Drama-Rama with Elon Musk & Dario Amodei: The Enemy of My Enemy Is My Friend

Elon Musk & Dario Amodei: The Enemy of My Enemy Is My Friend

Anthropic has struck a deal with SpaceX, giving the AI company access to the entire computing capacity of the Memphis data center. CEO Dario Amodei announced at the Code with Claude conference that the 300 megawatts of additional computing power will directly translate into higher usage limits for Pro and Max subscribers. The Opus API limits have been raised, and Claude Code now has double five-hour windows without peak-hour throttling. SpaceX highlights the 220,000 NVIDIA GPUs of its Colossus-1 supercomputer, while Anthropic has already signaled interest in “several gigawatts” of orbital computing capacity. Elon Musk, who just in February called Anthropic an “enemy of Western civilization,” now seems impressed after personal meetings: “Nobody set off my bad-guy detector.” → arstechnica.com

Synthszr Take: Memphis is becoming the new Panama Canal of the AI industry. Anthropic is using SpaceX's data center like a lock between scarce GPU supply and exploding developer demand. The real innovation isn't the 220,000 GPUs, but the business model: infrastructure deals are replacing model sales. Musk's sudden turnaround (from “enemy of civilization” to “impressed”) shows how computing power is becoming the hardest currency of the AI era. Anthropic's interest in orbital data centers sounds like science fiction, but it follows the same logic as the shift of factories to Asia in the 1980s: where terrestrial resources (power, cooling, land) become scarce, industry expands into new spaces. Whoever controls the infrastructure dictates the terms of the AI future. And the old rule applies in the parallel Musk vs. Altman lawsuit: the enemy of my enemy is my friend.

Anthropic Is Becoming the Wall Street API

Anthropic is rolling out ten pre-configured AI agents for the financial industry, designed to automate routine tasks at investment banks, asset managers, and insurance companies. The templates cover research, risk and compliance checks, and financial accounting. These include a “Pitch Builder” for target company lists and pitchbooks, an “Earnings Reviewer” for annual reports, and a “KYC Screener” for compliance escalations. The agents run as plugins in Claude Cowork and Claude Code directly at the workplace or as “Claude Managed Agents” autonomously on Anthropic's platform—for instance, for multi-hour deal closings with a complete audit log. New data connectors to Moody's, Dun & Bradstreet, and SS&C IntraLinks expand the partner ecosystem. Goldman Sachs, Citadel, and AIG are already among its clients. In parallel, Anthropic announced a $1.5 billion joint venture with Blackstone, Hellman & Friedman, and Goldman Sachs, while OpenAI is making a similar move with “The Deployment Company”. → Techpresso

Synthszr Take: Anthropic is no longer selling intelligence, but access to the nervous system of the financial world. The ten agent templates are Trojan horses: they look like harmless productivity tools, but they establish Anthropic as an indispensable infrastructure layer between Wall Street and its data. Once you run KYC checks, deal closings, and compliance workflows through Claude, there's no going back—the switching costs explode with every integrated system. The joint venture with Blackstone is the real coup: $1.5 billion not for technology, but for distribution through the portfolio companies of the largest private equity firms. OpenAI and Anthropic are no longer building better models; they are buying their way into the capillaries of the global economy.

The Anthropics Are Becoming the New Accentures

Silicon Valley is currently undergoing a strategic pivot: instead of continuing to sell models, the major AI labs are now building consulting firms. Anthropic is forming a joint venture with Blackstone, Hellman & Friedman, and Goldman Sachs, receiving $1.5 billion. OpenAI is following suit with “The Deployment Company,” backed by 19 investors like TPG and Bain Capital, who have already invested $4 billion at a $10 billion valuation. The new formula: a small team works closely with the client, understands their processes, and develops tailored AI systems. Brad Lightcap, previously COO at OpenAI, will now lead these “Special Projects”. Finance is already Anthropic's second-largest revenue segment, a fact highlighted by a financial services event in New York with high-profile guests. → AINews

Synthszr Take: The AI industry is currently repeating the development of management consulting in the 1960s. McKinsey emerged when companies realized they didn't just need computers, but also someone to show them how to rethink business processes with them. Anthropic and OpenAI are now building the Accenture of the AI era: implementation armies that mediate between the abstract intelligence of the models and the entrenched IT systems of the Fortune 500. The key insight here: the models themselves are becoming a commodity, while the ability to integrate them into existing workflows becomes the real moat. Private equity is getting in because they know the pattern: high margins on recurring revenue, scalable through junior consultants (or in this case, specialized agents). The irony: the smarter the models get, the more human labor it takes to make that intelligence usable.

