Teahose.
SIGN IN
NEW HERE — WHAT TEAHOSE DOES
We read the entire AI & tech firehose — so you don't have to.
PODPodcastsAll-In, No Priors, Acquired…
NEWNewslettersStratechery, Newcomer…
PAPPapersPhysical AI research
PHProduct Huntdaily launches
VCInvestor ScoutSequoia, a16z, Benchmark…
CLAUDE DISTILLS →
7 reads, 30 sec each — free, 6 AM ET.
+ a live graph of the companies, people & themes underneath.
HOME/20VC/20VC: The State of Compute: What…
POD
// EPISODE
20VC

20VC: The State of Compute: What Everyone Needs to Know | Why Neo-Clouds are Great Businesses | What Everyone Is Missing About GPU Prices? | How AI Infrastructure Investors Will Win with Kush Bavaria, Co-Founder @ Ornn

DATE November 21, 2026SOURCE 20VCPARTICIPANTS HARRY STEBBINGS, KUSH BAVARIA
// KEY TAKEAWAYS6 ITEMS
  1. 01Infrastructure investors are landlords, and landlords collect rent whether the tenants win or lose
  2. 02Compute prices are rising and the forward curve has flipped from backwardated to upward-sloping
  3. 03Supply is badly constrained and lead times are lengthening
  4. 04Old GPUs are not depreciating on the schedule bears assumed
  5. 05Inference will fragment into specialized verticals rather than consolidate into a few horizontal winners
  6. 06Compute needs a standardized index and futures market, modeled on Brent, S&P and Moody's

1. Key Themes

Infrastructure investors are landlords, and landlords collect rent whether the tenants win or lose

Kush frames data centers as apartment buildings and GPU buyers (Fireworks, Modal, FAL, Baseten, OpenAI, Anthropic) as renters. The landlord's cash flow doesn't depend on any single tenant winning. The risk is only a collapse in aggregate demand.

Kush Bavaria: "data centers are just like apartment buildings where someone buys the apartment building and then the actual people that take the ownership of the GPUs are renting the GPUs from the apartment building" [00:09:22]. He concludes: "I think infrastructure investors actually like end up winning because they will always keep collecting rent as long as the tenants, there's tenants to live in the apartment building." [00:09:51]

Compute prices are rising and the forward curve has flipped from backwardated to upward-sloping

The early market consensus was that GPU hours would be cheaper in three to five years. Kush says that has reversed, and that the free renewal options once given away are now priced.

Kush Bavaria: "When we started the company, a lot of people thought like, oh, the forward curve for chip prices is actually like backward dated in the sense that like three years out from now or five years out from now, the price for GP hour is going to be less." [00:13:01] Now: "No one gives that for free because I think people in the market see that, oh, maybe like three years from now, the chip price are actually like higher than they are today. I would not be surprised if their Rubin costs are like $20, $30 per GPU hour." [00:13:30] Harry noted Gavin Baker's observation that prices have risen daily for roughly 60 days, and Kush confirmed: "literally chip prices have been increasing." [00:12:25]

Supply is badly constrained and lead times are lengthening

Buying capacity requires committing long before you know demand. Kush's lead-time data points are unusually specific.

Kush Bavaria: "if you want to go to CoreWeave and buy something, it's probably like you're probably quoted like Q4, 2027, Q1, 2028 at this point... the OEMs usually like lie to you... it's probably like 20, 22 weeks time period to actually get the chips. And then install takes like another like two to three weeks." [00:11:12] On Moderna's Eric, who buys 18 to 20 weeks ahead: "he should start, he's probably gonna start buying like 22 to like 26 weeks now, just lead times to increase for everything." [00:16:49] He argues you can't buy your way out: "You can spend as much money as you want. You can't make the fabrication line at TSMC move faster." [00:05:11 on the Anthropic/OpenAI question, 01:05:11]

Old GPUs are not depreciating on the schedule bears assumed

Last year's consensus was that chips lose value almost immediately. Real rental data contradicts that, and Kush argues older chips will cascade into lower-tier inference work.

