AMD announced a definitive agreement on August 6 to acquire Taalas, a Toronto startup founded in 2023 that builds what it calls model-specific integrated circuits — chips that hardwire a trained model's weights into the wiring of the silicon itself rather than streaming them from high-bandwidth memory. The startup's first test chip, the HC1, fabbed on TSMC's 6nm process, served Meta's Llama 3.1 8B at roughly 17,000 tokens a second — a figure Taalas says beat Nvidia GPUs by 48x and Cerebras accelerators by 8.5x when it was announced in February.
The trade is stark: a chip built this way loses all flexibility. Change the model and you need a new mask set, though Taalas claims only two metal layers need re-etching for a refresh. AMD plans to fold the technology into its Helios rack-scale platform alongside Instinct GPUs and EPYC CPUs, targeting fixed-model workloads like customer service bots and document pipelines where throughput per rack-watt matters more than versatility.
The deal lands seven months after Nvidia paid a reported $20 billion to license inference technology from Groq — a different bet on the same problem. Terms of the AMD-Taalas deal are undisclosed; close is expected in Q4 2026, pending regulatory review.
OpenAI said this week that free and Go-tier ChatGPT users will get unlimited text-based chats, with the change rolling out next week. Free and Go accounts are also switching to GPT-5.6 Luna as the default model, replacing GPT-5.5 Instant, and picking up a new "Think" button that lets users request higher reasoning on harder questions. OpenAI says Luna makes 62% fewer factual mistakes than its predecessor in internal evaluations.
The limits that remain are for file uploads, images, voice and image generation — text is the only thing going fully unmetered. Plus and Pro subscribers separately get an upgraded GPT-5.6 Sol model tuned for more concise, reliable everyday answers, plus a reasoning-effort slider.
The move quietly narrows the gap between the free tier and the $8/month ChatGPT Go plan, whose main selling point had been exactly this: fewer limits on the base model.
Following June's executive order, the administration briefed roughly a dozen AI companies — including OpenAI, Anthropic, Google, Meta, Microsoft and Nvidia — on the finalized framework for reviewing "covered frontier models" before release. Under the voluntary program, developers can give the government secure access to a model for up to 30 days before it reaches other trusted partners, aimed at assessing cyberattack and vulnerability-discovery capability.
The framework defines a covered model as closed-source with state-of-the-art capability and national-security risk — without defining either term precisely — and explicitly excludes open-weight models from review, even after release. The administration has said it will not publish the framework's contents.
Critics note the carve-out creates a two-track system: closed labs face oversight and delay, open-weight releases from anywhere, including China, do not. The White House says participation remains voluntary and it does not "approve" releases.
NVIDIA Labs open-sourced NOOA (NVIDIA Object-Oriented Agents), a model-agnostic Python framework that represents an entire AI agent as a single class — methods are its actions, fields are its state, docstrings are its prompts, and type annotations become contracts the runtime enforces. A method left with an empty body is completed at runtime by an LLM-driven loop; anything else stays ordinary, deterministic Python.
The collapse of prompt templates, tool schemas and workflow graphs into one inspectable object pays off on benchmarks: a 253-line agent hit 82.2% on SWE-bench Verified with GPT-5.5, ahead of comparable open harnesses, using roughly half the tokens by keeping tool results out of the context window via pass-by-reference instead of round-tripping them.
NOOA was released as part of Nvidia's Open Secure AI Alliance contribution, alongside open models, weights and agent-harness research the company says the community can build on and contribute back to.
In March, Rippling's CFO told the executive team the company was on pace to spend 40% of its R&D headcount budget on AI tokens — millions of dollars, growing 80% month over month, roughly equal to the pay of 40% of the engineering org. This week the workforce-management platform launched AI Spend Console, a product born from fixing its own bill.
The tool ties token spend across Claude, Cursor and Codex to employee identity and downstream output — pull requests, code velocity, whether colleagues had to redo an engineer's work — then routes requests through an internal AI gateway to the cheapest model that fits the task. Rippling says the console dropped its own token spend from 40% of headcount budget to about 15%, even as usage kept climbing to 600 billion tokens in July.
The launch is a signal in itself: after a year of unmetered AI experimentation, enterprises are moving from "how much are we spending" to "what is it actually producing."