Legendary Short Seller Chanos: AI Cloud Service Providers Are Leasing Middlemen, with Compute ROIC in Single Digits
Miles Bennett
Legendary short-seller Jim Chanos clashed with bull-side fund manager Val Zlatev over the AI capex boom — Chanos called compute lessors low-return financial intermediaries with single-digit pre-tax ROIC, while Zlatev countered that demand is real and trackable, and Nvidia trades at just 15× forward earnings.
Where is the depreciation bomb hiding?
Chip and equipment vendors book revenue the moment they ship. Hyperscale cloud providers, by contrast, capitalize AI capex and spread it over 4–7 years of depreciation. This means → today's income statements look clean, but the costs are deferred, not avoided.
Chanos cited the 1998–2000 dot-com bubble: S&P 500 operating profit rose ~30% over two years, then collapsed 40% in 2001 as orders dried up but depreciation kept hitting. In plain terms = the shovel-sellers got paid first; the gold-diggers' books only looked good because the shovels hadn't been fully expensed yet.
He stressed that spending on chips and data-center construction currently sits in "construction in progress" on the balance sheet. Once those assets go live and depreciation begins, the profit impact will be massive — even assuming a conservative 10-year physical GPU lifespan.
What business are compute lessors actually in?
Chanos offered a blunt verdict: buy chips from Nvidia, lease someone else's data center, sublease compute to Microsoft or Google — "You're an equipment-leasing company. You're not a tech company; you're a financial company."
He disclosed the real returns on compute infrastructure today: even in a severely supply-constrained, seller-favorable market, pre-tax ROIC runs just 5%–8%, all single digits. This means → if returns are this thin when demand far exceeds supply, margins will only compress as capacity comes online.
Zlatev partly agreed but noted that not all emerging cloud providers are identical — Lambda Labs, for example, derives roughly 40%–50% of revenue from spot pricing on real-time inference workloads, giving it more upside flexibility when GPU spot prices spike.
Why do the bulls say "this time is different"?
Zlatev's first argument: AI demand today can be tracked in real time through underlying token usage, unlike 1999 when investors had to rely on CFO promises. Since January, compute spot prices have risen 40%–50%, and "even 6-, 7-, 8-year-old GPUs are seeing higher rents."
His second argument is valuation: Nvidia trades at roughly 15× its expected 2027 EPS; Broadcom, after its recent pullback, sits at about 12× 2028 expected earnings. In plain terms = this is not the 1999 Cisco-at-160× era — the sector is nowhere near extreme bubble territory.
Yet Zlatev himself acknowledged one core risk: AI scaling laws are empirical patterns, not physical laws. If a non-Transformer architecture emerges that dramatically cuts unit compute costs, the logical foundation of the current compute arms race collapses.
Why can't memory-chip supply keep up?
Zlatev highlighted a physical bottleneck on the supply side: semiconductor equipment companies — ASML, Applied Materials and peers — are constrained by supply-chain complexity and can grow shipments by at most 30%–35% per year. This means → even if demand explodes, capacity cannot double quickly, and tight supply will persist longer than the market assumes.
The market currently values memory companies at just 6–7× forward earnings, pricing in a cyclical downturn within 6–9 months. Zlatev argues this is a mispricing — this demand peak will be higher and longer-lasting than any in the past 25 years.
Does putting data centers in space make sense?
Chanos dismantled Musk's space-datacenter vision on commercial logic: the biggest costs in space are radiation hardening and heat dissipation. Data-center hardware breaks frequently — on Earth you send a technician to swap parts; in space you launch a rocket.
He added that SpaceX's launch business is still operating at a loss. This reflects the fact that even the company with the lowest launch costs has not made the economics of space maintenance work.
Zlatev explained the demand-side reasoning behind Musk's thinking: Musk projects the world will need 1 terawatt (1,000 gigawatts) of compute within the next few years, while tech giants' $750 billion in capex currently translates to only about 15 gigawatts — as long as scaling laws hold, the compute gap is exponential.
AI capex — bubble on the brink, or a real demand-driven expansion?
BULL
Demand is verifiable in real time
Token usage is trackable; compute spot prices are up 40%–50% this year.
Valuations are far from extreme
Nvidia trades at just 15× forward PE — nothing like Cisco at 160× in 1999.
Supply has a physical ceiling
Equipment makers can grow shipments by at most 30%–35% a year.
BEAR
The depreciation bomb hasn't detonated
4–7 year depreciation means peak cost impact is still ahead — 2001's profit crash is the template.
Compute middlemen earn razor-thin returns
ROIC is just 5%–8% even in a shortage; margins will compress as supply arrives.
Scaling laws are not laws
A new architecture that disrupts Transformers would collapse the arms-race thesis overnight.
In plain terms = the bulls are betting demand growth outpaces depreciation coming due; the bears are betting depreciation arrives before demand is monetized. The ultimate proof point is when 'construction in progress' starts flowing into the income statement at scale.
Content is for reference only, not financial advice.