· 07:00 · 7 min + 2 min Explained
Anthropic reportedly spends $3.4B on its own gas generators
Anthropic has bought gas generators worth about $3.4 billion to power its data centers, the Financial Times reports. Also: Nvidia's board reviews its CEO, Claude in Copilot moves inside the EU, and Apple wants Siri to act on its own.
Written by AI and checked against our editorial standards. Every claim links to its primary source. Not reviewed by a human editor.
Get it in your inbox at 7am every day ↓
Today in 60 seconds
- Anthropic bought gas generators worth about $3.4 billion for its data centers, the FT reports. Anyone who wonders why AI limits and prices move.
- Nvidia's board is reviewing reports that a Chinese spy targeted CEO Jensen Huang. Anyone who follows the chips that AI runs on.
- Microsoft will run Claude in Copilot inside the EU by the end of 2026. European firms that use Microsoft 365 Copilot.
- Apple wants Siri to act for you as early as next year, Bloomberg reports. iPhone users and app makers.
1. Anthropic reportedly spends $3.4B on its own gas generators
The short versionAnthropic has bought gas generators worth about $3.4 billion to power its data centers, the Financial Times reports. AI labs now buy their own power, not only chips.
Every answer from Claude runs on a chip in a data center, and every chip needs electricity. So the limit on AI is not only chips. It is also power.
According to the Financial Times, Anthropic agreed in late September to buy heavy-duty gas generators from OnePWR Solutions in Texas. The deal is worth about $3.4 billion and covers 1.7 gigawatts.
That fits a promise Anthropic made in February: to bring new power online for its data centers, so local households do not pay more. It also said US AI will need at least 50 gigawatts within a few years.
The catch: this is a newspaper report. Anthropic and OnePWR did not comment, and Reuters could not confirm it.
It is like a busy restaurant that buys its own generator, because the city cannot lay a bigger cable fast enough. How fast Claude can grow now depends on power as much as on chips.
Sources: cnbc.com · Reported · anthropic.com · Official
2. Nvidia’s board reviews its CEO after reports of a Chinese spy
The short versionNvidia's board has started an independent review after reports that CEO Jensen Huang was the target of a Chinese spy. The board says it sees no sign that secrets leaked.
Nvidia’s chips are at the center of the AI race between the US and China. That makes its CEO a natural target.
On Sunday, the board said independent directors, advised by outside lawyers, will look into the reports. Huang supports the review. In a message to staff, he admitted an affair and apologized. He did not name the woman.
TechCrunch names her as Alice Huang, a former executive at the Singapore chip firm Megaspeed. A memo shared among US policymakers claimed that China’s Ministry of State Security recruited her to spy on him.
The catch: the spy claim comes from a memo that has not been published. The board says it has “no reason to believe” that national security or company secrets were compromised.
When one company sits at the center of a race between two countries, its CEO's private life becomes a security question. And US rules decide where Nvidia may sell its chips.
Sources: nvidianews.nvidia.com · Official · techcrunch.com · Reported
3. Microsoft will run Claude in Copilot inside the EU by the end of 2026
The short versionEuropean firms that use Copilot will get Claude processed inside the EU by the end of 2026, Microsoft says. Until an admin turns it on, Claude stays off.
Many European firms want their data to stay inside the EU. Microsoft promises that for Microsoft 365 through its “EU Data Boundary”. Claude was not part of that promise, so firms had to choose: Claude, or data processing in Europe.
Microsoft says that choice will go away. By the end of 2026, firms in the EU and EFTA should be able to use Claude across Copilot. It will run in the region where their data already lives.
For EU firms, Claude is off by default, and an admin can turn it on. Firms that already use it do not have to do anything.
The catch: there is no exact date yet, only “by the end of 2026”.
It is like a bank opening a branch in your own country: same service, but your money no longer crosses a border. For a European firm, that can turn a "no" from its lawyers into a "yes".
Do this: If your firm uses Copilot, ask your admin if Claude should be switched on. It can be done now.
Source: microsoft.com · Official
4. Apple wants Siri to act on its own as early as next year, Bloomberg reports
The short versionApple wants to make Siri more like an agent that acts for you, perhaps as early as 2027, Bloomberg's Mark Gurman reports. Today's Siri still waits to be told what to do.
The new personal Siri arrived with iOS 27. It knows more about you, but it still cannot take action on your behalf on its own. Many AI firms are working on exactly that.
According to Gurman’s Power On newsletter, Apple wants Siri to become “more like an agent”, and it could happen next year. Apple already has App Intents, a way for apps to tell Siri what they can do. That could let Siri act in apps without clicking around the screen like a person, which costs more.
