I’ve been spending a lot of time lately staring at the CapEx lines of big tech companies. The numbers are staggering. We are looking at hundreds of billions of dollars poured into hardware that, historically, loses its value faster than a new car driving off the lot.
Everyone is talking about Nvidia’s stock price. But as an accountant, I’m interested in something much more boring and potentially much more dangerous: depreciation.
Here is my look at the data, the accounting implications, and why your fixed asset ledger might be lying to you.
Training vs. Inference: The Expensive Reality
To understand the accounting, you have to understand the engineering. I used to think a “chip was a chip.” That is not the case anymore.
The AI lifecycle has two parts:
- Training: This is the university phase. You spend billions teaching the model everything on the internet. This requires massive, power-hungry clusters. Nvidia dominates here.
- Inference: This is the “day job” phase. This is what happens when you type a question into ChatGPT and it spits out an answer.
The course breaks this down clearly. Training is a one-time capital outlay. Inference is a perpetual cost. Every time you ask a bot to write an email, it costs Microsoft money.
Microsoft’s new Maia 200 chip is designed specifically for inference. I looked at the specs from the research report:
- Process: TSMC 3-nanometer (cutting edge).
- Transistors: 140 billion.
- Goal: Lowering the “cost per query.”
Why does this matter? Because general-purpose GPUs (like Nvidia’s) are overkill for simple questions. By building their own custom silicon, Microsoft is trying to protect its gross margins. They are moving from renting the infrastructure to owning the factory.
The “Obsolescence Trap”
This is the part of the course that really grabbed me. We are seeing is called the “Obsolescence Trap” in fixed asset accounting.
Standard accounting practice suggests depreciating computer hardware over 5 to 7 years. But look at the innovation cycle.
- 2024/25: Nvidia Blackwell architecture.
- 2026: Nvidia “Vera Rubin” architecture.
- 2026: Microsoft Maia 200.
The research shows that a chip considered “state of the art” today might be 50% less valuable in 12 months. If you are depreciating an asset over 5 years, but its economic utility drops to near zero in 3 years because a new chip does the work for half the power cost, your balance sheet is inflated.
We discuss this in the video: Auditors aren’t likely to force a change in useful life estimates without a “triggering event.” But the reality of the hardware market suggests that companies are holding billions in assets that are technically obsolete long before they are fully depreciated.
Shifts in the P&L: COGS vs. R&D
Another fascinating angle we cover is the classification of costs.
When Microsoft buys 10,000 GPUs from Nvidia, that’s a capital expenditure (CapEx) that hits the balance sheet and depreciates through Cost of Goods Sold (COGS) or OpEx.
But when Microsoft designs the Maia 200?
- The design costs move to Research & Development (R&D).
- Only the manufacturing costs hit the asset line.
This allows tech giants to shift expenses around their P&L, potentially making their gross margins look better while bloating their R&D lines. It’s a subtle shift, but when you are talking about $150 billion in infrastructure spending, those buckets matter.
Summary:
- Inference is the new battleground: Tech giants are building custom chips to lower the daily cost of running AI, moving away from relying solely on Nvidia.
- Useful life is shrinking: The 5-year depreciation standard is likely inaccurate due to the rapid pace of chip innovation (3nm to 2nm).
- Accounting classifications are moving: Vertical integration shifts costs from CapEx/COGS to R&D.
- The risk is real: The danger of technological obsolescence creates a significant risk of asset impairment for companies holding older chips.
Want to earn CPE for this topic?
- Compare Options: See how we stack up against others in our 2025 Flexible CPE Guide
- Understand the Format: Read how Nano-Learning works for CPAs.
- Check Your State: Ensure you are compliant with our State Requirements Guide.
- What is EverydayCPE?
Related Courses:
