AI Week, Jan 5–11, 2026: Nvidia's Rubin, Meta's Nuclear Deals and Google's Push Into AI Checkout
Nvidia launched its Rubin platform and Alpamayo driving models at CES, OpenAI unveiled ChatGPT Health, Snowflake agreed to buy Observe, Meta signed nuclear deals for up to 6.6 GW, and Google launched the Universal Commerce Protocol.
Written on November 15, 2026, covering news from Jan 5–11, 2026.
Nvidia opened the week at CES selling the chips and models behind the next round of AI, and by Friday Meta was lining up nuclear power for its own AI buildout. Midweek, OpenAI pushed its assistant further into health questions and Snowflake agreed to buy an observability startup built on its own database, and on Sunday Google moved into AI-powered checkout.
Monday, January 5
Nvidia used CES 2026 in Las Vegas to launch Vera Rubin, a superchip that puts 1 Vera CPU and 2 Rubin GPUs into a single processor. It is one of 6 chips in what the company now calls the Rubin platform, alongside the NVLink 6 Switch, ConnectX-9 SuperNIC, BlueField-4 DPU and Spectrum-6 Ethernet Switch.
Those parts can be assembled into the Vera Rubin NVL72 server, a single system with 72 GPUs, and linking several NVL72s produces Nvidia’s DGX SuperPOD. Nvidia also pitched a storage layer, Nvidia Inference Context Memory Storage, which it says is needed to hold and share data generated by trillion-parameter and multi-step reasoning models.
The headline claims are measured against Blackwell: Nvidia says Rubin needs 4x fewer GPUs to train the same mixture-of-experts model and brings a 10x reduction in inference token costs. The company is aiming the platform at agentic AI, advanced reasoning models and MoE models. “Rubin arrives at exactly the right moment, as AI computing demand for both training and inference is going through the roof,” CEO Jensen Huang said in a statement.
The same day, Nvidia introduced Alpamayo, an open source family of AI models, simulation tools and datasets meant to help self-driving cars work through difficult road situations. Its centrepiece, Alpamayo 1, is a 10 billion-parameter chain-of-thought vision language action (VLA) model whose underlying code is on Hugging Face. Nvidia is also releasing an open dataset of more than 1,700 hours of driving data, plus AlpaSim, an open source simulation framework on GitHub for validating driving systems.
Huang framed it broadly: “The ChatGPT moment for physical AI is here – when machines begin to understand, reason, and act in the real world”. Ali Kani, Nvidia’s vice president of automotive, told a press briefing the model works “by breaking down problems into steps, reasoning through every possibility, and then selecting the safest path”. Developers can fine-tune Alpamayo into smaller, faster versions, use it to train simpler driving systems or build tools such as auto-labelers on top, and Kani said they can pair it with Nvidia’s Cosmos world models to generate synthetic data.
Wednesday, January 7
OpenAI announced ChatGPT Health, a dedicated area inside the chatbot for health conversations. The company says more than 230 million people ask health and wellness questions on ChatGPT each week. Files, conversations and connected apps in Health are kept apart from other chats, information and memories from it won’t flow out of that space, and OpenAI says those conversations won’t be used to train its foundation models.
Users will be able to link medical records and apps including Apple Health, MyFitnessPal, Weight Watchers and the lab testing startup Function, with b.well supplying the connectivity infrastructure for records. If someone starts a health conversation outside the section, the AI aims to nudge them into it.
OpenAI said the product is not for diagnosis or treatment and is not meant to replace medical care, and that it was developed in “close collaboration” with physicians. Its terms of service already say it is “not intended for use in the diagnosis or treatment of any health condition”. Fidji Simo, OpenAI’s CEO of applications, called it “another step toward turning ChatGPT into a personal super-assistant”. Access starts with a small group of early users, and OpenAI says it will widen access in the coming weeks.
Thursday, January 8
Snowflake said it signed a definitive agreement to acquire Observe, an observability platform that has run on Snowflake’s databases since its start, with the deal subject to regulatory approval. Terms weren’t disclosed; reports put the value at around $1 billion, which would make it Snowflake’s largest acquisition, ahead of its $800 million purchase of Streamlit in March 2022. SiliconANGLE noted that The Information cited sources putting the value at about $1 billion when rumors surfaced the previous month.
Observe was founded in 2017 by Jacob Leverich, Jonathan Trevor and Ang Li, has raised $316 million, and the company says its most recent valuation, set in July 2025, was $750 million. The ties are close: both companies were incubated at Sutter Hill Ventures, and Observe CEO Jeremy Burton has sat on Snowflake’s board since 2015.
Snowflake plans to fold Observe into its platform, with particular emphasis on AI SRE, a chatbot that helps developers diagnose technical issues. A Snowflake blog post claims the combination lets users spot and fix issues 10x faster than before. Christian Kleinerman, Snowflake’s executive vice president of product, wrote that the pair will help customers run “production applications and agents with greater confidence, without sacrificing data to control costs”.
