The Depreciation Monster Haunting AI Data Center Profits

 Nebius is one of the healthier publicly traded AI data centers companies. Nonetheless, its 2Q2026 income statement illustrates the industry's depreciation challenge. Despite Nebius linearly depreciating GPUs and other server hardware over a generous five years, total depreciation equals 48% of revenues. This figure underestimates true depreciation because technological change and physical burn out means the true average life span is closer to three years. Facebook has provided detailed data on equipment failure during its 56 day estimation of the Llama 3.0 frontier model: https://ai.meta.com/research/publications/the-llama-3-herd-of-models/. For a 100,000 GPU cluster, equipment failure happens every 30 to 40 minutes. This halts estimation and requires starting over from the previous iteration.

Almost all publicly AI data centers exaggerate their profitability and net cash flows by pretending the server hardware lasts 60% longer than it is really does. Michael Burry, who successfully shorted US real estate prior to the Great Recession, justifies his current AI shorts in part due to these dubious accounting figures.

In defense of Nebius, depreciation as a percent of revenue should fall as revenues ramp up. But as we have repeatedly witnessed, revenue growth generally falls short of expectations and contractual requirements due to limited power and hardware constraints. So there is high uncertainty regarding future revenues.

Beyond these issues, it is questionable whether large language models are the future of artificial intelligence. These models are fundamentally nonlinear statistical black boxes, not reasoning engines where you can review the code line by line. This is why these models hallucinate or make stuff up. They are language prediction models using neural nets that capture language correlations. Obviously code can be added to make them better. But if the core is statistical and that approach is a dead end, then adding code does not fix the fundamental weaknesses. If other AI approaches like Jan LeCun's world models prove more fruitful, then most of the trillions of dollars in AI infrastructure spending may be waste.

Chart Show Nebius Costs As Percent of Revenues

Comments

Post a Comment

Popular posts from this blog

The British Empire's Resilient Subsea Telegraph Network

Image
The British empire had largely completed its Red Line cable network by 1902. This network allowed news and messages to be delivered in a few minutes or several hours at most depending on the message queue's length. It spanned the globe and formed a network ring so traffic could be routed in the opposite direction in case of disruption. It was, as Dr. Michael Delaunay has argued, a highly resilient network. Besides the ring configuration, the network relied on multiple cables between any pair of given end points to ensure uptime. The British military believed it would be impossible for an enemy to cut enough cables on any route to sever all communications between any given pair of end points. The Committee of Imperial Defense concluded that 57 cables must be shut down to isolate the British Isles from the Red Line network. The figure was 15 for Canada and 7 for South Africa. The Empire was self sufficient in terms of manufacturing the components for a subsea telegraph cable and re...
Read more

The Short Life Expectancy of Starlink's LEO Satellites

Image
According to FCC filings Starlink shut down almost 500 Starlink satellites during the first half of 2025. The company had them reenter the atmosphere where they burned up. What is striking is that these satellites were all less than 5 years old. The general consensus is that LEOs have a life expectancy ranging from 5 to 8 years. Shorter than expected life spans for the satellites will hit Starlink's income statement hard by increasing network depreciation and replacement needs. However, Starlink has managed to lower its LEO's manufacturing costs down to $500K versus initial figures around $1 million. So these production economies of scale might offset some of the higher than expected depreciation. However, there are also rocket launch costs as well. It costs Starlink about $3 million to put a satellite into orbit. The Falcon 9 costs $67 million per flight and delivers 23 LEOs into low Earth orbit. As a private company Starlink financials are a bit of mystery. The company press ...
Read more

Southern Cross Subsea Cable Network Deploys Ciena's Granular Layer 1 Wavelength Product

Image
Southern Cross subsea cable system is implementing ODUflex, which is a relatively new ITU standard that allows granular Layer 1 bandwidth. No longer are we limited to 100G, 400G, and 800G for either subsea or terrestrial networks. Interestingly enough, Hauwei proposed the new standard and was its primary champion. All wavelengths consist of optical containers and ODUflex allows optical containers to be stacked at 1.25 Gps intervals. So you can lease 1.25 Gbps wavelength up to 400G in 1.25 Gbps increments. Note that port sizes are still 10G, 100G, 400G or 800G. So to access a 150G transmission rate the customer needs 400G intefaces. Ciena is one of the the vendors to implement ODUflex along with Hauwei. More details here: https://lnkd.in/dZvNbubc. It is worth noting that granular bandwidth is being implemented on the newer cables which Southern Cross owns like Next.  The commercial motivation is poor take up of 400G wavelengths. The only real customers for 400Gs are very big bandwi...
Read more