SoftBank's $7bn AI power write-off, Claude gets spreadsheets, Nvidia's one-demo robots
SoftBank lost about $7 billion in a quarter, largely because it paid in advance for power equipment its US data centres no longer needed. On the same day Anthropic put a spreadsheet and a whiteboard inside Claude, and Nvidia showed a robot model that it says learns a new task from a single demonstration. The thread running through all three: AI is leaving the chat window for offices, factories and power yards, and the bills and the work are following it.
Key takeaways
- SoftBank's $3.7 billion write-off on prepaid power equipment shows that committing early to AI capacity can be as costly as committing late.
- Claude Sheets is deliberately simple for now, but a spreadsheet you build by describing it lowers the bar for every small business that still runs on trackers.
- Nvidia says GR00T 2 reaches 71% average success on 50 robot tasks after one demonstration each, a self-reported figure that still needs independent testing.
SoftBank books a $7 billion quarterly loss after writing off prepaid power equipment for US AI data centres
SoftBank Group reported a net loss of about 1.1 trillion yen ($7 billion) for July to September, including a $3.7 billion hit on advance payments for power equipment at US data-centre projects whose plans changed.
SoftBank Group, the Japanese investor that is one of OpenAI's biggest backers, reported a net loss of about 1.1 trillion yen, roughly $7 billion, for the three months to the end of September. A year earlier it made a profit of 2.5 trillion yen in the same quarter, helped by a jump in the value of its OpenAI stake. Reuters reported that the loss far exceeded forecasts: the three analysts polled by LSEG had expected anything from a 233 billion yen profit to a 113 billion yen loss. The results were published on Tuesday, November 10.
The largest single item was a $3.7 billion loss on advance payments for equipment at US data-centre projects, including a large development in Pike County, Ohio, after plans changed. Chief financial officer Yoshimitsu Goto said fierce competition for energy supply and power-generation equipment had pushed the company to buy before long-term specifications were final. According to Reuters, the company's chief accounting officer said the payments could not be refunded on cancellation and carried fees. SoftBank would be entitled to rebates of up to $2.1 billion on future purchases. Financing costs also more than doubled, to 462 billion yen.
The context is a company that has borrowed heavily to bet on AI. Reuters reported that SoftBank raised a record $56.5 billion in the quarter, including an $11.87 billion term loan, a $10 billion dollar bond and a 1 trillion yen bond, and that it has spent $64.6 billion on a 13% stake in OpenAI. MLQ, summarising SoftBank's financial report, said the OpenAI stake was not revalued this quarter, so it contributed no gain, and that hedging contracts tied to Intel shares added further losses. The Vision Fund recorded a $1.3 billion gain on ByteDance that was outweighed by falls in listed holdings.
Some caution is needed in reading this. A write-off on prepaid equipment is an accounting loss, not proof that the data centres are being abandoned; the Ohio project is still described as a large development, and the rebates could recover part of the money if SoftBank keeps buying. The Intel figures come from a secondary summary that we could not check against the filing line by line. The share price is also sending mixed signals: Reuters said the stock surged 14% on Tuesday before the results came out, yet it has almost halved since the start of June.
The practical lesson reaches well beyond Tokyo. For two years the fear in AI was being left without chips or power, so buyers signed early and paid up front. SoftBank's quarter shows the opposite risk: plans change, and non-refundable deposits become losses. Goto put it plainly, saying strategies that looked fine a year or two ago sometimes have to change, and that reviewing assets the company no longer needs may come up for discussion. Any business signing a multi-year cloud, GPU or AI software commitment faces a smaller version of the same question.
Why it mattersIt is the clearest public example so far of an AI infrastructure bet going wrong on timing, not on demand: money was paid for equipment before the plans were settled.
👀 What to watch
How the shares trade now that the loss is public, whether SoftBank sells assets, and the OpenAI funding announcement that Goto said is highly likely within the current quarter.
3 opportunities from this story
AI and cloud commitment check for small businesses
Offer a fixed-price review of a company's prepaid and multi-year technology commitments: cloud credits, reserved GPU capacity, annual AI software seats and minimum-spend contracts. The buyer is a founder or finance lead who signed deals during the capacity scramble and is not sure what can be cancelled, resized or resold. SoftBank's write-off makes the risk easy to explain, and most small firms have never listed these commitments in one place.
- Best for
- Freelance finance consultants, fractional CFOs and cloud-savvy developers
- First step this week
- List every prepaid or minimum-spend technology contract in your own business or one client's, with end date, exit terms and actual usage.