Claude Now Needs to Dream More

At its Code with Claude developer conference, Anthropic introduced a new feature called “Dreaming” for Claude Managed Agents. The process analyzes past events and identifies information worth saving for future tasks. Managed Agents are a pre-configured infrastructure that allows multiple agents to work on complex projects for hours at a time. The “Dreaming” takes place as a scheduled process where sessions and memory are reviewed and important memories are curated. Unlike the previous method of compacting individual conversations, Dreaming can recognize patterns across multiple agents: recurring errors, converging workflows, and shared preferences. The feature is in research preview and requires access authorization. Additionally, Anthropic is doubling the 5-hour usage limits for Pro and Max subscribers after the compute infrastructure struggled to keep up with demand. → Ars Technica

Synthszr Take: Anthropic solves the memory problem of AI systems like a brain in REM sleep: consolidation instead of accumulation. Biology shows us how: while we sleep, the brain sorts through the day's events, strengthens important connections, and lets unimportant ones fade. Claude's “Dreaming” follows this exact pattern, except here, multi-agent systems are distilling their collective experiences. This is reminiscent of ant colonies that reinforce or let pheromone trails fade depending on the success of the route. The crucial difference from previous compaction methods: this creates emergent knowledge across agent boundaries, like a company whose departments suddenly understand what the others are actually doing. Anthropic is no longer building better chatbots, but organizational nervous systems.

Google Copies OpenClaw

Google is developing a personal AI agent called “Remy” that can handle tasks for users around the clock. This was reported by Business Insider, citing internal documents and two people familiar with the matter. Remy runs in an internal version of the Gemini app and can interact with other Google services. The description reads: “Remy is your 24/7 personal agent for work, school, and life, powered by Gemini.” Google employees are already testing the system internally. A Google spokesperson declined to comment. The product is strongly reminiscent of OpenClaw, the viral AI agent whose creator, Peter Steinberger, was hired by OpenAI in February. → Business Insider

Synthszr Take: Google is reacting like a city planner who, after the success of a pop-up restaurant, quickly sets up an official franchise. OpenClaw was the phenomenon of a single individual showing what's possible when you just let AI agents do their thing. Now Google comes along with Remy, integrated into its in-house infrastructure, secured by compliance processes. This is reminiscent of the story of Napster and iTunes: the pirate shows the way, the corporation makes it respectable. But while Apple revolutionized the music industry back then, Google might be too late to the party here. OpenAI not only hired Steinberger but also seems to have a better understanding that AI agents need to function like pets: personal, willful, sometimes unpredictable. Google's “deeply integrated across Google” sounds like an attempt to turn a wolf into a German Shepherd.

Amazon: Why the E-commerce Giant Will Dominate AI Infrastructure

Amazon has largely ignored the spectacular GPU purchases for model training, instead investing for years in its own chips and cloud infrastructure. While OpenAI, Anthropic, and others spend billions on Nvidia hardware to train ever-larger models, Amazon is positioning itself for the next phase: the inference era, where trillions of AI requests will need to be processed. Ben Thompson shows in his analysis how Amazon's new “Supply Chain Services” repeat the same pattern that worked for AWS: first build it for yourself, then sell it as a service. Its own Graviton chips may not match the performance of Nvidia GPUs, but they are sufficient for inference workloads and are significantly cheaper. Amazon thinks in decades, not in quarters. → Ben Thompson

Synthszr Take: Amazon is playing chess while others are playing checkers. The company has understood that AI training is like the Gold Rush: spectacular, but temporary. The real profits are made by those who sell the shovels or, in this case, those who control the inference infrastructure. Graviton chips are like Toyotas in a world of Ferraris: less impressive, but when you need millions of them, efficiency matters more than peak performance. The integration of logistics services shows that Amazon sees physical and digital infrastructure as two sides of the same coin. Whoever controls the ports, warehouses, and data centers sets the rules of the AI economy.

