US Government Fights China — and Its Own People

Anthropic Secures Watts & Wafers

Google is pumping up to $40 billion into Anthropic after Amazon pledged $25 billion last week. The deal includes $10 billion immediately at a $350 billion valuation; another $30 billion will follow upon reaching certain milestones. Starting in 2027, Google will provide the AI startup with five gigawatts of computing power – enough electricity for all households in Minnesota. Anthropic's annual revenue run rate exploded from $9 billion to over $30 billion, driven by the success of its coding tool, Claude Code. The circular deals between tech giants and AI startups are fueling the US economic boom: Google, Amazon, Microsoft, and Nvidia are investing billions in OpenAI and Anthropic, who then spend that money on chips and computing power from the very same investors. → The New York Times

Synthszr Take: It reads like an investment round, but it's a supplier negotiation. The $65 billion from Google and Amazon aren't bets on Anthropic — they are down payments for compute that doesn't even exist yet. Five gigawatts from 2027: that's a power plant construction project with a penalty clause. What's happening here is familiar from early industrialization: if you needed coal, you didn't buy coal, you bought shares in the mine. Anthropic values its stock at $350 billion and trades it for the only things that are truly scarce in this phase — watts and wafers. Money is abundant. GPUs are not. And electricity certainly isn't. What remains is the most elegant trade of the AI era: equity for electrons.

Deepseek & Co: The White House Accuses China of Warfare

The White House is accusing Chinese AI companies of copying American models through coordinated 'distillation' campaigns. In a memo, special advisor Kratsios describes how firms like DeepSeek, Moonshot, and MiniMax use thousands of fake accounts to extract OpenAI and Anthropic models. The attackers engage in industrial-scale jailbreaking: they bombard the APIs with requests until they can reconstruct proprietary model architectures. Washington now wants to develop countermeasures, from information sharing with US companies to 'best practices' against model theft. The memo does not mention specific sanctions. The Chinese embassy counters by pointing to China's transformation from the 'world's factory' to an 'innovation lab.' Last year, DeepSeek claimed to have developed its model for just a few million dollars – a fraction of what US companies spend. → MIT Technology Review

Synthszr Take: DeepSeek is playing the marginal cost game to its bitter end: when inference becomes a commodity, the one who can copy cheapest wins. The architectural innovations that keep DeepSeek competitive despite the GPU embargo are reminiscent of the Soviet space program: resource scarcity forces more elegant solutions. The fact that the US State Department is now officially warning of 'IP theft' while DeepSeek is open-sourcing its models highlights the absurdity of the narrative. The real threat isn't espionage, but China demonstrating that modern AI development works like distilling whiskey, not enriching uranium. Anyone trying to protect model weights like nuclear secrets hasn't understood the internet.

US Government Fires Its Own AI Safety Chief After 96 Hours

The U.S. Commerce Department dismissed AI researcher Collin Burns after just four days as head of the Center for AI Standards and Innovation. Burns, previously with Anthropic, was pushed to resign by the White House on Thursday, according to four people familiar with the situation. The concerns: his connection to Anthropic, which had engaged in heated disputes with the Trump administration in recent months over the military use of its systems. Trump had publicly attacked the company as a 'RADICAL LEFT, WOKE COMPANY.' Chris Fall, a scientist with extensive government experience, will take Burns' place. The center, established under Biden as the AI Safety Institute and renamed under Trump, is supposed to assess the security risks of new models like Anthropic's Mythos system. → Washington Post

Synthszr Take: The Burns affair shows how artificial intelligence is becoming an ideological predetermined breaking point. Washington faces a classic principal-agent problem: the expertise for AI regulation lies with the very companies that are supposed to be regulated. It's reminiscent of the Atomic Energy Commission's situation in the 1950s, when only physicists from the Manhattan Project understood the technology. The difference: back then it was about national security; today, commercial interests are mixed with political symbolism. Anthropic's 'Mythos' system may have dangerous hacking capabilities, but the real danger is the politicization of technical standards. When AI safety becomes a loyalty test, it ultimately loses.

Amazon Provokes Google & Nvidia with CPU Deal for Meta

Amazon has won Meta as a customer for millions of its homegrown Graviton CPUs – a targeted blow against GPU dominance in the AI world. The ARM-based processors are set to power Meta's growing AI workloads, particularly the compute-intensive tasks of AI agents like real-time reasoning, code generation, and multi-step coordination. The deal brings Meta's spending back to AWS after the social media giant had signed a $10 billion, six-year contract with Google Cloud in August. Amazon timed the announcement perfectly for the end of the Google Cloud Next conference – a calculated provocation. In parallel, Anthropic has secured Amazon's Trainium GPUs with a $100 billion deal over ten years, while Amazon has increased its investment in Anthropic to $13 billion. CEO Andy Jassy had already announced in his shareholder letter his intention to compete with Nvidia and Intel by offering better price-performance ratios. → Techpresso

Synthszr Take: Amazon is playing chess here while everyone else is still playing checkers. Positioning Graviton CPUs for AI agents is like inventing the diesel engine alongside the gasoline one: less glamorous, but more efficient for certain workloads. Meta doesn't need Formula 1 engines (GPUs) for city traffic (inference); it needs robust truck engines. The simultaneous commitment from Anthropic to Trainium GPUs reveals Amazon's dual strategy: training remains GPU territory, but the real value creation happens in inference with millions of parallel agent tasks. Jassy's price-performance argument is reminiscent of Southwest Airlines versus legacy carriers: the winner isn't the one with the best service, but the one who optimizes compute power per dollar. Amazon is turning AI compute into a commodity – and profiting from both ends.

