Thank you to the Ornn team for their valuable feedback on this piece.
If you visit my website, you'll find a section called 'On My Radar'. This section lists companies that have caught my attention - projects I'm actively following. The projects currently listed on the website are from Q4 last year, so newer projects I've been watching haven't been updated yet, but I still have high expectations for all of them.
In an era where research has become such a public good, the only thing a hobbyist researcher can really offer is taste. As many researchers around me have noted, it's less about being right or wrong, and more about showing "this is what this person finds interesting" - sharing one's taste can be entertaining for readers.
Among the various projects on my roster, I debated which to introduce first, and decided to start with my favorite: Ornn. In one sentence:
Ornn is a company building financial products that allow AI companies, investors, and data center operators to manage their exposure to volatile compute prices.
Preparing this piece taught me a lot about GPUs and the AI infrastructure market, which was genuinely fun. Since this isn't my area of expertise, please let me know if anything is incorrect - I'd really appreciate the feedback!
Let's start by defining the problem. The issue is that AI companies, hyperscalers/neoclouds, and data center operators are all investing massive amounts of capital into the AI industry, yet there's no benchmark for the price of "compute" - the core of these investments.
For example:
Say you're the CFO of an AI company like OpenAI. You need 10,000 GPUs for the next 18 months to train your next SOTA model. You'd typically call hyperscalers like AWS or Azure, or neoclouds like CoreWeave, to get quotes. Each company will give you different prices, and there's no reference benchmark to negotiate against. You sign a contract, but five months later a new model comes out and GPU/hour prices plummet - and there's nothing you can do about it.
You want to build a new AI data center and go to a bank to borrow money. The bank asks about your contingency plans if compute prices fall, or whether you can hedge. Currently, there's no easy way to do this. Unlike financing a new power plant, banks can't find ways to de-risk the output of a data center, so they have no choice but to lend to you at relatively expensive rates.
You operate a hyperscaler like AWS. Your enterprise clients keep wanting to lock in GPU prices for the next three years. You'd like to offer that, but what benchmark would you use to set those prices?
I can't say how realistic these examples are, but the point is: relative to the scale of the AI industry, much of the underlying economics is conducted through bilateral phone negotiations - essentially vibes-based.
For another example, let's look at the Q3 10-Q of CoreWeave, a prominent neocloud (they own GPUs and rent them to AI companies).
As this tweet shows, CoreWeave received a $230M prepayment from MagAI Ventures last August in exchange for providing pre-negotiated, fixed hourly compute prices to MagAI's portfolio companies. Another condition: if MagAI's portfolio companies don't use the compute from CoreWeave, the prepayment must be returned at 12% annual interest.
Why would MagAI Ventures agree to such terms? Likely they wanted to provide value-add to their portfolio companies by securing fixed-price GPU compute. If compute futures or options existed, they could have easily hedged through those instruments. But since they don't exist, they had no choice but to enter into this unusual arrangement with CoreWeave.
Of course, large companies with capital and information advantages can predict utilization rates and negotiate cheaper deals through scale. But companies without these advantages are exposed to information asymmetry and have no way of knowing where their deal stands relative to the market.
When the Ornn team discovered this problem and started working on it, they naturally looked at oil first - oil is the classic commodity. But the problem is that compute and oil are fundamentally different. Oil can be stored; compute cannot. An idle H100 at 3am on Tuesday has zero value. So we call commodities like oil "stock commodities" and commodities like compute "flow commodities."
The quintessential flow commodity is electricity, which has these characteristics:
Although battery technology is advancing, electricity also can't be stored in warehouses, and the grid must balance supply and demand in real-time, every second. Same with compute.
Electricity futures settle based on the average price over the contract period, not a single price at expiry. This is called "Asian-style settlement."
Electricity prices necessarily vary by region. The same 1MWh of electricity has completely different prices depending on where it's produced and consumed. For example, prices in PJM (covering the US East) and ERCOT North (covering Texas) are different.
Understanding these characteristics of electricity helps when thinking about GPU compute as a commodity.
Before looking at Ornn's product lineup, let's examine the team and founding story.
When Kush Bavaria and Wayne Nelms were consulting for private equity firms that lend to data centers, they kept hearing that PE firms were extending credit to GPU infrastructure companies but had no way to hedge their exposure. Without a benchmark to reference, there were no derivatives to hedge with, and financial institutions had no way to properly manage risk. The two founders identified this problem and created Ornn. The team consists entirely of MIT alumni with backgrounds at Google, SIG, Optiver - people who understand both trading and technology.
Now let's finally look at what Ornn actually provides.
The core is OCPI (Ornn Compute Price Index) - the GPU compute index.
Everything starts with a benchmark. The detailed methodology isn't public, but separate indices exist for each GPU type, and they're built on actual executed transaction prices - not just quoted prices ("selling at X"). Regional weighting is also factored in, reflecting that GPUs trade differently across geographies. Following the electricity model, settlement is Asian-style.
