The Spark-Spread, Market Completeness, & DeAI

A Compute-Spread Cometh

Almost two years ago, I wrote about the technical possibility of the decentralized training of AI and its existential importance: super-intelligence and capital was heavily centralizing into a few firms, in a few (two) countries, concentrating immense levels of power into a very small slice of humanity, making the entire world highly vulnerable to their preferences and machinations. A competent alternative to this reality might be a necessity to avoid dystopic outcomes.

Since my last writing, decentralized training has improved by leaps and bounds. The earliest training runs were starting at 10 billion parameters in late 2024. They’ve since scaled to 100 billion-parameter runs in 2026.

These are not SOTA models that compete with the labs, yet, but frontier labs might be crossing into issues the decentralized players won’t have, like infrastructure constraints and extreme political attention. And many users and firms are coming to the realization that most tasks do not need the highest levels of intelligence available. You don’t need to pay a 160 IQ genius to draft a standard legal contract or create a simple financial model for a public company; 120 is probably fine. And to draft a basic email 100-110 is probably fine too. For many tasks in the economy, good enough is good enough, and there is no need to overpay. Paying for a new sledgehammer when your old hammer works fine for you task is asinine.

Despite the large obstacle of communication overhead of decentralized AI, which is very real and daunting, there is no reason to believe that it’s absolutely impossible for these more distributed methods to meet (more likely) or even exceed (less likely) the frontier labs at some point in the future. It is a difficult problem and will take some time, maybe a long time, but incumbents have a way of getting locked into a certain way of doing something (rational to do so) while the periphery has the freedom to explore the entire possibility space, in order to find a brilliant way to do something better at a better price point. In the long run, I wouldn’t bet against the exploration function of an open market.

In addition to the technical progress I mentioned previously, there are financial forces at work that make the possibility of a more distributed compute landscape highly probable.

In my last post, I drew some lessons from the natural gas pipeline buildout that were germane to some of our challenges and opportunities today. As a consequence of my deep dive, the development of natural gas financial instruments also came into view as a rich reservoir of experience we can pull from to glean what the future might entail for compute.

Spark Spread

Before natural gas financial markets existed, natural gas was dominated by vertically integrated supermajors and utilities. In compute world speak these were “hyperscalers”. Without a financial market to price and manage risk, owning the source of risk was the best option available. Can’t hedge gas exposure? Buy the pipeline. Can’t hedge electricity revenue? Own the utility. Vertical integration is what fills the void left by absent financial markets.

The emergence of the spark spread, the difference between locking in natural gas costs and electricity revenue, weakened the necessity for this vertical integration. When gas and power prices became simultaneously hedge-able, a developer could lock in a reasonable future profit, take it to a lender, and finance a power plant without owning anything else. The independent power producer (IPP) was born. Between 2000 and early 2010s, IPPs armed with hedge books and project finance loans were the main builders of over 200 GWs of new U.S. natural gas-fired generation. Like most early markets, there was a quick boom and bust cycle, but nonetheless these types of smaller players were the major catalyst for the buildout. The pretenders and unlucky got terfed out and the strong and lucky continue on to this day.

A similar story happened upstream. Small shale producers used Henry Hub futures as collateral for reserve-based loans. The Permian Basin transformation, the most consequential energy development of the last 30 years, was financed by an instrument that didn’t exist in 1985. New financial instruments don’t just help existing businesses manage risk, they help create businesses that couldn’t otherwise exist.

Economists would say that the emergence of these types of financial tools moves a market closer to Arrow-Debreu completeness: a world where every contingency has a trade-able instrument and every risk has a price. Full completeness is an idealized state that never fully arrives but the market becomes more complete when more states of the world are priced and trade-able. As financial markets move toward greater completeness, the logic for vertical integration (owning the whole stack) weakens. Smaller, independent specialists emerge. In a similar manner, the compute market is about to become more “complete”.

The Compute Spread Is Coming

The spread between electricity costs and compute revenue is almost the entirety of a compute operator’s economic position. Until recently, compute was mostly un-hedgeable unless you had the size and credibility for a major financial institution to create a bilateral OTC hedge on your behalf.

The company Ornn has built a benchmark index for GPU compute prices that the Intercontinental Exchange (ICE)--an exchange and clearing house where most of the world’s oil, European natural gas, and US electricity derivatives are traded--is preparing to list GPU compute futures against it. Likewise, the CME group has announced plans to launch their own compute futures market with another benchmark provider, Silicon Data.

Natural Sellers: Compute operators locking in revenue to service debt. 

Natural Buyers: AI native operations like AI wrapper companies and model fine tuners 
whose operations carry compute cost exposure the way a gas-fired power plant carries 
fuel exposure. 

