The Lyceum: AI Daily — May 07, 2026

The compute layer just ate the model layer. Anthropic is renting its rival's entire supercomputer, xAI is dissolving itself into a rocket company, and DeepSeek is taking state money for the first time — all in 24 hours. The companies that build the chatbots no longer control where the chatbots run, and the rules of independence are being rewritten in real time.

• Claude Code limit changes (Anthropic): Five-hour rate limits doubled for Pro, Max, Team, and seat-based Enterprise; peak-hour throttling removed for Pro and Max; Opus API rate limits substantially raised — effective immediately.
• Gemma 4 MTP drafters (Google): Multi-token prediction drafter checkpoints across the Gemma 4 family, claiming up to 3× faster decoding with no quality loss; day-zero support in vLLM, Ollama, MLX, and SGLang.
• Managed Agents primitives (Anthropic): A research-preview "Dreaming" feature that reviews agents' work between sessions and updates context files, plus broader availability for delegating subtasks to specialist agents.
• Agents SDK for TypeScript (OpenAI): Sandbox agents and an open-source evaluation harness for developers building agent workflows.
• Finance Search API (Perplexity): Licensed financial data and live market feeds added to the Agent API, with 35 dedicated workflows for analyst-style tasks.

The strangest business deal in AI right now isn't a funding round. It's Anthropic — whose CEO Dario Amodei has been the loudest voice in Washington warning about uncontrolled AI — signing an agreement to use all of the compute capacity at SpaceX's Colossus 1 data center in Memphis. That's more than 300 megawatts and over 220,000 NVIDIA GPUs coming online within the month, per Anthropic's own announcement.

The product effect is immediate and concrete. Claude Code's five-hour rate limits are doubled for Pro, Max, Team, and Enterprise subscribers. Peak-hour throttling is gone for Pro and Max. Opus API rate limits are substantially higher. Anthropic's own support documentation has been updated to reflect the new ceilings. According to PCWorld, Anthropic had recently been considering dropping Claude Code from its $20/month Pro plan entirely — an executive admitted the consumer plans weren't built for token-hungry agentic tools. The Colossus deal appears to function as an emergency pressure valve.

The politics are notable: the dissolution and leasing move follow a period of public dispute between Musk and Anthropic. The Colossus rental lands amid that backdrop.

What changes if this works: capacity becomes a publicly disclosed product feature, not a procurement detail, and every other lab faces pressure to publish similar capacity commitments. What failure looks like: Anthropic's usage numbers don't move materially in the next 30 days, suggesting the bottleneck was never compute. Watch the developer event metrics this week.

Musk announced Wednesday that xAI will be dissolved as a standalone company and folded into SpaceX, rebranded as SpaceXAI. "xAI will be dissolved as a separate company, so it will just be SpaceXAI, the AI products from SpaceX," he wrote on X.

Republic World reports that by April, 11 of the 12 original xAI co-founders had exited, leaving Musk as the only remaining founding member. The company Musk assembled from elite researchers at DeepMind, Google, and OpenAI has been hollowed out. The dissolution lands the same day Colossus 1 — the supercomputer xAI built to train frontier models — gets leased to Anthropic.

CNBC reports the new SpaceXAI roadmap centers on space-based data centers, with SpaceX recently filing FCC paperwork indicating intent to launch up to a million satellites for orbital compute. The timing matters: SpaceX filed confidentially with the SEC on April 1 for an IPO targeting a $1.75–$2 trillion valuation, with the public S-1 expected by late May. Adding Anthropic as a named compute customer right before the listing strengthens the pitch that SpaceX is no longer just launches and Starlink.

The signal to watch: whether Grok keeps shipping competitive model updates over the next two quarters, or whether SpaceXAI quietly pivots from frontier research to infrastructure landlord. The Colossus rental is consistent with either path.

For the past year, DeepSeek has been the most unusual company in AI: a hedge-fund-backed lab that published everything, took no outside money, and somehow kept pace with labs spending ten times more. According to the Financial Times, citing four people with knowledge of the discussions, the China Integrated Circuit Industry Investment Fund — China's biggest state-backed semiconductor vehicle, known as the "Big Fund" — is in talks to lead DeepSeek's first fundraising round at a roughly $45 billion valuation. Bloomberg reports Tencent and Alibaba are also in talks to participate.