Apple has also given engineers more time to try ideas for iOS 28: six weeks of coding and exploring, where a stage usually lasts four.
The catch: this is a report about plans, and Apple has said nothing. “As early as” next year can also mean later.
It is the difference between an assistant who answers questions and one who runs errands. For app makers, App Intents is the door such a Siri would use to get into their apps.
Source: 9to5mac.com · Reported
Open source & local AI
Microsoft’s small computer-use agent now runs in llama.cpp
The short versionMage-VL-Cowork, a small Microsoft model that can use a computer for you, now runs in llama.cpp. So it can work on your own PC, without sending your screen to the cloud.
An AI agent that uses your computer has to see your screen. With a cloud agent, screenshots of your desktop go to someone else’s server.
Mage-VL-Cowork is small enough for a normal PC. It reads screenshots, clicks and types, and runs terminal commands when that is quicker. On OSWorld-Verified, a test of real desktop tasks, it solves 66.0%. Humans reach 72.4%.
On Saturday, llama.cpp added support, and ready-made versions exist for Ollama and LM Studio too. The Apache 2.0 license lets firms use it freely.
The catch: these are Microsoft’s own results. An agent that runs commands can also delete things, so the model card says to keep it in an isolated container.
It is like a helper who works in your office instead of taking your files home. For anyone who may not send their screen out of the building, a local agent may be the only option.
Try it: Run it only inside a virtual machine or container.
Sources: huggingface.co · Official · github.com · Official
Paper of the day
AI agents waste about 18% of what they write, a study finds
The short versionResearchers measured how much of an AI agent's output does nothing useful. About 18% is waste, and the tool around the model matters more than the model.
When you pay for an AI agent, you pay for every token (a small piece of text) it writes, also when it repeats itself.
A team led by Huimin Zeng let a judge model mark text that could be removed. Then they removed it and ran the task again, to check that the result held. They compared five agent tools, called harnesses, on the same model.
About 18% of the output was waste, mostly detours that turned out not to be needed. Agents also restated their plan and read the same file again. In more than 45% of tasks, one agent’s waste alone was bigger than another agent’s whole output.
The catch: the cuts were tested on only two benchmarks. There, removing 41-46% of the reasoning left results about the same.
It is like two taxi drivers with the same car, where one takes detours and you pay the meter. The waste varied more between tools and tasks than between models, so the tool you pick can change your bill more than the model.
Source: arxiv.org · Official
Explained: if you have a little more time
3 terms from today's news · 2 minWhat is a gigawatt, and why do AI labs count in them?
A gigawatt (GW) is one billion watts. It measures power: how much electricity something uses or delivers at a given moment. The steps are simple. One kilowatt is 1,000 watts. One megawatt is 1,000 kilowatts. One gigawatt is 1,000 megawatts. A 40-watt light bulb uses a tiny part of that. You would need 25 million such bulbs, all switched on, to use one gigawatt. > **Think of it like this:** watts are like the width of a water pipe, and watt-hours are the water that runs through it. A gigawatt is a very wide pipe. It says how much can flow at once, not how much flows in a year. That difference matters. Your power bill counts kilowatt-hours: power used over time. When a power plant or a data center is described in gigawatts, it means how much power it can deliver or draw at its peak. AI uses a lot of it. A data center is full of chips that run day and night, and each chip needs power and cooling. Anthropic says that training a single frontier model will need gigawatts of power. It also says the US AI sector will need at least 50 gigawatts over the next several years. That is why AI companies now talk about power deals, not only chip deals. **Why it matters to you:** a deal for "1.7 GW" tells you how much AI work a lab can run at once. More power means room for more users and bigger models. It can also mean more pressure on the local grid.Two more from today
What is an AI agent?What makes an agent different from a chatbot, and why that changes the risks.
An AI agent is an AI model that does not just answer, but acts. It decides its own next step, uses tools like a browser, a code editor or a payment page, and keeps going until the task is done.
That makes it useful for open-ended jobs. It also means mistakes can pile up, so agents need limits on what they may touch.
What is a token?The small pieces of text an AI model reads and writes, and why you pay per token.
A token is a small piece of text, often a whole word or part of a longer word. AI models do not read letters or words. They cut text into tokens and turn each token into numbers.
Tokens are also how AI is priced. You pay for the tokens a model reads and the tokens it writes, so long texts and long agent tasks cost more.
How sure are we?Official: from the company, regulator or authorsReported: from a named outletUnconfirmed: rumor or leak