Open the full playbook
Launch steps
- Build a one-page template: vendor, commitment, usage, exit terms, renewal date
- Run it free for two businesses you know and record what you find
- Turn the findings into a short sample report
- Offer the review at a fixed price to startups and agencies
Tools
Risks
You are not giving legal advice, so say so and refer contract disputes to a lawyer. Savings are not guaranteed: some contracts simply cannot be changed, and you should be honest about that up front.
A plain-English tracker of AI infrastructure spending and write-offs
Publish a weekly brief that tracks what the big AI builders commit to, borrow and write off, in language a non-specialist can follow. Investors, suppliers, journalists and students all need this and most coverage is either too technical or too excited. Earnings season gives you a steady supply of primary documents, and SoftBank's quarter is a ready-made first issue.
- Best for
- Writers, finance graduates and analysts who can read an earnings report
- First step this week
- Write a 500-word explainer of SoftBank's quarter from its investor briefing and the Reuters report, and publish it.
Open the full playbook
Launch steps
- Pick 10 companies to follow and bookmark their investor pages
- Create a simple table of commitments, debt raised and impairments
- Publish one issue per week, always linking to the filing
- Offer the table as a paid download once you have a few months of data
Tools
Risks
Mistakes with financial figures damage trust quickly, so quote the source and check every number twice. This is information, not investment advice, and it takes months to build an audience.
Brokerage for surplus power equipment from cancelled data-centre plans
When large buyers change plans, generators, transformers and switchgear can become available outside the usual waiting lists. A broker who knows African or other emerging-market buyers, such as data-centre developers, factories and hospitals, could match that surplus to real demand. It is a relationship business with long sales cycles, and nothing in SoftBank's report says its own equipment is for sale; the opening is the wider pattern of over-ordering.
- Best for
- Electrical engineers, energy consultants and equipment traders with supplier contacts
- First step this week
- Speak to three local buyers of large power equipment and ask what lead times and prices they are being quoted today.
Open the full playbook
Launch steps
- Map buyers in your region and their specifications
- Build contacts at equipment resellers and liquidators
- Start with one small, well-documented transaction
- Add inspection and logistics partners before taking larger deals
Tools
Risks
Deals are large, slow and can collapse late. Equipment may not match local grid standards, import duties can erase the saving, and you should never take payment for goods you have not had independently inspected.
Anthropic adds Claude Sheets and Whiteboards, putting a spreadsheet and a shared board inside Claude
On November 10 Anthropic released Claude Sheets and Claude Whiteboards in beta on all paid plans, made Claude Data generally available, and showed a meeting assistant for Google Meet.
Anthropic used an event in San Francisco on Tuesday, November 10, to move Claude further into everyday office work. The headline addition is Claude Sheets, a spreadsheet that people and Claude edit together. You describe what you need, such as a campaign tracker or a hiring budget, and Claude builds it with tabs, formulas, formatting and drop-down menus. Alongside it came Claude Whiteboards, a shared board for brainstorming with Claude. Both are in beta on all paid plans across web, desktop and mobile, according to Anthropic's announcement.
The details matter for anyone planning to use it. Claude outlines each edit it makes to a sheet so changes can be tracked, and sheets can be shared for joint editing. Files export to .xlsx or CSV. On Team plans the new tools are switched on by default; on Enterprise plans, Whiteboards and Sheets are off until an administrator enables them. Claude Data, previously called Claude Dashboards, is now generally available: it connects to data warehouses such as BigQuery, Amazon Redshift and ClickHouse and produces dashboards, written data stories and narrated videos.
Anthropic also showed Claude Meetings, in which Claude joins a video call as a visible participant, follows the discussion, drafts the plan being talked about and answers questions in the chat. TechCrunch reported that it is launching in beta on Team and Enterprise plans and works with Google Meet first. Separately, connected apps from partners including Adobe and Atlassian can now open as full applications beside the conversation. TechCrunch noted that the launch comes a few weeks after OpenAI unveiled its own workplace tools, as both companies try to become office suites that compete with Microsoft and Google.
The caveats come from Anthropic itself. Its announcement says that, for now, sheets have simple spreadsheet functionality, and it points people who need full Excel features to export the file or use its existing Claude for Excel product. No source we read gave pricing changes, usage limits or size limits for sheets. Reports also disagree on Meetings: TechCrunch and Engadget describe it as available to Team and Enterprise customers, while iClarified says it is coming soon, and Anthropic's own article on Sheets does not mention it. Treat Meetings availability as unconfirmed until you see it in your account.
For small businesses the interesting part is not that another spreadsheet exists. It is that the hardest step, designing a tracker with the right columns, formulas and drop-downs, now takes a sentence instead of an afternoon. Many shops, schools, clinics and agencies run on spreadsheets that one person built years ago and nobody dares to change. A tool that builds and explains the structure lowers the skill needed to fix that, and it creates demand for people who know which trackers a business really needs and how to keep the data clean.
