The Week AI Grew Up, Broke Things, and Got Distilled

OpenAI Can't Build Fast Enough — And That's Their Biggest Flex

Let me be direct: OpenAI pausing Pro subscriptions because demand exceeded capacity is the most bullish signal in tech this week. Most companies dream of that problem. The trigger was Astra — the agentic capability embedded in GPT-6 — which is apparently burning through compute at rates OpenAI didn't anticipate. The same week, OpenAI launched ChatGPT for Financial Services, a verticalized Work experience targeting equity research and financial modeling, built directly on top of Astra reasoning. This isn't a coincidence. OpenAI is shifting from horizontal consumer scaling to vertical enterprise capture, and using the Pro pause to manage the transition. The Bloomberg Terminal disruption narrative is real — but the bigger story is distribution. OpenAI is bundling finance-specific data and workflows into a single subscription, which is exactly the playbook that made Excel sticky in the 1990s. If Astra delivers even half of what OpenAI promises for analysts, the $25,000-per-seat Bloomberg Terminal starts looking defensively overpriced. The subscription pause is a temporary speed bump; the Financial Services launch is the strategic play that matters. Watch which banks sign pilot deals first — that's the real signal.

Distillation Wars: Anthropic Names Names

Anthropic published a report this week accusing Alibaba, Moonshot AI, and DeepSeek of running systematic distillation campaigns to replicate its models. This is unprecedented transparency — and it's a strategic escalation. For months, the industry has whispered about Chinese labs training on outputs from Western frontier models. Anthropic just put it in writing with attribution. Here's why this matters beyond the IP theater: distillation is how open-weight models close the gap to frontier closed models without paying for the R&D. If Anthropic's claims hold, it means Chinese labs are systematically compressing months of safety research and alignment work into weeks — work that Anthropic and others invested billions to produce. The report lands as Mistral raised €3 billion explicitly to push open-weight to the frontier. The timing is not accidental. There's now a real geopolitical fault line forming: closed frontier labs (OpenAI, Anthropic) versus open-weight champions (Mistral, DeepSeek) versus sovereign AI efforts. Mistral's €21 billion valuation tells you investors believe open-weight is a durable category, not a transitional one. My read: Anthropic is trying to draw regulatory lines before distillation becomes the default development methodology. Whether Washington or Brussels has the appetite to enforce those lines is a different question entirely.

Salesforce's Six-Pack and the Enterprise AI Integration Tax

Salesforce unveiled its Enterprise AI Harness on Thursday, bundling six existing AI tools into one unified infrastructure. On the surface, this is a marketing announcement. Underneath, it's an admission that the enterprise AI market has an integration problem — and Salesforce is positioning itself as the solution. Every enterprise I talk to is drowning in AI point solutions: one vendor for summarization, another for agentic workflows, a third for RAG, a fourth for image generation. Stitching them together eats the productivity gains they're supposed to deliver. Salesforce's bet is that customers will pay a premium for unified orchestration rather than assembling their own stack. This is the same play Microsoft made with the Office suite in the 1990s — bundle the unsexy integration work and charge for the convenience. The risk for Salesforce: if the underlying tools are merely average, bundling just creates a single point of failure. But given that Salesforce ships with Agentforce, Einstein, and Data Cloud in the mix, they have enough depth to make the bundle defensible. The bigger question: does any enterprise want a single-vendor AI stack, or will they always prefer best-of-breed? I'm betting on bundles winning the mid-market and losing the Fortune 100.

The Robots Are Coming — On-Device and On Schedule

Three stories this week pointed to the same future: edge AI on Jetson devices running real LLM-powered robotics. The tutorials covered vision-language robots, on-device LLM assistants, and Docker packaging for ROS 2 on Jetson hardware. Individually, these are developer blog posts. Collectively, they're the early scaffolding of a robotics stack that doesn't need the cloud. The implications are bigger than they look. Cloud-dependent robotics has a latency floor that makes real-time human-robot interaction frustrating. Moving LLMs onto Jetson — even small ones — changes the architecture. A robot that interprets natural language locally can operate in environments without connectivity: warehouses, hospitals, disaster zones, homes. This week's tutorials are the equivalent of the 2008 iPhone SDK announcement — quiet, technical, and foundational. Combine this with Proxima Fusion's €140 million bet on European fusion-grade superconducting tape, and you see two sides of the same coin: compute infrastructure is decentralizing. The cloud isn't disappearing, but the assumption that everything must run there is finally dying. For developers, the message is clear: learn edge deployment now, because the robotics companies hiring in 2027 won't wait.

Software Is Breaking, and Nobody's Surprised

This week's bug list is staggering: September Windows Server updates broke Remote Desktop Services across 2019, 2022, and 2025. Microsoft's KB5002914 broke Excel copy-paste globally. Surfshark's VPN — a privacy company — got breached via a misconfiguration. Shopify spent years on React Native, then rebuilt everything native in 12 weeks. Shopify's story is the most damning. If a company with Shopify's engineering resources concluded that React Native wasn't delivering performance or developer velocity, what does that say about the cross-platform promise? The pivot validates every team that stuck with native development and burns the bridge for anyone considering React Native for performance-critical mobile apps. On the security side, the BYOD voice-call attack on Microsoft 365 is the scariest story of the week. Attackers used personal smartphones on corporate networks to infiltrate M365, then handed off access to ransomware crews. This is a new attack surface that most enterprises don't even know exists. The lesson: every security architecture built before 2024 needs to assume personal devices are inside the perimeter. Combined with Excel and RDP breaking, this was a week where the foundation cracked while everyone was busy building the roof.

The Privacy Backdoor Nobody's Talking About

Spirit Airlines' bankruptcy is about to become a case study in data ethics. The carrier is preparing to sell passenger data to Google as part of its asset liquidation, and privacy advocates are rightly alarmed. This sets a precedent that should terrify everyone: bankruptcy as a shortcut around data protection law. When a company files Chapter 11, its assets — including data assets — become saleable to satisfy creditors. But user data isn't really the company's to sell. Users never consented to having their travel history, payment patterns, and identity information auctioned off to the highest bidder. If this deal closes, expect a wave of copycat bankruptcies where data brokers acquire datasets they could never legally collect from a solvent company. Google's interest is obvious: AI models are starving for high-quality behavioral data, and airline passenger data is uniquely rich. It's origin-destination pairs, purchase timing, companion information, and payment fingerprints. For Google, this is worth far more than the bankruptcy auction price. For users, it's a reminder that the fine print in every terms-of-service agreement is a time-delayed privacy waiver. Expect EU regulators to intervene first — they have the legal tools and the political will. The US will be slower.

That's the week AI grew up, the week software kept breaking, and the week the robots got their first real job description. The pace isn't slowing — it's accelerating into the friction. See you Monday.