Kush Bavaria: "CoreWeave has also said like their A100s are still being rented out for like a three year period... If you asked this question last year, people would have said, oh, these chips are getting depreciated like tomorrow... The A100 is like six years old. And they're still being rented out." [00:19:47] The older chips "will have a use besides being like frontier model training. They'll be used for lower scale inference, smaller model inference." [00:20:15]

Inference will fragment into specialized verticals rather than consolidate into a few horizontal winners

Kush expects one horizontal provider for everyday inference plus many specialized ones: fast, cheap-and-slow, image/video, and scientific modalities.

Kush Bavaria: "there will be like a horizontal provider that is like very good at like providing like base level inference... But then there will be like a hundred different specialized inference companies that exist." [00:21:48] On specialization: "It's like DNA or like proteins or like some chemical formula to like some other chemical formula." [00:21:28] Even if Modal wins fast inference and image generation, "you own like one part of it." [00:21:10]

Compute needs a standardized index and futures market, modeled on Brent, S&P and Moody's

Ornn's thesis is that compute is a fungible commodity with many suppliers and many demanders, so it needs a reference price and derivatives. Kush argues you don't need to standardize everything, just one benchmark grade.

Kush Bavaria: "Brent... is like 0.8%, 0.7% of the actual global like oil volume... compute, there's all these different flavors, even gradings of quote unquote, like crude oil of a B300. But what you really need to figure out is like you need to track a very specific type of that compute. And that becomes the index." [00:31:19] The target: "we want to be at the S&P for compute... compute is a larger market than oil will be or any of the other commodities have been for the past like 100 years, 150 years." [00:33:07]

Hedging demand: who actually needs compute futures

Kush identifies lenders and Neo-Clouds as exposed to two things: tenant credit risk and end-of-life price risk. He compares this to farmers selling corn futures before harvest.

Kush Bavaria: "I need to basically get to four years to pay back my loan or I need to get to five years, but I only have a three-year contract. So you're kind of exposed at the end of it." [00:37:04] And: "The same way farmers used to do this like back in the day... they needed to sell futures back in the market because I was at plant today and I only realized my revenue a while back." [00:37:31]

Standardization drives down credit discount rates

Harry relayed a point from Mars (Ornn's investor circle): an index drives standardization of clusters. Kush agreed and extended it to credit markets.

Kush Bavaria: "If the product is standardized, that means like if one customer default, someone else can move in. If I have an apartment that all looks the same for every single building, that means it doesn't matter if like I lose it, someone else can just move in." [00:39:14] Harry's synthesis: "I guess it would remove the discount rate."

Compute demand is real and inference demand is not speculative; training is the speculative part

Kush separates inference (customers pay for tokens, someone is on the other side) from training (a biotech-like bet).

Kush Bavaria: "the good thing about the inference demand is it's not speculative. It is like real. The tokens have to be used by someone... Training is much more speculative, right? It is like kind of like biotech where you spend all this money developing this drug and you don't actually know if it works out or doesn't work out." [00:42:22]

Culture and team composition as a scaling strategy

Kush argues that to 3x headcount without losing communication, you hire competitive college athletes, and he tests people in uncomfortable dinners rather than just technical rounds.

Kush Bavaria: "It turns out the best like communicators are like college athletes, because what do you do when you play a sport? You're just like communicating every single time. And the second thing is like, you are so competitive that you hate like losing." [01:00:21] On interviews: "We'll take an engineer with go-to-market people and have a dinner with them. And then we'll just see, do they vibe with the team?" [01:01:58]

Boutique over brand-name funds, and partner over logo

Kush picked Vine (a boutique) and Thrive's Miles for the people, not the logo, and advises other founders to do the same.

Kush Bavaria: "I'd go for like the boutique... The boutique investors like want to care about you a lot more because their success is tied inherently to your success." [00:52:59] On Miles: "the first like 45 minutes, he literally didn't care about the business. He just talked about like us as people." [00:50:20]

2. Contrarian Perspectives

Neo-Clouds have no moat and are still great long positions

Most investors want a moat; Kush says the moat is owning the infrastructure, and that the sell-off in CoreWeave and Nebius is a misread of price data.

