How to Invest in GPUs
How I learned to stop worrying and love high tech equipment rental.
Critical tasks across defense, cybersecurity, and medicine once deemed too sensitive for robots are now too sensitive for humans. Self-driving cars move passengers through our cities with remarkably better safety records for passengers, motorists, and pedestrians. Platforms like ChatGPT and agents like Manus offer everyone with an internet connection a designer, researcher, content writer, quantitative analyst, and (depending on their level of emotional intelligence) confidant. Thanks to exponential improvements in model quality, the same GPUs that produced low quality surrealisms three years ago produce content and audio-video today sometimes indistinguishable from reality (prompting some serious conversations with my older relatives).
2023
2025 (and they have five fingers now!)
We clearly live in a different world today than we did three years ago, and AI is today in its lifecycle where the mobile internet was in 2002, explaining global governments’ mad dash to offer the power, precious natural resources, real estate, and incentives needed to remain globally relevant in the AI era.
This revolution has a voracious appetite for “bare metal” compute, or blank-slate computational resources without the expensive software add-ons and nightmarish contracts of public cloud companies. Bare metal has already produced a crop of unicorns like CoreWeave and Crusoe Energy, which built themselves off a technologically complex but ultimately simple business: buy and rent out GPUs to customers who don’t want to own them.
Their business models and internal asset-level returns on capital are being achieved in parallel by a growing crop of independent cluster owners. Instead of buying in at a multiple and trusting management to grow discounted free cash flow, investors wanting direct exposure over equity appreciation are making off with 30%+ levered IRRs, 1.5x+ DPI, <3 year paybacks, and substantial depreciation tax offsets (100% bonus appreciation today - thanks BBB) by investing in the assets directly.
How is this accomplished?
GPU Economics: The Simplified Rundown
A latest gen deployment with mid-teens 70% LTV financing, mid-priced colocation, and cheap power might look like this:
…with the following underlying lifecycle:
Revenue
Let’s talk about the actual source of returns first.
Rental Rates
Your core economic unit is the GPU/hour, or the cost to rent a single GPU for one hour. This hour is spent conducting computational work on behalf of a customer. The underlying unit of computational work is the token and different GPUs produce different amounts of tokens, and thus different levels of economic value, per hour.
GPUs are typically rented out at the server level, with servers primarily comprising 8 GPUs. If a B200 GPU is paying you $4.00/GPU/hour, expect to make $32.00/server/hour.
Customers
Unless you have a AAA anchor customer with a quality long-term contract, seek out high levels of end customer-level diversification so you can bet on the AI demand trend, not on one company’s specific outcome.
AI is slashing the marginal cost of software to that of the compute resources required to generate it, pushing more clouds, neoclouds, and AI platform-as-a-service companies onto rented bare metal. Survival requires neoclouds (software companies at their core) to use their runway for R&D and marketing while avoiding long-term fixed payment obligations, like debt or lease payments on GPUs. It’s only when a cloud has reached a massive scale that it can afford to own GPUs to serve a stabilized customer-base.
Term Structure
You can rent out your GPUs for hours, days, weeks, months, or years.
Shorter term contracts (<1 year) are more lucrative if you have the relationships or offtake partners to get them consistently (easier said than done) and are the way to go in general. Shorter term contracts are the deepest part of the market, and cracking short-term contracts with minimal downtime in between is your recipe for sustainable returns.
Longer term contracts are less lucrative but more certain and do provide cash upfront. However, they are less common, and you should ask yourself how a first time cluster operator got so lucky in the first place. Waiting around for a long-term contract is a bad idea: you will get left behind.
If you secure a long-term contract before you have equipment, there is a reason they are choosing you over someone who is ready to go today. Unless you’re CoreWeave and have already established yourself as a satellite of a trillion-dollar public company, those reasons are:
1. Your customer may be a less reliable end credit with uncertain final monetization (no customers).
2. Your customer is getting a lowball price from you that nobody else would accept because they know you’re an amateur who they are putting into business.
Costs
Colocation
Every GPU deserves a loving home. You’re going to put your servers into a colocation facility, where you will need to ensure you’re paying less than $200/kW/per month and aiming for <10 cents per kilowatt-hour of electricity.
Colocation space is typically billed in terms of the electrical consumption of the underlying assets, and not square footage. “kW per month” means that, if you had one B200 server rated at 12-14 kW all-in per month, you’d be paying $1,400 of rent for a space that billed $100/kW. It’s like if a parking garage billed you per horsepower.
Some colocation centers have “all-in” pricing including metered electricity. Others charge for space and electricity usage. Vetting a colocation space and what is included in your monthly rent is an article unto itself.
Risks
Risk 1: Rental Decay Rates and Obsolescence
Most people who are new to the asset class ask “what if there’s a new, even faster GPU coming out?.” There is one coming out. There are several, and this multi-billion dollar industry knows how to soundly forecast the impact of supply, demand, and innovation on the foundational economic unit by which it lives and dies across obsolescence cycles.
The rental rate of a given GPU model will go down over time as NVIDIA and, to a lesser but growing extent, AMD and Huawei, continue their planned rollouts. A conservative investor should assume new models are coming out sooner and with better capabilities than assumed by consulting groups like SemiAnalysis, who forecast for hedge funds and hyperscalers with a remarkable sub-4% margin of error.
