GT Axis Daily, August 2, 2026

The argument is a loop: cheaper intelligence means more work is worth doing, more work done means more revenue, more revenue funds the next round of research. The numbers she put behind it are the news. OpenAI's models now reach more than a billion active users and over two million businesses. Six months after signing up, the average person sends roughly 50% more messages per day and uses the tool for about twice as many kinds of work. Inside OpenAI itself, agentic work through its coding tool now accounts for 99.8% of weekly output tokens, meaning almost none of the work is a human typing a question anymore. The line I'd underline for any leader: the right measure is not the price per unit, it's the cost of a successful outcome including the retries, the oversight, and the errors.

On a hard reasoning test called ARC-AGI-3, the score went from 13.3% to 38.3% while using six times fewer output tokens, purely by turning on retained reasoning, meaning the model keeps its own working notes between steps instead of starting fresh, and better context management. The model was identical before and after. This is the most practical AI story of the week and the least glamorous. If your AI pilot is producing mediocre results, the odds are strong that the problem is in how you've wired it up, not in which model you picked.

Lilian Weng, who co-founded the startup Thinking Machines after leaving OpenAI, stepped away from that company citing health reasons and rejoined OpenAI days later. She'll lead a team focused on recursive self-improvement, which means AI systems that speed up their own research and get better at building the next version of themselves. Two things to take from this: the talent in this industry moves in circles and rarely leaves it, and the most consequential research direction at the biggest lab is now openly named.

Under an executive order signed June 2, Treasury, the NSA, and CISA had 60 days to build a classified benchmarking process that designates certain frontier models as "covered" based on advanced cyber capability, and to set up a voluntary framework where developers can hand the government early access for up to 30 days before public release. To be clear about what did and didn't happen: August 1 was a deadline for the government, not for the companies. Nothing became mandatory. But once a scoring system exists and the NSA director holds the pen on designations, you're one bad incident away from voluntary becoming required.

It announced Project Perception, a system of specialized agents working in teams: red agents that attack your systems looking for weaknesses, blue agents that investigate real threats, and green agents that go fix what the other two found. They run in continuous loops rather than waiting for a human to kick them off. The honest framing is that attackers already automated their side of this, so defense automating too isn't optional, it's catching up. For a business leader, the question this raises is not whether to buy it, it's who on your team is accountable for reviewing what an autonomous agent changed in your environment at 3am.

Defender added prompt injection protection in preview, which identifies and isolates emails containing hidden malicious instructions before they reach the inbox. Prompt injection means text written to hijack an AI system rather than to fool a person, for example an email with buried commands telling your assistant to forward the contract folder. It's the phishing attack of the agent era, except the target is your software rather than your employee, and no amount of staff training helps. If your company has connected an AI assistant to email, this is the threat model you should be briefing your board on.

It closed its acquisition of Modular, whose work amounts to a common software foundation that lets AI models run across different chip brands instead of being locked to one. That matters because the hardest part of competing with Nvidia has never been the silicon, it's the decade of software everyone has already written for Nvidia's chips. Qualcomm just bought a shortcut around that wall.

Gemini Robotics ER 2 acts as the high-level brain, watching continuous video to track its own progress and hand off the physical motion to whatever lower-level system actually moves the arms. It hits 91.3% accuracy on identifying the exact moment a critical event happens, like when to stop pouring, at about four times the speed of larger models. It also lets different robots collaborate, so a wheeled robot indoors and a humanoid on rough ground can split a job through shared understanding. Knowing when a task is finished sounds trivial and has been one of the hardest unsolved problems in robotics.

CEO Cristiano Amon put it plainly on the earnings call: costs went up, so prices are going up. The driver is a memory and wafer squeeze running through the whole supply chain, from fabrication to packaging to test. Snapdragon chips sit inside most premium Android phones, plus tablets, smartwatches, Windows laptops, smart glasses, and connected industrial hardware, so this filters into a very long list of things businesses buy. The uncomfortable read: the AI buildout is consuming so much memory that it's now raising the price of devices that have nothing to do with AI.

Etched raised at a $10.3 billion valuation, led by Sequoia with Andreessen Horowitz, Jane Street, and memory maker SK Hynix joining. Its chip runs transformers and nothing else, transformers being the specific architecture underneath essentially every major AI model in use today. A general-purpose chip is a Swiss Army knife; this is a scalpel. The company says it has more than $1 billion in customer contracts, a Taiwan factory, and a 10-megawatt test lab fifteen minutes from its office, which tells you it's shipping rather than prototyping. The risk is right there in the design: bet the entire company on one architecture and you're brilliant until the architecture changes.

Applied Digital, which builds and leases the data centers this all runs in, closed the week up 0.7% at $27.39 and is down 22.9% over the past month. GlobalFoundries, which took the largest award in last week's $874 million federal chip research package, is down 35.3% over the same month. Both are down hard while Microsoft, Amazon, and Alphabet posted double-digit weeks. That gap is either the market correctly identifying who captures the value in this buildout, or it's the setup for a rotation back. Worth watching rather than acting on.