Anthropic Admits Intelligence Loss in Claude

Anthropic had to admit that Claude Code delivered poorer results for weeks. Three independent errors – a reduced reasoning setting from 'high' to 'medium' (March 4), a caching bug that cleared the context after each interaction (March 26), and an overly restrictive prompt instruction (April 16) – added up to a noticeable loss in quality. The errors only affected Claude Code, not the API itself. Anthropic has fixed all three problems with version 2.1.116 and promises stricter quality controls: broader evaluation suites, gradual rollouts, and more transparency via the new X account @ClaudeDevs. As compensation, the usage limits for all subscribers have been reset. → Techpresso

Synthszr Take: Claude Code is like a Formula 1 car where someone secretly changed the tires, throttled the engine, and altered the aerodynamics. What users perceive as 'the model is getting dumber' turns out to be a cascade of infrastructure decisions: Anthropic wanted to save latency (reasoning from 'high' to 'medium'), optimize memory (caching bug), and curb verbosity (word limit). Each individual change was rationally justified, but together they created a Frankenstein effect. The real drama is that the GPU shortage (98.95% availability instead of the industry standard 99.99%) forces such trade-offs. The AI industry is no longer optimizing for the best results, but for maximum utilization of scarce resources – a bit like airlines installing ever-tighter seats until the passengers revolt.

'Computer Use' is Coming to ChatGPT Pro

OpenAI has announced Computer Use for ChatGPT Pro: the model can now take screenshots, click on them, and operate programs. Anthropic had introduced the feature in October as a 'Public Beta,' but only via the API. OpenAI is integrating it directly into the chat interface for its $200 subscribers. The model sees the screen, understands the user interface, and performs actions – from Excel macros to booking flights. The latency is 22 seconds between screenshot and action, which is still too slow for productive use. Microsoft Teams and the Windows desktop remain blocked for security reasons. → Marcus Schuler

Synthszr Take: Computer Use is the first real step toward dissolving the boundary between thought and action in AI systems. The model becomes an intern who looks over your shoulder and then takes over the mouse itself. The 22-second latency makes it clear: this is still crawling, not walking. But the direction is right. When AI models can interact directly with software, APIs and interfaces will become relics (why build an API when the AI can just click?). OpenAI is democratizing a technology that Anthropic invented – a classic fast-follower play. The real test will come when the first users unleash their Computer Use agents on enterprise applications.

Cloudflare Turns AI Agents into Email Recipients

Cloudflare is launching an email service in public beta that gives every AI agent its own inbox. In addition to a simple mailbox, the service enables asynchronous work, a persistent state across multiple replies, and uses a communication channel that everyone already knows how to use. This is a clever alternative to the standard assumption that agents need special chat interfaces. Additionally, Cloudflare provides an email MCP server, CLI tools, and an open-source 'Agentic Inbox' that can be deployed with a single click. The news comes at a time when the industry is searching for practical interfaces between humans and AI systems. → Unwind AI

Synthszr Take: Cloudflare is turning email into the universal API for AI agents. This is reminiscent of the early days of the internet when email was the first truly interoperable protocol: SMTP worked everywhere, while proprietary systems remained trapped in their silos. The move is brilliantly simple: instead of forcing users to learn new interfaces, it taps into 50 years of email muscle memory. Asynchronicity and persistence are thrown in for free, both features that chat UIs have to painstakingly replicate. While everyone is betting on spectacular new interfaces, the most boring technology in the world might just win. Email is becoming the Trojan horse for agent adoption.

Mistral Introduces Voice Mode for Terminal Coding: Speak Instead of Type

Mistral has given its terminal-native coding agent, Vibe, a voice mode update. Developers can now interact with their codebase using voice instead of typing commands. The tool uses Mistral's own models and can also read out the output via text-to-speech upon request. The functionality allows for giving natural language instructions directly in the terminal to explore, modify, and understand code. Mistral is positioning Vibe as an alternative to GUI-based coding assistants, betting on the familiar terminal environment of many developers. → Unwind AI

Synthszr Take: Mistral is turning the terminal into a speech lab, hitting an interesting sweet spot between old-school developer culture and modern AI interaction. The decision not to package voice interaction into a fancy GUI but to integrate it directly into the terminal is reminiscent of early Unix philosophies: do one thing well. While Microsoft and GitHub are bloating the IDE with Copilot, Mistral is taking the opposite approach, turning the minimalist interface into a feature. This could particularly resonate with the growing number of developers who are increasingly centralizing their workflows in the terminal (tmux, vim, neovim). The real innovation isn't the speech recognition itself, but the bet that developers would rather stay in their familiar environment than switch tools for AI features.