compute capital markets
lately, there's been a lot of chatter, including from sam altman, around how compute will be metered as utilities. another analogy people like to make is comparing compute to commodities like oil. from some dimensions, the analogies make sense. like utilities, delivering intelligence requires massive up-front capital expenditure. like commodities, ai compute is becoming a universal input to a vast range of products and services. however, there are a few key differences. one, the foundation labs are seeking full verticalization: they want to produce the commodity product and own the distribution to end customers. two, unlike utilities, the value of compute is being driven by fundamental increases in capabilities much faster than other utilities (i.e, the models themselves are getting smarter and smarter), rather than just being driven by policy or market dynamics. third, unlike commodities, compute generates ongoing cashflows and is consumed directly by the end customer.
the purpose of capital markets is to allocate funds to economically productive activity. this means that the structure of capital markets must match the needs and characteristics of the economic activity. i believe that because of the uniqueness of how compute is created and delivered, there will be a build out of new financial infrastructure in parallel with the delivery of physical ai infrastructure. we need to create new financial rails for datacenter project finance and to transform compute into a investible asset class.
the complexities of datacenter project finance
datacenter project finance has a complex financing structure and balance sheet. first, you have equity. this part of the stack is straighforward; you typically have the primary stakeholder (like the hyperscaler or the foundation lab) and some massive infrastructure private equity firm who focuses on commercial real estate like brookfield. they hold the most risk but also the greatest return. after equity, you have your debt investors. you typically will have multiple debt investors, with senior and mezzanine tranches. senior debt will typically have a construction loan to finance the build out and a term loan(s). historically, in project finance, debt is usually syndicated through a bank, meaning it can either be sold directly to various institutional investors or structured into a commercial mortgage backed security. after your senior debt, you have mezzanine debt to fill the gap between equity and the senior debt to make sure the entire project is capitalized. mezzanine debt is subordinate to senior debt, which means that the senior debt gets paid out from the datacenter income first and that mezzanine debt is riskier. mezzanine lenders are typically firms with higher risk tolerance than banks, like private credit asset managers. i think with the growth of private credit and increasing capital requirements for banks, private credit investors will increasingly start trying to fund senior debt more and more.
now if that seems mildly complicated, just know that there are a lots of things i left out: typically the senior and mezzanine debt will have multiple loans and funding vehicles, there could be a rating agency involved if cmbs or an insurance asset manager is providing some funding. there are complex terms that govern the relationship between senior lenders, mezzanine lenders, and equity holders, and sometimes incentives aren't aligned. a lot of this looks like your standard cre financing structure. inevitably, constructing a datacenter is even more complicated due to two major characteristics: the power and other utilities to run the datacenter and the actual compute infrastructure like chips, server racks, networking, and cooling.
because these datacenters are 1) built in more remote locations and 2) extremely energy intensive, it means that you need to build them "batteries included." you need to add new energy production for the datacenter, and if you are integrating with the broader power grid, you need to account for interconnection costs and load balancing. beyond just a standard cre financing stack, you need to account for power equipment and tools like batteries and transformers in your construction and maintenance cost. you need to account for fluctuating energy prices and demands, including weather and seasonality, which impacts how you underwrite the debt since it impacts cashflow and margins.
you also have massive amounts of actual compute infrastructure and equipment that you need to procure. this compute infrastructure is just as complex as the power infrastructure, and you need to match your compute capabilities with the offtake agreements the debt is being underwritten against. with supply limitations on different parts of the compute infrastructure, such as gpus, liquid cooling, and memory chips, you need to secure agreements for these assets months in advance from chip manufacturers and other firms involved in the supply chain.
this introduces another part of the credit stack: equipment loans and leases used to procure all of the physical stuff you need to actually run the datacenter. this is where the project finance stack draws from asset-backed finance, where the receivables of loans and leases can be used as collateral to borrow money at a lower cost of capital. the equipment itself can also be used as collateral to lower the cost of funds because if the borrower defaults, the lender can claim the equipment and sell it for recoveries (though there is varying depreciation and homogeneity of the equipment, which affects recovery rates).
alright, with gross simplification, that is the basics of financing a datacenter. so what? why does that matter?
the reason it matters is because there is very little financial infrastructure developed to support the complexity of these transactions, which means there is a massive bottleneck in how efficiently capital can be deployed. in order to develop a mature, new asset class, the asset class needs to be legible to a broad set of investors so that they can adequately reason about the risk-return profile. how does an investor actually evaluate that a datacenter is the best way to generate a return compared to other investment opportunities, beyond just the narrative and hype being thrown behind artificial intelligence?
