Recently, Exa Labs, the developer of a next-generation search engine and a recipient of $111M in funding, deployed an Nvidia-based cluster of H200 GPUs. The cluster consists of 14 servers, each equipped with 8 GPUs, for a total of 144 GPUs, at an estimated cost of $5M.
Meanwhile, Together AI has deployed what appears to be a GB300 rack with 72 GPUs, estimated at ~$3M. Together AI operates as an inference platform, similar to Baseten and Fireworks AI.
Building GPU cloud infrastructure is a capital-intensive undertaking. Hyperscalers and specialized AI cloud providers are investing billions in NVIDIA GPUs, often financing these purchases through debt or equity. How these companies account for the useful life of their hardware, particularly the depreciation period, directly affects reported margins, cash flow optics, and investor perception. Two companies, CoreWeave and Nebius, illustrate this contrast in approach.
Disclaimer: We are not licensed financial advisors. This analysis is for informational and research purposes only and should not be construed as investment advice.
Depreciation Policies: A Tale of Two Clouds
-
- CoreWeave (CRWV)
In its S-1 IPO filing, CoreWeave disclosed that it depreciates “technology equipment” (including NVIDIA GPUs) on a straight-line basis over six years. For every GPU purchased, the company spreads the cost evenly over 72 months. That lowers near-term expenses, inflates operating margins, and makes the balance sheet look stronger. - Nebius (NBIS)
By contrast, Nebius, a rival cloud provider, depreciates GPUs over four years. This accelerates expense recognition, which can depress reported margins in the short term but may provide a more realistic reflection of GPU economic life, given rapid innovation cycles. - IREN (Iris Energy)
IREN, a public data center and AI infrastructure company, has not disclosed a fixed GPU depreciation period in filings. However, some of its financing arrangements involve 36-month leases, hinting that internally it views GPU hardware as effectively three-year assets.
- CoreWeave (CRWV)
Why Depreciation Matters
In capital-heavy businesses, depreciation is more than an accounting formality. It affects:
-
- Margins: Longer depreciation periods lower the annual expense, artificially inflating EBITDA and operating margin.
- Asset Values: The balance sheet retains higher book values for older hardware, even if secondary market prices collapse.
- Cash Flow Timing: Although depreciation is non-cash, it influences investor models and perceived profitability.
The Risk of Overstating Useful Life
Extending depreciation schedules, let’s say six years instead of four, has risks:
-
- Technological Obsolescence
NVIDIA introduces new architectures on an aggressive cadence (roughly every 2–3 years):
-
- Ampere (2020) → A100
- Hopper (2022) → H100
- Blackwell (2024) → GB200, GB300
- Rubin (expected 2026)
In practice, by the time a GPU reaches year three or four, demand shifts to newer generations. Assets can lose economic value far faster than their accounting life.
-
- Balance Sheet Distortions
If book values assume six years of useful life, but resale or market value drops 50% by year three, companies face the prospect of impairment charges—one-time hits that can surprise investors and damage credibility. - Competitive Pressure
AI workloads are hypersensitive to performance gains. A customer who can access 2× or 4× efficiency on a Rubin or Blackwell GPU will rarely accept older hardware, except at steep discounts. This accelerates obsolescence relative to depreciation schedules. - Financing Risk
Many GPU clouds finance hardware purchases with debt. If repayment schedules are structured around a six-year life but revenues fall after three, debt servicing can become a liquidity squeeze.
- Technological Obsolescence
A Conservative Approach: Nebius’ Four Years
Nebius’ four-year depreciation period may better align with industry realities:
-
- It anticipates the ~3-year cadence of NVIDIA architectures.
- It cushions against impairment risk if resale markets collapse.
- It gives investors a more transparent view of real asset turnover.
While it pressures short-term margins, it reduces the risk of future write-downs and credibility issues.
CoreWeave’s Gamble
CoreWeave’s six-year straight-line method reflects a bet: that GPUs will retain revenue-generating usefulness long enough to justify the accounting life. There’s some rationale; older GPUs can still be monetized for inference, lower-priority workloads, or emerging markets. However, if AI development continues at its current velocity, the gap between book life and economic life could widen.
Conclusion
Depreciation is more than accounting trivia; it’s a forward-looking statement about how long management expects GPUs to generate returns.
-
- Shorter schedules (like Nebius’ four years) acknowledge the reality of rapid obsolescence but weigh on near-term margins.
- Longer schedules (like CoreWeave’s six years) improve profitability optics today but risk painful impairments tomorrow.
As investors and customers evaluate GPU cloud providers, the key question isn’t just how much GPU capacity a company has, but how long that capacity remains relevant. In a market where NVIDIA redefines performance every 2–3 years, aggressive depreciation may be less about optimism—and more about papering over an uncomfortable truth.