30 秒でわかる
AI companies must commit to scarce computing capacity months before they know what it will cost or whether their customers will pay. Ornn co-founder Kush Bavaria wants an index and eventually futures contracts that let them manage that uncertainty. To get reliable price data, his company arranges physical GPU capacity financed by others. He expects older chips to remain useful for running smaller models, but says a slump in enterprise demand or a customer default could expose infrastructure lenders.
- 1GPU capacity can take months to secureBavaria says buyers face long chip delivery and installation lead times, while some cloud capacity is booked much further ahead. He sees continued demand keeping prices high, though weaker demand would change that.
- 2Buyers need protection against a changing GPU priceWhen customers must commit 22–26 weeks ahead, the eventual market price can differ. Bavaria argues that contracts and hedges could reduce that exposure.
- 3Older GPUs can still serve smaller AI modelsBavaria points to continuing rentals of earlier GPU generations and expects them to find uses in less demanding inference.
- 4A fragmented market needs a common reference priceDifferent customers need different locations and configurations, but Bavaria argues one standardized configuration could serve as a benchmark, as a particular grade of oil does.
- 5Ornn handles physical capacity to obtain pricing dataBecause providers would not readily share transaction records, Bavaria says Ornn arranges financing and manages chips for owners, gathering the data it needs for an index.
- 6Lenders face both price and customer credit riskA provider can be left with capacity to resell if a renter defaults; it also cannot know the chip rental rate after an initial contract ends.
- 7Inference demand is already influencing compute pricesBavaria distinguishes paying customers who use models from the more speculative expense of training a new model.
ポイント
- Bavaria says long lead times force GPU buyers to commit before they know future demand or prices.
- A standardized benchmark for one type of compute could provide a reference price without making every chip and data center identical.
- Providers and lenders face both changing GPU prices and the risk that their customers cannot pay.
- Older GPUs may retain value running smaller models even when they are no longer suitable for training frontier systems, Bavaria argues.
- Bavaria expects the business of running AI models to include specialized providers alongside general-purpose ones.
チャプターごとに整えた全文を読む
From student bets to Ornn
Bavaria describes building a small sports-betting app and learning to negotiate more firmly. He recounts running a marathon without the training he had planned.
Stebbings introduces Bavaria, who says that his first project with a college friend was a sports-betting app built after they failed to get internships at a similar company. They could code but lacked commercial and legal experience; the app found some use among MIT fraternities without spreading much further.
Bavaria says he looks to Uber founder Travis Kalanick as an example of greater ambition and assertiveness. He recalls selling data to trading firms for $5,000 or $10,000 a month, then realizing some were willing to pay substantially more.
A marathon he had entered before starting Ornn arrived when he had no time to train. Having run before, he finished in roughly three and a half hours, but calls it a painful experience rather than a model training plan.
Why compute prices could keep rising
Bavaria compares infrastructure owners with landlords and says scarce GPU supply supports rental prices. He also describes the demand reversal that could hurt lenders.
Stebbings asks whether AI can thrive while infrastructure investors lose. Bavaria compares data-center owners to landlords: demand for their buildings and GPUs sustains rent, although a weak AI market or tenants building their own capacity would change the picture.
Stebbings points out that enterprise customers still question returns and costs. Bavaria agrees that weakening demand would lower GPU-hour prices through the same supply-and-demand mechanism that affects property rents.
For buyers, he describes a tight market: some providers quote availability years ahead, while ordering chips and installing them can take about half a year. His figures are an account of the market he sees, not a delivery guarantee.
Bavaria is relatively sanguine about debt financing because he sees actual use for AI products and demand for their output. He says prices for some older and newer GPU generations have risen as buyers compete for capacity.
Longer-term contracts once assumed future GPU hours would be cheaper, Bavaria says. A renewal option at a fixed price, once offered freely, is now valuable because suppliers think future prices might instead be higher.
Asked why investors worry about companies such as CoreWeave and Nebius, Bavaria says their borrowing depends on both chip values and the ability of AI-company tenants to pay. He remains optimistic on the providers because he expects demand for compute to continue.
The bearish scenario is that large enterprises fail to find enough productive use for AI. If capacity has been bought ahead of demand and customers cannot pay, defaults could leave creditors holding equipment worth less than the price assumed in the financing.
Bavaria says the lead time for buying chips may require planning 22–26 weeks ahead. That creates uncertainty about the price at delivery and provides a reason, in his view, for buyers to hedge their exposure.
Managing uncertain GPU costs
Long purchase lead times make hedging important, in his view. He discusses Ornn’s own AI tools and the remaining usefulness of older GPUs.
For cash-rich companies committed to AI, Bavaria favors securing capacity early. He believes surplus chips could be rented to others, although that bet depends on his expectation that demand and prices will remain strong.
He describes using AI assistants for everyday work and coding products including Cursor and Factory. He says the ability to use the company’s own model credits through Factory has been useful; Stebbings notes that he is an investor in that company.
