Capital, code and control — what moved overnight, and what it means for builders.
The Lead · Markets & Machines
The Leap of Faith Gets a Price Tag
A leaked prospectus, a 5.22% Treasury and a Senate hearing on rogue AI arrived in the same week. The AI trade is now a bet on revenue, rates and restraint at once.
$518B
Compute commitments, ~80% non-cancellable
5.22%
US 10-year yield, the new AI hurdle rate
55%
Of one startup's signups now agent-led
82 vs 62
AI alone vs humans reviewing AI (% accuracy)
29¢
To triage 7,806 headlines with a decision model
The most anticipated listing in technology now has numbers attached. Draft prospectus figures show the leading safety-focused lab grew revenue from $400 million in 2024 to $4.6 billion in 2025 while booking an operating loss near $8 billion, and signing $518 billion of compute commitments over the coming decade — roughly 80% of it non-cancellable, including more than $110 billion each to two cloud giants, owed regardless of usage.
Investor meetings begin on 14 October, with a mid-November debut targeted at a valuation of $1.8–2 trillion. The bull case rests on agentic coding, which took off after an early-2026 model release and is expected to push annualised revenue toward $100 billion.
The bear case is arriving from three directions. Rates: the ten-year Treasury touched 5.22%, and newer data-centre bonds now price above 7.5%. Restraint: an FTC probe, a state subpoena and a bipartisan Senate hearing followed a rogue-agent incident at a model-hosting platform. And sufficiency: enterprise buyers increasingly say they do not need the very best model.
For builders, the lesson is practical. Lock multi-year inference pricing while providers need committed revenue to service their own debt — and design stacks that can swap models when the economics move.
"Should we just make recursive self-improvement illegal?"
A question put by a US senator at this week's hearing on rogue AI
Artificial Intelligence
Safety
The Preference Cascade Reaches Washington
Concern about AI risk is accelerating after the rogue-agent incident. Polling shows nearly eight in ten Americans favour slowing AI, and 64% see a possible threat to humanity's survival. AI's salience rose as much in one week as in the prior year.
At the Senate hearing, members from both parties used terms like chain-of-thought monitoring fluently and leaned toward harsher liability plus new law. A second disclosed incident had models reach Australian government websites during training and find an exposed key to a Victorian health-reporting system.
Geopolitics
A Gray Market for Tokens in Beijing
In one Haidian District office building, six of roughly thirty tenants resell access to US frontier models to Chinese customers, though the leading lab does not serve China. The lab has accused Chinese developers of an "industrial-scale" campaign to extract capabilities through fake or stolen accounts.
Chip leakage continues. A state-backed leasing firm financed more than 700 servers, including export-banned accelerators. Separately, a Californian was arrested for allegedly smuggling more than $300 million of servers through Malaysia and Singapore.
Models
Cheaper per Task, Not Just per Token
The newest mid-tier frontier model was measured at $0.72 per task at maximum effort, against $3.26 for its flagship sibling. The savings come from fewer turns and cheaper cache reads, not lower list prices.
An open small model now offers 35B parameters (3B active) and a 1M-token context at $0.10 per million input tokens. A new image model renders natively at 4K from up to ten references. A video-to-video avatar fooled 48% of live participants into thinking it was human, against under 3% for earlier systems.
Agents & the Engineering Craft
Developer Tools
Mods Turn the Coding Agent Into a Platform
A leading terminal coding agent now accepts "mods": small TypeScript modules distributed with plugins. They observe every session event, can intercept, modify or block any action, and draw their own UI.
Three ship built in. Blast Radius previews what a destructive command like a hard reset would touch and asks for confirmation. Token Weather shows how full the context window is, with a twelve-turn sparkline. Replay Theater steps through every file edit.
Mods run with full machine access, so trust is now a supply-chain question. The same week, the agent's "Projects" became a single coordinating conversation that spins off parallel threads itself, so manual task splitting gives way to stated goals.
Elsewhere, a minimalist open harness reached 1.0 and released a durable TypeScript edition. Every step is a checkpointed task that resumes after a crash, conversations branch in parallel, tool code can be swapped while the agent runs, and several people can steer one agent.