Why it mattersThe spreadsheet is the default business tool almost everywhere, and building one by describing it puts structured record-keeping within reach of people who never learned formulas.
👀 What to watch
How quickly Sheets gains more advanced features, whether Meetings reaches more plans and video platforms, and how Microsoft and Google respond in Excel and Google Sheets.
3 opportunities from this story
Done-for-you business trackers built in a day
Sell small businesses a set of working trackers for stock, cash flow, customer follow-ups and staff rotas, built with an AI spreadsheet tool and tested with their real data. The buyer is an owner who knows the business is leaking money through messy records but has no time to design a system. Because the build is now fast, you can charge for understanding the business and setting things up properly, not for typing formulas.
- Best for
- Freelancers, bookkeepers and virtual assistants who know how small businesses run
- First step this week
- Build one stock or cash-flow tracker for a business you know, using a prompt, then check every formula by hand.
Open the full playbook
Launch steps
- Choose one niche, such as salons, hardware shops or small schools
- Interview two owners about the records they keep today
- Build a pack of three or four trackers and test them with real figures
- Deliver with a 30-minute handover call and a one-page guide
- Ask each client for a referral
Tools
Risks
AI-built formulas can be wrong in ways that look right, so you must test every one. Claude Sheets is a beta with simple features and needs a paid plan, so export to .xlsx for clients who do not subscribe.
AI spreadsheet and whiteboard training for teams
Run short, practical workshops that teach a team to build and maintain its own trackers and plans with AI tools. The buyer is a manager at an NGO, school, SACCO or agency whose staff use spreadsheets daily but only know the basics. New tools arriving in products they already pay for create a clear reason to train now, and a workshop built around their own files is far more useful than a generic course.
- Best for
- Trainers, Excel-skilled professionals and consultants
- First step this week
- Design a two-hour session around three tasks: build a tracker from a prompt, check its formulas, and share it safely.
Open the full playbook
Launch steps
- Write the session using real examples from one sector
- Run it free once for a friendly team and collect feedback
- Record short clips for marketing
- Offer a follow-up clinic a month later
Tools
Risks
The products are changing quickly, so material goes out of date; keep it focused on habits such as checking formulas and protecting data. Remind teams not to paste sensitive personal or financial data into tools their organisation has not approved.
Build a connector that brings a local business tool into AI assistants
Anthropic says connected apps can now open as full applications beside the conversation, and Engadget reported more than 9,000 connectors in its marketplace. Most of those serve global software, not the accounting, payments and logistics tools that businesses in Kenya and similar markets rely on. A developer who builds and maintains a well-made connector for one widely used local system could sell it to that vendor or to its customers.
- Best for
- Developers comfortable with APIs and authentication
- First step this week
- Pick one local software product with a public API and read the Model Context Protocol documentation to scope a minimal read-only connector.
Open the full playbook
Launch steps
- Confirm the vendor's API terms allow third-party integrations
- Build a read-only version first: balances, invoices or orders
- Add write actions only with clear user approval steps
- Pilot with two customers and document the security model
- Approach the vendor with the working pilot
Tools
Risks
You handle business and financial data, so security mistakes are serious; start read-only and get a review before adding write access. The vendor may build its own connector, and marketplace acceptance is not guaranteed.
Nvidia's GR00T 2 robot model hits 71% average success on 50 tasks after one demonstration each, the company says
At the Conference on Robot Learning in Austin on November 9, Nvidia introduced GR00T 2, a model that predicts future video and robot actions together, along with version 3.0 of its open-source Isaac Lab training framework.
Nvidia introduced GR00T 2 on Monday, November 9, at the Conference on Robot Learning in Austin, Texas. The company calls it an open frontier world action model built on its Cosmos platform. In plain terms, it is a single model that predicts what the camera will see next and what the robot should do next, together. Interesting Engineering reported that it works across simple two-finger grippers, five-fingered hands and humanoid robots. Nvidia released it alongside Isaac Lab 3.0, the latest version of its open-source framework for training robots in simulation.
The headline numbers are Nvidia's own. Without any task-specific demonstrations, GR00T 2 completed 40 of 50 evaluated tasks at least once. With one demonstration per task it reached 71% average success across the 50 tasks, rising to 95% with five. Interesting Engineering added that the five-demonstration result followed a 30-minute fine-tune on a single Nvidia RTX 5090 graphics card, using under three minutes of robot data per task. A smaller version, GR00T 2 Turbo, runs on a Unitree G1 humanoid with an onboard Jetson Thor computer at 124 milliseconds per prediction.