Older GPUs are not necessarily obsolete when a new generation arrives, Bavaria says. He points to continued rentals of A100 and H100 chips and expects older hardware to run smaller models or less demanding inference, the process of generating answers from a trained model.
He expects demand to split between fast and inexpensive inference and workloads such as images, video or biological research. A general provider may coexist with specialized services; when pressed on consolidation, he allows that the number of competitors may shrink.
An overheated funding market and physical bottlenecks
Bavaria sees fear of missing out in fast funding rounds but would invest in data-center equipment. He and Stebbings contrast their impressions of New York and Silicon Valley.
Stebbings says venture rounds can now close extremely quickly. Bavaria interprets this as fear of missing the next major AI company and doubts that every investor in such a round can have examined the business carefully.
Bavaria considers some obvious application categories crowded. He would instead look at the physical equipment needed to build data centers, including electrical transformers, where he believes the size of the opportunity can be overlooked.
Asked whether Ornn misses opportunities by being based in New York, he says he values loyalty and camaraderie there. Stebbings adds his own criticism of groupthink and transactional relationships in Silicon Valley.
For Bavaria, working with four close friends from college is a central attraction of founding Ornn. He says the company is as much about the people he works with as the market it is trying to build.
Why compute needs a price benchmark
Bavaria says fragmented supply and many buyers make a market useful. A standardized reference configuration could anchor prices despite differences among clusters.
During a consulting internship, he used ChatGPT to generate spreadsheet formulas and wrote Python to complete assignments quickly. The experience helped convince him that he did not want a conventional corporate career.
Why should compute have a market? Bavaria says many suppliers offer broadly similar NVIDIA GPUs and many customers seek them, even though location, memory and other requirements vary. That combination makes prices a matter of matching supply and demand.
Stebbings asks whether suppliers have a defensible advantage if capacity is interchangeable. Bavaria says neoclouds are expanding into services such as model inference to differentiate, while owning scarce infrastructure can still be an attractive business.
He contrasts a historically concentrated cloud market with a larger number of specialized GPU suppliers and buyers today. Auctions for chips already resemble the price discovery he expects a more formal market to provide.
A price benchmark need not describe every type of compute. Bavaria compares the idea with oil: different grades trade at different prices, yet a standard reference grade can anchor pricing for a much broader market.
The essential step, he says, is to choose and standardize one specific computing configuration. Stebbings asks about trust; Bavaria says a pricing and index business also has to build the long-standing relationships that make customers trust its numbers.
Ornn began with an index because Bavaria believes buyers will eventually need futures and options on GPU prices. To construct a robust index, it also needs evidence of actual transactions in physical capacity.
Getting data by arranging physical capacity
Ornn moved into managing chips financed by others to observe transactions. Bavaria explains who might use futures and how credit and price risks interact.
The company found that established providers were reluctant to hand over valuable pricing data. Bavaria says Ornn therefore arranges for others to invest in chips or data-center projects while it controls and monetizes the capacity, an arrangement he likens to a managed rental market.
This approach is more involved than the software-data business he originally imagined. He says investors can fund the assets while Ornn manages them, though the underlying financing and operations are more complicated than the analogy suggests.
The first users of compute futures, he expects, will be parties already exposed to GPU prices. A lender financing capacity for an AI startup is also exposed to that startup’s ability to raise money or generate sufficient revenue to meet its obligations.
If the tenant defaults, the owner must resell the capacity at whatever price is then available. Separately, a chip may need to earn money beyond the term of its first contract to repay its financing, yet the later rental rate is uncertain.
He compares this uncertainty with a farmer planting a crop before knowing the harvest price. A future contract can provide a price ahead of time, rather than leaving the seller entirely exposed until the crop or computing capacity is delivered.
Stebbings asks about the cost of borrowing for Meta and AI infrastructure companies. Bavaria says investors are uneasy about the scale of spending on data centers and about whether the end demand will support that spending.
Standardization and inference demand
Bavaria argues that a reference index can encourage capacity that can be reassigned between tenants. He expects inference to be an increasingly important source of demand.
An index could encourage more standardized GPU clusters, Bavaria says. If a tenant leaves or defaults, a replacement could use the same capacity, much as another renter can occupy a standard apartment.
For a market to be liquid, he says, it needs more than one buyer. Ornn sometimes has to distribute capacity rather than sell everything to a single buyer willing to pay more, sacrificing near-term revenue to support a broader market.
Bavaria sees Ornn as a data and index business, with the physical market necessary to produce trustworthy data. He points to large commodity-data and exchange businesses as an indication of what a mature market might support, while forecasting an even larger role for compute.
He believes inference already helps determine compute demand because customers pay to use the resulting tokens. Training a new model is more speculative in his analogy: like developing a drug, it requires spending before the outcome is known.
One risk he sees is insufficient credit diligence by neoclouds serving startups. Even if a vacated cluster can eventually be rented again, a provider can still be exposed to an individual customer’s failure and uncertain resale terms.