TechCrunch notes that founder Liang Wenfeng controls nearly 90% of DeepSeek and had previously refused outside capital. The reason for the turn, per FT sources: competitors are poaching DeepSeek's researchers, and Liang needs equity to retain them.

The Next Web frames the Big Fund's involvement as a strategic statement: if China cannot acquire Nvidia's leading-edge GPUs at the volume required, it will finance the model labs that have demonstrated they can produce frontier results without them.

What changes if this closes: DeepSeek becomes a de facto national champion, and the open-weight transparency that made it globally influential gets squeezed by state-aligned shareholders. The observable signal: whether the next DeepSeek model release ships with the same technical report, the same weights, and the same permissive license as V3 and R1. If any of those slip, the lab everyone learned from has quietly become something else.

Buried in NBC News's coverage of Anthropic's developer day was a feature most outlets skipped. Anthropic introduced a research preview called "Dreaming" for its agent management software. Per NBC's reporting, the feature reviews agents' work between sessions, surfaces patterns, and updates files that store user preferences and context. Anthropic also announced wider availability for a feature that lets Claude break down tasks and delegate them to specialist agents.

This is a different category from rate-limit increases. An agent that reviews its own prior work and updates its own context files between sessions is taking a step toward persistent, self-modifying behavior. Combined with multi-agent delegation, the architecture starts to look less like a chatbot and more like a long-running autonomous system with memory.

The signal worth watching: whether Anthropic publishes safety documentation and red-team evaluations for Dreaming before it ships broadly, or after. That sequencing will tell you more about how the lab is actually balancing capability and caution than any policy paper.

Google released multi-token prediction drafter checkpoints for the Gemma 4 family — gemma-4-31B-it-assistant, 26B-A4B, E4B, and E2B — claiming up to 3× faster decoding with no quality degradation. The E2B drafter is just 78M parameters. Day-zero support landed in Transformers, vLLM, MLX, SGLang, Ollama, and AI Edge.

Speculative decoding — where a small draft model proposes tokens that a larger target model verifies in parallel — has been a research technique for two years. What's new is the open ecosystem getting a coordinated drop with quality-preserving speedups built in. llama.cpp shipped beta multi-token prediction support the same week, with one tester reporting roughly 75% steady-state acceptance and >2× throughput on Qwen3.6 27B.

What changes if this lands: local agents become genuinely practical on consumer GPUs, and the latency gap between hosted frontier models and self-hosted open ones narrows materially. The observable signal: whether the next wave of agent frameworks ships Gemma 4 drafter integration as a default rather than an opt-in.

Datacenter Dynamics reports Cerebras signed a 10-year lease for 40MW of capacity at a Digi Power X data center in Alabama, with an initial contract value around $1.1B and potential to reach $2.5B, staged through 2026–27. A separate Cerebras SEC filing discloses an agreement with OpenAI to deploy hundreds of megawatts of Cerebras capacity under multi-year commercial terms.

The boring contracts are the interesting story. Decade-long leases and binding term sheets are how labs lock in the physics — power, cooling, transmission rights — that none of them actually own. Anthropic stacking SpaceX on top of an existing Amazon expansion (up to 5 gigawatts) is the same pattern. The frontier labs are quietly becoming utility customers.

What changes if these commitments hold: the AI capex cycle stops looking like cloud and starts looking like industrial real estate. What failure looks like: any one of these multi-year contracts gets restructured or written down, signaling demand softer than the leases assumed.

Electrek reports that Musk admitted under oath in his ongoing lawsuit against OpenAI that xAI distilled OpenAI's models to train Grok — a practice that would violate OpenAI's terms of service. The contradictions are stacking up in real time: Musk dissolves xAI, rents its supercomputer to Anthropic, publicly praises Anthropic's safety culture after months of attacks, and has been reported to have admitted to model distillation in court. The legal exposure on the distillation question alone could reshape what AI companies are allowed to do with each other's outputs.

What to watch: whether OpenAI files a counterclaim or amends its complaint to address the distillation admission. If they do, the trial stops being about Musk's grievances and starts setting precedent on derivative training data.

A man dissolved his AI company into a rocket company on the same day his rocket company rented its supercomputer to the AI company he keeps suing, while admitting in court that he copies their homework. The new vibe in AI is "frenemies with benefits, payable in megawatts." Stay caffeinated.

If someone you know still thinks the model layer is where the action is, forward this — gently.