Playbook
Tasteful Skills
“Tasteful Skills” argues that the best agent skills are not documentation or best-practice lists.
Back to Podcast Digest
Anthropic just…wait what
TL;DR
Anthropic’s real problem wasn’t greed — it was a brutal compute shortfall — Theo says Dario admitted Anthropic planned for 10x annual growth and instead saw an annualized 80x surge in Q1, which explains the Claude Code limits, peak-hour throttling, and push to free up GPUs.
The SpaceX-Anthropic deal is shocking precisely because the two sides have history — despite Elon repeatedly mocking Anthropic as “misanthropic,” SpaceX is now leasing Colossus 1 capacity to Anthropic, a sign that Anthropic’s NVIDIA shortage got bad enough to override old beef.
Claude users get real relief, but not everyone benefits equally — Anthropic removed peak-hour reductions and doubled Claude Code 5-hour limits, but Theo notes heavy daily users who hit weekly caps may feel little improvement, while API customers see much bigger wins like Tier 3 input limits jumping from 800k to 5M tokens per minute.
Theo’s big XAI theory: Cursor was really about data, not just product — he argues XAI already had compute but lacked the chat-history data needed to train coding models, and that Cursor’s logs of real developer-agent back-and-forth may be valuable enough to justify the reported $10B collaboration fee or eventual $60B acquisition.
The most revealing number in the whole story is how much idle compute XAI seems to have had — according to figures Theo cites, Anthropic gets access to 300+ megawatts and 220,000 NVIDIA GPUs, which is most of Colossus 1 and implies Grok inference was using only a small fraction of the cluster.
OpenAI looks scary here because it’s the only player Theo thinks has all three essentials at once — his framework is research, data, and compute: Anthropic has research and data but not enough compute, XAI had compute but not enough data, Cursor has data but not compute, while OpenAI is the only one currently checking every box.
The Breakdown
The weirdest partnership in AI this week
Theo opens by victory-lapping a prior take: Anthropic’s painful pricing and limits were about capacity, not pricing power. The headline news is the wild part — Anthropic partnering with SpaceX/XAI, despite public hostility and Elon’s long-running “misanthropic” jokes about the company. That this deal happened at all, he says, tells you how severe the compute crunch really is.
Dario basically admits they got demand catastrophically wrong
At Anthropic’s Code with Claude event, Dario says they planned for 10x yearly growth and instead saw an annualized 80x jump in revenue and usage in Q1. Theo’s read is simple: nobody plans cleanly for that, but OpenAI was more aggressive in buying compute early while Anthropic stayed conservative and got caught short. That shortage, not some scheme to upsell everyone, drove the recent Claude Code restrictions.
What this actually changes for Claude users
Theo explains Anthropic’s subscription economics like airline seat management for GPUs: idle servers lose money, but if demand spikes past capacity, you’re dead. That’s why Claude had peak-hour reductions and weird usage behavior, and why those limits are now loosening as more compute comes online. The catch is important: doubling 5-hour Claude Code limits helps bursty users, but people running hard every day and hitting weekly caps may barely notice.
The API increases are the real enterprise story
He’s especially excited about API rate-limit bumps, because old limits made serious app usage painful. Tier 1 was comically low at 30,000 input and 8,000 output tokens per minute; Tier 3 jumps from 800k to 5M input, and Tier 4 from 2M to 10M. For anyone hosting products on Anthropic, Theo frames this as a much bigger deal than the consumer-plan changes.
Anthropic’s hardware stack is messy, and researchers probably still want CUDA
Theo walks through Anthropic’s fragmented infra: Amazon Trainium, Google/Broadcom TPUs, and Microsoft/NVIDIA Azure capacity. His hypothesis is that researchers still overwhelmingly want CUDA, so Anthropic is pushing as much inference as possible onto non-NVIDIA hardware to free NVIDIA clusters for training. That makes SpaceX’s giant pile of NVIDIA GPUs uniquely attractive.
A side theory: Azure may be proxying Anthropic rather than truly hosting it
Looking at throughput charts, Theo notices Amazon and Google perform materially better, while Azure and Anthropic’s direct API look almost suspiciously identical. He stops short of claiming certainty, but says the mirrored performance makes him suspect Azure may be piping requests through Anthropic rather than running a meaningfully separate deployment. He’s careful to say Amazon and Google clearly are not, because the performance profiles are too different.
Why Theo thinks XAI wanted Cursor: the data exhaust from coding chats
This is the heart of his speculation. Frontier success needs three things — research, data, and compute — and Theo argues XAI had compute but lacked the right kind of data, while Twitter/X data is noisy and badly suited for coding or reasoning. Cursor, by contrast, has exactly the goldmine labs want: millions of real developer interactions where users correct models step by step, producing training signals that can be shifted into better RL-style examples.
The numbers on Colossus make Grok look underused
Theo ends with the most concrete and funniest implication of the deal: Anthropic reportedly gets 300+ megawatts and over 220,000 NVIDIA GPUs, which is most of Colossus 1. If that’s true, then XAI only needed roughly 100 megawatts for Grok inference, suggesting the giant cluster was massively underutilized because hardly anyone uses Grok at that scale. His closing framework is the whole video in one line: XAI is plugging a data gap, Anthropic is plugging a compute gap, and both are making strange alliances because OpenAI is the one company they’re all scared of.
Share
Keep Reading
Two Ways to Follow
The Weekly Echo. The inbox-shaped summary of what mattered.
New editorials announced here.