Shadow AI: To Stay Productive, Employees Bypass Internal IT

Corporate America is burning through billions of dollars trying to develop AI tools that office workers actually want to use. Most of these efforts will fail because the gap between what a well-funded IT department can offer and what an inspired intern can “vibe” together over a weekend is closing, or has already closed. Shadow IT emerged 15 years ago when employees used their iPhones and cloud applications for everyday tasks, side-stepping the clunky internal networks their employers had spent a fortune on. Today's version is even more potent: the benefit-to-risk ratio has shifted so far in favor of benefit that companies have already lost control. Employees aren't thinking about data governance, security, or the 100 other things that can go wrong when freelancing in software development. The only way out for companies is to abandon walled gardens and build gateways through which a sea of vibe-coded apps can safely dock with corporate data. → Semafor Technology

Synthszr Take: Enterprise AI isn't failing because of the technology, but because of the organizational structure of companies themselves. The phenomenon is reminiscent of Prohibition: the more you forbid something, the more creative the workarounds become. If an intern with ChatGPT and Claude can generate more productivity in one night than the in-house IT department can in a quarter, the problem isn't the intern. The “gateways instead of walled gardens” solution sounds elegant but overlooks the real dynamic: companies must encourage their employees to cheat in order to stay innovative. The most productive employee is the one who creatively interprets the rules, not the one who follows them. Enterprise AI will only work if it legitimizes the institutionalized cheating that successful knowledge workers are already engaged in.

China: GLM-5V-Turbo Model Excels in Agentic Workflows

The GLM-V team from China presents GLM-5V-Turbo, a multimodal AI model specifically optimized for computer agents. With over 50 researchers involved, the model demonstrates performance on par with Western competitors on benchmarks like ScreenSpot and Agent-Bench. The architecture combines visual and textual processing for complex tasks such as web navigation, GUI interaction, and document analysis. Particularly interesting: the model was explicitly trained for low latency in agent tasks, not primarily for general language capabilities. The release coincides with the hype around Western agent startups like Cognition Labs. → Techpresso

Synthszr Take: While Silicon Valley philosophizes about Computer Use, China is already building the next generation. GLM-5V-Turbo isn't a research project, it's industrial product development: 50+ researchers, specialized architecture, clear optimization for agent performance instead of benchmark chasing. This is reminiscent of China's approach to electric cars: first observe, then dominate a specific market with massive resources. The Western AI industry systematically underestimates how quickly Chinese teams move from paper to product. Anyone who thinks Anthropic's Computer Use is unrivaled doesn't understand that the agent market has long been a two-front war.

Meta: Agentic Assistant to Become Part of WhatsApp & Co.

According to the Financial Times, Meta is planning to introduce an advanced “agentic” AI assistant that can independently perform tasks for users. The assistant is intended not only to respond to requests but also to act proactively—for example, making restaurant reservations or planning trips. The timing seems defensive: OpenAI, Google, and Anthropic have already announced or are developing similar systems. Meta is leveraging its strength: over a billion users on WhatsApp, Instagram, and Facebook who could use the assistant without installing an additional app. Integration into existing messaging services would transform Meta's AI strategy from a tech toy into an everyday tool. → Techpresso

Synthszr Take: Meta is making a virtue of necessity—like a supermarket that's late to switch to delivery services but suddenly realizes it already has stores on every corner. While OpenAI and Google compete for the smartest AI, Meta controls the digital living rooms of three billion people. An agentic assistant in WhatsApp is like a butler who already has the house key. The question isn't whether Meta's AI will be the best, but whether it will be the most invisible—embedded in daily communication without users ever having to open a new app. Meta is betting that infrastructure is more important than innovation.

Meta Trained AI with 150,000 Stolen Books – Now Publishers Are Suing Back

Five major publishers and author Scott Turow are suing Meta for “one of the most massive copyright infringements in history.” The lawsuit accuses Meta of systematically copying books and scientific articles from pirate sites like LibGen and Sci-Hub to train its Llama models. If you feed Llama two sentences from a textbook, it spits out the continuation verbatim—clear proof of training on copyrighted works. Internal Meta documents show that the company discussed how to handle “media reports about the use of datasets that we know are pirated.” The publishers are demanding damages and a complete list of all works used for training. → The Verge

Synthszr Take: The AI industry is currently going through its Napster phase, except this time the pirates are sitting in glass towers. Meta argues fair use while simultaneously hiding its own content behind ever-higher API walls—a classic case of “preaching water and drinking wine.” The business model resembles medieval enclosures: first, privatize the commons, then charge access fees. The Anthropic $1.5 billion settlement shows where this is headed: AI companies will become involuntary collecting societies, forced to retroactively pay for their training corpora. The real irony is that publishers have been fobbing off authors with meager royalties for decades and are now suddenly playing the copyright cavalry—not out of principle, but because they've finally found an opponent with deep pockets.