 3-9 months with a rising ceiling; the dominant title fades to 10% of concurrent players while total genre volume compounds each cycle documents the cycle pattern across five dominant titles from December 2023 through July 2026: each one captures approximately 50% of total genre concurrent players at peak, then fades to around 10% within 3-9 months. That fade is not audience loss. The total concurrent ceiling rises each cycle: 197,000 (December 2023, Lethal Company), 281,000 (March 2025, R.E.P.O.), 350,000 (July 2026, MECCHA CHAMELEON).
The market-size implication runs both directions. At this total scale, a sub-dominant title at 11.75% peak share (Zukowski's data for Mage Arena, attributed to GameDiscoverCo) sold 624,000 copies in one week for approximately $3,698,080 in net revenue. The structural gap that forms after each dominant title fades is commercial opportunity, not a warning sign.
"there is a fanbase that wants a new thing every couple months to play with their friends. It is a beast that is hungry and wants to be fed."
The execution risk is cycle timing, not genre saturation. A developer who misses the gap and launches into a 50%-share incumbent faces a harder market. The genre itself is not full.
- Chris Zukowski, Co-op party game demand resets every 3-9 months with a rising ceiling; the dominant title fades to 10% of concurrent players while total genre volume compounds each cycle. Co-op party game demand compounds each cycle rather than saturating; the dominant title fades from 50% to 10% market share in 3-9 months while total genre concurrent players rose from 197K to 350K across three cycles from 2023 to 2026.