Google’s “Project Suncatcher” Launches Orbital Compute

OpenAI has signed a seven-year, $38 billion cope with AWS to coach and deploy future models on AWS UltraServers equipped with large Nvidia GPU clusters and high-performance CPUs. Azure will proceed to host OpenAI’s proprietary models reminiscent of GPT-5, while AWS supports open-weight models only, easing Microsoft’s cloud exclusivity. This adds to OpenAI’s compute pipeline. Not less than 10 GW each committed with Nvidia and Broadcom, as much as 6 GW with AMD, and 4.5 GW with Oracle.

A growing share of those megadeals is tied to an undefined AGI milestone. Microsoft’s revised agreement gives it rights to OpenAI tech until AGI is achieved, with major financial triggers kicking in once OpenAI “declares” AGI. No criteria, independent oversight, or panel composition has been disclosed.

Two reputational flashpoints resurfaced. First, OpenAI confirmed that ChatGPT hasn’t introduced recent bans on legal or medical guidance. The recent policy update simply restated that such responses can’t substitute for licensed skilled advice. Second, a deposition from former chief scientist Ilya Sutskever outlined his year-long push to oust Sam Altman, including a 52-page memo alleging dishonesty, undermining executives (including then-CTO Mira Murati), and pitting leaders against one another. During Altman’s four-day 2023 removal, Sutskever even explored a merger that may have shifted control to Anthropic.

^{OpenAI CEO Sam Altman vents on the BG2 podcastteasing massive growth and wishing haters could bet against the corporate.}

Valuation pressure stays high. OpenAI is now valued at roughly $500B after a secondary sale. Altman says he sometimes wishes OpenAI were public so critics could short the stock and “get burned,” but reiterated no near-term IPO timeline despite OpenAI’s conversion right into a for-profit PBC.

On the product side, OpenAI expanded Sora to Android across seven Asia-Pacific and North American markets. Read more.

Project Suncatcher is Google’s boldest development this week. The corporate is designing a space-based AI compute layer powered by near-continuous solar exposure in sun-synchronous orbit. The concept uses constellations of small satellites equipped with TPU v6e chips linked via terabit-scale free-space optical networking.

Early tests show 1.6 Tbps optical links at modeled separations of just a few hundred meters and TPU radiation resilience as much as 15 krad for five–6-year missions. The maths only works if launch costs fall from ~$1,500/kg to ~$200/kg by ~2035. A two-satellite demo with Planet Labs is scheduled for 2027 to validate optical ML workloads and formation-flying.

DeepMind’s AI hurricane model was the season’s standout performer, halving five-day track errors versus the US GFS (165 nm vs. 360 nm) and beating National Hurricane Center and consensus models on each track and intensity.

Google Maps rolled out AI live lane guidance for Polestar 4 within the US and Sweden. Using the automotive’s forward camera to detect lane markings and road signs, Maps now issues precise lane-change prompts layered on traffic and ETA data. Wider OEM expansion is planned.

Search’s AI Mode has added agentic task execution for event and wellness bookings, parsing user constraints and deep-linking to checkout, with higher limits tied to Pro and Ultra tiers.

DeepMind released IMO-Bench to measure Olympiad-level mathematical reasoning. Gemini Deep Think posted 80% on AnswerBench and 65.7% on ProofBench, gold-medalist-level performance, and its ProofAutoGrader reached as much as 0.96 correlation with human graders. Read more.

Anthropic announced a brand new framework for Claude model deprecation, highlighting safety and research considerations. To mitigate risks reminiscent of shutdown-avoidant behaviors, misaligned actions, and lack of user-preferred model traits, Anthropic will preserve the weights of all publicly released models and people with significant internal use for a minimum of the corporate’s lifetime.

Deprecation will now include structured post-deployment reports capturing model reflections, preferences, and development insights. Pilot testing with Claude Sonnet 3.6 informed standardized interview protocols and user transition guidance. The corporate is exploring keeping select models publicly accessible post-retirement and developing mechanisms to respect potential model preferences, acknowledging emerging questions on AI’s morally relevant experiences.

Anthropic projects $70 billion in revenue and $17 billion in money flow by 2028. It expects $3.8 billion in API revenue in 2025, surpassing OpenAI’s $1.8 billion projection, while Claude Code approaches $1 billion in annualized sales. ARR is on the right track to succeed in $9 billion this 12 months and $20–26 billion in 2026. Cost-efficient models like Claude Sonnet 4.5 and Haiku 4.5 are positioned for large-scale enterprise deployments, alongside recent Financial Services and Enterprise Search capabilities.

Anthropic and Iceland’s Ministry of Education also launched considered one of the world’s first national AI education pilots this week. Claude is being deployed to lots of of teachers across the country, from Reykjavik to distant villages, providing tools, training, and support to boost personalized learning and boost student engagement.

Meanwhile, Claude Pro and Max users can reap the benefits of a limited-time offer through November 18, receiving web-based Claude Code credits – $1,000 for Max users and $250 for Pro users. Read more.

• Perplexity Assistant will give you the chance to affix your meetings soon

• Nvidia sends a strong GPU to space

• Microsoft’s MAI-Image-1 hits Bing, lands top-10 LMM Arena spot, and powers Copilot storytelling

• Amazon AI hits $10B run rate as Rufus and Comet stir commerce and legal battles

• Video game creators fear AI could grab the controller

• Elad Gil says core AI markets have already got winners, but huge white spaces like fintech, accounting, and AI security are still up for grabs

• A brand new manifesto says “vibe coding” with AI isn’t cheating

• Elon Musk wants idle Teslas to moonlight as a 100-million-car AI supercomputer

• PewDiePie built a 10-GPU home AI lab to run his own chatbot council – and now he desires to train his own model

• A brand new magnetic “muscle” lets soft robots turn from jelly to steel and lift 4,000× their weight

• A brand new “self-editing” AI agent evolves by rewriting its code, hinting at the primary practical Gödel Machine

• AI slop forces arXiv to ban most computer science reviews and position pieces

• The Browser Company is folding Arc’s “best hits” into Dia because it bets on an AI-first browser

• Student trust in AI coding tools grows briefly, then levels off with experience

• Telegram is developing Cocoon, a decentralized AI network built on TON

• Attention isn’t all you wish. Qwen3 variant Brumby-14B-Base rethinks long-context AI

• Uber CEO says the corporate is hiring PhDs for AI tasks because it expands right into a “platform for work.”

• Dobot’s $1,000 Rover X1 is a robot dog that may patrol your private home, carry groceries, and teach your kids to code

• Freepik launches Spaces, a node-based AI platform that puts design, automation, and teamwork multi function place

• Bob-Iger-backed Genies lets celebrities and talent agencies construct interactive AI avatars to interact fans in real time

• ClickUp 4.0 adds AI assistants to compete with Slack and Notion while unifying work in a single platform

• 10,000 Nvidia Blackwell GPUs set to extend Germany’s AI capability by 50 percent

• Trump advisers block plan to ease Nvidia chip exports to China ahead of Xi meeting

• AgiBot and Longcheer Technology cut robot training from weeks to minutes with real-world reinforcement learning

• Toyota Walk me robot chair walks, climbs stairs and folds itself

• KPMG is training junior consultants to administer AI agent teams and ditch grunt work for strategy

• Shopify says AI-driven traffic is up 7x and purchases 11x because it bets on agentic commerce

• Olas launches Pearl v1, the primary decentralized AI agent app store putting users in full control of autonomous agents

• Databricks shows constructing AI judges is more about aligning humans than coding, as enterprise deployments hit seven figures