Are GPUs Essentially a Subscription?

Are graphics processing units more akin to a subscription than “capital investment?” 

Think about your own smartphone purchases. Yes, it is a hardware purchase. But it is also a hardware purchase with a relatively-short useful life. You plan to replace the device regularly. 

So data center shells are one thing, GPUs and other accelerators possibly quite another. The shell might be depreciated over 15 to 25 years. Processors might be depreciated over six years. 

So processors are akin to a subscription: they are capital, but also capital that must regularly be replaced. 

In other words, processors are formally capex, but also resemble operating expense. So unless you believe artificial intelligence essentially is a fad, the continuing demand for processors is at least the size of the current and projected installed base. 

That assumes, of course, that revenue earned by using all that infrastructure produces a profit. 

So a GPU cluster's economics are a race between two curves:

• The depreciation curve (how fast the asset's value erodes)

• The monetization curve (how fast the cluster recovers revenue against its capex).

The relevance for current debates about chip infrastructure are only partly about hyperscaler investment levels (whether they are overestimating demand). 

If demand exists, then processor capex is essentially a recurring function, and hence similar to a subscription. 

Assuming there is demand, infra outlays then have to be compared to monetization curves. 

If the latter grows faster than the former, there is no real problem. And there lies the friction and uncertainty.

Nvidia's shift to an annual product release schedule creates a two-year to three-year frontier processor obsolescence.

Where Hopper (2022), Blackwell (2024) and Rubin (2026) releases happened on a two-year schedule, Rubin Ultra (2027) is headed for a annual cycle.

Some will argue that Blackwell's efficiency gains over Hopper are large enough that older hardware becomes non-competitive for frontier training within 18 months to 36 months. 

But data center depreciation schedules for such gear now sit at six years. 

That gap between "accounting life" and "economic life" is at the heart of skeptical views on AI capex. 

The monetization picture also is dynamic. 

Per-unit prices are collapsing fast, as inference costs have dropped roughly 1,000 times in three years, with GPT-4-equivalent performance costing about $0.40 per million tokens in 2026 versus $20 in late 2022.

Goldman Sachs researchers project total token consumption growing 24 times between 2026 and 2030, so volume growth offsets price decay.

Total inference spending grew 320 percent even as per-token costs fell roughly 280-fold.

So the real question for any given cluster isn't "is the accounting depreciation schedule realistic" — it's whether cumulative revenue recovery clears the capex bar before the hardware's real economic obsolescence catches up to it. 

Is it "Different This Time?"

If you worked at any venture-capital-funded startup during the dot-com bubble, you might recall hearing one of the most-dangerous phrases in equity markets: “it’s different this time.

Maybe you recall being told “you don’t get it,” or any of the variants of the idea that traditional valuation metrics no longer apply:

• “it’s a new era” 

• “the old rules no longer apply”

• “valuations don’t matter.” 

In fact, there were all sorts of phrases suggesting old investment rules were essentially useless:

• "It’s a New Economy" (The claim was that the internet had fundamentally changed how economic value gets created, so old metrics didn't capture it anymore)

• "Get big fast"/ "Get large or get lost" (market share and growth mattered more than profit)

• "Eyeballs" and "eyeballs over earnings" (attention and traffic became a proxy for value)

• "First-mover advantage" (burning cash to grab a market before anyone else, on the theory that being first was worth more than being profitable)

• "Clicks, not bricks" (dismissing physical/traditional retail as legacy infrastructure that the internet would simply route around)

• "Old economy" (anything industrial, physical, or profit-focused got dismissed as backward-looking)

• "Network effects" (often invoked loosely to claim that a company's value would compound in ways traditional accounting couldn't measure)

• "Burn rate is a feature, not a bug" (the idea that losing money fast was actually evidence of aggressive growth, not weakness)

• "P/E ratios don't matter anymore" (only "eyeballs" or "mindshare"). 

One hears those things in asset bubbles. We heard it quite a lot during the dot-com or internet bubble. 

