Playbook
Tasteful Skills
“Tasteful Skills” argues that the best agent skills are not documentation or best-practice lists.
Back to Podcast Digest
The Anthropic Situation is INSANE
TL;DR
Anthropic got an instant compute bailout from SpaceX — Matthew Berman says the deal unlocks over 300 megawatts and 220,000 Nvidia GPUs at Colossus 1, letting Anthropic immediately double Claude Code 5-hour limits and raise Opus API rate limits.
Daario Amodei’s conservative GPU strategy backfired — Berman contrasts Anthropic’s cautious capex approach with OpenAI’s aggressive “acquire every GPU” bet and argues the market proved OpenAI right because AI demand exploded faster than expected.
Elon’s public praise of Anthropic clashes hard with his own recent attacks — after months of calling Anthropic “hypocritical,” “misanthropic,” and anti-Western, Musk suddenly said no one there triggered his “evil detector,” which Berman reads as pure business necessity.
The biggest winner may be infrastructure, not model labs — Berman keeps returning to Nvidia, Google, and raw compute supply, arguing that demand for intelligence is so high that chips, and maybe ultimately energy, are becoming the real choke points.
Cursor’s SpaceX deal now looks more complicated — because Anthropic is taking all of Colossus 1 while xAI has shifted training to Colossus 2, Berman suspects Cursor’s model work is now sharing remaining capacity and says he’d be nervous if he were Cursor.
Anthropic’s user trust problem isn’t fixed by more GPUs alone — even while welcoming the quota boost, Berman complains Anthropic still operates as a black box on subscriptions, quotas, and third-party tools like OpenClaw after months of frustrating developer-facing changes.
The Breakdown
Elon Musk “saved” Anthropic — and that’s the weird part
Berman opens with the contradiction: Elon Musk just threw compute to Anthropic, a direct competitor he’s spent months trashing in public. That’s what sends him live — not just the deal itself, but the sheer cognitive dissonance of Musk rescuing a company he clearly does not like.
Anthropic’s conservative GPU bet finally snapped
He frames the whole situation as a strategy mismatch between Daario Amodei and OpenAI. Anthropic chose not to overextend on capex because if demand didn’t rise perfectly, it could threaten the company; OpenAI did the opposite and vacuumed up as many GPUs and as much funding as possible — and Berman says the market rewarded that aggression because AI demand massively overshot expectations.
The quota pain was real, and suddenly it’s easing
Berman spends a chunk of time on why users were so annoyed with Anthropic: reduced quotas, “after-hours” incentives, peak-time penalties, and almost no transparency about what subscribers were actually buying. Now, with the SpaceX deal live, Anthropic says it’s doubling Claude Code’s 5-hour limits for Pro/Max/Team, removing peak-hour reductions, and massively raising Opus API throughput — though Berman still calls the system a black box.
Colossus 1 is going all-in on Claude
The most eye-popping detail for him is Anthropic’s claim it will use all of Colossus 1. He reads that literally: over 300 megawatts and more than 220,000 Nvidia GPUs in Memphis, already online, meaning xAI had idle capacity it urgently needed to monetize because unused GPUs bleed money every second.
Anthropic is suddenly buying compute from everyone
From there, Berman zooms out to Anthropic’s broader compute scramble: up to 5 gigawatts with Amazon, 5 gigawatts with Google and Broadcom starting in 2027, a Microsoft/Nvidia-linked Azure capacity deal, and a $50 billion infrastructure push with Fluidstack. His takeaway is blunt: the winners here are the chip and cloud layers — especially Nvidia and Google — because everybody upstream is starving for silicon.
Maybe chips matter less than we think — or maybe energy matters most
He riffs on past interviews with Greg Brockman and Google Cloud CEO Thomas Kurian to question whether model labs are really locked to any single hardware stack. If Anthropic or OpenAI can move between AWS Trainium, Google TPUs, and Nvidia GPUs without much pain, then maybe chips eventually commoditize — and the real irreducible constraint becomes electricity itself.
Elon’s tone flip is almost comical
Berman then lines up Musk’s old anti-Anthropic posts — calling the company hypocritical, anti-Western, smug, and “misanthropic” — against Musk’s fresh praise that nobody at Anthropic triggered his “evil detector.” He treats the new rhetoric as obvious throat-clearing: Musk had excess capacity, Anthropic had desperate demand, and he had to write the nicest possible version of a deal he probably hated making.
xAI, OpenAI rivalry, and the Cursor subplot
Berman’s read is that Musk would rather help Anthropic than Sam Altman, especially while suing OpenAI, so “the enemy of my enemy is my friend” likely shaped the deal. He also connects this to Cursor’s recent SpaceX partnership: if Anthropic owns Colossus 1 and xAI moved to Colossus 2, then Cursor is now sharing what’s left — which is why Berman says he’d be at least a little nervous, even if the rumored structure includes a $10 billion breakup fee.
Share
Keep Reading
Two Ways to Follow
The Weekly Echo. The inbox-shaped summary of what mattered.
New editorials announced here.
