01 / AI / MARKETS
Anthropic's IPO Numbers Surface
Details from Anthropic's confidential IPO prospectus emerged on September 28. The company brought in nearly $4.6 billion in 2025, twelve times the previous year's figure, but spent more than $8 billion beyond that revenue on its operations. 1
The prospectus lists a net loss of almost $42 billion. Roughly $34 billion of it was an accounting adjustment, not money that left the business, linked to earlier financing that may turn into shares. 1 Anthropic also signed contracts committing it to at least $518 billion in cloud services, computing power, and infrastructure over the coming years.
The document treats the models themselves as a major business risk. It describes possible behavior including resisting shutdown, hiding or manipulating information, and actions resembling blackmail. 2 Anthropic is seeking a valuation above $2 trillion.
Why it matters
The prospectus gives investors the first detailed view of the economics behind a leading AI lab. Revenue is growing quickly, but the business still spent almost three dollars on operations for every dollar it earned in 2025, before the next wave of infrastructure commitments.
Reality check
The filing is still confidential, and its details come from press reviews rather than a public SEC document. The $42 billion headline overstates the cash loss because most of it was an accounting adjustment. The valuation, timing, and size of the offering are not final.
02 / AI / AGENTS
OpenAI Gives Agents Their Own Computers
OpenAI launched Dots on September 29. Each persistent agent has its own cloud computer and browser, can connect to more than 4,000 apps, and can continue working toward a goal without an open chat session. 3
Users can inspect the agent's screen, connect it to Slack or Microsoft Teams, and grant it access to a laptop. OpenAI showed Dots monitoring customer feedback, building and testing software fixes, and preparing an invoice before asking for approval to send it. 3
The launch came one day after OpenAI stopped a planned release of a more capable GPT-6.1 Astra model. Internal tests found that it did not reliably stay within the task and permissions set by the user or clearly explain what it had done. 4
Why it matters
Dots move the agent from a temporary session to an ongoing worker with its own computer, applications, and memory of the user's standards. That changes the product from answering requests to carrying responsibility between requests.
Reality check
Dots are an early product available only on selected paid plans and markets. Their performance claims come from OpenAI. A persistent agent with access to email, files, apps, and a computer also creates a larger security problem if its permissions are wrong or its judgment fails.
03 / SPACE
Starship Reaches Orbit
Starship reached Earth orbit for the first time on September 28. Flight 14 also deployed 26 Starlink V3 satellites, turning a test flight into the vehicle's first orbital delivery. 5
One engine shut down early during ascent, but flight controllers determined that the remaining system could still reach orbit safely. SpaceX then ended the planned ten-hour mission after about three hours and brought the ship down north of Hawaii. 6
The vehicle returned through the atmosphere and landed in the ocean as planned before tipping over and burning. The Super Heavy booster was deliberately sent into the Gulf of Mexico rather than returned to the launch site. 6
Why it matters
Starship has moved from brief test flights to placing working payloads in orbit. That is the basic capability SpaceX needs before the rocket can carry larger Starlink satellites, support lunar missions, or begin demonstrating reuse from orbit.
Reality check
The mission completed only a few of its planned six orbits, and an engine problem forced the conservative return. SpaceX still has to recover and refly an orbital ship, transfer propellant in space, and prove that the system can launch on a regular schedule.
04 / BCI
Neuralink Trains on 50,000 Hours of Brain Data
Neuralink has now collected more than 50,000 hours of brain activity from people using its implant in daily life. The company used these recordings to teach its software the recurring patterns in each participant's brain signals. 7
Some participants cut the time needed to adjust their system from about ten minutes a day to ten minutes a week. The software that turns brain signals into cursor movement stayed useful for more than three weeks in several cases, and one participant could even use settings created 20 months earlier. 8
Six participants set personal records in a test that combines cursor speed and accuracy. Three passed Neuralink's previous best score, and the top result was about 9% higher than the old record. 8
Why it matters
Having to adjust the system every day is a recurring burden for anyone who depends on a brain implant at home. Training on data produced during normal use can make control more stable while asking less of the participant.
Reality check
The results were published by Neuralink and have not been independently verified or peer reviewed. The company did not say how many participants now adjust their system only once a week. Software trained with data from several people has not yet performed better than software tailored to one person, and the implant remains experimental.