Nvidia's Boardroom Crisis Meets an AI Infrastructure Reckoning

Nvidia's Leadership Crisis and the Chip Industry's Big Bets

The Wall Street Journal and New York Times report that Nvidia's board is preparing to remove CEO Jensen Huang following revelations of a relationship with an executive who reports to him, while Ars Technica adds that Nvidia is separately investigating possible Chinese espionage tied to the same executive, framing the matter as a national-security exposure rather than just a governance lapse. If accurate, it would be one of the most consequential leadership changes in chip-industry history given Nvidia's central role in AI compute supply. Meanwhile TSMC's board approved roughly $40.9 billion in capital appropriations for advanced-node production, packaging, and fab buildout, signaling that foundry capacity remains the binding constraint on AI hardware supply, even as SoftBank swung to a $7 billion quarterly loss after writing down advance payments for power equipment tied to its US AI datacenter buildout shows power procurement can turn into real balance-sheet pain.

Away from Nvidia's boardroom drama, other chipmakers are refining non-GPU paths. AMD's upcoming Zen 6 CPUs focus on redesigning memory and cache subsystems to keep wider cores fed, according to a technical deep dive, offering a useful counterpoint to the GPU-centric AI hardware narrative. IBM, meanwhile, says its new Nighthawk quantum processor demonstrates improved speed and circuit quality over prior generations, with a stated roadmap to scale beyond 1,000 qubits by 2027 - incremental progress toward quantum systems useful for real workloads rather than a leap to fault tolerance.

Faster Models and Smarter Agent Workflows

On the research side, PyTorch's ExecuTorch team detailed a new WebGPU backend that runs models directly on a browser's GPU instead of falling back to WASM or CPU, a concrete path to desktop-class inference speed for client-side ML products without shipping data to a server. Apple's ML research group published a method for distilling knowledge from trained attention-based transformers into Mamba-style state-space models, offering a cheaper path to Mamba's lower inference cost without full retraining. And Goodfire's research explores treating individual neurons as a usable, interpretable basis for understanding model internals, laying out when raw neurons are informative versus entangled and misleading - a foundation for debugging model behavior in production.

Coding agents are getting infrastructure built around reliability and collaboration too. Google's Antigravity project documented Swarm, a system for coordinating many agents on a task and cross-checking their outputs instead of trusting a single agent's trace, targeting the reliability gap that shows up on longer, open-ended tasks. Cursor announced that Projects now supports multiplayer, letting multiple developers work within the same agentic coding workspace simultaneously rather than running isolated sessions. And Atuin's new Threads feature extends its shell-history syncing approach to AI agent sessions, making coding-agent context available across every machine an engineer works from.

Forecasts and Arguments About What AI Means for Work

The AI Futures Project published an updated timelines analysis arguing that, based on recent model progress, an 'intelligence explosion' scenario they previously modeled further out now looks plausible by 2027, walking through the reasoning rather than just asserting a date. That kind of compressed timeline feeds into sharper debates about what AI tools are already doing to engineering work: one essay argues AI-assisted development is pushing individual output so high that '100x developer' framing deserves serious re-examination, while Tom Tunguz argues the traditional experience-based moat between junior and senior talent is eroding fast enough that everyone is effectively 'junior' relative to the tools' capabilities. A longer piece from General Intelligence Company pushes the logic further, laying out what organizational and financial scaffolding would be needed for agent networks to handle most of a company's operational decisions with minimal human involvement.

Security Basics Worth Rechecking

On the defensive side, a practical guide walks through pinning GitHub Actions to commit SHAs, restricting third-party action usage, and auditing workflow permissions to prevent CI/CD supply-chain compromise, a quick audit worth running given how often compromised Actions become an attack vector. Wiz detailed an active exploitation case where Qubole's cross-account IAM access pattern, trusted by default for legitimate integration, was abused by attackers to gain unauthorized access, with detection and remediation guidance for teams running similar third-party integrations. And ESET published a technical breakdown of tooling used by the Turla APT group to maintain persistence and exfiltrate data from Microsoft Exchange environments, useful for security teams updating detection rules against advanced persistent threats.