Goldman Sachs: Compute Scarcity Grants Valuation Premium to Hyperscale Cloud Providers
nashnova research
Goldman analyst Rich Privorotsky argues that powered-up compute — not model leadership — is the most durable edge in the AI cycle, and hyperscalers who control this scarce resource deserve a valuation premium.
Why does Goldman say "compute matters more than models"?
Leadership in frontier AI models can change every quarter. Polymarket data on "best model by end-2026" shows Anthropic's lead collapsing within days, with Google surging to a 43% probability lead.
This means → no single model company has a real moat. Built and powered-up compute is the hard asset that doesn't get disrupted every quarter.
In plain terms = models are like smartphone apps — a new version can unseat the leader overnight. Data centers are more like power plants: whoever builds first earns first, and latecomers face years of waiting just to get grid connections.
What happened with OpenAI's "$20 billion revenue shortfall"?
The Financial Times reported that OpenAI disclosed annualized revenue near $50 billion to investors — roughly $20 billion below previously circulated figures. The Nasdaq fell on the news.
Privorotsky calls the gap largely an accounting-definition issue: whether partner and cloud sales are booked gross (as Anthropic does) or net (as OpenAI does) shifts the headline number dramatically.
What concerns him more is the market reaction itself: it reveals heavily concentrated long positioning, with investors growing increasingly sensitive to any negative signal.
This means → the real risk isn't whose accounting is more reasonable — it's that everyone holds the same basket of assets, so any tremor triggers an outsized stampede.
Why is compute scarce — isn't the bottleneck power, not chips?
Goldman data shows roughly half of U.S. data centers planned for 2026 will be cancelled or delayed. Of roughly 16 GW of planned capacity, only about 5 GW is actually under construction.
The core bottleneck is power supply, not chip availability.
Goldman's commodities team forecasts that by 2030, 25% of data centers will run on self-generated power ("behind-the-meter generation") — up from effectively zero in 2025.
In plain terms = the hardest part of building a data center isn't buying GPUs — it's plugging them in. The grid queue is so long that operators are resorting to building their own power plants.
How does this "power paradox" lift hyperscaler valuations?
The power bottleneck depresses valuations of data-center supply-chain stocks, but the same constraint makes already-built, already-powered compute increasingly precious.
Hyperscalers invest most aggressively in compute infrastructure and can therefore capture premium returns from the scarcity.
This reflects a structural dynamic: the force that constrains supply is the same force that inflates the value of existing capacity.
How long can this logic hold — what is the key checkpoint?
Goldman's previously stated precondition: token demand growth must continue to outpace the decline in token pricing.
This means → if AI inference usage grows fast enough, total hyperscaler revenue can still support ever-expanding capital expenditure even as unit prices fall.
In plain terms = think of ride-hailing — fares keep dropping, but as long as riders grow faster, platform revenue still climbs. Whether "volume can outrun price" is the core observation window for judging if the hyperscaler valuation premium will be vindicated.
市场有风险，内容仅供研究参考，不构成投资建议。