Morgan Stanley: Generative AI Inference-Stage ROIC Could Reach 25%-50%
Claire Weston
Morgan Stanley estimates that generative AI's inference stage can deliver incremental ROIC of 25% to 50%, arguing the market has underestimated hyperscalers' AI capex returns and maintaining bullish ratings on Amazon, Google, Microsoft, and Meta.
What was the market worried about?
Over the past year, surging AI capital spending by the four major hyperscalers left investors asking: can they actually earn this money back?
That doubt weighed directly on internet and cloud-sector valuations.
Morgan Stanley analysts Brian Nowak, Stephen C Byrd, and Adam Wood built three bottom-up quantitative frameworks to answer — and their conclusion is that returns are not just positive, but substantial.
What do the three frameworks show?
Framework 1: GPU leasing (IaaS). A single 1 GW data center runs roughly 410,000 Nvidia GB300 GPUs at 75% utilization and $8.50/hour. That yields about $22.9 billion in annual revenue per GW, incremental ROIC of roughly 31%. Across a $7–$10 price range, ROIC spans 23% to 39%.
Framework 2: API services (own compute). Covers Google Gemini, Meta API, and similar models. With 65% of compute allocated to inference and blended token pricing at $1.75 per million tokens, annual revenue reaches about $30.4 billion per GW, incremental ROIC of roughly 46%. As token pricing moves from $1 to $2.50, ROIC ranges from 19% to 63%.
Framework 3: API services (leased compute). Because operators pay third-party compute rental, returns are the lowest — baseline NOPAT margin around 25%. But at token pricing as low as $1 per million, the business turns loss-making (margin roughly −9%). This means → token pricing is the make-or-break variable for this model.
Why is the return gap so wide?
The core difference across the three frameworks is who owns the compute. Operators with their own data centers (Frameworks 1 and 2) spread fixed costs and keep wider margins; those renting compute (Framework 3) face an extra layer of rental expense that compresses profit.
In plain terms = building a house and renting it out versus subletting someone else's house — the infrastructure owner captures a structurally larger share.
This reflects a deeper signal: in the AI era, data-center capacity itself is a scarce asset. Whoever holds the compute holds the deeper moat.
Will strong returns invite more competition?
Morgan Stanley says clearly: healthy returns will attract more entrants, including newcomers and open-source alternatives.
Yet the report argues that rising competition actually reinforces two things: the value of continuous product and compute-efficiency innovation, and the scarcity premium on existing infrastructure. This means → the first-mover advantage of scaled players is unlikely to erode easily — it may widen under competition.
What does BlackRock think — and where do both firms converge?
BlackRock also favors AI infrastructure, arguing that cheaper AI models will reshape how profits are distributed but will not undermine the long-term AI investment case.
BlackRock frames this earnings season's key question as: where along the value chain will profits accumulate?
Both firms converge on the same validation point: can hyperscalers deliver scaled inference-stage returns while capex continues to expand? Put simply = the money has been spent — the next test is whether revenue keeps pace.
Content is for reference only, not financial advice.