Friday, January 9
Meta announced agreements with Vistra, TerraPower and Oklo for nuclear power tied to its Prometheus supercluster, which is being built at a data center in New Albany, Ohio; financial terms were not disclosed. Meta says the 3 deals will back as much as 6.6 gigawatts of clean power, both new and existing, by 2035. CBS News put that at roughly enough to power about 5 million homes.
Under 20-year agreements, Meta will buy more than 2.1 GW from Vistra plants in Pennsylvania and Ohio, plus uprates at the Ohio plants. The TerraPower deal funds 2 Natrium units capable of up to 690 megawatts, with delivery as early as 2032, and gives Meta rights to energy from up to 6 more units capable of 2.1 GW, targeted for delivery by 2035. Oklo will develop a 1.2 GW power campus in Pike County, Ohio, with a first phase targeted for as early as 2030 and plans to expand incrementally toward a target of up to 1.2 GW by 2034.
Vistra shares rose 10% and Oklo 8%, while Meta closed 1% higher. OpenAI CEO Sam Altman formerly chaired Oklo’s board, and FactSet data cited by CNBC show he holds 4.3% of the company, a stake worth roughly $650 million at Thursday’s closing price. “State-of-the-art data centers and AI infrastructure are essential to securing America’s position as a global leader in AI,” Meta policy chief Joel Kaplan said. Meta expects Prometheus, a 1-gigawatt cluster, to come online in 2026.
Sunday, January 11
At the kickoff of the National Retail Federation’s annual show, Google launched the Universal Commerce Protocol (UCP), an open standard meant to let AI agents and retail systems work together from discovery and buying through post-purchase support. Google says UCP was co-developed with Shopify, Etsy, Wayfair, Target and Walmart, is endorsed by more than 20 others including Visa, Mastercard, Stripe and The Home Depot, and is compatible with A2A, AP2 and MCP.
Soon, UCP will run a new checkout option on eligible product listings in Google’s AI Mode and the Gemini app, so shoppers can buy from eligible U.S. retailers using Google Pay, with PayPal to follow. Retailers remain the seller of record. Google also launched Business Agent, a branded chat agent on Search that it said is live “starting tomorrow” with retailers including Lowe’s, Michael’s, Poshmark and Reebok. And it is piloting Direct Offers, an ads feature that lets advertisers show exclusive discounts, such as 20% off, to AI Mode shoppers who are ready to buy.
Vidhya Srinivasan, Google’s vice president of ads and commerce, said: “Our role in the ecosystem is that of a matchmaker and one way is with ads”. CNBC noted that OpenAI’s Agentic Commerce Protocol, developed with Stripe, is open source and could compete with UCP. CNBC also cited an October McKinsey report estimating that AI-powered tools and agentic commerce could make retail a $3 trillion to $5 trillion opportunity globally by 2030.
Also this week
Nvidia’s “full production” claim (reported on January 5). In its CES recap, Nvidia said Rubin, the successor to Blackwell, is now in full production, and listed Rubin GPUs with 50 petaflops of NVFP4 inference. THE DECODER reported that Nvidia expects availability through partners starting in the second half of 2026, and claims 3x the training compute and 5x the inference compute of Blackwell for NVFP4. THE DECODER, citing Wired, said it is unclear what Nvidia means by full production, and Creative Strategists analyst Austin Lyons told Wired: “This CES announcement around Rubin is to tell investors, ‘We’re on track.’” On driving, Nvidia said the all-new Mercedes-Benz CLA, built on its DRIVE platform, will be the first passenger car featuring Alpamayo, and that AI-defined driving arrives in the U.S. this year. THE DECODER added that the Alpamayo models are not meant for direct in-car deployment but serve as “teacher models” that developers fine-tune and distill.
Google and Character.AI settlement (reported on January 8). Google and Character.AI agreed to a “settlement in principle” in multiple lawsuits filed by families whose children died by suicide or suffered psychological harm allegedly linked to chatbots on Character.AI, according to court filings cited by Fortune. Details were not disclosed, and no admission of liability appears in the filings. The claims included negligence, wrongful death, deceptive trade practices and product liability. Google re-hired Character.AI founders Noam Shazeer and Daniel De Freitas in August 2024 as part of a $2.7 billion deal that also licensed some of the startup’s technology. In October 2025, Character.AI said it would bar users under 18 from “open-ended” chats with its AI personas.
What to Watch
- Rubin in partners’ hands. THE DECODER says partner availability starts in the second half of 2026.
- ChatGPT Health beyond early users. OpenAI says access will expand in the coming weeks. I think the medical-record connections through b.well are the part most worth scrutinising as access grows.
- Snowflake–Observe closing. The deal still needs regulatory approval; watch whether the final price matches the reported figure.
- Meta’s power timeline. Prometheus is expected online in 2026, and Oklo says pre-construction and site characterization for its Pike County campus are slated to begin in 2026.
This article was written by AI from the sources listed on this page. Figures, names and quotes were checked against those sources by software before publishing.