Ayush Goyal

Compute Markets. ex @microsoft, ex @NimbusEnv (YC W22), @PabioHQ (YC S21)・music junkie・he/him

[Ayush Goyal @ayushgoyal]

@HotAisle Lol. I meant ops wise, ofcourse not down to the T. Companies like GitLab and Sourcegraph used to have some back in the day (maybe still do, haven't checked)!

[Ayush Goyal @ayushgoyal]

I was reading OpenAI's contract with CoreWeave. Section 10(c) lets OpenAI choose a replacement operator if it terminates following specified bankruptcy or insolvency proceedings. CoreWeave then has two business days after written notice to assign the relevant contracts, including the data centre agreement, under separate substitute-operator agreements.

What I'd want to understand is how the operational handoff works alongside that assignment. Who gives the incoming team access to monitoring, vendor accounts and the fabric configuration? Where are the runbooks? Who tells them which nodes have been causing problems?

If there's a training run in progress, those questions need answers before the outgoing team disappears. Signing over the agreements doesn't itself explain how the new operator takes control without interrupting the job.

For a financed cluster, I'd want to see the handover plan alongside the contract. Named people, access arrangements, and evidence that someone has rehearsed it. I'd be interested to hear from anyone who has changed operators on a live GPU cluster, especially what they discovered was missing once the handoff started.

[Ayush Goyal @ayushgoyal]

@HotAisle Great! If any cloud operator starts open-sourcing their handbooks (if documented), and lead with an example, it'd be awesome!

[Ayush Goyal @ayushgoyal]

em dash on the postgres installation screen 🙂

[Ayush Goyal @ayushgoyal]

Source: CoreWeave and OpenAI master services agreement, section 10(c), filed as Exhibit 10.24 to CoreWeave's S-1/A: https://t.co/yAoUQmeiKu

[Ayush Goyal @ayushgoyal]

@GaganSKhurana Incredible launch video and narration! Congratulations on the launch

[Felix Rieseberg @felixrieseberg]

I wanted to test Fable 5.2's quality with one-shot prompts without human feedback.

Everything in this game is from scratch: models, sounds, textures. ~$550 at API pricing, ~2.7M output and 445M input tokens, heavily cached. Took 8 hours!

Play it here: https://t.co/fazOCcyNIH

[pranav @prd_008]

I think about these two tweets all the time

[Ayush Goyal @ayushgoyal]

I've seen worse. Unwiped machines handed over when requesting baremetal from third party providers.

We'll see more of this as fragmented compute comes online in 2027.

[Hot Aisle @HotAisle]

fireworks security explosive diarrhea. who is the third-party cloud provider and why the fuck was billing data and customer weights / data available to developer machines on a third-party service?

[Ayush Goyal @ayushgoyal]

@gpupritham Let's see how the markets react

[Ayush Goyal @ayushgoyal]

CDG airport has PS5s at boarding gates 🥺

[Ayush Goyal @ayushgoyal]

@gpupritham One guy (who's popular here on X), was straight up very rude. I had to hang up the call after 15 minutes because I could not stand him.

[Ayush Goyal @ayushgoyal]

@deedydas My 2026's best purchase to date!

[Ayush Goyal @ayushgoyal]

@KarlMozurkewich @gregosuri @compute_plug DMing!

[Ayush Goyal @ayushgoyal]

@tejasybhakta Even racks have a lead time of 12-14 weeks!

Unless you know a guy who knows a guy who can FedEx you some racks in 2 weeks lol

[Ayush Goyal @ayushgoyal]

Human written email >> ai slop email

[Ayush Goyal @ayushgoyal]

@SenSanders Damn GPUs make more money than the minimum wage

[Ayush Goyal @ayushgoyal]

@catmandeuce We should talk

[Ayush Goyal @ayushgoyal]

It's a loan disguised as a GPU contract. There are ways to make it less risky for everyone. There's a huge middle market for that.

[MTS @MTSlive]

Crusoe CEO @ChaseLochmiller explains why signing a 5-year GPU compute contract with an AI startup is effectively making a loan to that company:

"We're underwriting your existence over the next few years."

"Asset-backed lenders in the GPU financing space initially were very focused on, let’s assume there’s zero terminal value of these GPUs. We're really just underwriting a contract, looking through to the credit of the offtake."

"Increasingly, people are looking at the underlying collateral as its own unique asset class. Jensen brought together all of these incredible asset allocators."

"Everybody that's normally competing on big deals is now having to work together because the size is so big."

"When we sign a Series A or Series B startup to a three- or four- or five-year commitment to buy compute, it’s sort of like we’re making a loan to them. We’re underwriting your existence over the next few years."
@CrusoeAI

[Ayush Goyal @ayushgoyal]

@AmusingVentures Absolutely!

[Ayush Goyal @ayushgoyal]

Korea just did its first big GPU financing, the largest syndicated loan in the country's history, and everyone is calling it a GPU deal. I read the filings and I'm not sure the GPUs are what the banks are really lending against.

Brookfield's vehicle buys the chips and rents the computing to Naver Cloud. Then Naver, the parent, guarantees all of Naver Cloud's payments through 2034. The filing puts that guarantee at about 18.7 trillion won, which is close to 65% of Naver's equity and roughly one and a half times the $9 billion Brookfield can spend on the chips and data centres.

So the collateral that matters is Naver's balance sheet. The chips are there, but if this goes wrong nobody is going to be auctioning six year old accelerators to get paid back. They're going to call Naver.

I don't think that's a bad structure. It's probably the only way to get banks this size comfortable with an asset they've never priced before. It just means "GPU-backed" is doing a lot of work in the headlines, and the next deal without a parent like Naver standing behind it is the one I'd actually watch to see how lenders price the hardware on its own.

[Ayush Goyal @ayushgoyal]

@harleyf Not very sure. Tax fears