NVIDIA has officially launched its Blackwell B200 GPU, delivering 4x the AI training performance of the previous-generation H100 while driving record quarterly revenue of $40 billion, yet a deeper accounting challenge threatens the economics of the massive data center buildout underway across the industry. The new chip, built on TSMC's custom 4NP process, packs 208 billion transistors and delivers 20 petaflops of FP4 AI compute, alongside 192GB of HBM3e memory with 8 TB/s bandwidth. When paired with NVIDIA's Grace CPU to form the GB200 Superchip, the system delivers 40 petaflops of AI performance, with major cloud providers including AWS, Google Cloud, Microsoft Azure, and Oracle Cloud announcing immediate availability.

The technical leap is undeniable. The Blackwell B200's second-generation Transformer Engine supports FP4 and FP6 precision, enabling 30x faster inference for trillion-parameter models compared to the Hopper generation. Training a GPT-5-class model now takes one-quarter of the time and cost, while improved inference efficiency could reduce the cost of serving AI applications by 70 percent. NVIDIA's stock surged 15 percent in after-hours trading following the announcement.

Why Are Data Center Companies Worried About Depreciation Schedules?

Behind the scenes, however, a critical accounting problem is emerging. Microsoft, Google, Amazon, Meta, Oracle, and other hyperscalers have collectively spent hundreds of billions of dollars on data centers, servers, memory, and NVIDIA GPU clusters. These costs do not hit the income statement all at once; instead, they are placed on the balance sheet and depreciated over time, directly affecting what companies report for earnings per share (EPS).

The problem is that these companies are not aligned on how long their AI hardware remains economically useful. Microsoft says servers and GPUs can be depreciated over 6 years. Google also uses a 6-year life for key technical infrastructure assets. Meta moved most servers and network assets to 5.5 years. Amazon, by contrast, reduced the useful life of a subset of servers and networking equipment from 6 years to 5 years, citing the faster pace of AI and machine-learning technology development.

The math matters enormously. A 6-year depreciation schedule recognizes only 16.7 percent of an asset's cost each year. A 5-year schedule recognizes 20 percent. A 4-year schedule recognizes 25 percent. When AI assets are measured in hundreds of billions of dollars, a longer depreciation life can be the difference between an EPS beat and an EPS miss. The issue is whether investors are paying a premium multiple for real earnings growth or flattered earnings.

How Fast Is the GPU Replacement Cycle Actually Moving?

The acceleration of NVIDIA's chip roadmap is the core of the problem. Hopper dominated 2023 into 2024. Blackwell is now being deployed at scale. NVIDIA has already positioned Rubin and Vera Rubin as the next major GPU generations, each promising higher throughput, better power efficiency, and lower cost per token. NVIDIA has stated that Rubin can reduce inference costs by as much as 10x while requiring 4x fewer GPUs for certain mixture-of-experts workloads than Blackwell.

That pace of improvement creates a cascading problem for data center operators. If Blackwell lowers cost per token versus Hopper, customers become less willing to pay premium prices for Hopper compute. If Rubin improves on Blackwell, Blackwell eventually faces the same pressure. The chips may remain useful, but their premium economics can deteriorate before the depreciation schedule is finished.

AI data-center companies should recover their original investment before the next GPU resets market pricing. Breaking even is not enough when you have hundreds of billions on the line. Shareholders do not give a company capital merely to get the same money back years later.

Steps to Evaluate AI Infrastructure Investment Risk

• Assess Depreciation Alignment: Compare the useful-life assumptions across Microsoft, Google, Amazon, and Meta in their SEC filings to determine which company's estimate best reflects the actual economic life of AI hardware in a rapidly accelerating chip cycle.
• Calculate Cost-Per-Token Trends: Track how inference costs decline with each new GPU generation and model whether current data center investments can generate attractive returns before the next generation arrives and resets pricing.
• Monitor Monetization Velocity: Examine whether AI product pricing and adoption rates are growing fast enough to offset the massive capital expenditure on infrastructure, since much AI usage remains free or heavily discounted.
• Review Hardware Treadmill Exposure: Identify which companies are most exposed to rapid hardware obsolescence by examining the percentage of their recent capex allocated to GPU clusters versus other infrastructure assets.

The deeper issue is that these companies are spending like they already have the next iPhone, but the monetization engine is still unclear. Under Steve Jobs, Apple's R&D was nearly $800 million before the iPhone launched in 2007. In 2008, Apple sold the product, earned back all of its R&D spend and then some, all within 12 months with $1.8 billion in iPhone revenue. Today's hyperscalers are doing the reverse: spending hundreds of billions of dollars first, with monetization coming later, hopefully.

AI products are not currently priced like the iPhone. Much of AI usage is free. The AI LLM (large language model) market is crowded with literally hundreds of free models, plus a massive open-source ecosystem. Model routing sends simpler tasks to cheaper models. The on-device AI market is experiencing an unprecedented surge, projected to skyrocket to $75.5 billion by 2033, shifting processing directly onto smartphones, laptops, and wearables without touching an AI data center.

Meanwhile, companies like Google are funding AI data centers while building tools designed to lower inference costs and route work more efficiently. Microsoft recently announced that customers can now use the cheaper Chinese model, DeepSeek. When demand is high and a new technology is expected to revolutionize the industry, it is unusual for companies to prioritize selling the cheapest model to recoup the $200 billion they have just invested in it.

That is great news for users, but it is bad news for investors in companies that need to make enormous profits to recoup their spending over the past two years and growing, which collectively exceed $1 trillion. If companies spend hundreds of billions building AI capacity, then create products that make each query cheaper, investors need to ask whether profits can grow fast enough to offset falling cost per unit. More AI usage does not automatically mean attractive return on invested capital.

The Blackwell launch represents a genuine technical achievement, and NVIDIA's record revenue reflects real demand for AI compute. But the depreciation trap reveals a deeper tension in the industry: the hardware is improving faster than the business models that justify the investment. Until data center operators can demonstrate that their massive infrastructure spending will generate strong cash returns before the next GPU generation arrives, the accounting estimates buried in SEC filings may tell a more cautious story than the headlines suggest.