And as always, the traders and speculators are in the market straddling both the natural 
buyers and sellers, making the market.

With an emerging compute spread, capital allocators will be much more willing to underwrite against this market. A data center operator can sell compute forward, use the forward contract as collateral, and get construction financing without a hyperscaler as anchor tenant. Reserve-based lending, applied to GPU-hours instead of proven oil or gas reserves.

For those getting the impression that this is all imminent, I’lll have to throw a little cold water on the excitement. There are still technical issues around the fungibility and verification of compute that need to be adjudicated before the possibilities I describe can manifest. My bet is there are large enough incentives for the general market to solve these difficult but not impossible problems.

The input side, electricity, already has hedging mechanisms, but they could be developed further. Medium-to-large operators and massive hyperscalers lock in power through Power Purchase Agreements, bilateral OTC swaps, and Nodal Exchange, which clears locational power futures at hundreds of hubs, zones, and nodes across every ISO. But Nodal lists the liquid head of the curve: the major locations, in monthly blocks, for participants big enough to clear and post margin. The thousands of individual nodes beneath that, the hour-to-hour shape a battery actually lives on, and the sub-50MW operator who can't justify the clearing overhead. None of them get served. Outside of the major nodes, zones, hubs, and participant sizes, these financial products stop making economic sense: minimum deal sizes, negotiation overhead, and credit requirements price out anyone who isn’t already large.

My prediction is this service gap won’t persist indefinitely. Compute hot spots like West Texas, Northern Virginia, Central Ohio, the Phoenix corridor, could conceivably concentrate enough small, mid, and large operators to generate enough aggregate electricity generation and demand locally to make even more localized and granular hedging products highly desirable and feasible. That economic concentration, combined with the low-cost financial infrastructure being built in DeFi and prediction markets, makes the creation of local electricity instruments around these discrete nodes more economically rational.

Greater Heterogeneity

The end state for compute is not pure decentralization nor the continuation of hyperscaler oligopoly. It’s what happened to every other open market that matured: a heterogeneous ecosystem where every scale finds its proper niche.

ExxonMobil didn’t disappear after gas deregulation. But Pioneer, Devon, and EOG built enormous businesses. Tens of thousands of small independents drilled wells the majors wouldn’t touch. Heterogeneity finds things the big guys miss or is a better fit to service.

Compute is poised to go in the same direction. Hyperscalers and frontier labs retain real advantages but they’ll lose their market share dominance while market entrants expand and the market itself continues growing.

What fills in around them: independent compute operators (ICPs) financed against hedged spread positions, running batch inference and fine-tuning at competitive prices. Decentralized training collectives like Bittensor, Prime Intellect, Pluralis, etc aggregating distributed hardware from ICPs, optimizing and experimenting with things the general labs aren’t.

And at the far edge of the compute spectrum, individual participants will run consumer and even advanced GPUs, contributing to all sorts of computing use cases, especially inference.

Virtual Compute Plants

Virtual Power Plants (VPPs) have been an interesting topic of discussion for years in energy circles. With the success of companies like Base Power, and others, VPPs have moved out of pure theory into practical implementation and scaling. Having the home become a node in a distributed power plant is cool but lurking beneath the surface is another opportunity that might be even bigger.

If a home can host a battery, it can also host a GPU.

That’s exactly the bet NVIDIA, Span, and PulteGroup are making. Span (the smart panel company), NVIDIA, and PulteGroup (major US homebuilder) launched XFRA, attempting to create a distributed cloud out of aggregating retail edge compute. A “Virtual Compute Plant” if you will.

This is a small pilot right now, 100 homes, but the CapEx outlay for deployment at a scale that matters is risky and expensive. A highly liquid compute futures market would be a huge help in financing and de-risking this type of effort. Equipping 100 homes vs 100,000 homes is a totally different animal. So, not only do more Independent Compute Producers (ICPs) become viable but aggregating the compute from individual contributors becomes much more viable thanks to the development of liquid compute futures.

A Virtuous Cycle

The tight coupling of compute and energy, increases the feasibility of the smaller end of the compute market which in turn affects the energy and power markets as well. A 1 MW solar installation with a battery, a micro-turbine on stranded wellhead gas in a county with no pipeline takeaway, and the home battery market all have another source of demand. The increased viability of ICPs and VCPs increase the viability of more IPPs and VPPs. This in turn creates cheaper and more plentiful opportunities for ICPs. Decentralizing compute also helps decentralize our energy and power markets and vice versa.

A virtuous cycle is poised to be unlocked and play itself out. The development of a Compute Spread might be the key that does it.