A 100 MW dedicated inference campus running GB200 NVL72 racks would need about $2.91 per GPU-hour to generate a 15% levered equity internal rate of return, according to a deep dive by Scotiabank Global Equity Research.

Scotia’s model places total project capital expenditure at $4.03 billion for 100 MW of critical IT load, or $40.3 million per megawatt. Of that total, 64% is allocated to IT equipment, priced at $3.0 million per NVL72 rack, with much of that spending flowing to NVDA. The remaining 36% is attributed to the facility itself, at $14.5 million per megawatt.

The bank assumes 70% of the project is financed at 8.0% over six years and includes a 15% residual value on IT capex after year six, underscoring how sensitive the return math is to both financing terms and equipment value at the end of the holding period.

Scotia said current GB200 rental indications clear the hurdle. At the $2.91 requirement, the model translates to about $0.81 per million output tokens. By comparison, current closed-model token output costs range from $1.50 for Gemini 3.1 Flash-Lite to $15 for GPT 5.4. DeepSeek V4 Flash is priced at $0.28 to $0.66, while Ox Alpha appears to be around $0.50 for a product that approaches Claude Opus 4.8.

The report also highlighted the scale of hyperscaler commitments. Scotia said off-balance-sheet numbers have reached $2.588 trillion, with another $779 billion already on the balance sheet. Those figures do not include $120 billion of guarantees and other contingent exposures, which the bank said are not firm commitments.

Another relevant figure is the operating expense estimate. Scotia pegged opex at $0.52 per GPU-hour, which it described as a floor because, below that level, it is better to shut the facility down. The report noted that commodity producers and factories have sometimes operated at a loss for a period of time, but said the same logic still sets a practical lower bound.

The bearish case for AI infrastructure, according to the report, is that one of the committed tenants to these data centers — such as OpenAI or Anthropic — could fail. In that scenario, a facility built for $4.03 billion and financed with $2.82 billion of debt would lose its tenant. Because a purpose-built AI campus has no quick alternative use, the lender would be left with immobile, single-purpose collateral. The equipment inside would also be the fastest-depreciating part of the asset. Scotia said the lender would likely take a writedown, find an operator, and sell compute at whatever the market would bear.

The bank pointed to the fiber-optics sector as a historical parallel, citing the period between 2001 and roughly 2010. In that case, a single default removed contracted demand and added spot supply at the same time, pushing secondary values lower for the same reason rents fall when supply exceeds demand.

Scotia’s cash-cost estimate maps to $0.145, while market-median API list pricing for output tokens is $1.94, though the report said that pricing can change.

The report said another major risk is power availability. Interconnect queues, transformer and turbine backlogs, and shortages of skilled electrical labor could limit how quickly announced capacity is physically delivered. That creates the possibility that facilities are completed before power is ready, leaving the GPUs to age before the sites can be fully utilized.

One benchmark to watch, excluding default risk, is contract pricing. Scotia wrote: “In our model, the difference between $2.50 and $3.50 separates a failed base case from a project that clears both return and coverage tests.”

The report also leaves open the question of how hyperscalers should trade if these large investments only barely meet capital hurdle rates. If the capex produces genuine incremental earnings, EPS could grow into a derating, leaving the stock flat for years. Scotia said that is a common outcome when a high-return business shifts toward a more capital-intensive model.

Scotia also emphasized the circular financing dynamic that has become a key feature of the AI buildout. Hyperscaler and model-developer commitments support project debt, creating correlated exposure across tenants, infrastructure owners, and lenders. Lower API pricing could weaken tenant credit, pressure residual values, widen project debt spreads, and slow refinancing and construction. The end result, the bank said, depends on whether rising demand intensity can offset declining token prices and the obsolescence of installed capacity.

The final question, however, cuts in the opposite direction: what if AI succeeds? In a world of superintelligence, models may quickly optimize, which could push token costs lower. A self-learning model could also design a chip that is far more efficient than those in use today.