▶ YouTube Summaries

• YT :: https://www.youtube.com/watch?v=FHGfSjDGUIw
• Original title :: pstack creator spends $132,000,000/year on ai????

A viral post claimed Lauren (poteto, creator of pstack and the React compiler, now at Cursor/SpaceX AI) burns $132M a year on tokens, based on 1.5 trillion tokens in 30 days. Theo takes the math apart: an $8/M blended rate ignores that most tokens are cheap cache reads, so the real cost is orders of magnitude lower. He then argues with David Cramer (Zeeg, Sentry founder), who claimed intelligence is not really getting cheaper in practice. Theo's data says cost at a fixed level of intelligence now drops close to 100x per year. The meta-lesson: learn to spot people who post confident, emotionally appealing but numerically wrong takes about AI, and verify the numbers yourself.

The $132M claim

• Premise: 1.5T tokens/30 days at an assumed $8 per million = $384k/day, $11M/month, $132M/year. Even with a 90% discount it "buys 26 staff engineers".
• Cursor profile counts are an aggregate of four token types (cache read, cache write, input, output), each priced very differently.
• Theo uses almost only Opus 5.5 ($5 in / $25 out). His actual blended cost is $0.41 per million, 20x below the $8 assumption. Cache reads are $0.20/M; plain input and output are under 8% of his spend.
• Lauren likely uses cheaper models (Grok), runs long agentic sessions with heavy cache hits, and Cursor almost certainly has a large Anthropic discount (SpaceX AI rents compute to Anthropic, Cursor was a huge Anthropic customer).
• Theo's estimate: roughly $0.10-0.20 per million blended, i.e. $15k-50k a month; worst case at list prices about $1-3M a year, not $132M.
• Even at the inflated number, Theo argues 26 staff engineers with zero token budget would be less productive than Lauren alone with unlimited tokens.

Why the spend is worth it anyway

• Companies spend over $1M on events that get less attention than a couple of Lauren's tweets; $25k/month is cheap marketing.
• Her tokens also ship product: Cursor got noticeably faster and less buggy after she joined, and she lands thousands of PRs a month on Grokbot.
• Value her spend as product, engineering, research, marketing and FTE work combined, not dollars-in vs software-out.
• Nobody is claiming this is normal. She is exploring what happens when cost of intelligence is ignored, because today's $10M of tokens will cost about $10k in a year or two.

Cost per intelligence is collapsing

• Artificial Analysis "cost per task at fixed intelligence": Sonnet 4.5 (Sep 2025, first to break 20) cost $0.56/task; a year later 6 Luna low matches it at $0.0045, about 100x cheaper.
• Score 30: Sonnet 4.6 Max at $2.50/task, then 6 Luna high at $0.03 within roughly 8 months.
• The rate itself is accelerating: it used to be ~10x per year, now ~100x in 6-8 months.
• Raw token prices: DeepSeek V4 Flash vs GPT-5 is 21x cheaper on input and 55x cheaper on output; GLM-5.3 Flash is 29x cheaper than GPT-5 in bench runs and nicer to work with.
• Theo's own Skate Bench: o1 (Dec 2024) scored 67.7% at $0.12 per question; Gemini 3 Flash (Dec 2025) scored higher for 58x less; GLM-5.3 Flash scored 70.8% at 680x less, in under two years.

The argument with Zeeg

• Zeeg's claims: lab people don't understand paying for tokens; old models get retired so there is no real price drop; 10x gains are only benchmarks; a task that cost 10 cents a year ago still costs about that.
• Theo's rebuttals:
• Complaining that intelligence isn't cheaper while always buying the newest model is like complaining iPhones aren't cheaper while refusing last year's model.
• Benchmarks are unreliable now because cheap models saturate them, which is evidence of progress, not of benchmark failure.
• A task "costs the same" only because the agent now does work the human used to do: reproducing the bug, editing, running the dev server, checking the browser. The ticket costs the same, but the human did nothing.
• GPT-4o in OpenCode couldn't even call tools. Three years ago models could not build a to-do app at all.
• Theo offered to turn Zeeg's real tasks from a year ago into a benchmark to prove the 10x; Zeeg replied with more essays.
• Theo admits his own past mistakes: assuming an intelligence ceiling and a price floor, mocking Windsurf for letting models run commands, and doubting Anthropic's RL.

Lessons

• Before sharing alarming numbers, paste the post into Claude or Codex and ask it to check the math.
• Distrust people who attack motives ("you're paid by the labs", "it's marketing", "they're normalising it") instead of engaging with the points.
• Anti-AI takes go viral because they confirm what people want to believe. If you dislike OpenAI and Anthropic, falling inference prices are the argument that their margins will vanish.
• At the current trend, Lauren's usage could cost about $3k a month within months, and about $100 a month in a year, if she kept using equivalent rather than frontier models.