Kush Bavaria: "I don't think neoclouds essentially have a moat... they realize like early on there is no moat. And so now everyone's trying to expand to find some moat." [00:29:46] Yet: "I want to be like the landlord. Like that's a great business to be in." [00:30:11] On the shorts: "the price per GPU hour is still increasing. So I like don't understand why like people are like, oh, we need to go like short, like CoreWeave or Nebius... I'm not allowed to publicly invest, but if I could, I would go like very long." [00:14:17, 00:15:07]

Better to have more compute than better talent

In a room full of talent-obsessed investors, Kush picks compute.

Kush Bavaria: "You can maybe even argue that having more compute is better than having better talent in the company itself. It's such a big moat." [00:43:51] Asked which he'd choose: "More compute. I think that the models are good enough to do like RSI. So you just kind of need compute. Or they will be good enough in like three months from now." [00:44:17] He'd buy OpenAI over Anthropic on the same logic: "They have more compute." [01:05:02]

Enterprises adopt AI slower than the world thinks, while capability moves exponentially

Kush rejects the idea that the world changes uniformly. The AI bubble in San Francisco compounds while enterprises stay linear, which he says creates volatility, not a bust.

Kush Bavaria: "I think the enterprises will take longer than people think to adjust. And then the world will change faster underneath them... Coca-Cola is probably wondering like, oh, how do I put like AI in my business?" [00:45:02] He concedes the bear case openly: "Enterprises don't see that AI is actually impacting their business... which I think is also partly true today." [00:15:23]

Vertical-integration risk for inference providers is overrated; Neo-Clouds will reach inference first

The popular view is that Fireworks, Modal and others will move down-stack into chips. Kush bets the other direction.

Kush Bavaria: "the neoclouds have like gone through this cycle, like I think two cycles now, deploying chips, learning things... I would take the side of like a CoreWeave or an EBS and say they're going to be able to get to the inference before you are going to be able to get to the chips." [00:48:04]

New York beats Silicon Valley because the Valley is "coin operated"

Kush says talent in the Valley moves for marginal comp and equity, while New York has loyalty and family-style culture.

Kush Bavaria: "We have this whole theory that people in the Valley are like coin operated... if you pay a researcher 100K more, they're joining your company. If you give them 50 bips more equity, they're joining your company. In New York, there's loyalty in people." [00:00:00]

3. Companies Identified

Ornn

Kush's company; builds price indices, a physical/spot compute exchange, and futures products for GPU compute. Mentioned as the subject. Backers include Vine, Thrive, a16z, AIX. Kush's goal: "we want to be at the S&P for compute" [00:33:07], with "debt ratings for like data centers" as a future revenue line [00:46:48].

CoreWeave

Neo-Cloud GPU provider. Mentioned for lead times and A100 rental durability, and as a long. "if you want to go to CoreWeave and buy something... you're probably quoted like Q4, 2027, Q1, 2028" [00:11:12]; "CoreWeave has also said like their A100s are still being rented out for like a three year period" [00:19:47]. Kush also notes CoreWeave is expanding into post-training and tokens to find a moat [00:30:05] and that its bond spreads are way higher than Meta's [00:38:13].

Nebius

Neo-Cloud. Mentioned as a long alongside CoreWeave. "if I could, I would go like very long" [00:15:07]. Also listed as a likely inference winner ("an EBS" in the transcript, read as Nebius) [00:48:31].

Crusoe

Vertically integrated data-center builder. "Crusoe is such a good company because they build out all of like the actual electricity infrastructure and they build out the transformers. And like they vertically integrate very well." [00:24:57]

Modal, Fireworks, FAL, Baseten

Inference providers that are Ornn's marquee tenants. FAL "just raised like 1.6 billion. Modal also raised... 750" [00:40:04]. Kush says their money "is just going to compute." [00:40:04] He argues Ornn must turn down their large orders to seed liquidity for small startups [00:40:56].

Factory

AI coding harness. Kush's strongest endorsement in the episode: "I would put all, if I had personal cash, I'd put it all into factory." [01:07:19] Why: bring-your-own-key economics ("$200 per month plan is basically unlimited tokens" [00:18:36]) and cost efficiency versus Devin ("the tokens used on Factory end up going like 10x higher than they do on like Devon" [00:19:05]). He also said the "underlying harness of Factory somehow is better." Harry, a Factory investor, confirmed.

