Compute Above the Clouds; DeepSeek: Hackers' Fave; Claude Testing on You

Today's AI Outlook: 🕺

This Data Center Has Plenty of Space

SpaceX is partnering with Nvidia to build its Starmind orbital data centers around Nvidia’s Vera Rubin NVL72 system, with Elon Musk targeting the first space-optimized racks for orbit by Q4 2027. The hardware will be redesigned for radiation, heat, weight and cost, giving SpaceX a purpose-built AI computer for workloads including Grok and its satellite operations.

Why it matters

AI data centers need huge amounts of power, cooling and land, while new projects are facing rising costs and local opposition. SpaceX is betting reusable launches and specialized hardware can eventually make orbit economically competitive.

The Deets

• Nvidia’s NVL72 links 72 chips into one computing system.
• Nvidia says each Vera Rubin chip can deliver up to 25x the compute of an H100.
• Musk says the orbital version will be lighter, denser and cheaper than conventional racks.
• Orbital compute is estimated to cost more than 4x ground-based compute today.

Key takeaway

SpaceX wants to give AI infrastructure somewhere new to scale when terrestrial power grids and permitting become bottlenecks.

🧩 Jargon Buster - NVL72: Nvidia’s rack-scale system that connects 72 GPUs so they operate like one enormous AI computer.

♟️ Power Plays

Thomson Reuters Wants to Own More of Its AI

Thomson Reuters built its first homegrown AI model by adapting Alibaba’s Qwen and training it on decades of proprietary legal and professional content. The two-year project cost about $40M.

Why it matters

Companies with valuable data are starting to calculate whether owning specialized models makes more sense than paying frontier-model providers indefinitely.

The Deets

• The latest training run cost about $450,000.
• Internal benchmarks reportedly beat several frontier models in selected tasks.
• The model has used less than 10% of Thomson Reuters’ content library.
• An open-weights version is planned for researchers.

Key takeaway

Strong open models are making proprietary corporate AI far more attainable.

🧩 Jargon Buster - Open weights: A model whose trained parameters are released so others can run or modify it.

DeepSeek Gets an Unwanted Power User

Chinese state-linked hacking groups are carrying out more than twice as many attacks after integrating models such as DeepSeek into their operations, according to Taiwanese cybersecurity firm TeamT5.

Why it matters

Cheap, downloadable models can help skilled attackers write code, analyze targets and scale operations without relying on tightly controlled cloud services.

The Deets

• Researchers say attackers favor DeepSeek for its low cost and loose cyber safeguards.
• AI has reportedly been used to write intrusion code and analyze large target lists.
• One operation involved accessing a Taiwanese company’s email.
• Researchers warn that open models may be particularly difficult to police.

Key takeaway

AI is becoming a cheap force multiplier for experienced cyberattackers.

🧩 Jargon Buster - Guardrails: Restrictions designed to stop AI systems from assisting with dangerous or prohibited requests.

Anthropic’s A/B Test Touches Trust Nerve

Some Claude Code users were unknowingly placed into an Anthropic experiment that changed how reasoning-effort settings were represented. Anthropic said the underlying reasoning effort itself was not reduced.

Why it matters

Quiet experiments become more consequential when developers rely on AI tools inside production workflows and need to know whether a behavioral change comes from their code or the model.

The Deets

• A developer noticed an unexpected reasoning value in API logs.
• Anthropic confirmed an undisclosed A/B test.
• The company said the test changed effort-value mapping, not actual reasoning allocation.
• The episode sparked criticism over transparency.

Key takeaway

AI companies can iterate quickly, but invisible product changes make debugging and trust harder.

🧩 Jargon Buster - A/B test: An experiment where different users receive different product versions so a company can compare results.

🛠️ Tools & Products

Nvidia Gives AI Agents a Caffeine Injection

Nvidia says its Groq 3 LPX inference accelerator has entered full production, bringing technology acquired through Nvidia’s reported $20B Groq purchase into its broader AI hardware lineup. The specialized chip focuses on generating tokens quickly, with Nvidia positioning it for the increasingly demanding workloads created by AI agents.

Why it matters: AI agents often make many model calls during one task, so faster token generation can sharply reduce latency and cost.

The Deets

• A cited 5,000-token workload falls from about 50 seconds to 1.5 seconds.
• AI Secret describes the resulting economics as roughly a 35x improvement.
• Groq 3 LPX focuses specifically on inference.
• Nebius is the first announced cloud customer.

Key takeaway

Faster inference could make complex AI agents cheaper and much less painful to use.

🧩 Jargon Buster - Inference: Running a trained AI model to generate an answer, prediction or action.

💰 Funding & Startups

Nvidia + Perplexity; AI Valuations Keep Climbing

Perplexity is reportedly discussing a funding round involving Nvidia at a valuation as high as $30B, while General Intuition is raising money at a reported $6B valuation for world-model technology aimed at robotics.

Why it matters

Investors are still paying heavily for companies positioned around AI distribution, infrastructure and robotics.

The Deets

• Perplexity’s annual revenue reportedly topped $750M.
• Nvidia is discussing an equity investment and possible technology deal.
• General Intuition is developing models intended to help machines understand physical environments.

Key takeaway

Capital remains plentiful for startups that control strategic pieces of the AI stack.

🧩 Jargon Buster - World model: An AI system that builds an internal representation of an environment so it can predict what may happen next

🧪 Research & Models

Faraday Puts a Smaller Model in Charge

DeepMind alumni-founded Inherent introduced Faraday, an AI research agent that uses a 27B-parameter Qwen 3.6 model to plan scientific experiments while GPT-5.5 Codex handles coding.

Why it matters

Faraday suggests that specialized model orchestration can sometimes beat relying on one massive frontier model for every part of a task.

The Deets

• Faraday decides what experiments to run and when to stop.
• GPT-5.5 Codex handles coding work.
• Inherent says the system beat Claude Opus 4.8 and GPT-5.5 Codex on its internal research-replication benchmark.
• Those results still need broader independent validation.

Key takeaway

Smarter delegation may become as important as raw model size.

🧩 Jargon Buster - Reinforcement learning: A training method that rewards an AI system for behaviors that lead to better outcomes.

⚡ Quick Hits

• Meta reportedly plans to launch consumer AI agent platform Hatch within weeks.
• Einride ordered 500 Tesla Semis for an AI-optimized freight network across several U.S. states.
• Porsche signed a five-year, $1.46B TCS deal to integrate AI across customer, factory and engineering operations.
• Waymo revealed a custom 5nm chip that preprocesses data from 13 robotaxi cameras.
• The U.K. and Ukraine signed an AI defense partnership involving battlefield data.
• Descartes bought Tai for about $100M, adding AI-enabled freight brokerage software.
• Luke Metz reportedly joined Meta after previous stops at OpenAI and Thinking Machines.

🔧 Tools of the Day

• Open Design builds reusable AI design systems by extracting visual rules from websites, logos and other brand assets.
• Wan 3.0 is Alibaba’s latest generally available AI video-generation model.
• Antigravity is Google’s agentic development tool, now with remote-control capabilities.
• Firefly Audio generates commercially safe music, voiceovers and sound effects.
• Apodex 1.1 coordinates multiple AI teams working on tasks in parallel.
• Pipette benchmarks how AI models perform directly on phones and laptops.

Today’s Sources: The Internet, The Rundown AI, AI Secret