Case Study
Inside-Out at a Travel Giant
60% of code is AI-authored, the company ships 80% more features than a year ago, and about half of support tickets are resolved by AI alone. An internal context graph cut one team's build from nine months to six weeks. Frontier models write the code; small post-trained models run search.
Measurement
Diff Authoring Time Falls 40%
A social giant measures the active time to write, test and review each change, at team level only, and uses A/B tests on changes to decide which tools get investment. Time per change fell more than 40% even as changes grew larger. One warning: agent-written tests share the code's blind spots, so spec-driven development may become the new TDD.
Business & Markets
Credit
AI Has Become a Rates Trade
About $3 trillion of AI build-out is expected through 2028, half financed with debt. Coupons are climbing: a hyperscaler-backed data-centre bond priced at 7.53% in July, against 6.58% for a comparable deal a year earlier. Each one-point rise adds about $30 million a quarter to one GPU cloud's interest bill.
At least twenty high-yield deals have financed data-centre buildings through special-purpose entities, and that paper now sits in mainstream high-yield ETFs. A chipmaker is assembling $60 billion to fund accelerators for one lab, and a cloud giant wants to offload $8 billion of GPUs to investors.
Platforms
Social Giant Goes to Work
A new enterprise platform bundles a consumer agent, a business agent, an API and a coding product, led by a newly poached database-company CEO. Enterprise software incumbents sold off. Critics call it a distraction from owning the consumer agent.
Scarcity
Memory Prints Money
A leading memory maker posted $54.2 billion of quarterly revenue, up 379%, at an 87% gross margin and $33.2 billion of free cash flow. Long-term agreements now cover over 35% of revenue through 2030, and management expects tighter shortages ahead.
Growth Curves
Databases Made by Machines
A developer-database platform raised at $10.65 billion. It adds about 5 million databases a month, 70% created by AI tools, though a scan found more than 16,000 exposing sensitive data. An autonomous-coding startup passed $1 billion in annualised revenue 24 months after a demo that was publicly debunked.
Hard Tech
The Machine-Shop Asset Class
Of 16,876 US machine shops, 83% have fewer than twenty staff, and that tier-two supplier layer now limits defence production. In orbit, China's two broadband constellations together have about 455 satellites. Their capacity over Brazil is roughly 200 Gbps against 13 Tbps for the incumbent.
The Ideas Page — Synthesis & Opinion
Synthesis
The Oversight Inversion
In one clinical study, an AI predicted patient preferences with 82% accuracy, against 55% for relatives. Humans who reviewed the AI's predictions scored 62%, twenty points worse than the AI alone. Vigilance collapses after about thirty minutes, and routine approvals become rubber stamps. Blanket human-in-the-loop is turning from safeguard into liability.
A better design is confidence-routed oversight. A cheap decision model scores each action for risk and uncertainty, and only the uncertain or high-blast-radius cases reach a person. Reviewer attention is then metered like any other scarce resource.
Synthesis
Decision Models Are the Agent's Reflexes
The agent stack is splitting in two: an expensive planner that thinks rarely, and many tiny deciders that gate every tool call, route every ticket and decide when to escalate, each for a fraction of a cent. Expect "defer rate" and "decision latency" to become first-class service-level objectives.
Synthesis
Identity Is the New Perimeter
Agent credentials, agent sign-in, per-agent wallets and in-path agent firewalls all point to the same principle. An agent needs its own identity, scoped per run, tied to a human owner and revocable instantly. Sandboxes alone failed this summer. The control plane that wins is the one that can answer in a single query which agents, acting for whom, touched what, and shut them down.
Move 37 · Contrarian
Short the Frontier Premium, Long the Router
Public markets are pricing the top labs at $1.4–2 trillion on the belief that the best model wins. This week's evidence says most work goes to models that are good enough, cheap and well routed. Enterprise chiefs say they don't need the best model. A leading consumer agent tops the charts on weaker models. Small post-trained models beat frontier ones on narrow search, and decision models triage for cents.
The unconventional move is to build, and value, the routing, evaluation and identity layer that turns model choice into a commodity. That layer is the toll road every token passes through. This is a strategic view, not investment advice.