You can see the H100 index value on the website, but other metrics appear to require payment. While preparing this piece, I found two articles written using OCPI data, suggesting it's already being used as a meaningful indicator by investors:
Both articles using Ornn's data are behind paywalls. I personally paid to read the second one, which briefly summarized: the article uses data from Ornn covering A100 SXM4, H100 SXM, and H200 from October 2 to December 30 last year to examine how utilization affects price volatility for each GPU type.
From public information, partners contributing to OCPI include:
Hydra Host: Infrastructure company with 30,000+ GPUs across 50+ locations, supplying real-time operational data to OCPI
InfraSight Software: They define a "Workload compute unit" framework for measuring and comparing compute across different models, likely enabling index weighting across different GPU models
Built on top of OCPI is the actual hedging product: Compute Swaps. A compute swap is a contract where two parties agree on a fixed GPU hourly rate for a future period. At settlement, the losing party pays the difference in cash based on OCPI - no actual compute changes hands.
For example, if I enter a 30-day swap contract for 10,000 GPUs at $2/GPU-hour, at month's end Ornn calculates the average of daily OCPI values. If the average is $2.30, my counterparty pays me ($2.30 - $2) × 10,000 = $3,000.
The beauty of swaps is you don't need to change who you buy compute from. Keep using your existing hyperscaler or neocloud - if the GPU hourly price rises and your bill increases, the swap payment offsets it. A very convenient structure. Notably, the first-ever compute swap occurred through Ornn on December 11th last year!
Compute futures are compute swaps with standardized terms, contract sizes, and expiration dates, making them easier to trade on an exchange. Ornn currently operates under CFTC de minimis exemption and is pursuing a DCM (Designated Contract Market) license, so compute futures are not currently available.
However, Architect - the institutional exchange run by former FTX US President Brett Harrison - recently announced they'll support GPU perpetual futures contracts based on Ornn's data, suggesting compute futures trading will become available to institutions via OCPI.
Recently, Ornn announced they'll also launch futures on memory prices like HBM, beyond just hourly GPU rates.
I'm no expert, but apparently when running inference on large LLMs, the actual bottleneck isn't GPU computation but data transfer from memory. HBM is made by stacking memory vertically and connecting it with microscopic copper pillars - extremely difficult to manufacture, which is why only three companies make it: Samsung, SK Hynix, and Micron. Prices have fluctuated over 250% in the past two years. Despite this volatility, no standardized financial product existed to hedge HBM price risk.
Memory and GPU prices often move together during AI demand spikes, but respond to different supply constraints. By offering hedging products on both, Ornn enables participants to hedge more precisely.
Ornn also offers GPU Value Protection through Residual Value Swaps. You pay a quarterly premium, and at the contract's end, if you elect to sell your GPUs, you're guaranteed to receive an agreed-upon price.
This addresses a core problem in GPU ownership: hardware worth tens of millions can see resale values collapse overnight due to new GPU generations or shifting infrastructure requirements. Traditional financing treats GPU terminal value as uncertain, forcing operators to accept conservative terms. RVS transfers this end-of-life price risk to another party.
The key insight is that a GPU's residual value is closely tied to its future revenue potential and this is exactly what OCPI tracks. By establishing a reliable benchmark for compute prices over time, OCPI provides the foundation for pricing residual value risk. Without knowing where GPU hourly rates are heading, you can't meaningfully price what a GPU will be worth in three years.
For datacenter operators, this enables better financing terms, reduces equity volatility, and allows investment decisions based on expected performance rather than worst-case scenarios. For lenders, it provides a defined floor under collateral value, supporting higher advance rates and insulation from technology obsolescence.
Given how important this problem is, I looked for other players besides Ornn. I couldn't find anyone else building compute-based financial products, but Silicon Data is a notable player building compute price benchmarks.
Silicon Data, backed by DRW and Jump, aggregates price data from global sources to provide GPU price benchmarks. The detailed methodology isn't public, but you can see actual metrics by creating an account. I can't precisely compare OCPI and Silicon Data's benchmark without knowing their methodologies, but my understanding is: Silicon Data aggregates quoted prices ("selling at X") from various services and benchmarks them, while Ornn builds OCPI based on actually executed transaction prices. At the company level, Ornn provides financial products on top of OCPI, while Silicon Data stops at benchmark provision.
Not direct competitors to Ornn, but related companies include GPU compute spot marketplaces like SF Compute and Compute Exchange.
SF Compute is a real-time spot marketplace where users can purchase GPUs via market or limit orders. They also offer an auto-procurement feature called sf scale that continuously purchases short-term reservations to always maintain at least 1 hour of compute.
Compute Exchange is more auction-based. Sellers list compute configurations and prices, buyers submit bids, and prices are determined and executed based on this. Compute Exchange shares ownership with Silicon Data, so they appear to leverage Silicon Data's benchmarks and data.
