Beyond financial reports, NVIDIA's true risks and opportunities lie in areas beyond the financial statements.

Title: Some Thoughts Ahead of Nvidia Tonight
Author: @GavinSBaker
Translation: Peggy, BlockBeats

Editor’s Note: After Nvidia’s earnings report, market focus often centers on revenue, profit, and guidance ranges. However, author @GavinSBaker attempts to shift the discussion to a longer-term perspective: what truly determines Nvidia’s value is not just quarterly data, but how long AI demand can sustain and whether compute investments genuinely generate lasting returns.

Starting from historical experiences of technological cycles, the article discusses whether “bubbles and overbuilding” might recur, while noting that this AI cycle faces power and wafer supply bottlenecks that could temper expansion. On the other hand, high utilization rates of older GPU models and rental prices provide real-world validation of “AI ROI.”

Below is the original text:

Here are some personal observations that may be helpful for those interested in Nvidia. In my view, there are only two core variables worth discussing around this company: one is the sustainability of demand, and the other is AI’s investment return (ROI), which is closely related to the effective lifespan of GPUs.

Sustainability of Demand: Will History Repeat?

From the historical experience of technological waves, nearly all similar cycles have experienced financial bubbles and overcapacity expansion. Carlota Perez systematically discusses this in Technological Revolutions and Financial Capital. She points out that each technological revolution—whether railroads, broadcasting, or the internet—early on reveals its long-term potential to financial markets, which then often leads to capital frenzy and bubbles (this can also be explained by Mauboussin’s concept of “diversity collapse”). Bubbles cause overbuilding, which then triggers phased demand drops and market crashes; excess supply of foundational technologies eventually lays the groundwork for a “golden age.” The development trajectory of the internet is a classic example.

Therefore, for Nvidia, the key isn’t quarterly results or next quarter’s guidance—these are often well anticipated by institutional investors. The real focus should be on the sustainability of earnings per share (EPS), not just the growth rate in a given year.

From current valuation expectations, the market seems to be signaling that Nvidia’s profits may be approaching a cyclical peak, with underlying concerns about excessive capital expenditure. It’s important to emphasize that the market’s worry isn’t a “valuation bubble,” but a “fundamental bubble”—a potential overbuilding risk driven by capex. If the market can develop confidence that Nvidia can maintain high single-digit revenue compound annual growth rate (CAGR) beyond fiscal year 2027, then valuation levels could be supported.

Is This Time Truly Different?

The phrase “this time is different” is often a dangerous assumption. However, this AI cycle does have differences: significant bottlenecks exist in both power (watts) and advanced wafer supply, and easing these constraints could take years.

These supply-side constraints may actually suppress overcapacity expansion. Large cloud providers might theoretically continue to scale up if conditions permit, but in reality, power and wafer limitations restrict their pace. Unlike the historical technological revolutions described by Perez, where no such supply bottlenecks existed to limit deployment speed, today’s constraints could prevent overbuilding.

Without overbuilding, crashes are less likely—especially given that overall tech stock valuations are not at extreme highs.

Among these bottlenecks, wafers may be more critical than power. Controlling wafer capacity growth could be a key factor in extending the AI cycle. TSMC’s management has always been cautious, emphasizing industry stability and long-term value rather than short-term aggressive expansion. Without power and wafer constraints, Nvidia’s growth over the next 24 months might be faster, but the risk of overbuilding would also rise significantly.

In a sense, supply constraints might be “slowing down” the entire AI cycle to a more stable pace. The high dependence on advanced process wafers could actually be a key factor in avoiding sharp fluctuations in this cycle.

To achieve some extreme hypothetical scenarios, compute capacity might need to increase by hundreds or even thousands of times. The time required for such expansion provides a buffer for societal adjustments and regulatory adaptation.

Historical experience offers a reference: after James Watt invented the rotary steam engine, it took decades for railroads to fully replace horses. AI’s iteration speed may be faster, but it’s unlikely that social restructuring will happen in a very short period.

More importantly, humans only need 20–30 watts of power to achieve “general intelligence.” In a world with limited electricity, this efficiency advantage could persist long-term. Therefore, a smoother, more sustained AI cycle might actually be beneficial for society.

GPU Lifespan and True AI ROI

GPU rental prices fundamentally reflect the token’s economic value and are a core indicator of “AI ROI.” Theoretically, as higher-performance chips are continuously introduced, older GPU models’ rental prices should gradually decline—even if AI ROI remains positive.

However, over the past two months, rental prices for the nearly four-year-old H100 have risen significantly. This indicates that, especially in agentic AI and code generation scenarios, compute power is creating real and substantial economic value.

Meanwhile, even with the launch of Blackwell, the 6-year-old A100 still maintains high utilization, with rental prices not showing obvious decline. This strongly suggests that GPU effective lifespan could be at least six years or longer—possibly exceeding most customers’ depreciation cycles.

This has structural implications: if residual value is higher than previously expected, the cost of financing GPUs will decrease further. In contrast, ASICs tailored for specific models or uses tend to have shorter or less flexible lifespans. In fast-paced environments, specialized chips have higher capital costs and are harder to finance.

To some extent, the versatility of GPUs is their moat. As functions like prefill and decode are decoupled and supporting chip ecosystems mature, compute architectures are evolving from “single-chip logic” to “multi-chip collaborative systems.” AI infrastructure no longer relies on a single device but on a highly integrated system.

With prefill and decode decoupled, Nvidia’s ecosystem may complete structural adjustments earlier than TPU’s. Coupled with design choices by different vendors, customers’ relative advantages in inference costs are changing.

If some vendors previously relied on cost advantages to lower token prices and gain market share, then as these advantages diminish, market behavior will become more rational. In the long run, this will positively impact AI ROI—especially during the transition from training to inference.

This shift may be more noteworthy than any quarterly earnings report.

A final lighthearted wish: I hope Nvidia will someday bring back superheroes as chip codenames. Surprisingly, the “Green Camp” has never used the name “Banner” (the real name of Marvel’s Hulk).