The Quarter That Changed Everything

A Chinese startup just did what Silicon Valley said was impossible. DeepSeek released V3 and R1—models that match OpenAI's GPT-4o performance at roughly 5% of the cost. The market noticed: Nvidia shed billions in market cap as investors suddenly questioned whether more efficient software might reduce hardware demand.

R1 introduced "reasoning" capabilities similar to OpenAI's o1—but as an open-weights model anyone can run. The geopolitical implications are staggering: state-of-the-art AI is now accessible without US-centric infrastructure or budgets.

Forget chatbots. The theme of CES 2025 was unmistakable: AI agents that do things, not just talk about them. Microsoft, Nvidia, and a swarm of startups showcased software that codes, shops, automates—anything a human might do at a keyboard.

The hardware followed suit. "AI PCs" with dedicated neural processing units took center stage, alongside purpose-built agent devices that live on your desk, not in your browser. The pitch: 2024 was about generating text; 2025 is about completing tasks.

This isn't just marketing. When Anthropic gave Claude the ability to control a mouse cursor in October, they crossed a threshold: from "AI that answers questions" to "AI that executes workflows." CES confirmed the industry is all-in on that transition.

Google Quantum AI unveiled Willow, a chip that does something previously theoretical: error rates that decrease as you add more qubits. In quantum computing, that's the Holy Grail—the breakthrough required to build useful machines instead of lab curiosities.

The benchmark numbers border on absurd. Willow completed a computation in minutes that would take a classical supercomputer 10 septillion years. More importantly, it demonstrated that scaling up quantum systems is now possible without drowning in noise.

OpenAI staged a 12-day launch blitz that felt less like a product announcement and more like a flex. The finale: o3, their most capable reasoning model yet, blowing past o1 on coding and math benchmarks (including the notoriously difficult Arc-AGI tests).

The supporting cast was just as significant: Sora video generation went live (turbo mode), Canvas debuted as a collaborative interface, and a $200/month "Pro" tier signaled that enterprise-grade AI is no longer cheap. The pace from o1 to o3 was months, not years.

The message: the "reasoning model" race is accelerating faster than anyone predicted. If you thought GPT-4 was the plateau, you weren't paying attention.

AMD's Instinct MI300A chips now power the fastest computer on Earth. El Capitan, built for Lawrence Livermore National Laboratory, hit 1.742 exaflops—crushing the previous record and cementing AMD as a legitimate rival to Nvidia in high-performance computing.

The machine's purpose is nuclear weapons simulation, but the implications extend far beyond defense. AMD has proven its silicon can compete at the absolute frontier of compute. That changes the conversation for anyone building AI infrastructure at scale.

Claude 3.5 Sonnet gained a new capability in October: controlling a computer cursor, clicking buttons, typing text, navigating software. Not in a sandboxed demo—on actual desktops, executing real workflows.

This is the bridge between "AI that reasons" and "AI that acts." Developers can now build agents that use browsers, IDEs, spreadsheets—any tool a human would. The beta release was cautious, but the direction is clear: chatbots are evolving into digital workers.

Combined with CES's agent-everywhere message, Anthropic's move marked the philosophical inflection point. The question is no longer "can AI understand?" but "what can AI do?"

Google signed the world's first corporate deal to buy power from small modular nuclear reactors. Microsoft is restarting Three Mile Island. Amazon is shopping for nuclear capacity. Meta issued RFPs for 1-4 gigawatts of nuclear power.

The message is unmistakable: AI's growth is now constrained by electricity, not software. Training frontier models and running inference at scale requires power that existing grids can't reliably provide. Big Tech's response? Vertical integration into energy infrastructure.