Welcome back! This is where money, metal, and memory all lead. Huge checks are being written for servers that most people will never see, a rocket company wants to take cheap computing off the planet, data warehouses are trying to become the homes of your agents, and a quiet model reads screenshots like plans and sends back code that works. The main idea is simple: intelligence is becoming infrastructure, and the real choices now are where it runs, who pays for it, and which middlemen you let be there.
In today’s Generative AI Newsletter:
Oracle raises $50B for AI buildout.
SpaceX folds xAI into space compute plan.
OpenAI puts GPT-5.2 inside Snowflake data.
Kimi K2.5 turns screenshots into working code.
Latest Developments
Oracle’s $50B AI Powerhouse for Meta and OpenAI Models
Oracle told investors it wants to raise around $50 billion in 2026 to build more cloud and AI capacity, and the stock dipped as people priced bigger revenue potential but with a bigger bill. Following Oracle's recent AI-driven hype cycle, the company is now assigning a cost to the buildout and seeking participation from shareholders and bond buyers to fund it. Failure to acquire sufficient chips and data centers promptly could result in Oracle missing out on contracted demand and ceding momentum to larger competitors.
Here are the details:
Split: Roughly half debt, half equity or equity-like, including investment-grade bonds plus “convertible preferred” and an at-the-market program.
Bond: Oracle already raised $25B in bonds across 3 to 40 years, with demand peaking at a $127B order book.
Concerns: Oracle's long-term debt is around $100B, and CDS costs have gone up, which is a strong sign from the credit markets.
Clients: The customer list includes OpenAI, Meta, NVIDIA, AMD, TikTok, and xAI, which can mean dependence.
Oracle is merely attempting to finance an AI takeover without looking like it has lost focus. One key advantage is Oracle's ability to secure high-margin cloud growth with guaranteed demand and potential early access as its customer base expands. On the downside, Oracle faces risks of dilution, mounting debt, and increased interest burdens. Amid Big Tech's push to market AI capabilities as an inevitable progression, Oracle has strategically embraced the AI boom, aligning itself with industry trends.
Special highlight from our network
Labs in Front is a new, weekly under-the-hood series from AI21 for builders shipping agentic systems. This time we look into:
1️⃣ How an orchestrated test-time compute shifts the cost vs accuracy curve for coding agents on SWE-bench-verified.
2️⃣ A workspace layer for MCP that supports safe parallel attempts for state-modifying agents.
Read more here.
Special highlight from our network
Steve Nouri maps what’s next.
GenAI Academy presents Fresh Start: 2026 AI Predictions, Applied, a free, live three-day course for people who want to make better choices early in the year.
These are live sessions focused on turning understanding into action.
You’ll learn where to focus in AI and how to move forward with confidence.
Free · Live
Feb 17–19 · 9-10 AM PT
SpaceX Integrates with xAI as IPO Talks Reach $50B
SpaceX has put xAI inside its orbit, turning a deal between Elon Musk's companies into one financial sheet. The move ties rockets, satellites, a chatbot and a social platform into a single strategy of owning the pipes, data and compute narratives. Reports estimate the combined value at $1.25 trillion, with plans to move the cheapest AI compute off the Earth in 2 to 3 years. With plans to move the cheapest AI compute off the Earth in the near future, the possibilities for advancements in space technology are seemingly endless.
Here’s what the evidence shows:
Deal: SpaceX acquired xAI as IPO talk swirls, while teams stay separate for now.
Plan: Musk sells space-based AI data centers powered by solar and moved by Starship.
Scale: Starship targets one flight per hour and 200-tonne payloads, plus a million satellites over time.
Risk: Grok’s content scandals and SpaceX defense work (Starshield) raise the access question.
This could make AI cheaper in the same way Amazon made computing feel infinite by controlling the infrastructure and selling the excess. With more affordable launches and universal connectivity, the compute challenge may shift from a land-and-power issue to a logistics problem. SpaceX's Starlink project aims to provide global internet coverage through a network of satellites in space, potentially disrupting traditional telecom companies. If regulations treat SpaceX as a hybrid of telecom, defense contractor, and media platform, it could set a tone for how other space computer companies are regulated in the future.
OpenAI Brings GPT-5.2 Directly to Your Company’s Private Data
OpenAI and Snowflake have announced they signed a $200M, multi-year deal to put OpenAI models directly inside Snowflake’s data platform. It explains that companies store sensitive data in Snowflake so they can now run GPT-5.2 and build AI agents without sending data out to yet another tool. OpenAI frames it as enterprise-ready AI delivered inside the warehouse, but the risk is that it can potentially raise concerns about data privacy and security, as the integration of AI models directly into a data platform may pose new vulnerabilities.
Here's what the deal involves:
Access: OpenAI models show up natively via Snowflake Cortex AI and Snowflake Intelligence, aimed at business users and builders.
Workflow: Teams can call models from SQL using Cortex AI Functions across structured and unstructured data.
Reliability: Snowflake wraps the pitch in a 99.99% uptime SLA, so agents do not die when dashboards go dark.
Power: Snowflake turns into the control room for agents that can read company data and take actions.
Data platforms want to become the default home for models, agents, spending and governance. The upside is fewer exports, tools, ways to leak data and tighter governance with a shorter path from data to action. The disadvantage is that one vendor starts setting the rules, the pricing, the weaker accountability and the limits of what safe means. Snowflake has already placed a $200 million Anthropic bet, therefore it is certainly purchasing a model mall. Customers will enjoy the convenience till renewal season begins.
Kimi K2.5: Vision to Code From Screenshots and PDFs
Kimi K2.5 is a multimodal model that can read images, text and then turn them into usable outputs like UI code, cleaned tables, and refactored components. It’s handy when your input is visual or messy and you want something you can paste into your project without a long setup.
Core functions (and how to use them):
Screenshot: Upload a screenshot of any page. Ask for a responsive React + Tailwind rebuild that matches spacing, typography, and components.
Corrections: Share a screenshot of the broken layout plus the expected result. Ask for a small diff patch that changes only the CSS or component causing the issue.
Conversion: Upload tabled PDF report. Request a CSV export, column cleanup rules, and totals, rates, and pivot formulas.
Refactor: Ask it to break down, remove duplicate styles, and simplify state without changing behavior.
Execution: Say “extract data, clean it, summarize it, format it.” Request the final document and intermediate files for reuse.
Try this yourself:
Grab a screenshot of a pricing table or settings page. Upload it and prompt: “Recreate this as a responsive React component in Tailwind. Match spacing and font sizes. Use Card, Button, and Input components. Return the full code and a short list of assumptions.” Then open the page in your browser and reply with one concrete correction, like “increase left padding by 8px and make buttons 44px tall," so it updates only what you flagged.
