Four throughlines run today, starting with who pays for AI’s inference boom. Cerebras’ new CS-4 pairs with AMD and AWS chips instead of replacing GPUs, Jane Street tested Etched’s chip before leading a $700 million round at a $21 billion valuation, and Nvidia increasingly finances the demand it supplies.
Follow the Money: The Inference Hardware Fight
Three stories, one question: who is actually paying for the chips that run inference. Cerebras concedes GPUs aren’t going anywhere, a trading firm turned a pilot into a $21 billion bet, and Nvidia bankrolls the customers buying its own chips.
- Cerebras Ships CS-4, and Quietly Concedes GPUs Aren’t Going Away. Cerebras’ new wafer-scale chip pairs with AMD and AWS silicon for prompt processing, an admission that GPUs still own roughly half of inference work.
- Jane Street Tested Etched’s Chip, Then Led Its $700M Round. The trading firm ran Etched’s hardware in its own datacenter before pricing the company at $21 billion, a customer-as-investor structure worth watching.
- As AMD and Google Close the Gap, Nvidia Leans on Its Balance Sheet. CNBC reports Nvidia is now financing the AI buildout it once merely supplied, a strategy that flatters chip demand while parking the risk on its own books.
The Restraint Playbook: Choosing Control Over Speed
Four separate actors chose to slow down, lock down, or defend against misuse this week rather than ship faster. The throughline is control, over training, over governance, and over what a stripped-down model is allowed to say.
- OpenAI Publishes Its Own Account of Slowing Frontier AI. OpenAI says it paused RL training and froze its largest frontier run over an unreleased model’s cyber risk, though the broader safety framework it promises is still unbuilt.
- Anthropic Plans Founder Supervoting Shares Ahead of IPO. Reuters reports Anthropic is designing supervoting stock for founder Dario Amodei, whose own stake sits near 2 percent of the company.
- A Runtime That Checks AI Agent Permissions at Every Step. The runtime blocked or altered 94.8 percent of policy-violating agent actions but finished only 86.9 percent of legitimate tasks, a tradeoff worth pricing before deployment.
- Fool’s Gold: A Defense That Makes Stripped Models Lie. A Microsoft Azure researcher’s defense trains open-weight models to answer hazardous prompts wrong, and with confidence, once an attacker strips their safety tuning.
Shrinking the Frontier: Big Models on Small Machines
Frontier-scale performance keeps landing on hardware that used to be too small for it. The tradeoffs between local and cloud inference, and between a flat price and a real bill, are getting measured instead of assumed.
- A 753B-Parameter Model Now Fits on One Workstation GPU. FreeToken, a new serving system, claims a 753 billion parameter GLM model can run on a single workstation GPU by re-deciding what stays in fast memory.
- A Laptop-Sized Model Ties the Cloud on Quality, Not Speed. Tomasz Tunguz’s own test found a 27B local model matched a cloud model’s answer quality but took nearly seven times longer to respond.
- Z.ai Keeps GLM-5.3’s API Price Flat. The Real Bill Isn’t.. The per-token rate matches GLM-5.2, but Artificial Analysis puts the cost per finished task about 55 percent higher because the new model talks more.
Agents Grow Up: From Demo to Infrastructure
Warp, Harvey, and Liquid AI are all publishing what actually happens when agents run in production instead of a demo. The pattern is agents becoming infrastructure, with the failure modes, memory systems, and scaling problems that implies.
- Warp Turns Its Software Factory Pitch Into a Shipping Product. Warp Factories packages agent orchestration, memory, and evals into one product, though pricing, customer numbers, and defect rates are still missing from the launch.
- Harvey Gives Legal AI Agents Memory That Persists Across Matters. Harvey II lets agents carry matter context and a lawyer’s preferences between tasks, and the launch post leaves cross-client confidentiality mostly unaddressed.
- Liquid AI’s Coding Agents Passed Every Test, Then Broke at Scale. Every failure that mattered was invisible to the agents’ own tests and only surfaced once real production data and an outside verifier entered the loop.
- The Packfile Problem: Why Git Doesn’t Scale Like a Normal Server. Cursor’s engineering blog explains why Git’s core storage format defeats standard scaling tricks, and ranks the three architectures host builders have tried, including its own.
- Miles v0.1 Fixes the Stall That Wastes GPU Time in Agent RL. RadixArk’s RL framework now updates models mid-rollout instead of freezing the pipeline, and adds the sandboxing and hardware support production runs need.
Quick Hits
The rest of what moved today, in one line each.
- Why Thinking Machines Built Inkling Without a Vision Encoder. Thinking Machines built Inkling to train images and audio inside the same 66 layers as text, with no separately pretrained vision encoder bolted on, a bet on retraining over raw benchmark scores, per ByteByteGo.
- Why Two AI Researchers Doubt the Data Moat Thesis. Ryan Greenblatt and Shuchao Bi argue in MBI Deep Dives that algorithmic progress and smarter data distributions, not proprietary data volume, are what is actually driving AI gains industry-wide.
- Vercel Offers Up to $1 Million to Researchers Who Can Escape Its Sandbox. A two-week bug bounty, tiered up to $50,000 per report, targets the Firecracker boundary that isolates AI agents’ untrusted code inside Vercel’s infrastructure, a direct incentive to find the flaw before an attacker does.