the answer is reliable, transparent data and the requisite tools to analyze the data. in order to scale datacenter financing, the market needs to understand all the documents and underwriting information. the market needs to have ongoing, high quality information about how the construction progress is going during build out and how the datacenter is performing once it is live. investors need to project cashflows based off performance assumptions for the datacenter: as input costs change and market demand changes, how will the revenue profile of the datacenter look, and how does that impact the flow of funds to different debt and equity investors? these are the questions investors need to be able to answer in order to decide that a datacenter is a better use of capital than another asset, and we are in the very early stages of them being able to answer those questions reliably and repeatably.
compute as an investable asset class
for the most part, i think the evolution of project finance when it comes to datacenters is fairly obvious if you understand how new asset classes emerge. however, that is just one part of capital markets evolutions that i think will happen with compute. the second is the transformation of compute into a investable fixed-income product at scale. this one is particularly interesting to me because of my experience in structured products and asset backed finance. above, i talk about how markets will need clean and standardized data around datacenter performance in order to understand the return on various investments. this data will become the foundation of securitized compute cashflows, creating a new funding vehicle for the firms involved in these projects.
securitization is a form of financial engineering that takes some set of cashflows and transforms it into tradable asset with various risk return profiles. in the context of mortgages, a mortgage backed security takes the principal and interest payments of the loans (cashflows) and splits them up into different bonds with different coupons and risk profiles (the tradable asset). investors can then buy different notes at different prices. in a mortgage backed security, you will have an AAA bond; this bond gets paid first from the collection of principal and interest payments, but at a lower interest rate. then you'll have the AA bond, which will get paid out second. if people stop making payments on their mortgages, the AA bond will be the first to lose losses. there will be a cascading waterfall of bonds, with different coupons and risk profiles.
conceptually, you could securitize any set of cashflows. most securitization happens with some sort of loan: mortgages, corporate loans, consumer loans, etc. in the project finance section, i talked about equipment loans and leases; you can securitize those payments too. i've seen some more novel stuff lately; a couple of years ago, i bought some shares in product that securitized royalties from the shrek soundtrack (that security was eventually decommisioned and i haven't seen anything like it since, likely because there wasn't enough investor demand).
i believe that the massive cashflows that datacenters generate will securitized into tradable assets and create a more efficient way for firms building them to fund their businesses. i also believe that securitized compute cashflows will be an incredibly attractive asset to fixed income investors because it would be one of the few fungible assets they will be able to invest in to capture returns from the ai infrastructure build out.
as big as the project finance build out is, the capital structure of each is so unique, i think the secondary market for the debt will lack liquidity. i don't think securitized compute will have that problem for a few reasons. one, for the most part, compute in one datacenter is swappable for compute in another datacenter, especially for inference and cpu's. that means the underlying asset is more commodified. second, the data to track the performance of a datacenter once it is more comparable: overall capacity, utilization, reliability, power consumption and costs, etc. when you compare one datacenter to another, you can compare these metrics. you can also directly correlate the cashflows of the datacenter to these metrics, allowing you to, with enough data, build better assumptions on how datacenter performance impacts the cashflows it generates.
in order to support securitized compute, we need to build a lot of financial infrastructure. first, like i highlighted in the project finance section, you need reliable and clean data around datacenter performance and operations. then, you need someone to do the financial engineering: you need to actually defined the notes that will be issued and create the rules of how cashflows will be distributed. you need trustees to administer the transaction: collect payments and distributed them. you need ratings agencies to rate the transactions so that institutional grade investors can buy the bonds you issue. you need to build trade settlement and accounting systems for these new assets. all that's to say, in order to offer securitized compute products to help datacenter builders finance their projects, you need to build a lot of stuff that is technically and financially complex.
conclusion
the current ai infrastructure buildout will require new innovation in financial services. first, we need to build the data infrastructure, tools, and systems that enable more efficient capital allocation for project finance. markets need to build workflows for repeatable investments and have high quality data for decision making. then, i believe compute can be transformed into a securitized product, allowing compute producers to access more efficient debt capital by offering bonds paid out by the cashflows from their compute. these are two structural evolutions i believe need to happen in order to enable the ongoing ai infrastructure build out. there is a massive opportunity to build out (the likely ai native) tools for datacenters to access more capital.