AI价值链：GPU真的能运行6年吗？

AI Value Chain: Can you really run a GPU for 6 years? Amid increasingly cautious sentiment around the AI debate, investors have reopened thedebate around whether GPUs can really run for 6–7 years, and, by extension, whether Stacy A. Rasgon, Ph.D.+1 213 559 5917stacy.rasgon@bernsteinsg.com Daniel Zhu+1 917 344 8309daniel.zhu@bernsteinsg.com In short: Yes. GPUs can profitably run for ~6 years, and that the depreciationaccounting of most major hyperscalars is reasonable.We observe that cash costsof operating a GPU are very low compared to market prices for GPU rental, making thecontribution margins of running old GPUs for longer quite high. Even with meaningfulimprovements in price/performance with each GPU generation, vendors can makecomfortable margins on 5-year old A100s, in turn implying a 5-6 year depreciation lifespan Mark L. Moerdler, Ph.D.+1 917 344 8506mark.moerdler@bernsteinsg.com Gautam Chhugani+91 226 842 1416gautam.chhugani@bernsteinsg.com Mark C. Newman+1 212 845 7822mark.newman@bernsteinsg.com One nuance is that GPUs probably lose more value after the first year than a linear6-year depreciation would imply, but that they appear to retain value fairly wellbeyond that point.We observe that data center operators often lose a significant numberof GPUs to “burn-in”, as configurations based on the previous generation of GPU may notbe quite right for newer hardware and operators take time to figure out the configuration.Likewise, users often prefer to run demanding workloads like AI training on the latest Arpad von Nemes+1 917 344 8461arpad.vonnemes@bernsteinsg.com Alrick Shaw+1 917 344 8454alrick.shaw@bernsteinsg.com Firoz Valliji, CFA+1 917 344 8316firoz.valliji@bernsteinsg.com We also observe that, given the prevalence of long term contracts, even if GPUsdepreciate faster than companies are modeling, the cost may be born by end-usersin the form of artificially high prices.For instance, if OpenAI signs a 5-year contract forCoreweave H100 capacity, even if the H100s are worth less than a 5-year depreciation Shelly Tang, CFA+1 917 344 8342shelly.tang@bernsteinsg.com Mahika Sapra+91 226 842 1408mahika.sapra@bernsteinsg.com In contrast to memory/storage, accelerated compute does not appear to price asa commodity, with older GPUs commanding higher prices than price/performanceparity would suggest.This suggests that end users are still running legacy workloads that Sanskar Chindalia+91 226 842 1445sanskar.chindalia@bernsteinsg.com Aman Jain+91 226 842 1486aman.jain@bernsteinsg.com While in theory this would be negative for GPU vendors, given it implies lowerdemand for GPU replacements, in practice much of the dynamic is driven by thefact that compute demand is so overwhelming that it still makes sense to keeprunning older, lower-efficiency hardware. Our main takeaway instead is that despiteconcerns that GPU lifespans are overstated, understating the real cost of GPU April Li+1 917 344 8339april.li@bernsteinsg.com BERNSTEIN TICKER TABLE INVESTMENT IMPLICATIONS NVDA (OP, $225):The datacenter opportunity is enormous, and still early, with material upside still possible AMD (MP, $200):AI expectations remain high, but a new deal with OpenAI has the prospect to drive further (possiblysubstantial) growth. AVGO (Outperform, $400 PT):A strong 2025 AI trajectory seems set to accelerate into 2026, bolstered by software, cashdeployment, and superb margins & FCF. IREN (OP, PT $125), CORZ (PT $24), RIOT (OP, PT$25), and CLSK (OP, PT$24):We see enormous re-rating potentialfor miners pivoting into AI datacenter assets. IREN is our top pick as vertically integrated datacenter opportunity owning over~3GW of power assets. RIOT, CORZ and CLSK with over ~4.5GW of combined power portfolio continue pursuing AI co-location DETAILS Investor sentiment is growing increasingly cautious around the AI trade (as evidenced by surging CDS spreads on Coreweave andOracle debt). Against this backdrop, investors have resurfaced the debate around whether GPUs can really run for 6–7 years and, In short: Yes. GPUs can profitably run for ~6 years, and that the depreciation accounting of most major hyperscalarsis reasonable.We observe that cash costs of operating a GPU are very low compared to market prices for GPU rental, whichin turn means that the contribution margins of running old GPUs for longer are quite high. Even with meaningful improvementsin price/performance with each GPU generation, vendors can make comfortable margins on 5-year old A100s; it’s only lookingat 7-year old Volta GPUs when we start to reach cash break-even. This in turn implies a 5-6 year depreciation lifespan is •Old GPUs generally still function beyond ~6 years.Through our conversations with industry participants, we havegenerally received consistent feedback that GPUs generally still function at 6-7 years or more. While there are some highprofile stories of GPUs burning out after ~6 months, most of these are more attributable to burn-in: as