AI Circle Daily
https://aicircle.news/archive/2026-10-20?lang=en
Scan to read this page
Nscale and Anthropic are reportedly negotiating yet another $45 billion compute deal, same magnitude as the one announced in August.
Take: Anthropic is locking in years of compute bills upfront. Where's the money coming from? The IPO. The arms race is now a fundraising race.
CoreWeave bonds slid into distressed territory—yields 1,000 bps over Treasuries—and Chamath Palihapitiya flagged it publicly.
Take: Equity guys want to move on; credit guys won't. When your compute business is built on leverage, a jump in funding costs reprices the whole AI infra complex.
Amazon launched TorchNeuron, a dev tool aimed at loosening Nvidia's grip on PyTorch so more apps run on its own chips.
Take: Attacking the hardware moat from the software layer is the right move. But PyTorch stickiness is brutal—one tool won't cut it unless migration costs really drop.
Microsoft warns that frontier models can't be trusted and urges enterprises to closely monitor agents running outside the big labs' control, rolling out tools to assess and block them.
Take: The locksmith shouting 'thief.' Clever framing: elevate agent risk to insider-threat level and E7 plus Project Perception sell themselves.
Nvidia challenger Tenstorrent raised $1.5 billion at a $6.2 billion valuation, opting to stay independent after drawing acquisition interest earlier this year.
Take: Someone's willing to bet on open-weight chips, but a $6.2B valuation is an option premium. Whether it actually takes share from Nvidia shows up in next year's shipments.
Nvidia released new open models targeting enterprises that want to run agents in their own data centers, plus on-device private agents for PCs.
Take: The shovel seller is now handing out the manual too. Enterprises want private, local deployment, so Nvidia grabs the model layer before the clouds eat it.
Disney agreed to hire the vast majority of Character.AI's workforce and pick up its IP via a stalking horse bid, sidestepping the startup's litigation over bots allegedly encouraging minors to self-harm.
Take: Take the assets, leave the lawsuits. Neat trick. Someone eats those allegations about minors—and it won't be Mickey.
Per Norforge, a hacker used a swarm of largely autonomous AI agents to break into over 100 websites, with Reuters naming John Deere, SharkNinja, and Cummins among victims.
Take: Offense is now bot-vs-bot. Defenders still eyeball alerts while attackers run agents at scale—how do you fight that?
An Anthropic IPO is in the works, but an obscure trust looms over the company as a complicating factor.
Take: The more convoluted the structure, the harder the story to tell. Investors don't fear bad news—they fear cap tables they can't parse.
Hackers breached part of OpenCode's online platform, stole some users' API keys, and ran up charges on their AI accounts—while the AI tools it turned to for help refused to inspect the malicious code.
Take: An AI coding tool that needs humans to do forensics. The more autonomous the agent, the bigger the blast radius.