Arun

Some solid Q&A on seat vs usage vs outcome pricing for AI. Clearly, no perfect answers but it def made me fine-tune my POV. Views: > marginal cost of s/w = 0; marginal cost of AI is not. That alone implies you HAVE to find SOME way to correlate price to usage/COGS. > "but AI labs all do tiered pricing" - actually all AI-lab rev is token metered which is why every tier has rate limits, caps, model gating, etc. Labs are actually the strongest case _for_ token/usage pricing. > variable cost / task, but customers want predictability; agree - but true only for tasks that have median & mode usage within striking distance; in AI, power users can use 1000x median and seat pricing on variable COGS is a disaster. Needs to be _very_ carefully planned. > I recall "runaway" usage as a huge issue when we launch bigQuery at GCP; knee-jerk reaction was to limit use; what made BQ a $10B biz was still usage pricing but paired with account budgets, threshold alerts, dry-run estimates. That's a better approach (but harder for long-horizon / unbounded tasks I'd admit) > "why not do outcome based pricing" - imho, this _is_ usage/token based pricing, just metered at a higher order unit. It works best, again, when median vs mode are similar. e.g. say sierra charges 2$ / resolution, and resolution takes 2min (cost = 0.2$) or 2hrs ($12), the lose on the 2hr call (v low vol) and make money on the 2min call (v high vol). Unbounded agentic tasks need to be metered on usage. My best guess is that the real uncapped promise & value in AI is on agentic tasks. There is this idea of agentic frontier where if you are more human-assist / co-pilot, you'd end up preferring seat-based but over time as you start solving agentic problems, which is the true value of AI lies, you'd end up above the frontier and shift to usage based. The more unbounded & agentic the problem, the higher the preference for usage pricing. Harvey may be sitting below the frontier today, but will eventually move above.

Thesis: the problem with AI working in every domain = all the edge cases. Antithesis: domains with lots of edge cases = difficult & time consuming to practically impossible for error-prone people. Synthesis: such domains = where AI agents will do best. (Such as SAAS migration…)

if you thought saas-pocalypse was bad just wait for computer use to get really good later this year the implications for incumbents are 100x more than coding agents because computer use asymmetrically benefits “hostile” integrators & expect a race to commoditize complements

I am converging on the fact that unless you are an athlete in training you can in fact, meet all your macro needs with simple straightforward simple meals. Here is a simple mean that cost between 60-80 to make at home and honestly does provide a pretty good ROI without needing to be overtly fancy. I can tune that protein up by adding some more curd on the side. I do need to add that ghee because fat is delicious. I have started measuring food on the basis of now much willpower it costs to eat. This one costs zero because this is genuinely nourishing and has the flavor I grew up with. I am not trying to optimize calories per gram of protein because any day I can get 60-70 gm of protein in a day staying under 1800-2000 calories, that day is a win. the long term answer to eating well is always to be thoughtful about what you do in the kitchen. Eat you dal chawal anon.

Enjoyed this piece on Stripe's Minions: an internal built end-to-end background coding agent “Over 1,300 Stripe pull requests merged each week are completely minion-produced, human-reviewed, but containing no human-written code”

yeah so this is insane 24 year old turned $225M into $5.5B in <12 months. dug into his recent investments and… holy fuck - MASSIVE $885M position in Bloom Energy (specialises in portable energy turbines for… you guessed it - ai data centers) - this 1 position is 20% of the entire fund lol - massive SHORT position on Infosys. he’s betting claude code, codex are going to replace outsourced IT work (he’s right) - added $300M to his corweave position totalling $700M (someones gotta run those gpus) - aggressively pivoted investments into electrical and energy infra (aka AI’s biggest constraint right now) - dumped $100Ms of NVIDIA and Intel positions. - aggressively buying bitcoin mining companies and re-purposing them for ai data centers. (cipher, bitdeer) - fund up $1.5B in the last 3 months (+35% last quarter) - now owns 10% of core scientific (levered bet on coreweave 😂) He outperformed the S&P500 8X in the funds first 6 months. fucking goated.

Snowflake highlighting that companies are deleting software vendors and rebuilding workflows on top of their data in Snowflake Not sure they actually win this opportunity long-term, but interesting nonetheless