Cognition (Devin)

Coding agent product. Kush used it and moved off it on cost: "why does it cost me so much to like use the product itself?" [00:19:05]

Cursor

Part of Ornn's stack: "we use like Cursor and GrokBot." [00:18:09]

xAI / GrokBot

Kush uses it for daily tasks with no human assistant: "I don't have an assistant or any of these things. Just pure GrokBot. And it works well." [00:18:09] He also praised Elon's practice of lending excess chips: "he goes to like SpaceX and he says, oh, I have extra chips" [00:17:48].

Rogo, Harvey, Legora

Vertical AI incumbents. "Rogo is like killing it and like Harvey's killing it and Legora... the easy categories are all done." [00:24:03]

Novig, Kalshi, Polymarket

Prediction/sports exchange competitors; Kush built a college-scale precursor and almost interned at Novig. [00:05:04]

Thrive Capital, Vine, a16z, AIX

Investors. Vine's Alex and Adam Valkin invested on relationship terms [00:51:02]; a16z moved in a week [00:53:39]; AIX (Nick Krantz) is a long-term mentor [01:03:41].

ICE, CME, S&P, Moody's

Reference comps for Ornn's end state. "ICE is a $100 billion business. CME is like $120 billion business. S&P is worth more than both" [00:41:21].

Meta

Cited for its bond pricing as an AI infrastructure credit: "I think it's getting more expensive just because like people are just like a little scared in the market." [00:38:13]

Eight Sleep, MongoDB, Framer

Episode sponsors, mentioned in ads; not relevant to the substance.

Anthropic, OpenAI

Cited for valuations and compute. Kush prefers OpenAI at the stated prices ($1.2 trillion versus Anthropic's $1.5 trillion) because of compute access [01:05:02].

TSMC

Named as the true production bottleneck: "You can't make the fabrication line at TSMC move faster." [01:05:15]

BCG

Kush's former employer; source of consulting studies he cites on weak enterprise AI ROI [00:15:23].

Shopify

Used by Harry as an example of volatility: "dropped 15% after a tiny miss in earnings" [00:44:28].

Blackstone

Mentioned by Harry as a provider of Anthropic's bond financing [00:38:13].

SpaceX

Referenced as the vehicle through which Elon rents out chips [00:17:48].

4. People Identified

Myles Grimshaw (Thrive Capital)

Introduced Harry to Kush. Harry: "Myles Grimshaw, who I think is one of the best investors in venture capital at Thrive" [00:00:37]. Kush on why he chose Thrive: "I think Miles. I also think... the partner matters like way more." [00:54:08]

Travis Kalanick

Kush's idol: "he's a Steve Jobs of my generation" [00:06:23]. Used as a mental model: "Way, way more like aggressive and ambitious." [00:06:47] He would add Travis to his board, and his question would be: "How do you handle emotional volatility?" [01:04:50]

Nick Krantz (AIX)

Mentor since Kush was 17. "he's mentored me since I was 17 years old... I call him like once a day... he cares about me as a person rather than like the company." [01:03:41]

Wayne

Kush's co-founder and pro tennis player; co-built the college betting app and pushed to move fast on the a16z term sheet [00:05:04, 00:53:39].

Adam Valkin and Alex (Vine)

Investors who prioritized the founders as people. Kush: "Adam and Alex like really care about like us as like people... for them, this will matter like way more." [00:51:59]

Chris Dixon (a16z)

Kush's a16z round came through his team: "talked to them for like a week, got the term sheet." [00:53:39]

Eric (Moderna)

Quoted on buying compute 18 to 20 weeks ahead; Kush also noted Eric's comment that compute has many flavors [00:16:49, 00:31:46].

Gavin Baker

Harry cited him for the claim that chip prices rose daily for roughly 60 days [00:12:25].

Jensen Huang

Referenced via his "we are not a car" podcast comment on the morning of Kush's marathon [00:08:47].