The model learns mostly from people, not robots. Interesting Engineering reported a training mix of 402,730 hours of first-person human video, 10,184 hours of handheld-gripper demonstrations and 23,500 hours of robot teleoperation. Nvidia's blog highlights 100,000 hours of human video that it has released as open data. Early users named by Nvidia include Nexuni for robotic laundry, Foxconn for server manufacturing and Siemens, which is evaluating industrial uses. Isaac Lab 3.0 adds a choice of physics engines and a library for further training of models such as GR00T, with up to 2.5 times the training throughput of the previous version in one benchmark.
Several things are not yet known. The results have not been independently reproduced, Nvidia has not named the 50 tasks or the test robot in the material we read, and completing a task at least once is a much weaker standard than doing it reliably. The two figures for human video, 100,000 hours and 402,730 hours, appear to describe the open portion and the full training set, but the sources do not spell that out. Nvidia calls the model open, yet the announcement gave no licence terms or release date for the weights. Lab success rates also tend to fall in messy real workplaces.
If the claims hold up, the economics of teaching robots change. Today the costly part of a robot project is collecting hundreds of demonstrations for each task. A model that needs one to five, and that learns from ordinary first-person video of people working, shifts the bottleneck from robot hours to good recordings of human skill and good simulated practice environments. That favours small specialist teams: people who can film tasks properly, build simulation scenes, and test whether a robot really does a job well enough to trust on a factory floor.
Why it mattersCutting the demonstrations a robot needs from hundreds to a handful would make it affordable to automate small, varied tasks that were never worth programming.
👀 What to watch
The release of model weights and their licence, independent results on public benchmarks, and whether pilots at firms such as Foxconn and Siemens turn into deployments.
3 opportunities from this story
First-person task video collection for robot training
Robot models now learn from first-person video of people doing real work, and the reported training mix for GR00T 2 leans heavily on it. Robotics companies need footage of specific tasks in specific settings, such as kitchens, workshops, farms and warehouses, recorded consistently and with consent. A small team with head-mounted cameras and a clear process could supply that, and varied real-world settings outside the usual labs are in short supply.
- Best for
- Video teams, data-labelling firms and operations people who can manage field crews
- First step this week
- Record 30 minutes of one repetitive manual task with a head-mounted camera and review whether hands and objects stay clearly in view.
Open the full playbook
Launch steps
- Study the format of Nvidia's open human video dataset
- Write a recording protocol and a consent form
- Produce a sample set of 10 hours across three tasks
- Pitch robotics startups and research labs with the sample
- Add labelling and quality checks as a second service
Tools
Risks
Buyers are few and have strict format requirements, so sell a pilot before investing in equipment. Filming workers raises privacy and labour concerns: get written consent, pay fairly and blur faces and anything confidential.
Freelance simulation scenes and tasks in Isaac Lab 3.0
Robots rehearse in simulation before they touch real objects, and each new task needs a scene, objects and a way to score success. Isaac Lab 3.0 is open source and newly released, so few people know it well yet. A developer who learns it now can sell environment building and training runs to robotics startups and university labs that lack the time to do it themselves.
- Best for
- Developers with Python and some 3D or game-engine experience
- First step this week
- Install Isaac Lab 3.0 from GitHub and get one of the bundled example tasks training on your own machine or a rented GPU.
Open the full playbook
Launch steps
- Work through the official examples
- Build one original task, such as sorting or cable insertion
- Publish it with a short video and write-up
- Offer custom environments to labs and startups you find through robotics forums
- Specialise in one industry's tasks
Tools
Risks
The learning curve is steep and GPU time costs money before you earn anything. Skills that work in simulation do not always transfer to real robots, so be clear with clients about what you deliver.
Beginner tutorials and courses on modern robot learning
Each major robotics release brings a wave of students and engineers looking for clear, working tutorials, and official documentation rarely serves beginners well. A creator who publishes step-by-step guides to Isaac Lab 3.0 and to fine-tuning robot models can build an audience early. This is the lowest-cost way into the field and it builds the reputation that later sells consulting.
- Best for
- Technical writers, YouTubers and engineering students
- First step this week
- Publish a tested guide to installing Isaac Lab 3.0 and running a first training job, with every command shown.
Open the full playbook
Launch steps
- Document your own setup, including the errors you hit
- Release one tutorial per week for six weeks
- Collect questions from comments to plan the next topics
- Package the series into a paid beginner course
Tools
Risks
Tools change fast and tutorials break, so date each one and update the popular ones. Income from content is slow and uncertain, and you should only teach what you have actually run yourself.
Get these as a PDF every 3 days
Free. One email every 3 days. Unsubscribe any time.