Why are people worried about the massive capital expenditures of U.S. tech giants?

Because these expenses generate depreciation, tech giants now generally adjust the depreciation period for these hardware devices to 5-5.5 years. This means that if these investments cannot bring substantial profit growth, the money is essentially wasted.

So here's the question: Bitcoin miners can still use RX580 (a 2017 graphics card) for mining today, so why can't the giants' H100s be used for 5.5 years?

Not really. The answer lies in: "Physical lifespan" does not equal "economic lifespan."

For the giants' hyperscale data centers, hardware becomes "an economic liability" long before it physically breaks down.

The main reasons are as follows:
First, cost inversion: "Electricity costs more than gold."

Bitcoin miners use old graphics cards because they often rely on cheap electricity (abandoned hydropower, stolen electricity, etc.) and don't consider rack space.

But the giants' data centers are constrained by two physical limits: power capacity and rack space.

Assume a flagship card from 5 years ago (e.g., V100) consumes 300W of power, with a computing power of X.
Today's flagship card (e.g., H100) consumes 700W of power, with a computing power of 30X.

The harsh reality: If the giants want to continue running that old card 5 years later, it occupies valuable rack space and power outlets but can only provide 1/30th of the output.

Opportunity cost: Against the backdrop of tight power supply (U.S. data centers are competing for nuclear power), every watt of electricity is a scarce resource. Wasting electricity on inefficient old equipment equals actively giving up the opportunity to deploy more efficient new equipment.

Conclusion: Long before the 5th year, the electricity cost + maintenance cost + space rental cost of running the old equipment, allocated per unit of computing power, might be more expensive than simply discarding it and buying new cards. This is a form of "economic obsolescence."

Second, AI has the "bucket effect."
Bitcoin mining involves parallel and independent computations. One slow miner doesn't affect another; it just needs to contribute its meager share of hash value.

But AI training and inference are cluster operations.

Training large models requires tens of thousands of cards to work collaboratively through high-speed interconnects, forming a whole.

Synchronization bottleneck (Straggler Problem): In a cluster, if a batch of old cards processes slowly, the entire cluster of tens of thousands of new cards has to stop and wait for them. You cannot mix H100s (2023) and V100s (2017) in the same training cluster.

Conclusion: Once a new generation of architecture becomes mainstream, the old architecture cannot be integrated into new clusters. It can only be downgraded to run edge tasks (e.g., cold data processing), leading to a cliff-like drop in value, unable to support its original high book value.

Third, the physical barrier of the memory wall.
This is the hardest physical limitation. The development speed of AI large models has surpassed Moore's Law.

VRAM capacity and bandwidth:
V100 VRAM in 2017: 16GB/32GB HBM2.
H100 VRAM in 2024: 80GB HBM3.
Rubin VRAM in 2026: Possibly reaching 192GB+ HBM4.

If it can't run, it can't run: Models like Llama 3 405B simply cannot fit into the VRAM of old cards. It's like trying to run "Black Myth: Wukong" on an iPhone 6 from 10 years ago.

Conclusion:
The recent sharp decline in U.S. stocks was precisely due to people's financial concerns about the giants.
The giants are shifting their current cost pressures to the future through accounting methods. When that day comes (e.g., 2030), if these 5.5-year-old devices truly become a pile of scrap metal, the giants will have to take massive asset impairments, and a multi-billion-dollar hole will suddenly appear on their financial statements.

$Amazon(AMZN.US) $Microsoft(MSFT.US) $Tesla(TSLA.US) $NVIDIA(NVDA.US) $Alphabet - C(GOOG.US) $Meta Platforms(META.US)