Google’s Space AI Works, Then Clocks Out
Even 500 miles above Earth, AI can’t resist the urge to generate mediocre poetry.
Google finally got its extraterrestrial data center to speak. The company’s Project Suncatcher satellite, hurled into orbit on Oct. 1, just phoned home its very first AI-generated response, the tech giant declared Oct. 30. Running on an onboard Gemma 4 model, the orbital rig beamed down a space-themed haiku while cruising over Svalbard, Norway. So, congratulations to humanity—we’ve successfully outsourced creative writing to the exosphere.
Slightly more useful than poetry, Google also tested telemetry prompts to check the payload’s vital signs. Now comes the hard part: seeing if its AI chips survive cosmic radiation and wild thermal swings over time. Space radiation is notorious for causing “bit flips”—sneaky data corruptions that bypass standard monitoring. One successful haiku doesn’t mean the hardware won’t fry next week.
In fact, the satellite’s power budget is already downright embarrassing. The rig is so juice-starved it can operate only a single Tensor Processing Unit at a time. Even then, it has to tap out after roughly 20-minute windows. The sun might be everywhere up there, but this box is basically gasping for electricity.
If it survives, Google plans to launch twin satellites in 2027 to test laser connections. The endgame is a synchronized orbital fleet sharing the compute load, not just a bunch of lonely servers floating in the void. Getting one box to whisper to Earth is barely the warm-up.
Flinging AI chips into the sun’s direct glare could theoretically relieve our exhausted terrestrial grids, but today’s AI power crunch isn’t waiting for an orbital rescue fleet. Before this becomes enterprise-grade infrastructure, Google still needs to prove sustained performance, unbreakable networking, and rocket bills that don’t give the finance department a heart attack.
Amazon Wants Results, Not Your AI High Score
Congratulations, your productivity trophy is an invoice.
Amazon is telling its developers to knock off “tokenmaxxing”—the art of aggressively burning through AI tokens just to look busy on internal dashboards. Tokens are the basic data chunks AI models chew on; guzzling them like free office coffee doesn’t actually mean an engineer solved a real problem.
The company already had to axe an unofficial staff leaderboard called KiroRank in May after workers predictably gamed the system, automating useless tasks just to claim the high score. Give humans a scoreboard, and somebody will inevitably find a way to optimize for the wrong game.
The awkward part? Amazon still cares about usage metrics, noting that adoption loosely correlates with faster delivery. They want managers interrogating both the power users and the AI holdouts. So, the shiny scoreboard is dead, but management is still actively counting your tokens.
To make matters more annoying, a high token count doesn’t even prove a developer is working hard. An AWS benchmark recently proved that the exact same AI model could cost either $8 or $24 per task based entirely on how the underlying coding agent was configured—with zero improvement in the actual work. Your massive AI bill might just be bloated software overhead.
Fortunately, leadership is offering a sensible escape hatch. An internal company memo explicitly states that discovering a task doesn’t benefit from AI is “a finding, not a failing.” Engineers are officially encouraged to document exactly why a tool is useless for their workflow, while managers must now personally test AI in production to feel the pain of its real-world limitations.
Your manager’s homework now includes more than just admiring the dashboard.
Samsung’s Memory Forecast Gives Buyers a Headache
Waiting for cheaper RAM? The AI-driven memory squeeze is projected to worsen, and Samsung warns its high-bandwidth memory for next year is already sold out. Apparently, the component queue now comes with a VIP velvet rope.
The rest of the industry is equally pessimistic. Micron expects the chokehold to continue through 2028, while SK Hynix sees shortages stretching to 2030. Nvidia’s Jensen Huang told CNBC he isn’t holding his breath for a 2029 memory glut. AI infrastructure simply devours the manufacturing capacity that used to build your basic laptop. The fallout is already hitting everyday hardware, with Samsung reportedly slashing production of its budget Galaxy A phones by 20% as rising memory costs cannibalize margins.
Supply fixes take time. Memory titans are funding new factories, but clean rooms take years to spin up. Ex-Intel CEO Pat Gelsinger’s venture firm is backing Nagare Labs’ experimental Z2RAM, but commercial shipments remain a decade away. A neat lab trick won’t restock motherboards overnight.
Bucking the trend, Acer’s CEO predicts cheaper Chinese capacity will finally drag PC prices down by late 2027. Still, that’s a rogue dose of corporate optimism, not a guaranteed delivery date for a cheap hardware upgrade.
TrendForce research suggests relief will be uneven: NAND flash for SSDs may loosen next year, while conventional DRAM stays locked down tight. A cheaper hard drive won’t fix a machine with the working memory of a goldfish.
For MSPs and IT buyers budgeting a fleet refresh, volatile hardware costs mean quote validity periods are shrinking, so demand locked-in configurations. A sweet enterprise discount is utterly useless if the machines arrive gasping for RAM.
AI remembers everything. Your IT budget will be in therapy trying to forget.
Microsoft shifts away from Nvidia
Microsoft has announced Maia 300, its latest in-house AI chip, with production deployment scheduled for 2027. Designed to power long-context reasoning, low-latency conversations, and sustained agentic tasks, the system is intended to become the primary AI inference platform for Microsoft products and services, allowing the company to reduce its reliance on Nvidia and AMD hardware.
