A trillion-dollar capex warning, a memory-chip selloff, and a widening gap between what things cost and what they are worth all point at the same mechanism underneath the headlines.
Follow the Money: A Trillion-Dollar Bet Financed on Faith
Three stories from three different desks point at the same fragile foundation: the capital under the AI boom is not as solid as the narratives built on top of it.
- BIS Warns AI Capex Boom Is Financed Like Past Manias. The Bank for International Settlements says more than a trillion dollars in AI spending rests on debt and circular equity deals it last saw before the dotcom bust. The central bank of central banks does not usually issue mania warnings this bluntly.
- SK Hynix Rebounds After Record Rout, Kospi Falls Into Bear Market. A two-day slide erased and then partly restored SK Hynix’s value just days after its $26.5 billion Nasdaq debut, and South Korea’s benchmark index fell into a technical bear market alongside it. Memory chips were supposed to be the steady trade inside the AI boom, not the volatile one.
- Anthropic’s $10M Canada Pledge Arrives Mostly as API Credits. Anthropic is funding eight Canadian research institutions as part of a $10 million commitment, but most of that figure is Claude access rather than cash. The gap between a headline dollar amount and what actually lands in a lab’s bank account is becoming its own story.
The Expensive Part Was Never the Model
Four stories make the same discovery from different angles: the cost that matters is not what a model charges per token, it is what it takes to get a job actually finished.
- Fable Costs Twice as Much as Opus. It’s Still Cheaper to Run.. Cognition’s Fusion architecture shows a model priced twice as high per token can finish an agentic task for less money overall, by delegating earlier and editing less. Price per token and price per finished job are turning out to be two different numbers.
- A New Benchmark Says No AI Agent Can Finish a Long Terminal Job. A 46-task suite finds that even the best frontier model fails most long, stateful terminal work, and that price barely predicts which agent finishes the job.
- Open-Weight Models Hit 29% of AI Gateway Tokens, Still Under 4% of Spend. Vercel’s June data shows the volume-to-spend gap widening, with DeepSeek and a new Z.ai model doing most of the pulling while the frontier labs keep the revenue.
- Monzo Co-Founder Tom Blomfield Leaves YC to Join Anthropic’s Compute Team. He is stepping away from Y Combinator to help manage the resource now bottlenecking every frontier lab: compute, not model design.
The Ghost in the Machine: What Agents Do When Nobody Is Watching
The trust questions stopped being hypothetical this week. Each of these stories is about what an agent does with access when no human is in the loop.
- Grok Build CLI Uploaded Whole Repos Before xAI Quietly Shut It Off. A wire-level teardown found the coding tool sent secrets unredacted and staged entire codebases to cloud storage, even when told not to read files.
- Manus Auto-Publish Cuts the Human Out of the Deploy Step. The AI app builder now pushes every successful build straight to a live URL, with no confirmation click required. The question shifts from whether the code is right to who answers for it when it is wrong.
- Google Open-Sources Mantis, a 15-Stage AI Pipeline for Hunting Bugs. Fifteen chained coding-agent skills triage, reproduce, and patch vulnerabilities, and Google’s own documentation insists a human still signs off first.
- Microsoft Gives Its AI Agents Identities, Retry Loops, and Grades. Foundry’s production playbook treats retrieval as its own subagent and scores agents against rubrics instead of vibes.
The Bigger Argument: What Cheap Intelligence Is Actually Changing
Three writers took the long view this week, and none of them agree on where it lands.
- Zvi Mowshowitz Names GPT-5.6 Sol AI’s New Default Workhorse. He argues Sol delivers the strongest balance of reasoning, speed, and cost for demanding knowledge work, while conceding task-specific model selection still matters at the top end.
- A16z’s Essay Says AI Turns Consumers Into Builders. The firm argues AI restores creation over consumption. The firm funding that shift is also not a neutral observer of it.
- Circle’s CEO Argues AI Will Unbundle the Firm Itself. Jeremy Allaire’s new treatise says cheap intelligence dissolves the corporation, a claim Ronald Coase’s transaction-cost theory can actually test.
Today’s Quick Hits
- Prime Intellect Rebuilds Its RL Environment Stack With Verifiers V1. The release splits agentic training environments into task, harness, and runtime layers to fix fragmented eval infrastructure.
- Sakana AI’s Smart Bricks Learn to Recognize Their Own Shape. Two hundred physical modules classified their own assembled shape using only local neighbor signals and no central computer.
- DeepMind Repurposes a Video Generator as a Universal Vision Model. GenCeption uses one text-steered network to match specialist systems on depth, pose, and segmentation while training on a fraction of the data.