The AI cost curve just snapped. A Chinese lab trained a frontier model for less than most Series A rounds, and Wall Street lost its mind. Meanwhile, the reasoning race went global and Google rebuilt Search from the inside out.
In this edition:
A $6 million model wipes $600 billion off Nvidia
Half a trillion dollars in AI infrastructure gets a White House announcement
Google puts an AI engine inside Search
Three labs ship reasoning models in 90 days
OpenAI raises the largest private funding round in history
DeepSeek R1: The $6 Million Model That Shook Wall Street
A Chinese AI lab proved you don't need a billion-dollar budget to build a frontier model. On January 20, DeepSeek released R1, an open-source reasoning model that matched or beat OpenAI and Google on key benchmarks. The claimed training cost: approximately $6 million.
Need to know:
DeepSeek R1 uses a technique called inference-time computing, activating only relevant portions of the model for each query. The result is dramatically lower computational and energy costs compared to traditional approaches. Marc Andreessen called it "one of the most amazing and impressive breakthroughs" he'd witnessed.
On January 27, Nvidia shares fell 17% in a single day, erasing $600 billion in market value. It was the largest single-day loss for any company in stock market history. The Nasdaq dropped 3%. Energy stocks tumbled on fears that efficient AI could reduce demand for power infrastructure.
The model is fully open-source. Anyone can download and run it. Unlike Western frontier models locked behind API paywalls, R1 handed the blueprints to the world.
Nvidia's stock recovered within a month. CEO Jensen Huang argued the market overreacted, and the company's next earnings showed demand for its chips remained strong. But the question DeepSeek planted won't go away: is Silicon Valley overspending on AI development?
The $6 million figure is almost certainly incomplete. It doesn't include prior research, failed experiments, or the cost of the data. But even with generous adjustments, DeepSeek exposed a gap between what frontier AI costs and what the market assumed it costs. That gap repriced the entire sector in a single afternoon.
Signals
Stargate: Half a Trillion Dollars in AI Infrastructure
On January 21, President Trump stood in the White House alongside Sam Altman, Masayoshi Son, and Larry Ellison to announce the Stargate Project, a $500 billion AI infrastructure venture.
Need to know:
Stargate is a joint venture between OpenAI, SoftBank, Oracle, and Abu Dhabi's MGX. SoftBank holds financial responsibility. OpenAI holds operational responsibility. The initial commitment is $100 billion deployed immediately.
The venture will build data centres across the United States, with plans for more than 100,000 jobs. The scale has drawn comparisons to the Manhattan Project.
SoftBank's Son is chairman. The project represents the largest single infrastructure commitment to AI in history, dwarfing anything previously announced by Google, Microsoft, or Amazon.
The announcement landed one day after Trump signed an executive order revoking Biden's AI safety framework and replacing it with a pro-innovation directive. The message was unmistakable: the U.S. government is betting on acceleration, not caution.
Product
Google Puts an AI Engine Inside Search
In March, Google launched AI Mode, a new experimental feature that puts a full conversational AI directly inside Google Search.
Need to know:
Powered by Gemini 2.0, AI Mode lets users ask complex, multi-part questions and dig deeper with follow-ups, all without leaving Search. It goes well beyond AI Overviews, which summarise results. AI Mode reasons, compares, and explores.
The feature launched in Search Labs for Google One AI Premium subscribers in the US at $19.99 per month. Google's stated goal is to open it to everyone over time.
AI Mode pulls from Google's Knowledge Graph, real-time web content, and shopping data for billions of products. It combines the model's reasoning with Google's proprietary information systems.
This is Google's direct answer to ChatGPT Search and Perplexity. For anyone watching how businesses get found, the signal is clear: AI-mediated search is no longer an experiment. It's the main product roadmap for the world's dominant search engine.
The Roundup: The Reasoning Race
Claude 3.7 Sonnet Anthropic's hybrid reasoning model shipped February 25 with toggleable "extended thinking mode." Developers can set a thinking budget to control how long Claude reasons before responding. It set new highs on SWE-bench Verified for real-world software engineering tasks. Breakthrough
GPT-4.5 OpenAI's largest pre-trained model arrived February 27 with reduced hallucinations and improved pattern recognition. Altman called it "the first model that feels like talking to a thoughtful person." At $75 per million input tokens, it costs 30x more than GPT-4o. Product
Gemini 2.5 Pro Google's most intelligent model shipped in March with specialised agent capabilities for interacting directly with computer interfaces. It powers the new AI Mode in Search. Google's clearest statement yet that models and products are converging. Breakthrough
OpenAI Operator OpenAI's first AI agent launched January 23 for Pro users. It uses a Computer-Using Agent model to browse the web autonomously, clicking, typing, and scrolling to complete tasks like ordering groceries. Success rates: 38% on complex tasks, 87% on simpler ones. Product
Alibaba Wan 2.1 The Chinese tech giant open-sourced its video and image generation model in February, joining a growing wave of capable open models from Chinese labs that are shifting the competitive balance. Build
The Roundup: Power Moves
OpenAI closes $40 billion The largest private tech raise in history. SoftBank led with $30 billion, joined by Microsoft, Coatue, Altimeter, and Thrive. The round values OpenAI at $300 billion, but SoftBank can cut its commitment to $20 billion if the for-profit conversion isn't done by year-end. Money
Anthropic raises $3.5 billion Lightspeed, Fidelity, and Salesforce Ventures led the Series E at a $61.5 billion valuation. Anthropic has now raised more than $10 billion total, with Amazon as anchor investor. The safety-first lab now rivals Databricks in valuation. Money
Trump revokes Biden AI order On January 20, Trump signed an executive order revoking Biden's October 2023 AI framework. The replacement, "Removing Barriers to American Leadership in Artificial Intelligence," mandates a new action plan within 180 days. A clear shift from risk mitigation to innovation acceleration. Trust & Power
Apple delays Siri AI features Context-aware Siri, on-screen interactions, and smarter AI responses have all been pushed past 2025 with no new date. Apple's AI strategy, once framed as privacy-first and cautious, increasingly looks like privacy-first and late. Industry
AI captures 53% of all VC AI startups raised $59.6 billion globally in Q1 2025, more than half of all venture capital deployed. Strip out OpenAI's $40 billion and it's still $19.6 billion, more than any other sector by a wide margin. Money
Quick Bytes
AI Overviews now appear in over 50% of Google searches, up from 18% in early March, reaching 1.5 billion monthly users globally.
The UK released its AI Opportunities Action Plan on January 13, confirming a regulatory approach built around enabling safe development rather than restricting it.
Amazon launched Alexa+, its first major overhaul in a decade, integrating generative AI for personalisation and home automation.
Microsoft shipped Phi-4-mini and Phi-4-multimodal, compact models that process text, images, audio, and video on mobile devices with under 6 billion parameters.
Elon Musk's lawsuit challenging OpenAI's for-profit conversion gained traction as the California attorney general weighed in, adding legal pressure to a $300 billion restructuring.
Google Search ad spending grew 9% year-over-year in Q1, but click growth was only 4%. That 5% gap represents more dollars chasing fewer clicks as AI answers absorb intent.