Citi: NVIDIA's Revenue per GW to Rise from $18B to $40B
nashnova research
Citi reiterated a Buy on Nvidia with a $315 target, projecting per-GW revenue opportunity from $18 billion to $40 billion across three chip generations — a sign Nvidia is shifting from selling GPUs to selling entire AI factories, multiplying what it earns per project.
Per-GW revenue more than doubles — where does the money come from?
Citi maps three generations: Hopper $18B → Blackwell $25B → Rubin $40B, all measured as revenue opportunity per gigawatt of compute.
GW (gigawatt) measures how much power a data center draws — more power means more compute, and more products Nvidia can sell into each site.
This means → growth is not just GPU price hikes. Nvidia now bundles CPUs, networking gear, and AI-factory infrastructure into a single solution, capturing more spend per project.
In plain terms = it used to sell the engine alone; now it sells the chassis, gearbox, and dashboard too — so the price per vehicle doubles.
How big are the orders, and who is buying?
One frontier AI lab has directly contracted 2.6 GW of Nvidia AI infrastructure, with delivery scheduled before 2028.
Multiple cloud providers and emerging cloud players have signed indirect contracts totaling more than $180 billion.
This means → demand is not a slide-deck forecast — it is booked orders with multi-year delivery timelines, locking in future revenue.
What does a $150 billion buyback signal?
Nvidia added $150 billion in new share-repurchase authorization, bringing total buyback capacity through FY28 to $235 billion.
The company's current market cap sits at roughly $5.52 trillion; the $235B program is about 4.3% of that.
This reflects strong management confidence in sustained free-cash-flow growth — you don't commit that much capital to buybacks unless you expect the cash to be there.
Why is Nvidia moving into AI safety?
Nvidia launched the Open Agent Safety Platform, targeting security, compliance, and governance for AI agents — autonomous programs that execute tasks on their own.
The platform includes OpenShell, an open-source runtime, and Sentry, a hardware monitoring system running on the BlueField-4 DPU — a chip purpose-built for data-center security tasks.
This means → Nvidia's business boundary has expanded from "selling compute" to "governing how compute is used," carving into customers' security and compliance budgets.
Open models are surging — what does Nvidia gain?
Open models — AI models anyone can download and run — now account for roughly 75% of token generation, up from about 40% a year ago. Total token consumption has grown 25× year-over-year.
In plain terms = more people running open-source models means more AI applications, and every new application consumes more compute — which ultimately means buying more Nvidia chips.
This reflects a reinforcing loop: more open models → more applications → more compute demand → more Nvidia revenue.
What is the key verification point for this story?
Whether per-GW revenue actually tracks the $18B → $25B → $40B trajectory from Hopper to Blackwell to Rubin is the core metric to watch.
If it does, it validates Nvidia's "full-stack platform" strategy — chips, software, and safety, all under one roof — as genuinely lifting the value of each project.
If it falls short, the most likely reason is that customers chose rival suppliers for the non-GPU layers, meaning Nvidia's bundling strategy failed to capture the full budget.
市场有风险，内容仅供研究参考，不构成投资建议。