I see two main implications of these spot marketplaces for Ornn:
Data providers: They can supply real-time transaction data to Ornn's indices. These marketplaces are the source of the "actual transaction-based prices" that Ornn emphasizes.
Potential customers: They can use Ornn's products to hedge their GPU price exposure. For example, if SF Compute enters long-term GPU contracts, they could use Ornn futures to manage future price fluctuation risk.
Groups that need Ornn's financial products can be divided into three categories.
First: AI companies wanting to stabilize costs. When training large models costs enormous compute, and that cost can swing dramatically within months, it's inherently stressful. Second: data center operators wanting to stabilize revenue. GPU infrastructure construction is a multi-year investment, and if you can sell future capacity at a known price, investment risk decreases.
The third group is financial institutions that need to manage exposure - let's examine this group more closely.
Currently, when data center operators request project financing from banks, banks naturally ask about revenue projections. Operators present their scenarios, but banks are understandably anxious.
Compare this to power plant financing. If a power plant signs a 10-year Power Purchase Agreement (PPA) and additionally hedges price risk with electricity futures, banks can treat that cash flow as "nearly guaranteed." Collateral value is clear, risk is quantified, so they can provide large-scale funding at low rates. Through Ornn's services, compute can achieve a similar structure, capital costs decrease, and more projects become economically viable.
In ship and aircraft leasing markets, Residual Value Insurance exists - when a company buys an aircraft to resell in 10 years, insurers guarantee a minimum price for that aircraft after 10 years, compensating the difference if it falls below.
With OCPI now established, Ornn has launched exactly this: Residual Value Swaps (RVS) for GPUs. Companies buying or leasing GPUs can now use RVS to transfer risk and secure funding on better terms.
Beyond this, things like GPU-based Asset-Backed Securities become possible. Ornn already works with USD.ai on exactly this kind of structure: USD.ai provides the loan and financing for GPU infrastructure, while Ornn reduces financing costs by offering a put option on the hardware in year 3 or 4. This gives lenders confidence in the terminal value, enabling better terms for borrowers.
It's all one story: once a price benchmark that market participants can agree on emerges, risk becomes measurable, and capital can be deployed more comfortably.
But the path to this future obviously isn't easy. The biggest challenge is creating a proper index. Without a well-constructed index, hedging won't work perfectly, and if market participants don't trust it, the index becomes just a number with no greater value. The problem is GPUs may be even trickier than electricity - beyond GPU model and region, performance varies based on workload type, setup configuration, and more. Properly measuring these factors and creating a reasonable index seems critical.
Beyond this, there are liquidity and regulatory issues, but if the above risk is solved, the market need appears so strong that these should resolve quickly.
I had the chance to speak with the Ornn team and asked them some questions about how OCPI works under the hood and where they're seeing traction:
Q. What's the minimum transaction volume needed for a price point to be included in OCPI?
A. There's no minimum transaction size, the index is volume-weighted, so larger transactions naturally carry more weight.
Q. How do you handle the heterogeneity problem, same GPU model but different performance based on workload type, network topology, or cooling setup?
A. Through weighting. Different configurations are weighted appropriately to reflect their relative performance characteristics.
Q. What's your approach to regional weighting? How granular does geography get?
A. City-specific. Prices can vary significantly by location, so the index accounts for geography at that level of granularity.
Q. How many data partners contribute to OCPI?
A. Over 10 data partners currently contribute to the index.
Q. How do you verify that reported "executed prices" are real and not manipulated?
A. We have contracts with our data providers and actually parse their invoices directly. This ensures we're working with verified transaction data rather than self-reported figures.
Q. Which customer segment is moving fastest: AI companies, data center operators, or financial institutions?
A. For the hedging products, it's operators. For the Architect partnership, we expect it will attract more speculators looking to take positions on compute prices.
Compute, which may become the most important commodity going forward, deserves a real market. Just as oil has WTI and electricity has LMP, compute needs a price everyone can reference. Ornn is building the index (OCPI) that will become that market's foundation, and derivatives that let market participants trade risk based on that index.
Until now, buying compute has been like buying a used car. You can't tell if the seller's price is fair, comparing to the shop next door is difficult, and knowing what prices will be in six months is even harder. The more widely Ornn's products are used, the more transparent prices become, the more efficient capital deployment becomes, and ultimately, the lower compute costs will be. This is good for companies building AI and everyone using AI.
On a separate note, there's one more thing I'm excited about. In a past piece, I wrote about how we need to expand supply of critical infrastructure to lower costs, and redesign that expansion as investment opportunities accessible to younger generations. As infrastructure like Ornn matures, exactly that future becomes possible. Of course, data center REITs and infrastructure funds already exist, but those provide indirect exposure to real estate or company valuations. Products like GPU price-tracking ETFs, yield products linked to data center cash flows, or tokenized compute exposure would provide pure exposure to compute prices themselves. Institutions will be the main customers for now, but if the day comes when anyone can bet on the 21st century's most important commodity - that would be quite an interesting world, wouldn't it?
Moyed
Support dialog
Comments