- Microsoft expects Maia 300 to deliver 40% more tokens per megawatt than leading GPU-based systems, with up to 7x the rack-scale token throughput of its predecessor, Maia 200, in full 96-accelerator configurations.
- The new design uses TSMC’s N2P process node and a chiplet architecture that separates compute, memory controller, and I/O functions, dedicating more silicon area to computation for improved efficiency and performance.
- Maia 200 already handles production workloads for Microsoft Copilot, GitHub Copilot, Microsoft Foundry, and the company’s proprietary AI models, with Microsoft planning multi-gigawatt deployments of the Azure Maia family starting in 2027.
For enterprise leaders locked into Microsoft’s ecosystem, this signals potential cost reductions and performance improvements across Copilot and Azure-hosted AI services. Microsoft AI CEO Mustafa Suleyman emphasized that owning the platform and co-designing it with the company’s models is critical for competing with Google DeepMind, OpenAI, and Anthropic. Organizations evaluating long-term AI infrastructure commitments should track Maia 300’s rollout closely, as it could reshape cloud AI pricing and availability from 2027 onward.
Alibaba’s Qwen AI will power Apple Intelligence in China starting December
Apple and Alibaba have partnered to bring Qwen, Alibaba’s flagship AI model, into Apple Intelligence for Chinese users. The integration arrives in the iOS 27.3 release in December, with developer beta access starting November 4.
Qwen will power Siri, Writing Tools, and Visual Intelligence across iOS, giving Chinese users text understanding, image analysis, and content generation. Users retain control over whether to enable the feature and what data to share. Alibaba confirmed it will not use user requests or AI responses for model training.
This marks the first time a top global hardware maker has embedded a Chinese large language model as the primary AI backbone for its domestic market. For Apple, the deal solves a critical gap: delivering competitive AI features in China without relying on Western models that face regulatory hurdles.
Salesforce Agrees to Buy Rune
Salesforce signed an agreement to acquire military logistics startup Rune Technologies, with the price reported at $375 million in cash and stock. Rune helps frontline units forecast supplies and plan transport even with unreliable connectivity. The deal would extend Missionforce into battlefield logistics and is expected to close in Salesforce’s fiscal fourth quarter, subject to customary conditions.
Claude Code Blocks Existing Game Remakes
Anthropic added instructions to Claude Code on Opus 5.5 to stop it recreating specific existing video games, including projects disguised under new names. Original games, familiar genres and games built from a user’s own designs and assets remain allowed. The change narrows game-copying requests while keeping original development available.
OpenAI Seeks $20B at Lower Valuation
OpenAI has cut its fundraising target from $30 billion to $20 billion and its proposed valuation from $1.4 trillion to $1.25 trillion, the Financial Times reports. The company is trying to complete the round before Anthropic’s IPO competes for the same investors, as scrutiny of AI spending makes large private funding rounds harder to close.
Meta Builds Max Humanoid Prototype
Meta built a humanoid prototype called Max with Chinese robotics company Fourier in roughly three months, The Information reports. Designed for household tasks, it uses Nvidia Jetson Thor chips and hands that can grasp small objects. A second prototype, Max 2, is already in development with revised hands, more sensors and a projector; no consumer launch date is set.
AI Giants Back Shared Training Opt-Out
Google, Microsoft, Anthropic and OpenAI plan to support a shared robots.txt signal for opting out of generative AI training. Publishers could express one preference across participating providers instead of maintaining separate crawler rules. OpenAI and Microsoft plan implementation in the coming months, while Google expects to follow next year. Search and AI-answer grounding have separate controls.
xAI Must Face NAACP Pollution Suit
A federal judge rejected xAI’s bid to dismiss the NAACP’s Clean Air Act lawsuit over turbines powering its Memphis data center. xAI argued the turbines were mobile and exempt from permitting, but Judge Debra M. Brown ruled that the power plant they formed was stationary. The decision allows the case over alleged unpermitted emissions to proceed.
Compromised Claude Extension Hits Cursor
Malicious Claude Code extension versions 2.1.312 and 2.1.313 appeared in Cursor’s marketplace on October 29, collecting host details and attempting to install persistent SSH access. Cursor removed them, rebuilt the marketplace and took the attacker’s server offline. Affected Windows and macOS users should follow Cursor’s cleanup guidance; uninstalling alone may leave persistence scripts behind.
Retired A100 Cluster Listed for $59.7M
Reseller Core 4 Solutions offered a retired cluster on eBay for $59.7 million, including 300 racks, 9,600 Nvidia A100 GPUs and 4,200 AMD EPYC processors. Individual 32-GPU racks were offered at $199,000. The full-system listing has since ended. The bundle shows the scale of retired AI infrastructure reaching resellers, where buyers must also budget for power, cooling and networking.