When Stebbings asks what will determine which AI labs win, Bavaria names talent and compute, ultimately choosing compute if forced to pick one. He expects models to improve rapidly, while acknowledging that enterprises may take much longer to incorporate them.
Uncertainty, AI adoption and credit ratings
Bavaria expects established enterprises to adopt AI more slowly than the frontier changes. He also outlines his ambitions for data-center credit ratings.
Stebbings says uncertainty makes it difficult to price businesses five years out. Bavaria argues that companies may still be figuring out how to apply AI while the capabilities available to them change much faster.
Bavaria offers highly speculative ideas about AI and longevity, including the possibility of much longer lives. Later, when asked which 2030 belief is most likely wrong, he acknowledges that his more extreme predictions may not come true on that timetable.
One adjacent business he hopes Ornn can develop is credit ratings for data-center debt. A specialist, he argues, could assess tenant and infrastructure risks that general ratings firms may not understand in the same detail.
If resources were unlimited, he says, he would explore building data centers too. He expects inference providers and neoclouds to move into each other’s businesses, but thinks operators with several hardware deployment cycles have a head start.
Choosing investors for a long-term company
Bavaria explains how he chose backers, named the company and sought advice. He discusses the psychological strain of founding a company.
Bavaria says the company raised money partly to get the right partners, not merely because it needed cash. He wants a durable business built with friends, rather than one designed for a quick sale.
The name Ornn came from a League of Legends character and an inexpensive four-letter domain, he says. A short name appealed to him more than an elaborate origin story.
He says a conversation with investor Miles stood out because it focused first on the founders as people. Stebbings asks how that fits with another dinner where Bavaria seemed eager to move past small talk.
Bavaria explains that others at the table were discussing tennis, a subject he could not easily join, while he had another investor meeting afterward. A smaller fund later offered similar terms to a larger one, and he valued how much the investment would matter to its partners.
He favors boutique investors when he believes their incentives are closely tied to a startup’s success. He also describes raising from a larger firm after another potential investor took longer to decide, then choosing Thrive for the partner relationship.
Because he is young and has limited industry experience, Bavaria seeks advice from older founders and investors who have already made mistakes. Their judgment on people and decisions, he says, is difficult to recover from books alone.
He says he underestimated the psychological strain of starting a company. The uncertainty and constant swings between good and bad developments can make the work mentally consuming; Stebbings asks how he manages that pressure.
Asked about relationships, he jokes about staying single until the company reaches its next funding milestone. He says relationships can distract from work but may also provide emotional stability.
He says working with friends matters more to him than personal wealth, provided he has enough to live comfortably. He shares a sixth-floor walk-up apartment close to the office and spends much of his time at work.
Stebbings suggests that founders with high company valuations might sell a small amount of stock to improve their living conditions. Bavaria says the team already spends on things such as a gym and healthy food, and questions how much a nicer apartment would change his routine.
Hiring for a high-intensity team
He describes the communication and competitiveness he looks for in colleagues. Stebbings challenges him on his lifestyle and the pace of work.
Bavaria says he initially underestimated company culture but changed his mind after spending long weeks alongside colleagues. He now considers it important that people can enjoy working together through sustained pressure.
He says many hires have played competitive sports. He associates that background with communication and a desire to win, qualities he thinks will matter as the team grows from roughly 30 people toward 100.
Stebbings values the discipline of repeated exercise without an immediate payoff; Bavaria adds that long-term benefits can motivate people. They discuss how such habits might translate into persistence at work.
The company also brings candidates to dinner with people from other functions to see how they interact outside a familiar technical setting. Bavaria says technical skill matters, but he also wants people who can navigate uncertainty.
Asked how he avoids believing publicity about his company, Bavaria points to revenue, profitability and a disciplined approach to valuation. He says Ornn declined some higher-priced offers when choosing investors.
Investors, coding tools and hopes for 2030
Bavaria discusses his mentors, compute constraints and changing view of coding tools. He closes with speculative hopes for longevity and AI.
For a favorite investor meeting, he recalls meeting a mentor at 17 who continued to advise him over several years. He values an investor who pays attention to the founder as a person, not only the company.
If he could add one person to the board, he would pick Travis Kalanick, whom he has not met. He says he would ask how Kalanick handles emotional volatility.
Asked to choose between OpenAI and Anthropic at hypothetical valuations, Bavaria picks OpenAI because he believes access to compute is a major constraint. He argues that money alone cannot immediately accelerate chip fabrication.
He has changed his view of independent coding tools that sit around AI models. He now values the ability to switch models for different tasks and the experience built around them.
Bavaria says he particularly likes Factory’s coding product, though he does not claim to know all the company’s financial fundamentals. Stebbings repeats his interest as an investor in it.
On the future, Bavaria concedes that his more dramatic 2030 predictions could be wrong. He is excited by biotechnology and longevity, and hopes more accessible computing capacity can support work that improves human lives.
同じチャンネルの記事
1:15:12EN
1:15:58EN
50:17EN
57:17EN
1:08:33EN
1:02:12EN