SpaceRock

xAI has officially switched on Colossus 2, a record-breaking AI training cluster consuming one gigawatt of power. That is more electricity than the entire city of San Francisco uses at peak times. As the first operational system of this scale, it allows xAI to train massive frontier models at a single location rather than splitting the workload across different data centers. The facility is expected to scale up to 1.5 gigawatts by April, with plans to reach 2 gigawatts soon after. The system is powered by approximately 555,000 GPUs, representing an estimated hardware investment of $18 billion.

OpenAI is officially “on track” to launch its first AI hardware device in 2026. Executive Chris Lehane confirmed at Axios House Davos that the company is shifting focus from experimental concepts to stable, everyday consumer products. He described the upcoming hardware as a more natural way to interact with AI, moving away from reliance on traditional screens like phones and laptops. This timeline aligns with recent comments from CFO Sarah Friar regarding 2026 as a pivotal year for monetization. It also fits CEO Sam Altman’s vision to define a new category of AI-first devices, developed in collaboration with former Apple design chief Jony Ive.

OpenAI has begun testing ads for free users in the US, offering increased usage limits in exchange for viewing sponsored content. At the same time, the company is expanding “ChatGPT Go” globally. This is an $8/mo plan that includes ads but unlocks extra features, while Plus and Enterprise tiers remain ad-free. OpenAI assures users that ads will be clearly labeled, won’t influence generated answers, and won’t rely on sold user data. Additionally, ads will not be shown to minors or appear alongside sensitive topics like health or politics. While the move is designed to fund broader access to advanced AI, critics warn it could compromise user trust if the chatbot begins promoting products too aggressively.

New Jersey’s outgoing governor Phil Murphy signed a deal with NVIDIA on his way out, committing $25 million to build a state-run supercomputer for universities, community colleges, and an AI hub. It’s pitched as shared infrastructure for students, researchers, and startups—essentially an industrial policy bet on building local AI capacity rather than relying on rented cloud services. The timing locks it in before the next administration takes over, making it harder to abandon. While other states have done similar deals, this one bundles education, workforce training, and compute into one package. The real test is whether it becomes genuinely shared infrastructure that keeps AI talent in-state, or just fades after the press conference.

The U.S. has imposed a 25% tariff on advanced AI chips like Nvidia’s H200 and AMD’s MI325X that are produced abroad but exported from America to China—a strategic move aimed at limiting China’s access to cutting-edge AI technology while supporting domestic chipmakers who prefer selling with tariffs over losing market access entirely. The tariff exempts chips used domestically for research or defense, and comes as Chinese companies rush to place early orders while potentially reconsidering their import restrictions to avoid falling behind in AI development. While this signals America’s intent to reduce dependence on foreign semiconductor production—the U.S. currently manufactures only 10% of its own chips—the tariff itself doesn’t address the fundamental bottleneck: America’s limited domestic chip manufacturing capacity, which remains the core challenge in achieving true technological independence.