Surprise Elon Anthropic Team Up Reshapes the AI Race
Anthropic's developer day 'Code with Claude' focused on agentic features including memory management ('dreaming'), outcome-based quality review, and multi-agent orchestration. The event was overshadowed by a surprise partnership between Anthropic and SpaceX/Elon Musk, giving Anthropic access to Colossus 1's 220,000 NVIDIA GPUs. The host argues this represents Elon's strategic pivot from model builder to compute infrastructure provider.
Summary
The episode covers two major stories: Anthropic's developer day announcements and a surprise Anthropic-SpaceX compute partnership.
Anthropix's 'Code with Claude' developer day featured no major model releases, instead focusing entirely on agent infrastructure. The centerpiece announcements were around 'managed agents' — Anthropic's harness-as-a-service platform. Key new features included 'dreaming,' a scheduled memory review process that runs between agent sessions to extract patterns, curate memories, and improve agent performance over time. This was compared favorably to similar functionality in open-source tools like Hermes, with commentators noting the open-source ecosystem has led on agent primitives for nearly a year. A second feature called 'outcomes' allows users to define success rubrics, with a separate grading agent scoring task outputs and kicking back subpar results for revision — Anthropic reported 8.4% and 10.1% quality improvements for Word and PowerPoint documents respectively. Multi-agent orchestration was also formalized in the platform, allowing lead agents to delegate to specialist subagents working in parallel. Anthropic also launched Claude Finance, a suite of 10 predefined financial services agents, and previewed future model capabilities including 'infinite' context windows, higher code judgment, and better multi-agent coordination. Dario Amadei revealed Anthropic saw 80x annualized growth in Q1, far exceeding their planned 10x annual growth.
The dominant story, however, was the Anthropic-SpaceX partnership announced mid-day. Elon Musk tweeted that after spending time with senior Anthropic staff and finding them competent and safety-focused, he agreed to lease Colossus 1 — XAI's Memphis data center housing 220,000 NVIDIA GPUs at 300 megawatt capacity — to Anthropic. Immediate user-facing changes included doubling Claude Code's 5-hour rate limit, eliminating peak-hour usage reductions for Pro/Max users, and increasing Opus API rate limits by 2-10x depending on tier. Elon also announced XAI would be dissolved as a separate entity into SpaceX AI.
The host frames this as 'Elon's AI play 3.0' — moving from OpenAI funder (1.0) to model builder (2.0) to compute infrastructure provider (3.0). The argument is that Elon's proven strengths lie in executing known-but-hard physical infrastructure challenges, not in achieving unknown research breakthroughs required for frontier model development. XAI had stalled on models, lost co-founders, and had no meaningful agentic product, but had excess compute. Anthropic had strong models and harnesses but was severely compute-constrained. The host speculates Elon's long-term vision involves orbital data centers, and that SpaceX as a vertically integrated AI compute company positions him as a Jensen Huang-style infrastructure kingmaker rather than a Sam Altman-style model competitor. The partnership was seen as near-inevitable given the Sam Altman feud made OpenAI an impossible partner for Elon.
Key Insights
- The host argues that the key AI competition in 2026 has shifted from model benchmarks (Opus vs GPT) to agentic harness ecosystems (Claude Code vs Codex), with Anthropic's Dev Day reflecting this by having no model releases and focusing entirely on agent infrastructure.
- The host claims the open-source agent ecosystem (Hermes, G-Brain) has led the closed labs on agent primitives like persistent memory and self-improving orchestration by nearly a year, with Anthropic's 'dreaming' feature essentially productizing what open-source builders already shipped.
- Dario Amadei revealed Anthropic experienced 80x annualized growth in Q1 2025, dramatically exceeding their internal planning assumption of 10x annual growth, with compute scarcity being the primary constraint on further scaling.
- The host argues Elon Musk's pivot to compute infrastructure (SpaceX AI) represents 'AI play 3.0,' contending that his proven talent for executing known-but-hard physical builds makes him better suited to be a Jensen Huang-style infrastructure kingmaker than a frontier model competitor.
- The host speculates that the XAI-SpaceX merger was never really about giving SpaceX a model in Grok, but rather about giving SpaceX a terrestrial compute footprint to build upon toward Elon's vision of orbital data centers.
- Anthropic's 'outcomes' feature applies external grading agents to subjective knowledge-work outputs using user-defined rubrics — the host notes this is a meaningful extension beyond the coding domain where automated grading (unit tests) was already well-established.
- Boris Cherny, Claude Code's creator, stated there is literally no manually written code anywhere at Anthropic anymore, with Claudes coordinating over Slack and resolving issues autonomously, and argued the term 'vibe coding' dramatically undersells the rigor of modern agentic engineering workflows.
- The host contends that Anthropic was the only viable compute partner for Elon given the depth of his feud with Sam Altman, making the partnership almost structurally inevitable once Elon reframed his AI strategy around infrastructure rather than model competition.
Topics
Full transcript available for MurmurCast members
Sign Up to Access