Mars

An investor whose question Harry relayed about index-driven standardization [00:39:04].

Matan (Factory co-founder)

Founder of Factory; Harry called him and Ina "amazing" [00:19:05]. Kush wants to invest: "Tell Matan I want it." [01:07:19]

Jacob (Novig)

Founder Kush and Wayne wanted to intern for; inspired the college app [00:05:04].

Gabe (Rogo)

Mentioned as someone Kush spoke with; Rogo is "killing it." [00:24:03]

Will

Ornn engineer; "we spend like a hundred hours a week together." [00:59:25]

Elon Musk

Cited for monetizing spare compute through SpaceX [00:17:48].

Brian Johnson

Harry's reference for the longevity thesis [00:46:39].

Pat Grady

Harry named him alongside Myles as one of the smartest in venture [00:04:29].

Rich Socher

Founder of AIX, Nick Krantz's fund [01:04:24].

Diego

Referenced by Kush regarding the Jeff question (a founder Harry has coming on the show) [01:05:50].

5. Operating Insights

Pricing discipline: you're probably underpricing by 10x to 50x

Kush's early data deals were closed at $5,000 and $10,000 a month with instant acceptance, a signal he now reads as a pricing failure. Kush Bavaria: "we ended up selling it for like 5,000 and like 10,000 a month. And I was like, are we like losing here? Because they would immediately say yes. Then I realized like they're willing to pay like $100,000 a month, like 500k a month." [00:07:11] Tactic: a customer who says yes immediately means you quoted too low.

Use customer concentration as a liquidity tool, not a revenue goal

Running a marketplace means turning away your largest customers. Kush Bavaria: "You have to say no to a modal. You have to say no to a fireworks. You have to say no to a FAL, which is willing to pay you gazillions of dollars in revenue... It's like short term game versus like long term profitability." [00:40:56] Without that, "If modal is the only one that buys all the compute, it's just going to consolidate. No one else gets to access it." [00:40:45]

Go down-stack when no one will give you the data

Ornn tried for six months to get Neo-Clouds to share transaction data and failed, so they started controlling chips via a capital-light Airbnb-style model. Kush Bavaria: "We tried for like six months... we have this concept of like, no one's going to come save us. We got to save ourselves." [00:34:45] The mechanism: third-party investors fund the chips, Ornn controls and monetizes them, "I will control the chips. I will manage it. I'll run it for you. You just put the money here." [00:35:42]

Hire for culture and communication, not just technical skill

Because AI tools have flattened baseline technical ability, Kush filters for mobility, uncertainty tolerance, and competitiveness through awkward group dinners and athletic background. [01:01:50]

Test your own valuation against profitability and multiples

When asked how he avoids believing the hype: Kush Bavaria: "Revenue numbers, profitability, look under the books... our numbers do check out with what the valuation of the company is. Our multiples are much lower than any of the other companies." [01:02:46] He declined term sheets at $1.5 to 2 billion to protect that discipline [01:03:14].

6. Overlooked Insights

Ornn's data is a byproduct of the market, so the physical exchange is the real moat

Kush says only five or six people work on the data and index product, yet it only exists because roughly 30 others run the physical market. Kush Bavaria: "it's a data and index company. That's true. And that's like five people on the team, six people on the team. But in order to get that data, that is like from the other like 30 people on the team." [00:41:21] The implication: Ornn's real asset is a proprietary transaction-level dataset (billions of dollars of compute traded) that Coreweave and others refused to hand over for six months. That scarcity, not the index methodology, is what Moody's and S&P comps would rest on.

Chip credit rating is a standalone business, and the existing rating agencies may not understand the asset

Mentioned in a few sentences but potentially the largest long-term revenue line. Kush Bavaria: "Moody's for data center debt... do you think Moody's understands how data center like debt works? Who knows? Like do they really understand like the credit risk that modal is having on their people that they're buying the infrastructure from? I don't know." [00:47:00] Combined with the point that tenant credit quality (AI startups) is the real exposure for Neo-Cloud lenders [00:14:45], there is an unfilled gap in rating tenant risk and residual GPU value, and Ornn's pricing data is the natural input.