Tuesday, September 22, 2026

Your commerce portfolio now splits cleanly: ad-dependent retailers will block agents, transaction-fee platforms like Shopify will court them.

Ask every incumbent software vendor for an AI discount now; they are cutting price to keep you from shifting spend to Anthropic and OpenAI.

Route classification workloads to Jev and stop paying frontier-model prices for yes/no calls; skip Grok 4.7 until the token economics improve.

Before your portfolio companies buy another eval vendor, make them show you 100 reviewed traces; the metric dashboards are measuring the wrong failures.

Anthropic just proved a Western lab's usage telemetry can trigger a Chinese regulatory probe — assume your API logs are geopolitical evidence.

Alibaba's 3x chip gain and 20GW plan mean your China compute-scarcity thesis has maybe four quarters left before it stops holding.

The physical-delivery layer is the durable asset in agentic commerce; price logistics-owning retailers as moats, not as disruption targets.

The EA framework underpinning your AI-safety-adjacent bets is now a press liability; expect founders to start distancing themselves publicly.

Tworek's first public outing as Core Automation CEO is the reason to attend; book the room before Wednesday.

Meta at $741 already prices in Muse working; the transaction-cut monetization model is the assumption to stress-test before you chase it.

Discount world-model pitches that promise near-term physical reasoning; the category's own leading founder says the approach is still missing something.

Leverage in AI vendor negotiations has flipped toward buyers this quarter; reopen your contracts before the IPO window reopens and it flips back.

The consulting economics of transformation work are being repriced downward — relevant if you hold anything in professional services.

The going rate for four weeks of agent training is $2,500 a seat with sold-out cohorts — that pricing signal is the useful datapoint here.

That $145B-to-$770B infrastructure curve is the number to bring to your LPs when defending compute-adjacent exposure.

Cursor's cost-per-task chart contradicts Ben's Bites' token-burn complaint; run your own workload before believing either number.

Huang's liability-over-regulation framing is the argument chipmakers will push in Washington; expect it in every policy conversation you enter this quarter.

Run one weekend building a ten-task personal benchmark and you will likely find a model hundreds of times cheaper does most of your routine work.

Instrument human interactions per PR across your engineering orgs this quarter; it exposes which agent deployments are actually throughput gains.

Copy Warp's public-Slack agent channel next week; it is the cheapest way to move your power-law agent users' skills to everyone else.

OpenAI is behind on consumer agents and its compute overhang is now the accelerant — reprice your assumption that ChatGPT distribution is a durable moat.