Everyone wants to own the stack.
Nine moves from one fortnight: who builds the chip, who sits in your Slack, and who gets to keep the best model.
If you run a business, you can skip most of this week's AI headlines. Three forces underneath them actually touch your operation: AI keeps getting cheaper, it keeps getting more independent, and your company data is the prize everyone is chasing. The teams pulling ahead start small, get one task reliable, and measure what they actually keep. Here is the fortnight through that lens.
- 01OpenAI now builds its own chips.
- 02The AI teammate moves into your Slack channel.
- 03The government pulled the plug on the best model.
- 04Anthropic says someone already copied the homework.
- 05Someone is selling Claude at 93% off.
- 06A model whose only job is picking other models.
- 07The agent now writes its own org chart.
- 08Give it a goal, let it run on a timer.
- 09Everyone wants off Nvidia.
OpenAI now builds its own chips.
OpenAI revealed its first custom chip, Jalapeño, built with Broadcom to run its models more cheaply than the Nvidia hardware it rents today. Nvidia still owns the harder job of training models, so OpenAI stays a major Nvidia customer for now.
The price of running AI keeps falling. When the biggest buyer in the market builds its own chip to cut costs, those savings reach your monthly bill eventually. A workflow that was too expensive to automate last year is worth pricing again this year.
The AI teammate moves into your Slack channel.
Anthropic replaced its Slack app with Claude Tag. You type @Claude in any channel and it takes the task, does it with your connected tools, and replies in the thread. It is shared: one Claude works with the whole channel and builds up that channel's context over time. Anthropic says it already writes 65% of the code on its own product team.
This is what an "AI employee" actually looks like day to day. It lives in the tool your team already uses, so adoption stops being a fight you have to win. Before you switch it on, settle the boring question first: which channels it can read, and what that exposes.
The government pulled the plug on the best model.
The US government ordered Anthropic to switch off its two most powerful models worldwide, for every customer, reportedly within 90 minutes. Officials feared the models were good enough at finding security flaws in code to be dangerous in the wrong hands. Anthropic disagreed in public, noting that other widely available models do the same task.
A tool your business runs on can disappear overnight, even when the vendor did nothing wrong. If one model is wired into a core process with no fallback, you are carrying a continuity risk you cannot see. Pick tools you can swap out, and keep a plan B for the one that matters most.
Anthropic says someone already copied the homework.
Anthropic told the US government that a group tied to Alibaba ran the largest copying attack it has seen on Claude: around 28 million conversations through fake accounts, used to train a cheaper rival model on Claude's best skills. Alibaba has not confirmed it, and an accusation is not yet a finding.
Your prompts and documents have real value, and someone is always trying to collect them. A few basic habits cover most of the risk. Know where your company data goes the moment your team opens an AI tool, and keep sensitive work on an account you control.
Someone is selling Claude at 93% off.
Resellers in China sell Claude access at up to 93% off the official price. They afford it by sharing pooled accounts, committing payment fraud, and quietly harvesting customers' prompts and answers to resell as training data. The cheap access is the bait; the data is the real business.
When AI access looks too cheap to be true, you are usually paying with your data. The same trap shows up closer to home, in free tools and no-name plugins your team installs. Buy from the source, read what a tool does with your inputs, and treat any 90%-off AI deal as a red flag.
A model whose only job is picking other models.
Sakana AI released Fugu, a system that sits on top of several existing models, sends each task to whichever one is best, and combines the answers. Sakana says it beats the leaders on almost every benchmark, though all of those scores are its own and still unchecked by anyone else.
Ignore the leaderboard wars. The only benchmark that matters for your business is whether a tool does your task right, on your data, every time. Test it on one real job before you trust the chart, because the company selling a tool always wins its own benchmark.
The agent now writes its own org chart.
The newest AI tools can now run a whole job on their own. Given an objective, a model plans the work, splits it across several copies of itself, runs them at once, then checks the results before replying. The catch is cost: a run like this can use 15 times the resources of a single chat.
AI is moving from answering questions to finishing jobs. That means bigger tasks can come off your team's plate, the multi-step ones that used to need a person to babysit them. It also means the bill grows with the ambition, so the real skill now is choosing which jobs are worth handing over.
Give it a goal, let it run on a timer.
A "loop" is an AI agent you give a goal and a schedule. It plans, does the work, checks its own result, and stops when the work passes. The pattern teams keep landing on is a ladder: get one run working by hand, save those steps as a reusable skill, wrap that skill in a loop with a clear pass-or-fail check, and only then put it on a timer. Skip a rung and the agent quietly burns through budget with nobody watching.
This is how you adopt AI without it turning into another expensive disappointment. Start with one task done by hand until it is reliable, then automate it. And track one number: cost per accepted change, the money spent per result a human actually kept. Tokens used and tasks attempted tell you nothing.
Everyone wants off Nvidia.
Qualcomm is in talks to buy the AI chip startup Tenstorrent for up to $10 billion. Tenstorrent builds chips meant to compete with Nvidia, whose hardware nearly every AI company depends on today. The talks are early and could still fall apart.
This is the second move in two weeks aimed at loosening Nvidia's grip, right after OpenAI's own chip. The headline for you is simple: more competition in AI hardware keeps pushing the cost of running AI down. Plan for it getting cheaper, and avoid locking into long contracts at today's prices.