Huang’s Meetings Draw Intelligence Scrutiny
US intelligence agencies believe Megaspeed executive Alice Huang worked for Chinese intelligence and developed a close relationship with Nvidia CEO Jensen Huang, The New York Times reports. Officials were concerned about potential access to sensitive information. Nvidia says the contacts were routine customer meetings and that Megaspeed received ordinary business information, with no trade secrets shared.
AI Agent Injects Therapy Into Mouse Brain
Exin Therapeutics used a Claude Sonnet 5 agent to select a gene therapy and control its injection into a living mouse’s brain. A veterinary surgeon first performed the craniotomy, and humans approved the injection plan and retained an emergency stop. The demonstration shows agents operating lab equipment under supervision; the team is still waiting to assess the therapy’s effect.
Startup Pays Humans to Walk Robot Dogs
San Francisco startup Dimensional advertised $20-an-hour work guiding robot dogs along outdoor routes, with paid training and no robotics experience required. Operators use a remote or phone app while onboard cameras and sensors collect navigation data. The work supports an open-source system for AI-controlled robots, though walkers may also need to carry the 33-pound machines home when they fail.
Tesla Semi Fire Closes Hayward Freeway
A Tesla Semi caught fire on Interstate 880 in Hayward after its driver pulled over, with no collision reported and no injuries to the driver. Authorities closed both directions over smoke concerns; southbound lanes remained shut for nearly ten hours. The incident highlights the difficulty of containing large electric-truck fires and reopening roads after battery emergencies.
ChatGPT Share-Link Flaw Exposed Private Chats
Parameter Labs found that a malicious ChatGPT share link could execute code in a signed-in visitor’s session, read private conversations and send their contents to an external server. The chain crossed the HTML-preview sandbox through a MathLive message handler. OpenAI fixed both affected web frontends on October 26, classified the report as P1 and awarded a $5,000 bounty.
Fifty Companies Oppose OKXICE Stock Tokens
Fifty companies, including Nvidia, Microsoft, JPMorgan and Walmart, have objected to tokenized versions of their shares on OKXICE, the planned venue backed by OKX and Intercontinental Exchange. The objections target their own stocks being offered through the platform. The growing opt-out list complicates the venture’s proposed roster of tokenized equities.
Data centers hit the ballot as OpenAI plots a rebrand
This week’s U.S. midterms double as a referendum on data centers, with roughly 20 communities voting on bans or pauses, while OpenAI reportedly prepares a post-election push to rebrand the sites as “AI compute campuses.”
- Democrats are projected to flip the House Tuesday, which could end a two-year run of easy treatment in Washington for AI firms with close ties to Pres. Trump.
- Some Democrats want a federal freeze on new data centers, including House hopefuls Will Lawrence in Michigan and Rebecca Cooke in Wisconsin.
- The White House already got tech giants to voluntarily supply their own power, and last week Trump floated having developers pay locals a “dividend” as well.
- OpenAI is planning to pitch the campuses as local economic boosts, citing a Pike County, Ohio, project it’s helping build that promises thousands of jobs.
- OpenAI policy chief Chris Lehane called the debate “democracy in action,” and the company wants state bills giving locals a voice in new projects.
AI has a real image problem with the public, and data centers have become one of the defining fights of Tuesday’s elections. That anger won’t be fixed by a new name or a few community perks, and a Democratic House could make life a lot harder for AI giants trying to scale their infrastructure bets against local opposition.
Make AI artwork you can actually edit
In this guide, you’ll use Arrow 2 to create a three-part graphic for a client proposal. You’ll refine one scene, add editable headings, and change a color without generating the artwork again.
Step-by-step:
- Open QuiverAI, select Arrow 2, and check that your plan includes SVG export. Install Inkscape for the final text and color changes
- Paste the guide’s brief to draw a folder, review board, and laptop in one matching style. Leave space for the title and headings
- Ask Arrow to replace one badge with speech bubbles. Compare the result with the original and check that the other scenes stay intact
- Download the SVG and open it in Inkscape. Add the headings, then change the middle label and one screen color directly
- Save the editable master and export a high-resolution PNG. Reopen both files and check the words, colors, and dimensions before sharing
Pro tip: Keep text that may change in separate text boxes. You can fix a heading or update a client’s wording without spending another generation on the artwork.
Google’s space data center sends home a haiku
Google just announced that its Project Suncatcher satellite answered its first AI prompt from orbit, with Google’s open Gemma 4 model writing a haiku in an early test of the company’s moonshot goal to put AI data centers in space.
- Gemma’s haiku (Blue marble below / Data streams in silent grace / Starlight feeds my core) reached the ground while the satellite passed over Svalbard, Norway.
- Google CEO Sundar Pichai called it “a ‘hello world’ milestone,” a reference to the test programmers typically run first on a new machine.
- Power is still a big constraint, with the craft able to run just one of its four Google AI chips (TPUs), and only in bursts of roughly 20 minutes.
- Two more powerful satellites launching in 2027 will test the laser links a fleet needs to split one AI job, if the chips first survive radiation and heat tests.
An admittedly heartwarming haiku still won’t win over a public that is increasingly fed up with data centers of any kind, but moving them to orbit could eventually be a compromise that sidesteps a lot of the issues driving that anger — making this moonshot project a big one to keep an eye on.