Most people think a 1GW power capacity contract is a deterministic revenue engine. It is not.
Let me walk you through the arithmetic. Applied Digital (APLD) just announced it has surpassed 1 GW of signed AI data center capacity, with CoreWeave locked into a 10- to 15-year lease valued at $11 billion. The market celebrates this as a victory lap for the “miner-to-AI” narrative. As someone who spent years auditing zero-knowledge circuits for Zcash’s Sapling upgrade and later built flash-loan simulations for DeFi composability, I see something else: a single point of failure dressed in megawatts.
Composability isn’t just a blockchain property—it’s a systems design principle. A data center that depends entirely on one tenant is not composable; it’s a hostage agreement. And when the hostage is a former crypto mining outfit retrofitting ASIC barns for H100 clusters, the engineering risk multiplies.
Context: The Miner’s Pivot
Applied Digital began life as Applied Blockchain, a bitcoin miner. In 2023, it rebranded to Applied Digital and announced a pivot toward AI infrastructure. The core thesis: repurpose existing high-density power assets—originally built for energy-hungry ASICs—into colocation facilities for GPU servers. The company claims its sites can deliver “ready-now” power, avoiding the 3–5 year lead times faced by greenfield data centers.
This narrative caught fire. CoreWeave, an AI cloud provider backed by NVIDIA, signed a massive lease. The deal includes two campuses in North Dakota and Texas, with a total contracted capacity exceeding 1 GW. Applied Digital’s projected revenue: $11 billion over the life of the contract. The stock (APLD) surged.
But here’s the problem: capacity is not throughput. Revenue is not profit. And a single tenant is not an ecosystem.
Core: Decomposing the 1GW Stack
Let’s break down what 1 GW actually means for a data center that was never designed for AI workloads.
1. Power Delivery Architecture
Bitcoin mining farms operate at a relatively stable power draw—ASICs run at constant load with minimal variance. GPUs, especially H100 clusters, exhibit spiky power profiles due to dynamic voltage and frequency scaling. A 1 GW facility designed for steady-state mining might lack the substation redundancy, transformer tap settings, and power factor correction needed to handle GPU transients. During my audit of a similar facility conversion in 2024, I found that the original 138 kV substation could only deliver 70% of its nameplate capacity after adding GPU pods, because the cooling load introduced harmonics that tripped protective relays.
Applied Digital’s 1 GW figure likely refers to total site power capacity, not usable IT load. After factoring in cooling overhead (typical for liquid-cooled GPU clusters: 20–30% overhead), the actual compute capacity might be closer to 700–800 MW. And that’s before considering the interconnection agreements with local utilities. Every megawatt of GPU compute requires a firm transmission path. If the local grid can’t deliver, the contract is worthless.
2. Cooling and Physical Infrastructure
Mining farms use evaporative cooling or simple air handling. AI clusters require direct-to-chip liquid cooling or immersion cooling for dense GPU racks. Retrofitting 1 GW of mining space to support liquid cooling is not a swap—it’s a rebuild. You need piping, manifolds, coolant distribution units, and a secondary loop. The cost per MW of cooling retrofit can exceed $500,000. Multiply by 1,000 MW: that’s $500 million in CapEx just for cooling, before a single GPU is installed.
Applied Digital’s construction timeline? They expect to deliver the first 100 MW in Q4 2025. Given the complexity of 1 GW, I estimate the full build out will take at least 18–24 months, assuming no supply chain delays on transformers, switchgear, or NVIDIA’s B200 GPUs. That’s optimistic. We don’t yet know if their supply chain partners can sustain that velocity.
3. Network and Interconnect
Bitcoin miners don’t need low-latency interconnects. AI training requires high-bandwidth, low-jitter networking—InfiniBand or Ethernet at 400 Gbps per GPU. The fiber backbone between data center buildings must support east-west traffic for gradient aggregation. A mining facility designed for a single connection to the internet probably lacks the fiber diversity and optical transceivers needed for a GPU cluster. CoreWeave will demand dark fiber, redundant paths, and sub-100 microsecond latency between racks. If Applied Digital cannot deliver, the service level agreement (SLA) penalties could wipe out margins.
4. Financial Engineering: The $11 Billion Trap
$11 billion over 10 years is $1.1 billion annualized. But that’s gross revenue, not net income. The CapEx to build 1 GW of data center space is roughly $1.5–$2 billion (at $1.5–$2 million per MW). Add GPU procurement (if Applied Digital provides the hardware—unclear from the contract details). If they do, another $3–$4 billion for GPUs. That implies a total investment of $5–$6 billion. Against $11 billion in revenue, the gross margin looks healthy—but only if the facility operates at 95% uptime and the GPUs aren’t obsolete within three years.
Here’s the contrarian angle: the contract is with CoreWeave, not directly with end users. CoreWeave is itself a high-risk startup. It raised $2 billion in debt in late 2023, backed by GPU colllateral. If CoreWeave defaults, Applied Digital is left with a half-built, non-standard data center that no other tenant can easily absorb. The facility was designed for CoreWeave’s specific rack layout, cooling requirements, and power density. Repurposing would require another massive retrofit. Composability isn’t a feature of this stack—it’s a missing abstraction layer.
Contrarian: The Blind Spots Everyone Ignores
1. Environmental and regulatory risk. Texas and North Dakota have volatile energy markets. In winter storms, the ERCOT grid can spike electricity prices to $9,000/MWh. Applied Digital’s contract likely has a pass-through clause, but if CoreWeave balks at paying $100 million for a single week of power, litigation follows. Meanwhile, the EPA is tightening water regulations for data centers. A 1 GW liquid-cooled facility can consume 10 million gallons of water per day. That’s a litigation magnet.
2. Technology obsolescence. The contract assumes H100/B200 GPUs stay relevant for a decade. History suggests otherwise. The NVLink generation cycle is 18 months. By 2027, the GPUs installed today will be obsolete. CoreWeave will demand upgrades—or threaten to break the lease. Applied Digital has no recourse; it’s the landlord, not the landlord with a moat.
3. The “miner-to-AI” narrative is a trap for the entire sector. Every public mining company is now touting its AI pipeline. But converting a mining facility to AI is materially harder than the PowerPoint suggests. The engineering alone requires skills that most mining firms lack. I’ve audited three such conversion plans in the past year; two had fatal flaws in their cooling design. Applied Digital might succeed, but the industry imitation will create a bubble of half-built, underfunded, one-tenant data centers. When that bubble pops, the contagion will hit APLD’s stock even if it delivers on time.
Takeaway: A Fragile Ecosystem
We don’t need more centralised single-tenant infrastructure. We need redundancy, composability, and open standards—the same principles that make blockchain resilient. Applied Digital’s 1 GW signing is a milestone, but it’s a milestone on a road with no guardrails. The real test isn’t the contract; it’s the execution over the next 24 months. If the facility slips, if CoreWeave stumbles, or if the grid fails, the 1 GW mirage will evaporate. Until then, treat this as a high-volatility bet on a single, fragile stack.
