welcome to my thought repository @Mercor_ai | prev restructuring @Wharton alum
new york
Joined August 2025
- Tweets236
- Following1.5K
- Followers188
- Likes479
for all the fair criticism that x lost its town square feel because of elon's shitposting, the platform deserves credit for cutting down the bots from a few months back
3 ways enterprise ai plays out imo
1. models get good enough to justify the spend (muse, grok bot, starting in small biz where context needs are low)
2. big enterprises posttrain on proprietary data and keep hill climbing
3. model agnostic players with rich data access beat 1/2 (i expect consolidation)
Ishan retweeted
I like Blackstone and JG but I don't think this is quite right. While Gray's description of general conditions in spot market for compute is correct, that isn't super relevant to Coreweave since a vanishingly small % of Coreweave's compute can harvest these lucrative spot market rates. 98% of $CRWV's compute is locked up in contracts whose average life is 5 years.
The irony is that if Coreweave was more "speculative" and wanted to build compute before having customers first, they would be gushing cash now. But because they were "responsible" and secured offtake partners that gave Blackstone and other lenders the confidence to lend, they are locked in at "low" compute rates.
Blackstone's Jon Gray says neoclouds that can deliver compute get 50% more for it than a year ago, and old H100s lease near their original rates
[ CoreWeave, a company that you work with... is seeing its yields on some of its debt go to 15%. ]
"Well, it is creating a challenge. And it's interesting when you talk about CoreWeave and the neo clouds, there's a real contrast between delivering compute and demand for compute."
"And so if you're in a neo cloud and you can deliver compute, the price you can get today is up 50% from where it was a year ago. Your older chips, the H100s, are leasing today close to their original rates, where people thought three or four years ago they wouldn't have any value."
if a company earns returns above its CoC, competitors pile in and squeeze them back to equilibrium
ai adoption runs the same play. early movers earn the spread, everyone else catches up, and the savings get passed on as lower prices. unless you own data nobody else can buy
something consuming large amounts of your bandwidth for an extended enough period of time becomes your identity
among other things new york city is walkable because of the quick reward cycle of making up each ave and street
Ishan retweeted
Most evals are bullsh*t.
A few companies have asked me for help with evals. The pattern is consistent: teams grade whether an agent sounds good, not whether it does the job.
I wrote up what I’ve learned. If your team needs help, I’d love to chat.
feeling grateful is the art of imagining yourself in a version of your existence where you could have been be worse off - and then comparing that version
to reality
