Friday, August 28, 2026

Add ZDR passthrough to every model vendor contract now, or architect for portability — your own security questionnaires will inherit your provider's retention policy.

Nvidia's 2027 price increases mean your hardware and inference cost curves stop falling — budget for that before you sign multi-year compute commitments.

If your portfolio includes neoclouds or offshore GPU capacity in Southeast Asia, September comment drafts will tell you whose customer base is about to be audited.

Anthropic's Model Hardware Standard makes wet-lab automation a platform play — the instrument vendors and CRO middleware in your pipeline just got a new dependency.

The three-variants-then-annotate loop is a cheap standard to impose on your design-adjacent agent work; so is banning model-invoked skills.

Anthropic's regulatory overhang clears just as it prices an IPO — and GLM-5.3 Flash at $0.15 in means your cheap-tier routing should be re-benchmarked this week.

Run the competitor test against your portfolio companies' homepages this week; most will fail it, and the repo gives you a free remediation path.

The bio-uplift argument is the strongest case for Anthropic's retention posture — weigh it before you treat mandatory logging as pure customer hostility.

The software repricing was sentiment, not fundamentals — if you shorted the SaaSpocalypse thesis, cover; if you're selling a SaaS asset, the window is now.

The inference layer, not the model layer, is where Chinese open weights monetize — Modal, Baseten and Fireworks are the tradeable expression of that thesis.

Nvidia now guarantees OpenAI's credit and owns Hugging Face — your Nvidia exposure is increasingly a levered bet on OpenAI's ability to service that paper.

Use the two-questions split in board meetings: valuation belongs to you as investor, throughput belongs to your operators, and conflating them stalls deployment.

A $45B mark on $900M ARR with unresolved gross margins is the comp every coding-agent company in your pipeline will now anchor to.

Nothing operational here; read it in two minutes on a Friday and move on.

If you hold consumer-facing assets, the sentiment-spend divergence argues against trimming Q4 inventory or marketing on macro pessimism alone.

Harness engineering is now a defensible cost lever — a 60% reduction beats any vendor negotiation, and Visa just gave you the reference implementation free.

Maintenance load, not capability, is what kills per-person agents — consolidate to one shared agent with per-user auth before your team scales its Claw count.

If you hold Anthropic secondaries, liquidity may come at the IPO rather than six months later — but plan for a longer lockup on whatever you keep.

The most-responded-to post of her career identifies an underserved enablement segment — worth noting if you back AI training or upskilling companies.

Tworek and Zelikman on one stage is the actual reason to go — both are running new companies and will be pitchable in person.

Self-regulation, not statute, is the shape federal AI oversight is taking — position your compliance posture around industry-body membership, not agency rulemaking.