Era / Bubble	Example of “This Time Is Different” Reasoning	Outcome / Context	Source Link
Tulip Mania (Netherlands, 1630s)	Speculators treated rare tulip bulbs as a new, superior form of wealth whose prices could only rise; traditional notions of intrinsic value were set aside.	Prices collapsed ~99% in 1637.	NST article on historical examples
South Sea Bubble (UK, 1720)	Investors believed a new trading monopoly would unlock unprecedented riches, justifying extreme share prices detached from fundamentals.	Shares rose dramatically then collapsed; widespread losses.	NST historical summary
Roaring Twenties / 1929 Crash (US)	Belief in a “new era” of endless prosperity driven by technology, consumerism, and industrial profits; margin buying and high valuations were rationalized as sustainable. Business Week noted the recurring “new era” illusion.	Market crash; prolonged depression.	FT “New eras, same bubbles”; Irish Times on 1929 parallels
Japanese Asset Bubble (late 1980s)	Decades of strong growth led some to predict Japan would eclipse the US economy under unique structural advantages that rendered prior cyclical risks obsolete.	Real-estate and equity collapse; multi-decade stagnation (“Lost Decades”).	Financial Post on historical peaks
Dot-Com / Internet Bubble (late 1990s–2000)	The internet was said to transform the economy so thoroughly that traditional P/E ratios and profitability no longer applied; companies with little or no revenue received multi-billion valuations.	Nasdaq fell ~78% peak-to-trough; many pure-play firms failed.	NST on Dot-Com narrative; GMO / Templeton reference
US Housing Bubble (mid-2000s)	Widespread conviction that national real-estate prices “could never fall” and that new financial engineering (securitization, subprime lending) had permanently reduced risk.	Housing crash; global financial crisis of 2008.	NST housing example; Reinhart-Rogoff framework
AI / Tech Boom (2020s, ongoing discussion)	Claims that AI is so transformative, or that hyperscaler balance sheets and cash flows make the cycle fundamentally safer than prior tech bubbles, so that elevated valuations and massive CapEx can be sustained under new rules. Parallel debates occur around crypto.	Still unfolding; critics note the classic narrative while supporters emphasize differences in profitability and financing.	Grantham quote coverage; GMO AI analysis

But there are important differences between dot-com financing (venture capital; public equity IPOs of unproven firms; vendor financing) and artificial intelligence financing in the compute infrastructure part of the value chain. 

AI infrastructure spending is dominated by the “Magnificent seven” hyperscalers (Microsoft, Alphabet/Google, Amazon, Meta, and often Nvidia, Apple, Tesla; sometimes Oracle). 

These firms generate enormous free cash flow and operating profits from established, diversified businesses (cloud, advertising, e-commerce, software, chips). 

Early-to-mid phases of the buildout were largely self-funded from internal cash flows and strong margins rather than pure external speculative capital. 

That provides some resilience to credit tightening, which stopped the dot-com bubble in its tracks. 

Many recipients had limited or no profits, weak balance sheets, and business models centered on “eyeballs” or future monetization. When capital markets tightened or growth disappointed, cascading failures ensued; overbuilt fiber and equipment sat underutilized for years.

Venture capital still plays a large role in pure-play AI startups, and there are circular elements (investments, capacity commitments, and vendor-like arrangements involving Nvidia, OpenAI, CoreWeave, etc.). 

Still, the bulk of physical infrastructure investment by hyperscalers has been anchored by operating profits.

Operational (internal cash flow) financing provides more bubble resilience than pure VC financing:

• Hyperscalers can slow spending, absorb write-downs or lower returns on data centers/chips without bankruptcy and continue funding core non-AI businesses. Lower equity valuations, delayed returns, or margin pressure can happen, but there is much less danger of widespread defaults.

• VC funding is less stable. When sentiment shifts, capital can dry up quickly, leading to mass failures of non-viable firms.

• Hybrid/circular financing sits in between: it can inflate activity and create reflexive loops (spending supports valuations that support more financing), but is anchored by solvent, profitable buyers.

That reliance on operating earnings rather than venture capital reduces the probability of a broad credit crunch or mass bankruptcies.

A full dot-com-style multi-year tech bear market with trillions in equity destroyed is less likely precisely because the financing base and profitability differ. 

There are lots of real risks from energy constraints, component costs, regulatory issues, monetization and overbuilding. 

So even if every bubble has some common elements, that does not mean they are identical. Despite the risks, a devastating financing crash on the pattern of the internet bubble seems less likely.