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
· Summarized
Key takeaways
- Kush sees data-center owners as landlords who can keep collecting rent, but their real downside is falling AI demand and tenants defaulting on GPU contracts.
- GPU capacity is currently undersupplied: new projects can take roughly half a year to build, while some CoreWeave capacity is reportedly booked into late 2027 or early 2028.
- He believes GPU prices will continue rising, with Rubin potentially reaching $20–$30 per GPU hour versus roughly $7–$8 for today’s five-year pricing.
- Older A100 and H100 GPUs are not automatically obsolete; they can serve smaller models, lower-scale inference, and specialized workloads.
- The inference market may combine one horizontal provider with many vertical specialists focused on speed, price, images, biology, physics, or other domains.
- Orn is building compute indexes, physical markets, and eventually futures, options, and debt ratings because GPU pricing, credit exposure, and depreciation need standardized benchmarks.
- Kush views rapid AI financings as FOMO-driven and considers physical infrastructure such as transformers and electricity systems more underheated than customer support software.
- For Kush, long-term relationships, founder-focused investors, and an athlete-oriented culture matter more than maximizing short-term valuation or compensation.
Summary
Kush’s dorm-room sports betting app failed beyond MIT
Kush Bavaria said he and his friend Wayne tried to intern at Novig during their MIT freshman or sophomore year, were rejected, and responded by building their own sports-betting marketplace. The app let users buy and sell contracts on player statistics, resembling products later built by Novig, Kalshi, and Polymarket. Kush worked with MIT fraternities to distribute it, and it became popular within the school but never expanded because the founders could code while knowing almost nothing about marketing, legal work, or go-to-market. He also said his admiration for Travis Kalanick helps him counter his naturally agreeable personality: early data deals were sold for $5,000–$10,000 per month even though buyers would have paid $100,000 or $500,000.
The marathon he ran without training
Kush had previously run and completed a marathon, then signed up for another before starting Orn. Once the company began, he had no time for the roughly two hours a day required for training, but refused to quit when race day arrived. Recalling Jensen Huang’s “we are not a car” comment, Kush woke up determined that he and the company would still run the marathon. He finished in roughly three and a half hours, calling it painful and a terrible experience, though Harry noted that his own trained marathon time was about 3:45. Kush presented the episode as evidence of his tendency to commit first and endure the consequences.
Infrastructure investors resemble landlords, but tenants can still default
Kush compared data centers with apartment buildings: infrastructure investors own the buildings and rent GPU capacity to inference providers such as Fireworks, Modal, FAL, and Base10, as well as companies including OpenAI and Anthropic depending on deal structure. He believes infrastructure investors generally win because they keep collecting rent as long as tenants need the buildings. Harry pushed back that weak enterprise ROI or falling demand could hurt providers and therefore infrastructure owners. Kush agreed that GPU-hour prices would fall if demand weakened, but maintained that the business ultimately follows supply and demand rather than a permanently guaranteed return.
GPU capacity is booked years out and prices may keep rising
Kush described the compute market as severely undersupplied. A customer asking CoreWeave for capacity could reportedly be quoted late 2027 or early 2028; obtaining chips directly from an OEM takes about 20–22 weeks, followed by two or three weeks of installation, even for a relatively small build. He said H100 and newer B200/B300 prices have risen because AI companies sell more tokens and must either become more efficient or buy more compute. He expects Rubin capacity could cost $20–$30 per GPU hour, compared with roughly $7–$8 today on a five-year arrangement. Earlier contracts sometimes gave buyers free options to renew at today’s price, a concession the market no longer offers.
Why Kush would go long CoreWeave and Nebius
Kush said the bearish view of CoreWeave and Nebius reflects leverage and doubts that AI demand is real, but he would go long both companies if he were permitted to invest publicly. Their risk is not simply chip pricing: lenders must also ask whether the AI companies renting the GPUs can keep paying. Kush’s bear case is that enterprises such as Coca-Cola or McDonald’s discover AI does not materially change their businesses, leaving companies with more compute than they need and causing defaults. He nevertheless believes current token demand is useful, that chip prices continue rising, and that the public markets are too quick to dismiss products people already use.
