AI News: Google’s Space Data Center Study, AI Deals and New Media Models

Google is studying a proposed space-based AI infrastructure system called Project Suncatcher. The concept combines solar-powered computing satellites and links between them. This is an engineering study of a possible future system, not a deployed orbital data center.

The research suggests a solar panel in a suitable orbit could be up to eight times as productive as one on Earth. I see abundant solar energy as appealing because power availability is a constraint on terrestrial data centers, though the comparison depends on the conditions and design.

Training and inference also require rapid communication between chips. A distributed space system cannot simply connect distant satellites with conventional cables, so Google proposes tightly coordinated satellites communicating through high-bandwidth links.

I compare the proposal with earlier long-term bets such as Waymo, while acknowledging that Suncatcher could face many obstacles. The team has begun using mathematical and physical models to study whether a constellation can maintain the necessary formation.

The concept calls for satellites separated by only hundreds of meters, a demanding formation for fast-moving spacecraft. Google’s illustration models how a free-fall satellite formation evolves under Earth’s gravity; it is a simulation, not a photograph of hardware in orbit.

Radiation is another challenge. Google tested its Trillium TPU using a proton beam to study ionizing dose and single-event effects, and reported promising results. A laboratory test does not by itself prove the lifetime or reliability of a complete orbiting data center.

Launch cost may determine whether the concept is economic. Google’s analysis projects that a sustained learning rate could bring launch prices below $200 per kilogram by the mid-2030s. I discuss reusable rockets and possible further reductions, but those numbers are projections, not current prices or a guaranteed outcome.

Turning to infrastructure on Earth, I discuss a social-media compilation of OpenAI-related compute deals and proposed investments, including figures associated with Stargate, NVIDIA and AMD. The headline amounts are presented as reported claims; different commitments, ambitions and proposed investments should not be added together as if all were completed spending.

The post also raises a possible trillion-dollar IPO, and I compare the scale of these plans with earlier reports of a much larger fundraising ambition. An IPO and future financing are speculative in this discussion, not confirmed transactions.

I discuss OpenAI’s corporate restructuring and its changed Microsoft arrangement. I describe the end of a prior right-of-first-refusal arrangement and a subsequent AWS partnership as part of a broader effort to secure compute. The reported terms and timing should be understood in their 2025 context.

The Verge reported that Apple was planning to use a custom Google Gemini model for Apple Intelligence, with an estimated payment of about $1 billion per year. The report described the model running on Apple’s private cloud infrastructure for tasks such as summarization and planning; this is a reported plan, not a confirmed launch in the video.

I note that Apple had reportedly considered other model providers and already worked with Google and OpenAI in different ways. I welcome the idea of a more integrated Siri experience; the report says Tim Cook had indicated a new Siri was planned for the following spring, a schedule described at the time.

OpenAI announces a quality-of-life update for long-running ChatGPT requests: a user can add context while the task is running rather than start over. I see particular value for long research or reasoning tasks where restarting can waste considerable time.

Microsoft releases MAI-Image-1, an image generation model it describes as developed in-house and not open-weight. At the time, Microsoft says it entered the top ten for text-to-image on LM Arena and describes efforts to avoid repetitive, generic styles.

The company highlights data selection, evaluation and feedback from creative professionals. I welcome its focus on more distinctive results and mention photorealistic imagery as a stated strength. These are company claims and my impressions, not an independent assessment of all use cases.

Vidu Q2 appears on Artificial Analysis’s text-to-video leaderboard, with the source reporting a number-eight position at the time. It supports multiple reference images and variants producing clips up to eight seconds at 1080p, according to the information shown.

I cite prices of $4 per generated minute for Turbo and $6.10 for Pro as reported then; pricing and leaderboard placement can change. I compare the models using sample clips, starting with an animated elf riding a unicorn through a forest.

A second comparison shows a dolphin moving above and below the water. I prefer aspects of Vidu Q2’s lighting and water to one competing sample, while also praising Sora 2 Pro. These judgments are subjective observations of selected examples rather than a systematic evaluation.

Perplexity says it received a legal threat from Amazon concerning the use of the Comet assistant on Amazon’s site. Its statement argues users should be able to ask an agent to compare and purchase products while logged in. That is Perplexity’s account of the dispute, not a finding about Amazon’s legal position.

I speculate that advertising incentives help explain why platforms may resist agents browsing and buying on users’ behalf. Amazon has an advertising business, but the video does not establish that advertising was the reason for its action. I argue that users should be able to delegate tasks to an agent.

Finally, I mention Sora character cameos: users can upload a picture or design, including a pet or an original character, for use in generated videos. I present this as a new feature in the November 2025 update.