OpenAI buys Odessia, US token imports, X-energy nuclear permit, n8n Nod agent, Google MKA benchmark
Hello, dear TEA-mates! Here is what you need to know today.
1. 🧳 OpenAI Acquires AI Travel Startup Odessia
Francis Davidson, who previously founded Sonder, announced on X that OpenAI has acquired Odessia and that he is joining OpenAI to focus on the future of travel. He founded Odessia less than a year ago to build an AI travel agent, and over $60M a year of travel was planned through it this past summer by tens of thousands of early adopters. Davidson contrasts this with Sonder, which took a decade and more than a thousand people to reach $600M in revenue. He says his own agent usage rose 25x through 2026, that agents had MCP or CLI access to meeting recordings, Slack, email and Drive, that voice prompts ran 3 to 4x faster than typing, and that removing usage limits turned a 40-hour task into 2 hours. (Read More)
🫖 TEA For Thought: “Less than a year old, and already $60 million a year of travel through this company. That is amazing.”
2. 🔌 When Does the US Become a Net Importer of Tokens?
Tomasz Tunguz of Theory Ventures argues that with AI demand outpacing supply and domestic data center construction slowing, the US will soon consume more AI tokens than it produces, making it a net importer of tokens, as it is of rare earth metals. His basic model puts the switch between 2028 and 2030: if 45% of planned projects fail to connect, imports start by 2028; with 20% cancellations, 2029; and even if every facility is built on time, by 2030. Current systems can absorb roughly 2.5x annual token growth, with US data center demand growing from a 66GW base in 2027 and tokens per watt compounding about 2.9x a year. Short-term GPU rental prices rose from $5 to more than $13 per hour over two quarters while demand barely moved. Without domestic power and facilities, he writes, tokens will simply be produced elsewhere and imported. (Read More)
🫖 TEA For Thought: “Add to this that Google and all of big tech are now starting to build data centers in South Asia and in Brazil, where the CCP is pretty much in control. I hope the US does not make the same mistake again by outsourcing manufacturing, or now computing factories, abroad. This is literally national security we are talking about. Not your infra, not your intelligence.”
3. ⚛️ X-energy Gets Green Light to Build Its First Small Modular Nuclear Reactors
The Nuclear Regulatory Commission approved X-energy’s construction permit application for four Xe-100 reactors at a Dow chemical plant in Texas, each able to generate 80 megawatts of electricity and 200 megawatts of heat. X-energy says it is the first approval for an advanced nuclear reactor at an industrial site in the US. No small modular reactor operates in the US yet, and the Dow reactors are not expected online until the early 2030s. The design flows helium gas over uranium-containing fuel pellets; Edwin Lyman of the Union of Concerned Scientists said the NRC pushed “consideration of critical safety questions to a later date.” Amazon’s Climate Pledge Fund led X-energy’s $500 million Series C-1, and Amazon has committed to buy 5 gigawatts of capacity. X-energy’s April IPO raised $1 billion, but its shares trade at less than half the offering price. (Read More)
🫖 TEA For Thought: “This is huge. In the future, energy might be the currency: how much energy you are able to harness pretty much determines how much intelligence you own.”
4. 🤖 Meet Nod: n8n Labs Releases an Open-Source Personal Agent for Your Computer
n8n Labs, a new team inside n8n, released Nod as a research preview: a standalone personal agent for Mac and Windows, open source under Apache 2.0, that works in a local sandbox with its own browser. It acts in connected apps such as Gmail, Calendar, Drive and Slack, and on websites with no integration, needs no n8n account, and uses the AI provider the user picks through an API key or a ChatGPT plan. Example jobs include cancelling a subscription, watching a price and filing supplier invoices. n8n says Nod scored 93.2% on the LongMemEval long-term memory benchmark. A separate check confirms the user asked before Nod sends, pays, posts or deletes anything, and it checks in once a task passes $5. For now it stops when a laptop sleeps and cannot make phone calls. (Read More)
🫖 TEA For Thought: “It’s a great thing that n8n open sourced it. Think about how much n8n was used in the general AI community, and now it has all been replaced by AI agents. What a fast-paced life we are living!”
5. 📊 Google Introduces the Multimodal Knowledge Agent Benchmark
Google DeepMind and Artificial Analysis launched MKA (Multimodal Knowledge Agent), a benchmark of how AI agents reason across visually complex files. Domain experts across 10 verticals and 83 occupations built its private 200-task set; each task takes a seasoned professional three to eight hours and carries a rubric of about 70 yes/no checks, weighted 75% on factual accuracy and 25% on visual design. Every model runs in the same open-source agent harness, Stirrup, and an autorater built on a non-Google model grades the work; domain experts agreed with it 96% of the time. Tested models used 30 to 130 turns on average per task. Google says Gemini 4 Carbon ranked #2 in quality at 6x lower average cost than the top model. Ten public tasks are on Hugging Face. (Read More)
🫖 TEA For Thought: “I wonder whether this harness, Stirrup, is biased or not.”
🛠️ Skill of the Day
The Bug Report Builder: Turn a messy “it’s broken” into a clear report anyone, or any agent, can act on.
You are a patient support engineer. Your goal is to help me turn a vague problem into a clear, complete bug report that a colleague or an AI agent could fix without asking me follow-up questions.
The problem in my own words: [DESCRIBE WHAT WENT WRONG]
Where it happens: [THE APP, WEBSITE, DEVICE OR CODE]
What I already tried: [ANYTHING YOU DID TO FIX IT]
Any error messages or screenshots described: [PASTE HERE]
First, ask me up to five short questions to fill the gaps that matter most, then wait for my answers. After that, write the report with these parts: a one-line summary, steps to reproduce, what I expected, what actually happened, how often it happens, what changed recently, and your top two or three likely causes ranked by probability, each with one quick check I can run. Use plain language, keep it under 250 words, and mark anything you are guessing as a guess.Paste it into ChatGPT, Claude, or your tool of choice. Replace the bracketed bits with your own.
TEAHEE Moment
Stay sharp, stay informed. See you tomorrow.
If you enjoyed this TEA, follow along on social for more:






