AI compute demand is clear, but the race is now about speed to power.
At the recent Maxim Data Center Summit, the Bitcoin-to-AI pivot made this clear: the market is rewarding operators with existing power and infrastructure who can bring compute capacity online fastest.
A few themes stood out across the day.
- The Bitcoin-to-AI pivot: Former Bitcoin mining sites are becoming real AI infrastructure opportunities.
- Development starts with power: Power access remains the foundational constraint for data center development.
- GPU economics are more resilient than expected: Existing GPU capacity is holding value better than many anticipated.
- There is no single way to develop data centers: Business models and financing structures are evolving across powered shell, GPU cloud, and compute financing.
- Speed to market is the new moat: Faster deployment timelines are becoming a major competitive advantage.
- Smaller cities and non-hyperscale customers are underserved: The market between roughly 1 MW and 20 MW remains a meaningful gap.
For Comity, the summit was especially relevant because many of the challenges discussed sit directly in areas where we spend time: underwriting power and infrastructure opportunities, understanding the value of underlying assets, supporting creative financing structures, and helping transactions move faster from need to execution.
1. The Bitcoin-to-AI Pivot
One of the conference’s dominant themes was the migration of former Bitcoin mining infrastructure into AI and high-performance computing (HPC) data center use cases. Several operators are repurposing existing mining sites that already have contracted power, land, substations, and utility relationships in place.
These sites often already control the hardest-to-acquire assets: power, land, interconnection, and utility relationships. In some cases, existing power infrastructure can shorten build timelines materially, with certain projects targeting 10 to 12 month development windows compared with the more typical 24 month timeline for larger greenfield projects.
Some former mining sites have access to more than 1 GW of approved grid capacity, with hundreds of MW already committed to AI compute customers across phased deployments. This creates a real opportunity for operators that can convert power-heavy mining assets into AI-ready infrastructure.
2. Data Center Development Strategy Starts with Power
Every company at the conference treated power access as the foundational constraint. A major theme was the move toward power-first development, including behind-the-meter sites, stranded or underutilized grid capacity, substation-adjacent locations, and smaller markets where available infrastructure may be overlooked by larger players.
Some strategies focused on colocating near existing generation, including large sites adjacent to more than 1.5 GW of combined generating capacity. Others focused on sub-50 MW colocation opportunities with available grid power, deliberately avoiding the multi-year queues associated with 100 to 300 MW hyperscale projects.
3. GPU Economics Are More Resilient Than Expected
GPU economics appear more favorable than many expected. While hardware depreciation remains a key risk, demand for existing GPU capacity has remained strong, and certain prior-generation chips continue to be commercially relevant.
The key point is that GPU financing cannot be evaluated only through a traditional hardware depreciation lens. The value of the equipment depends on customer demand, workload fit, utilization, power cost, and the broader compute market. That creates complexity, but also opportunity for parties that understand the underlying asset value.
Older Hopper-class chips are still seeing demand because many workloads do not always require the newest silicon. Model architecture, inference requirements, and dollar-per-token economics can make prior-generation hardware attractive. There are also examples of older GPU fleets being sold at strong residual values several years after purchase, with proceeds reinvested into newer hardware.
4. There Is No Single Playbook for Data Center Development
The conference showcased a range of business models across the stack:
- HPC colocation, powered shell: approximately $10 to 15M/MW of CapEx, approximately $1.8M/MW/year of revenue, 10 to 20 year leases, and 95% to 99% EBITDA margins
- GPU cloud: approximately $40M/MW of total CapEx, approximately $30M/MW of revenue, higher margins, but shorter 1 to 5 year contracts
- Token-based compute financing: some companies are experimenting with tokens redeemable for bare metal GPU capacity, creating a form of non-dilutive financing
The powered shell model is about leasing the physical data center infrastructure, including power, cooling, and the building, while the customer often brings the compute hardware. GPU cloud is more vertically integrated, with the operator owning the facility and the GPUs, then renting compute capacity to customers. Token-based compute financing is a model where tokens act like prepaid GPU credits, making it a non-dilutive complement to traditional debt financing.
GPU hardware financing alone can be roughly 5x the data center facility cost by dollar value. One MW of data center capacity may cost around $12M, while filling that same MW with hardware can cost $50 to 60M.
As these deals become more complex, financing solutions need to evolve with them. The right structure may not always look like traditional project finance. It may instead be built around power contracts, equipment, customer offtake, collateral, or other project-level assets, with capital providers taking a more flexible view of underlying asset value.
5. Speed to Market Is the New Moat
Across the conference, speed of deployment came up as a major differentiator. The strongest-positioned projects tend to have some combination of existing power infrastructure, modular construction, smaller deployment tranches, and early access to long-lead equipment.
This matters because customer demand for compute is moving faster than traditional infrastructure development timelines. Large-scale projects can still take years, while many customers need capacity much sooner. As a result, smaller, faster, and more flexible deployments are becoming increasingly attractive.
Several strategies stood out: using pre-existing power infrastructure, deploying modular or prefabricated designs, targeting smaller 2 to 20 MW tranches, and securing long-lead equipment early enough to keep construction timelines on track.
6. The Market Gap: Smaller Cities and Non-Hyperscale Customers
A consistent thread across multiple discussions was that the market between roughly 1 MW and 20 MW remains underserved. Hyperscalers are generally focused on larger deployments, while many legacy data centers are not well equipped to handle AI workloads.
This leaves a meaningful gap for smaller customers, regional markets, and distributed inference use cases. Hospitals, schools, local governments, enterprises, and regional compute providers may not need hyperscale campuses, but they still need access to reliable AI-ready infrastructure.
There is also growing opportunity outside the largest US markets, including international markets such as Brazil and India, where regional compute providers may lack the structured finance, hardware procurement, and capital access available to larger platforms.
The opportunity is not only in the largest data center campuses. It is also in the smaller, more complex, and faster-moving projects where power, capital, collateral, and demand need to be brought together efficiently.
