Author：区块律动
原文标题：Some thoughts ahead of Nvidia tonight
Original author: @GavinSBaker
Compiled by: Peggy, BlockBeats
Editor's Note: After Nvidia's earnings release, the market's focus is often on revenue, profit, and guidance range. However, this article's author, @GavinSBaker, attempts to shift the discussion back to a longer-term perspective: what determines Nvidia's value is not a single quarter's data, but how long AI demand can continue, and whether computing power investments truly generate sustainable returns.
Drawing on historical experience with technology cycles, the article discusses whether "bubbles and over-construction" will repeat themselves. It also points out that the current AI cycle faces bottlenecks in power and wafer supply, potentially leading to a more restrained expansion pace. On the other hand, GPU leasing prices and the high utilization rate of older chip models provide real-world validation for "AI ROI."
The following is the original text:
The following are some personal observations that may be helpful to those who follow Nvidia. In my opinion, there are only two core variables that are truly worth discussing regarding this company: the sustainability of demand and the return on investment (ROI) of AI, the latter of which is closely related to the effective lifespan of GPUs.
Sustainability of demand: Will history repeat itself?
Historically, almost all similar cycles of technological revolutions have involved financial bubbles and overcapacity. Carlota Perez systematically discusses this in *Technological Revolutions and Financial Capital*. She points out that with each technological revolution—whether it's railroads, broadcasting, or the internet—financial markets are quick to recognize its long-term potential, and the ensuing capital frenzy often breeds bubbles (which can also be explained by Mauboussin's concept of "the collapse of diversity of ideas"). Bubbles lead to over-construction, which in turn causes a temporary drop in demand, resulting in a market crash; while the oversupply of basic technologies ultimately lays the foundation for a "golden age." The trajectory of the internet is a prime example.
Therefore, for Nvidia, the key is not this quarter's results or next quarter's guidance, as these are often already fully anticipated by buy-side institutions. What truly matters is the sustainability of earnings per share (EPS), rather than the annual growth rate.
Judging from the expectations implied by the current valuation, the market seems to be expressing the judgment that Nvidia's earnings may be nearing a cyclical peak, implying concerns about excessive capital expenditure. It's important to emphasize that the market's concern is not about a "valuation bubble," but rather a "fundamental bubble," namely the potential risk of over-construction driven by capex. If the market can build confidence that Nvidia can maintain a high single-digit revenue compound annual growth rate (CAGR) after fiscal year 2027, the valuation center may find support.
Is this time really different?
"This time is different" is often a dangerous judgment. But this AI cycle is indeed different: there are substantial bottlenecks globally in two key dimensions: electricity (watts) and advanced process wafers, and it may take years to alleviate these constraints.
This hard constraint on the supply side may actually curb excessive capacity expansion. Hyperscale cloud providers, theoretically, would continue to expand if conditions allowed, but in reality, electricity and wafer supply limit their expansion pace. Unlike the historical technological revolutions described in Perez's book, there were no similar supply bottlenecks limiting deployment speed at that time.
Without excessive construction, a collapse is unlikely to occur, especially given that the overall valuation of tech stocks is not currently at an extremely high level.
Of these two bottlenecks, wafer fabrication may be more critical than electricity. Controlling the pace of wafer production capacity could become a significant variable in extending the AI cycle. TSMC's management is known for its prudence, emphasizing industry stability and long-term value rather than short-term aggressive expansion. Without the constraints of electricity and wafer fabrication, Nvidia's growth over the next 24 months might be faster, but the risk of over-construction would also increase significantly.
In a sense, supply constraints may be slowing down the entire AI cycle to a steady state. AI's high dependence on advanced process wafers may actually be a key factor in avoiding drastic fluctuations in this cycle.
To realize some of the extreme hypothetical scenarios, the computing power might need to be increased to hundreds or even thousands of times its current level. The time required for this expansion itself provides a buffer for social adjustment and institutional adaptation.
Historical experience also provides a reference: after James Watt invented the steam engine, it took decades for the railway system to truly replace horses. AI may iterate faster, but it still won't be able to restructure society in a very short time.
More importantly, humans only need 20–30 watts of power to achieve “general intelligence.” In a world with limited electricity, this efficiency advantage will persist for a long time. Therefore, a smoother and more sustainable AI cycle may not necessarily be a bad thing for society itself.
GPU lifespan and the true ROI of AI
The rental price of GPUs essentially reflects the economic value of the token and is a core indicator of "AI ROI". Theoretically, as higher-performance chips continue to be released, the rental price of older GPU models should gradually decline, even if the AI investment return rate is positive.
However, in the past two months, the leasing price of the H100, which has been in service for nearly four years, has increased significantly. This means that, especially in the scenarios of agentic AI and code generation, computing power is creating real and considerable economic value.
Meanwhile, even with the release of Blackwell, the six-year-old A100 still maintains high utilization rates, and rental prices have not shown significant signs of softening. This strongly suggests that the effective lifespan of GPUs may be at least six years, or even longer than the depreciation cycle of most customers.
The impact is structural: if the residual value is higher than previously expected, the financing cost of GPUs will decrease further. In contrast, ASICs customized for a single model or specific application rarely have a similar lifecycle advantage. In a rapidly iterating environment, specialized chips have higher capital costs and are more difficult to finance.
To some extent, versatility is the moat of GPUs. With the separation of prefill and decode functions and the gradual formation of supporting chip systems, computing architecture is evolving from "single-chip logic" to "multi-chip collaborative system". AI infrastructure no longer depends on a single device, but is a whole set of highly coupled system engineering.
With the decoupling of prefill and decode, the NVIDIA ecosystem may complete its structural adjustment earlier than the TPU ecosystem. Coupled with the differences in design roadmaps among different vendors, the relative advantage customers have in inference costs is changing.
If some vendors previously relied on cost advantages to drive down token prices and gain market share, then as this advantage diminishes, market behavior will become more rational. In the long run, this will have a positive impact on AI ROI, especially during the transition from training to inference in computing power requirements.
This turnaround is perhaps more noteworthy than any quarterly performance.
One last lighthearted wish: that Nvidia will revert to using superheroes as chip codenames in the future. Surprisingly, the "green camp" has never used the name "Banner" (the real name of the Marvel character Hulk).
