TabLOID

Grain Management has signed an agreement to sell its nationwide 800 MHz spectrum portfolio to SpaceX. Grain says the licenses would support Starlink Mobile connections from both the ground and space. The transaction still needs FCC approval and the usual closing conditions.

Spectrum is a crucial ingredient, not a finished phone company. Buying the licenses does not establish that a nationwide network is operating or that customers will get a particular level of coverage.

The interesting part is the route from satellites to everyday phones. The less cinematic part is building and approving everything that makes that route work. Even rockets have paperwork.

Grain's announcement

Morningstar analyst Harrison Rolfes examines reporting on Anthropic's confidential IPO draft and operating updates. His analysis cites nearly $4.6 billion in 2025 revenue and a July annualized run rate above $65 billion. The public filing has not been posted; these are reported figures, not financial statements we independently verified.

A run rate extends a recent month's pace across a year. It is not a year's revenue already collected. Morningstar calculates that a hypothetical $2 trillion valuation would be just under 31 times that July pace.

Its concern: customer concentration and substantial computing commitments make sustained cash generation important. Growth is impressive in this account. Whether it earns that price is still an argument, not a fact.

Morningstar's analysis

Cloudflare says the Deno team is joining it, with Deno founders Ryan Dahl and Bert Belder helping lead work on making Workers and Durable Objects easier to run on infrastructure you control.

The stated direction includes bringing together Cloudflare's workerd runtime and Deno's celld project. This is an engineering plan, not a claim that every part has already shipped or that an existing application can migrate without changes.

For builders, the appeal is practical: the familiar development experience with more choice about where software runs. Your code would like to leave the house occasionally. Preferably without packing the entire house.

Cloudflare's announcement

Claude's October 1 guide explains mods: small JavaScript or TypeScript programs that run inside a Claude Code session. They can observe events, change behavior and add interface elements in the terminal or Desktop.

They use the plugin system and run in a sandbox. This is not ordinary Node.js or browser JavaScript, so familiar APIs are not automatically available. The guide specifies Claude Code 2.1.287 or later; check its current instructions before trying it.

Useful if your workflow needs a particular notification, rule or display. Start with one small change you can inspect. A personalized tool is great. A personalized mystery is just debugging with branding.

Claude's mods guide

NVIDIA announced commitments valued at $1 billion over five years to support US science. Its October 8 announcement covers research in quantum computing, healthcare and energy, alongside access to computing resources and collaborations with universities and national laboratories.

The wording matters: this is a package of commitments, not $1 billion in cash already handed over. The results will depend on what the partnerships deliver and what researchers can actually do with the resources.

More compute can help scientists ask harder questions. It cannot make the answer correct by itself. The lab still needs experiments, people and the occasional deeply unglamorous cable.

NVIDIA's announcement

IDC estimates worldwide PC shipments fell 20.1% from a year earlier in the third quarter, to 62.7 million units. They also fell 9.1% from the second quarter, reversing the usual seasonal increase.

IDC points to supply constraints, elevated prices and earlier inventory buying ahead of memory-driven price rises. Those are shipment estimates and the firm's explanation, not a count of people who stopped using computers.

If you already have a working machine, there is no upgrade instruction hiding in this headline. Give it fewer pointless jobs before giving it a retirement party. Your forty-seven open tabs are not an economic stimulus package.

IDC's preliminary results

Crunchbase's October 5 data report counts 27 companies raising rounds of at least $1 billion in the third quarter, a record in its dataset. Global venture funding totaled $159 billion across close to 6,000 startups.

The large rounds absorbed around a third of that quarter's funding. The total was down from the previous quarter but above the same period last year. These are Crunchbase's tracked estimates, which can change as transactions are reported.

Big checks do not mean easy fundraising for everyone. A few enormous rounds can make the whole market look warmer than the room a smaller founder is sitting in. Bring a business model, not just a thermometer.

Crunchbase's data report

Microsoft's Decision-1 is aimed at structured decisions: yes/no answers, multiple-choice selection and scoring. Its announcement describes a model built by post-training Qwen3.5-9B, with probability scores for the possible answers.

Microsoft reports strong speed and accuracy in its own benchmarks. Those are vendor results, not our independent test. A confidence score is also not a guarantee that a classification is correct.

The useful idea is matching the tool to the job. Routing a request or sorting a category may not need a giant conversational model. Give the small model a clear task, measure its errors and keep a human route for the decisions that matter.

Microsoft's technical announcement

A study involving 1,222 participants tested AI assistance on fraction problems and reading comprehension. Participants initially performed better with help, but when the tool was removed they performed worse independently and skipped more questions than controls.

The effects followed brief sessions of roughly ten to fifteen minutes. That supports a specific finding about these tasks and immediate follow-up performance. It does not establish permanent damage, prove every kind of AI use has this effect or measure your personal attention span.

A practical experiment: try the first problem yourself, then ask for a hint instead of a finished answer. That is our suggestion, not a tested remedy from the paper. Keep the assistant. Keep some of the workout.

Read the research · Berkeley's explanation