Sunday, October 4, 2026

TL;DR: The AI buildout's real fragility lives in the circular contracts and guarantees linking OpenAI, Nvidia, Oracle, and CoreWeave.

This maps the actual financing plumbing — vendor financing, backstops, take-or-pay commitments — connecting the handful of companies carrying the buildout. For an investor audience, the question of who holds the bag if demand slips is the central question of the cycle, and this piece names the specific contractual chokepoints rather than hand-waving about a bubble.

TL;DR: The veteran chip architect behind OpenAI's in-house silicon explains why a model lab is becoming a hardware company.

Cutress gets a rare long sit-down with the person actually running OpenAI's custom chip effort — a Google TPU and lightmatter alum now betting a lab can out-design merchant silicon. It's the clearest window yet into how OpenAI thinks about co-designing models and hardware, told through one engineer's career arc from EDA to the frontier.

TL;DR: AI is already improving itself, yet recursive self-improvement is producing diminishing returns, not a runaway takeoff.

Naam makes the sharpest empirical case against fast-takeoff: labs have had AI-assisted AI research for years and progress remains stubbornly incremental, bottlenecked by compute, data, and physical-world feedback loops. If he's right, the entire investment thesis shifts from 'singularity hedging' to a long, capital-intensive grind — a framing that will still matter in a year, not just a month.

TL;DR: China is doing to robotics what it did to solar and EVs: winning on manufacturing scale, not model quality.

A 9,000-word report arguing the humanoid race will be decided by supply chains, actuator costs, and factory deployment loops — all areas where China already dominates — rather than by who has the best foundation model. It's the counterweight to Silicon Valley's software-centric robotics narrative, with enough detail on specific companies to be genuinely useful.

TL;DR: Back-of-envelope math says 100M agent users need ~1–4GW of power — explaining Meta's rumored 7–10GW buildout.

A genuinely rigorous bottoms-up model — concurrency assumptions, peak-to-average ratios, CPU vs. inference split — that lands on a surprising result: the sandbox/VM layer is tiny (~$3B of CPU and DRAM) while inference dominates everything. You bookmarked it, and it rewards the attention: this is the kind of unit-economics scaffolding you can reuse on every consumer-agent announcement.

TL;DR: Gates now believes AI self-regulation is 'insane' and the people in charge are in denial about what's coming.

The arc here is the story: the original techno-optimist, now redirecting a $200B foundation, arguing government must act before cyberattacks and bio-risk materialize. Klein presses him on the tension between his fear and his continued investment in the technology, and Gates's visible discomfort is more revealing than any policy paper.

TL;DR: Instinct's founder on building an app-less agent users trust with credit cards — and the compute bill behind it.

Shinn is candid about the operational reality behind viral agent growth: proactive AI is brutally compute-hungry, and trust (not capability) is the real moat when an agent holds your payment credentials. His 'Instinct-to-Instinct network' vision — agents negotiating with agents, rewriting the internet's business model — is the most concrete articulation yet of where consumer AI demand for compute comes from.

TL;DR: OpenAI's 3–8 year compute contracts sit atop 15–20 year leases — someone else is holding the duration risk.

The piece's core insight is the duration mismatch: OpenAI has structured its commitments so Oracle and landlords like SB Energy hold the buildings long after the compute contracts expire — though at two campuses OpenAI has started signing decades-long leases itself. It's a precise anatomy of who actually bears the terminal-value risk in the buildout, grounded in filings rather than vibes.