AI's Power Strain Meets Org-Chart Doubts

Compute Keeps Growing, Power Pushes Back

Nvidia's board is reportedly weighing a move against CEO Jensen Huang, with tensions tied to the company's China strategy, according to CBS News. Because Nvidia sits at the center of nearly every AI infrastructure decision, any leadership upheaval there would ripple into chip supply and roadmap planning far beyond Santa Clara.

Even as compute demand grows, the money behind it keeps growing too: Anthropic has signed another $4.3bn in compute agreements with Microsoft, adding to an already deep infrastructure commitment and underscoring how capital-intensive frontier model development has become. That demand is colliding with physical limits, as data center energy needs are growing roughly ten times faster than grid capacity can expand, and the town of Ottawa, Kansas has approved a ten-year moratorium on new data center construction, a sign that local resistance is becoming a tangible siting risk. The constraints aren't only regulatory, either: a false smoke alarm at a Jane Street datacenter in New York nearly triggered water damage from its fire-suppression system, a reminder that physical failure modes remain a real operational risk no matter how sophisticated the software layer above them.

Agents and Generative Models Move Into Products

Several companies are pushing generative and agentic AI deeper into core workflows rather than treating them as bolt-on features. Shopify's engineering team detailed its generative discovery model, which reframes product search as a generative task instead of traditional keyword retrieval, with real architecture tradeoffs for anyone building search or recommendation systems at scale. Databricks expanded Genie Agents to analyze more data sources, pushing agentic querying further into the warehouse layer rather than leaving it to separate BI tools. Dropbox's engineering blog traced how two engineers' side projects became the agent now taking action inside Dropbox, an unusually candid account of exploratory work turning into a shipped feature. For teams building internal knowledge assistants, a deep dive argues GraphRAG improves on vanilla retrieval by encoding entity relationships explicitly, enabling multi-hop reasoning that flat vector search struggles with. Adobe is opening Creative Cloud apps to third-party AI assistants, letting external agents drive tools like Photoshop and Illustrator directly, a move read alongside its partnership with Google as a signal of a more AI-native future for Photoshop's workflow.

When Standard Checks Miss the Real Problem

A LessWrong write-up shows LLMs can learn multi-step reasoning using filler tokens rather than explicit visible chain-of-thought, meaning models can offload computation into channels that aren't captured in what they display, a real concern for anyone treating visible CoT as a safety or debugging signal. A similar lesson shows up in software engineering: in the second installment of a series on an email system called Barista, the author's passing test suite still had a bug that silently dropped emails, one that a TLA+ model checker caught before it could ship. And in security, researchers detail a technique attackers use to keep reading Microsoft 365 mailboxes even after an organization revokes sessions, undermining a standard first response to compromise. Each case shows that the checks teams rely on, whether model output, passing tests, or session revocation, don't always capture what's actually happening underneath.

Two other pieces push toward more rigor rather than relying on surface checks. A practical guide walks through locking GitHub Actions to specific commit SHAs rather than mutable tags, closing off a well-known supply-chain path where a compromised or re-tagged action can inject malicious code into a CI pipeline. And a long-form research piece argues for deriving GPU kernels systematically rather than hand-writing them, treating kernel construction as a derivation problem with correctness guarantees instead of ad hoc tuning, a dense but substantive reference for low-level performance work.

Rethinking Org Design as AI Shifts Value

As AI reshapes what individual contributors can do, several writers are questioning old assumptions about organizational structure. Amplify Partners published a long essay questioning the default justification for management layers, arguing that AI-leveraged individual contributors deserve a fresh look at whether manager-heavy structures still make sense. A CNCF blog post makes a related argument about platform teams, warning they are over-rotating on pure automation at the expense of understanding what developers actually need, while still needing to remove themselves as a bottleneck in day-to-day work. Tom Tunguz separately maps out where margin accrues across the AI stack, from chips to model providers to application layers, a framework useful for founders weighing where pricing power is likely to survive commoditization. And Jason Wei argues that as AI narrows the gap in what people can technically accomplish, who you know becomes a bigger determinant of career and company outcomes than raw capability.