Chip CEOs warn of an AI capital crunch

CNBC reports that Intel and Broadcom's CEOs see access to capital becoming a major constraint on AI expansion. Following our October 9 coverage of OpenAI's revenue, this is the supplier-side view of how AI infrastructure expansion gets funded.

Tencent plans a roughly $5 billion bond offering

Bloomberg via Techmeme

Bloomberg sources say Tencent plans to fund AI expenditure through a bond offering, with most of the proceeds going toward chip purchases.

Bridge seeks to expand its loan to about $6.6 billion

Bloomberg via Techmeme

Bain-owned Bridge Data Centres is in talks to increase an existing $2.8 billion loan to about $6.6 billion. The financing remains under discussion.

SWI raises €800 million for data-center upgrades

Wall Street Journal via Techmeme

According to The Wall Street Journal, SWI raised €800 million to upgrade data centers, including facilities from Genesis Digital, the bitcoin miner caught in FTX's bankruptcy.

Anthropic separates session history, the agent loop, and tools

AI Engineer

Anthropic engineers describe an append-only session log, a separate harness and model, and interchangeable execution environments. A proxy holds credentials outside the model and sandbox; a sandbox crash becomes a failed tool call, allowing the sandbox to be replaced.

Dreaming revises memory and prompts from past sessions

AI Engineer

Anthropic's research preview sends earlier session transcripts through a sub-agent analysis loop to identify skill gaps and revise memory and prompts. The presenter says those changes require evaluation because the optimization paths can diverge.

AWS tests agent sandboxes across 1,000 cases

AI Engineer

Amazon's team had a coding agent generate tests across ten domains and found separate problems in the MCP interface and sandbox environment. Browser workloads hit memory limits, while coding workloads exposed persistence and package-path problems.

Browserbase persists lessons from failed browser runs

AI Engineer

A task agent and a reviewing agent analyze traces and screenshots, then preserve successful workarounds as reusable instructions. In the presenter's OpenTable demo, reported run cost fell from $6.22 to $0.51.

Friendly AI routes requests to reusable prefix caches

AI Engineer

Friendly AI describes cache-aware routing and tiered storage for the key-value cache to avoid recomputing repeated input. The presenter reports customer benchmarks with more than sevenfold faster GLM inference than other providers.

Cerebras and Gimlet split prefill from decoding

AI Engineer

The presenters separate compute-intensive prompt processing from bandwidth-limited token generation and route those phases to different hardware. Gimlet reports three- to tenfold faster inference under the same power constraints.

US restores UK safety testers’ access to frontier models

The Telegraph via Techmeme

The Telegraph's sources say the US restored the UK AI Security Institute's access to the latest US models for safety testing. Last month, it had ordered OpenAI and Anthropic to stop sharing models.

Brad Smith adds a legal duty to victims of agent incidents

The Guardian

Following yesterday's Nadella item, Microsoft's president calls for kill switches and a legal obligation to victims whose systems are breached by rogue AI agents. The liability proposal adds to Microsoft's call for external controls.

AI-generated work creates more human verification

The New York Times via Techmeme

Checking AI output creates additional human work, The New York Times reports. Its account brings together economic research and corporate surveys reporting limited efficiency gains.

Stripe reports revenue growth alongside new billing models

AI Engineer

Stripe's presenter projects 175% AI-company revenue growth and describes hybrid pricing that combines base fees with usage charges. That projection measures sales growth, while the productivity research measures efficiency gains.