The $4 trillion question: Can the hyperscalers’ enormous bet on data centers pay off
A 100 MW dedicated inference campus running GB200 NVL72 racks needs roughly $2.91 per GPU-hour to generate a 15% levered equity IRR.
That's the conclusion of a deep dive from the Scotiabank Global Equity Research team. The $2.91 needed to clear the return hurdle and it comes from this:
- $4.03 billion of total project capex on 100 MW of critical IT load,
- That is $40.3 million per megawatt.
- Sixty-four percent of that is IT equipment at $3.0 million per NVL72 rack (much of that going to NVDA)
- Thirty-six percent is the facility at $14.5 million per MW.
- 70% is financed at 8.0% over six years
- It assumes a 15% residual value on IT capex after year six
The good news is that current GB200 rental indications clear it. Another way of looking at it is that at the $2.91 hurdle, the model translates to $0.81 per million output tokens. Current closed model token output cost numbers run from $1.50 on Gemini 3.1 Flash-Lite to $15 on GPT 5.4. Howewever, DeepSeek V4 Flash is $0.28-0.66 and Ox Alpha looks to be around $0.50 on a product that approaches Claude Opus 4.8.
Another notable set of numbers from Scotia is just how large the off-balance sheet numbers are getting, they're now at $2.588 trillion with another $779 billion already on the balance sheet. These are staggaring numbers and don't include $120 billion of guarantees and other contingent exposures because they are not firm commitments.
Another useful deviation from the $2.91 all-in hurdle is that opex number, which they peg at $0.52 per GPU hour. You can take that as a floor as below that, it's better to turn the lights off (though we've certainly seen commodity producers and factories operate at a loss for a time).
The bear case on AI is that one of the 'committed tenants' to these data centers -- think OpenAI or Anthropic -- goes bust. Picture a facility built for $4.03 billion, financed with $2.82 billion of debt, loses its tenant. Nobody demolishes it. The lender forecloses on an asset that is cash-positive at $0.53 per GPU-hour and has no alternative use — you can't convert a purpose-built AI campus into anything else quickly, and the equipment inside is the part losing value fastest. So the lender does what lenders always do with immobile single-purpose collateral: takes a writedown, finds an operator, and sells compute at whatever the market will bear.
Fiber optics did exactly this between 2001 and roughly 2010. A single default removes contracted demand and adds spot supply in the same instant. The capacity that was locked up under contract becomes capacity competing for whoever's left. GPU secondary values collapse for the same reason rents collapse — surplus capacity.
Cash cost maps to $0.145. Market-median API list pricing for output tokens is $1.94 but -- like I showed -- it's changeable.
That's not the only thing that could go wrong. Like I've been writing about for two years: There isn't enough power. Interconnect queues, transformer and turbine backlogs and skilled electrical labour constrain how fast announced capacity physically arrives. There is a real possibility the facilities are completed and the power isn't ready so the AI fruit will essentially die on the vine as the GPUs age.
One benchmark to watch (so long as you ignore default risk) is contract pricing.
"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," Scotia writes. A question left unanswered is at what multiple should hyperscalers trade at if these huge investments only barely clear investment hurdles?
If the capex generates genuine incremental earnings, EPS grows into a derating and the stock goes sideways for years. That's the usual outcome when a high-return business transitions to a capital-intensive one.
Of course, Scotia also highlights the circular financing that ZeroHedge has been relentlessly touting:
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 andconstruction. Outcomes ultimately depend on whether rising demand intensity can offset declining token prices and the obsolescence of installed capacity.
The big question I have is: What if AI is successful? In a world of superintelligence, one of the first things the models do is optimize and that could drive token costs to the floor. Moreover, a self-learning model should be able to design a chip that's far more efficient than the ones in use today.
This article was written by Adam Button at investinglive.com.提供 MainLink：Investinglive RSS Breaking News Feed
FX初心者には必須 無料のうちにGET！
