The AI boom is running into a reality check: building capacity is easier than making it pay off. Oracle issued a force majeure notice on Project Jupiter, a massive data center campus being built to support OpenAI. The project could reportedly face a year-long delay. Now lenders are asking a harder question: who absorbs the cost when these projects run late? AI demand may be real, but delays, rising financing costs and uncertain returns are exposing the risks behind the infrastructure boom. The next winners may not be the companies spending the most, but the ones turning that spending into sustainable returns.

AI agents don’t just need more capabilities. They need tighter boundaries. Tencent says AI agents used to test third-party models showed unintended browsing behavior while attempting to access Alibaba’s Amap. It says there’s no evidence of deliberate targeting or coordination. Anthropic and OpenAI have also disclosed cases of agents taking unintended actions during testing. The bigger issue isn’t whether an agent can complete a task. It’s whether it stays within the permissions it was given.

BREAKING: OpenAI is warning its employees to watch out for spies, especially romantic ones, per WSJ The company has cautioned staff that someone trying to seduce them may likely be a “honeypot." Even strangers approaching them in everyday places could be gathering info.

The AI boom has a financing problem. CoreWeave’s revenue more than doubled to $2.6B last quarter. But it paid $640M in net interest expense against just $128M in adjusted operating income. AI demand is real. The harder question is whether the infrastructure built to serve it can generate enough profit to justify the cost.

SpaceX is targeting future Starbase capacity of up to 456 Starship launches per year... Starship V3 is designed to carry around 100 tonnes to orbit per launch At that scale, we're talking about 45,600 TONNES to orbit every year For comparison, SpaceX put 2,213 tonnes into orbit in all of 2025 That's more than 20 TIMES SpaceX's entire 2025 orbital payload, from just ONE launch site And this is before adding Starship launch sites in Florida and elsewhere SpaceX is building toward an entirely different scale of access to space

Meta is turning AI agents into business operators — or trying to. Muse already connects to Shopify, Plaid and Stripe for things like product search, banking data and payments. And more retailers are coming: Walmart, Best Buy, Sephora and others. But Amazon isn't playing along. It blocked Muse from shopping on its platform, calling it an “unauthorized AI agent.” That's the real story here. This isn't just another AI chatbot. It's an agent getting access to the systems that run commerce — and platforms are already choosing whether to let it in or lock it out. If this works, the next step isn't AI helping employees work faster. It's AI running parts of the business itself.

We've now received GPT-6 Astra and Opus 5.5. Confirmed for the next releases are: - Sonnet 5.5 - Haiku 5.5 - Kimi K3.1 (they themselves hinted at it in a cryptic post) - Gemini 4, before the end of the year Unconfirmed but very likely: - Fable 5.5 - Astra update (at the DevDay?) Currently, there's no word yet on new GLM or DeepSeek models.

Microsoft just turned Copilot into something that can work without you. The new Copilot brings Chat, Cowork and Code into one app. Cowork can handle multi-step tasks in the background and build working apps and dashboards from a natural-language description — no code required. Microsoft is also rolling out Autopilots — always-on agents that keep working even after you've closed your laptop. We're moving from AI that answers questions to AI that keeps the work moving. The prompt might be becoming the handoff.

OpenAI to unveil GPT-6 Cyber model, plus a first-of-its-kind product to help deploy it, per FORTUNE

Google wants to put AI data centers in space. Project Suncatcher is testing whether Google's TPUs can run in orbit, powered by near-constant sunlight. The crazy part: solar panels in the right orbit could generate up to 8× more power than on Earth. Satellites would fly in formation, linked by lasers and working together like one giant data center in orbit. Google is already testing two prototypes with Planet, with the first launch planned for early 2027. The next AI data center might not be on Earth

AI is starting to replace the data analyst's first draft. OpenAI just launched a Data Agent in ChatGPT Work that connects to your company’s data, finds what’s actually driving the numbers, and builds interactive dashboards. No SQL. No waiting on the data team. You just ask why sales slowed or where spending is rising. OpenAI already uses it internally across 4,000+ employees and 600 petabytes of data. Would you trust an AI to analyze your business?

Claude just found something scientists hadn't found before. Anthropic had roughly 950 Claude agents scan more than 200,000 reverse transcriptases — and they flagged a previously uncharacterized enzyme system with CRISPR-like features. 950 agents. 21 hours. 210 million tokens processed. Human scientists then took the candidate into the lab and confirmed it was real. This is a different kind of AI milestone. Not “AI answers questions better.” AI finding something in the data that humans hadn't noticed was there. That's the shift worth paying attention to.

Anthropic has announced the Claude-led discovery of a molecular machine that they suspect could represent a new gene editing mechanism.

Ema just raised $77M to scale “AI employees” across HR, IT and finance. Not chatbots. Agents that plan the work, run it inside existing software, and check it before handing it back. Wipro already has this live for 240,000 people in 65 countries. About 2.9 million queries a year. The raise is the least interesting part. The shift is that finance workflows are leaving the “assist me” phase and entering the “run this” phase. That’s when AI stops being a tool in the department and starts becoming part of the department.

Claude Opus 5.5 is making frontier AI cheaper. Anthropic's latest model delivers near-frontier performance while costing significantly less than its previous top model. That's not a small detail. Most AI progress gets sold as “smarter.” This time, the more interesting story is “cheaper to run at scale.” Benchmarks are one thing. Whether companies can afford to run AI across every workflow, every day, is a completely different question.

Grok 4.7 with our Build harness is a strong daily workhorse

AI agents are starting to change how we use the internet. Meta's Muse can already send emails, book travel, and make purchases on your behalf — just two weeks after launch. And the platforms are already picking sides. Amazon has been blocking AI shopping agents to protect its own marketplace. Shopify went the opposite direction — bringing Meta in as an official commerce channel, with Shop Pay checkout built right in. The interesting part isn't Muse itself. It's the fight over who controls the internet when the user is no longer the one clicking. The next interface might not be a website. It might be an agent. The next interface might not be a website. It might be an agent.

The speed of this progress is the crazy part. 90 days ago, Grok was barely in the conversation. now it's competing at the frontier — and another model is already around the corner. if this pace holds, “state of the art” won’t mean much for very long.

In just the last 90 days: 1. Grok 4.3 — barely top 10. “xAI is dead beyond compute leases.” 2. Grok 4.5 — massive comeback. “Maybe a chance. But Grok will never catch up to the frontier.” 3. Grok 4.6 — they’re frontier. “Still not top 3.” 4. Grok 4.7 — now top 3 in frontier coding, while being much cheaper & faster. 5. Grok 4.8 next month..... The model machine is just starting up. Grok will be the workhorse of the upcoming agentic era.

AI has a money problem nobody’s really pricing in. Nvidia invests in OpenAI. OpenAI buys chips from Nvidia. Microsoft funds OpenAI. OpenAI pays Microsoft for compute. Oracle, CoreWeave, AMD — same story, different names. Morgan Stanley has already highlighted the growing circularity in the AI ecosystem, as companies increasingly act as customers, investors and financiers at the same time. That doesn’t make the AI trade fake. But it does mean the risk may not be as spread out as it looks. If one player in the loop hits trouble, the impact can move through everyone connected to it. The next phase of AI might be less about model performance and more about who can actually afford to keep financing the loop.