Older GPUs are staying useful through specialized inference
Kush rejected the assumption that older GPUs become worthless as soon as a new generation arrives. CoreWeave has said its A100s can still be rented for a three-year period, while H100s are also being rented for long periods; Kush characterized the chips as roughly six and four years old respectively. He expects older hardware to move from frontier-model training into lower-scale inference and smaller models. Inference itself may split into fast inference optimized for tokens per second, cheap and slow inference, and specialized workloads such as images, video, biology, physics, DNA, proteins, or chemical formulas. The chips therefore retain value beyond the frontier-training market.
Inference will have a horizontal base and many vertical specialists
Kush expects one horizontal provider to deliver everyday, base-level inference, but believes perhaps a hundred specialized inference businesses can survive by serving distinct workloads. Harry challenged whether the current field of roughly a thousand inference companies is simply a temporary bubble that will consolidate. Kush conceded the total number will shrink, but argued that the survivors will verticalize rather than one general-purpose provider winning everything. He cited FAL’s image and video focus as an example and suggested that companies may own one or two inference categories, such as speed, image generation, biology, or another domain-specific form of computation.
Kush calls rapid AI investing FOMO and sees physical infrastructure as underheated
Kush said a financing round completed in two hours is probably driven by fear of missing out rather than careful understanding. Investors who missed early OpenAI or Anthropic opportunities are now trying to buy the next company that could become a trillion-dollar business. He considered customer support particularly overheated, with many companies pursuing similar categories, while companies such as Harvey, Rogo, and Lagrange make the easiest legal and workflow categories difficult to attack. His preferred underheated area is the physical AI infrastructure stack: transformers, electricity systems, and other data-center components. He argued that even apparently small transformer markets matter when data centers could become a $7 trillion market.
New York’s loyalty appealed more than Silicon Valley’s transaction culture
Kush said New York is better for his company than Silicon Valley because he believes Valley employees are more “coin-operated”: a researcher may move for $100,000 more compensation or 50 basis points more equity. He sees stronger loyalty in New York, where employees identify with their colleagues and families rather than constantly optimizing transactions. Harry agreed that Silicon Valley has more groupthink and an incestuous movement among firms and startups, and criticized the American business habit of treating everything as a transaction. Kush said one reason to build Orn is that he works daily with four close friends from school, making the company a community rather than merely a financial project.
Orn had to control physical chips to create trustworthy price data
Kush explained that compute needs a market because many suppliers sell broadly comparable NVIDIA GPUs while many buyers demand them with different specifications, locations, memory, or timing. Orn began with index products and futures, but discovered that providers would not voluntarily share their transaction data: Kush said CoreWeave would not hand over its most valuable pricing information, despite six months of attempts. Orn therefore began arranging access to chips and data-center projects itself, using sale-leasebacks and similar structures in which other investors fund the equipment while Orn controls and monetizes it. This lets Orn build a physical market whose transactions can support robust indexes rather than relying on incomplete provider disclosures.
Standardized compute could unlock futures, lending, and ratings
Kush compared compute with oil: Brent represents only a small fraction of global oil, yet serves as a reference price. Orn similarly does not need to standardize every GPU configuration; it needs to define a specific, widely traded compute product and use it as a benchmark. He said standardization also helps lenders because a defaulted customer’s cluster can be reassigned to another tenant instead of being a bespoke stranded asset. The likely first users of compute futures are lenders and owners exposed to GPU prices, customer credit, and depreciation. Kush also wants Orn to become a trusted provider of data-center debt ratings, effectively a Moody’s or S&P for this infrastructure.