While you can achieve this level of accuracy scientifically by weighting the total cost of server ownership per GPU/hour (TCO) against the market cost-per-token scaled for tokens per GPU/hour based on forecasted supply and demand, you can do just as well by sensitizing your model to make sure your economics work when decaying your rental rate by 15-25%.
Obsolescence/revenue decay is the foundational risk of this asset class and, accordingly, the best researched.
With other risks boxed out or accepted, disruption is a valid and natural risk to continually revisit. The most common question I get is: what about DeepSeek?
My partners make millions of dollars running DeepSeek inference for consumers. DeepSeek is a distillation (in essence a copy) of existing models that required tens of millions of dollars of compute to train. First, DeepSeek had to generate synthetic data by spending hundreds of millions of dollars running every conceivable prompt through existing models ahead of its final, much cheaper training run. Per Rand’s Lennart Heim,
DeepSeek operated Asia's first 10,000 Nvidia A100 cluster, reportedly maintains 50,000 “Hoppers” (which could be Nvidia's H100, H800, or H20), and has additional unlimited access to Chinese and foreign cloud providers... This extensive compute access was likely crucial for developing their efficiency techniques through trial and error and for serving their models to customers.
Not only did DeepSeek’s final training run ride the coattails of substantially more compute-intensive models, but model training enables the real value and source of utilization: inference compute.
Risk 2: Depreciation and Liquidation
Public names like CoreWeave and Amazon tend to depreciate their servers over 5-7 year terms. I depreciate them conservatively over 4-year terms as the economics still work great on this schedule and I prefer to sandbag expectations than to defend them. Groups exist to guarantee residuals in this range (I work with a few of them) to secure a liquidation value at a known price ahead of acquisition.
Risk 3: Utilization
Utilization risk is for operators without a distribution plan and/or without a cluster spec that people want. Partner with a high utilization offtake partner to spec and distribute your build and sleep well at night. I work for a company that delivers >90% utilization on cards that are nearing the ends of their relevant lives and even greater for flagship cards. It can be done.
Transaction Structure and Financing Strategy
What to Buy
The best economics are achieved on GPUs that are new enough to see high adoption but have been on the market just long enough for the developer ecosystem to adjust to them. How you have configured your GPUs is just as important as the GPUs you’ve purchased.
How Long to Hold
I would not want to own a GPU greater than four years old, and I target a 3 year round trip, 3 year financing, and 3 year guaranteed residual value. There will come a point in every model’s life where your colocation cost/kW is less than your revenue/kW or newer models present a material opportunity cost. Don’t be drawn in by the siren song of high unit-level margins post-loan payoff and wake up with underwater business-level margins.
Here's an analogy: if I have 24 seats of prime restaurant space in Chelsea buying 3.5 drinks per hour, should I sell Miller High Life for a 1000% margin of $3.60 or wine for a 300% margin at $20 per glass?
Larger tech companies typically sell between the three year mark (where bare metal economics start making less sense with traditional setups) and the 7 year mark (where they no longer can depreciate them). An entire IT asset disposition (ITAD) ecosystem exists, which hoovers up V100s and A100s before they can hit the market and resells them to groups with an economic life-extending opex or software advantage.
A GPU is like a Mercedes S Class: its first and second owners are very different kinds of entrepreneurs (albeit both would be in high-margin commodities businesses).
Capex Management
There are many wrinkles in GPU financings, and I work with a small, select group of lenders and lessors who get it. Some quick capex management tips:
Lease Pricing
Consider the full cost of capital for a lease. You won’t own the asset at the end. You’ll typically only have the option to purchase it after effectively paying the full value, plus interest, already. If there’s a ramp-up period in your monetization, that money is out the door: no equity has been built.
For example: even if a 3 year lease charges an initially attractive, say, 8% yearly cash pay on capital provided compared to a 15% loan on the same term. A loan gives you the full asset. This means your actual cost of capital is astronomically higher considering you’re likely paying for 100% of a 7 year asset but only getting 3/7ths of the value (gross of any non-linearity in the depreciation curve).
Loans
I love loans, and not just because I used to originate them. You own the asset, you capture the depreciation, and, depending on your structure, you can accelerate payment to lower your interest obligations and restructure to lower fixed costs.
Acceleration is key. Years 1 and 2 are your most lucrative and a great opportunity to use your levered returns to deleverage and de-risk while margins are at their most forgiving.
Besides, If I’m charging my LPs a 10% pref but paying >10% interest, I should accelerate my payments anyway to return more cash.
I have designed a loan achieving all of the above and work closely with an originator who provides them.
Barriers to entry
Why isn’t every opportunistic investor and PE group doing this?
Short answer: if they aren’t one year from now, I’ll consider this a personal failure.
Longer answer – all you have to do is:
Design a cluster with the right switches and overall configuration and don’t overpay for it
Secure a financing
Ensure the financing structure makes sense
Build a sales team to monetize your equipment and support customers
Find a software that enables you to provision your compute and handle billing
Find a data center operator/colocation center that knows what they’re doing and doesn’t fleece you
Ensure consistent utilization
Eventually sell your assets for what you forecasted they would be worth
It’s like mining but harder. All of the above are a full-time job and expertise is scarce. After investing in GPUs as a private lender and designing structures for them in my capacity at Hydra Host, I’ve spent the past 6 months pre-packaging the above for family offices, neoclouds, and even colo operators looking to compete with their customers to get investors up and running in 2 months, not 12.
I have lots more to share and a financial model that breaks this all down month by month and across financing structures, hold periods, and GPU models.
If you’ve made it this far, let’s chat.