Meta buys an agent startup, NVIDIA pays $20B for inference, and CES sets the AI roadmap

Big M&A activity kicks off the year, CES 2026 reveals where the major players are pointing, and Claude Code is now writing 100% of its own code.

This week was all about who’s buying whom, what NVIDIA showed at CES, and a quiet milestone in AI coding capability.

Meta acquires Manus for over $2B

My take: Manus built general-purpose agents that perform tasks by writing and running Python scripts, and hit a $125M revenue run rate in just 8 months. Meta is moving fast to get agentic capability into WhatsApp and Ray-Ban smart glasses. The signal for the rest of us: agent startups are now the new acquisition target, just like AI assistants were two years ago.

NVIDIA's $20B deal with Groq for inference chips

My take: Structured as a licensing deal but functionally an acquisition. Groq makes high-speed inference chips, exactly what you need for real-time agents that can't wait three seconds to respond. NVIDIA is locking down the inference market as well as training. Inference costs are going to keep falling, which is good news for anyone building products on top of AI.

NVIDIA's Vera Rubin chips already in production

My take: Promising 5x faster inference and a 90% cut in token costs compared to current Blackwell chips. They're also redesigned to handle long-running agent tasks. If you're building anything that uses AI heavily, expect the unit economics to keep improving every six months. Don't lock in long contracts at today's prices.

Amazon revamps Alexa with a web app

My take: Alexa now has a proper chatbot interface at alexa.com that works on any device. The strategy is to make Alexa the family's AI command center, not just a smart speaker. Worth watching if you're in retail, hospitality, or anything where a household-level AI assistant could change how people discover and buy.

Jevons Paradox for knowledge work

My take: Box CEO Aaron Levie's point is the one to internalize. When something becomes cheaper and more efficient, demand for it usually goes up, not down. Knowledge work follows the same rule. The companies that win are the ones that take on more projects, not the ones that lay off and call it efficiency.