More compute beats better talent in Kush’s model of the AI race
Asked what determines which NeoClouds win, Kush answered “talent” and compute, then said more compute may be a larger advantage than better talent. His reasoning is that the tools and models are becoming capable enough to produce recursive improvement, so a company with greater capacity can train and improve faster. He used this to prefer OpenAI over Anthropic when forced to choose between the two: OpenAI has more compute, while Anthropic’s cash cannot make TSMC fabrication lines move faster. Kush expects enterprises to adopt AI more slowly than the San Francisco AI ecosystem changes, creating a widening gap between exponentially improving models and businesses still implementing technology linearly.
The founders chose people-focused investors and built an athlete-heavy culture
Kush said he chose Thrive because Myles Grimshaw focused on him and his co-founders as people rather than only evaluating the idea; he made the same distinction between boutique investors and large multi-stage funds. He raised from Andreessen after another large fund moved too slowly, and said Orn rejected term sheets implying $1.5 billion or $2 billion valuations in favor of disciplined financing. The team now has roughly 28–29 people and hires many former college athletes because they communicate under pressure, hate losing, and understand delayed rewards. Interviews include dinners and unusual cross-functional situations to see whether candidates handle uncertainty and connect with colleagues. Kush said founder life is mentally consuming and emotionally volatile, with confidence and panic alternating in the same day.
Kush expects longevity and AI-human relationships to redefine 2030
Kush said beliefs that currently sound “kooky” include humans becoming biological vessels for AI, with people teaching systems emotions and biological understanding rather than controlling them. He also believes biotechnology could make dying less real and raise both lifespan and quality of life from roughly 80 years to 150 or 200, without the added years being frail or inactive. He admitted these are among the beliefs most likely to be wrong by 2030. His broader optimism is that more compute will support biotechnology and longevity research, and that Orn’s role in financing and hedging compute can help make more capacity available for uses that improve human life.
Sponsors
- 8sleep · 8sleep’s Pod is a smart mattress cover that cools each side to 55°F, heats to 110°F, and tracks heart rate, HRV, and sleep stages without a wearable. The Pod 6 hub is 50% smaller and starts at $1,999. The system was described as being used by Mark Zuckerberg, Andrew Huberman, Charles Leclerc, and the Aston Martin F1 team.
- MongoDB · MongoDB presented its database as a data platform for AI agents, combining a JSON-native database, vector search, and Voyage AI embeddings so agents can access accurate, real-time context without maintaining multiple data pipelines. MongoDB for Startups offers Atlas and Voyage AI credits, technical support, and dedicated support.
- Framer · Framer described its product as an AI website builder for production-ready, editable sites. Users can prompt, inspect, edit, publish, manage CMS content, optimize SEO, add code components, and audit issues in one place. The offer mentioned was 30% off a Framer Pro annual plan.
Show notes
Kush Bavaria is the Co-founder and CEO of Ornn, building the financial market for AI compute. Ornn helps companies price, finance and hedge GPU capacity—bringing the tools of commodity markets to AI infrastructure. Kush is just 23 years old and has raised funding from some of the best in the world including Andreessen Horowitz, Thrive Capital and BoxGroup.
AGENDA:
05:00 Building a Viral Sports Betting Platform from a Dormroom at MIT
09:00 Can AI Win While Infrastructure Investors Lose?
13:00 What Is Everyone Missing About GPU Prices?
14:00 Why Is Everyone Betting Against CoreWeave and Nebius?
19:00 Are Older GPUs Really Becoming Obsolete?
24:00 Where Are AI's Biggest Untapped Opportunities?
29:00 What Would It Take for Compute to Trade Like Oil?
43:00 What Wins the AI Race: Better Talent or More Compute?
48:00 Why Raise Money When You Don't Need It?
59:00 How Do You Build a Team That Hates Losing?
01:05:00 Would You Buy OpenAI or Anthropic?
Follow this show on AudioBrevity and new episodes get summarized for you. Create a free account