# data center revenue — X 热门讨论 (2026-09-16 04:38 UTC)

## @HedgeyeComm (Andrew Freedman, CFA 🦅) · 09-16 01:28 · ♥38 ↻4 💬10 The Agent Layer Is the New Operating System https://x.com/HedgeyeComm/status/2100033989159600265

## @MilkRoadAI (Milk Road AI) · 09-15 20:13 · ♥33 ↻8 💬5 The world is about to need more computing power than ever before and this chart shows how quickly that demand is turning into revenue (Save this).

NVIDIA’s data center revenue increased from approximately $7 billion in fiscal 2021 to $194 billion in fiscal 2026, representing an almost 28 fold increase in five years.

BMO Capital Markets expects that figure to reach $369 billion in fiscal 2027 and $559 billion in fiscal 2028, which would nearly triple the company’s data center revenue in only two more years.

The AI computing demand is expanding beyond the initial training of large language models and is in post training, reasoning, video generation, scientific computing and continuous inference.

Traditional software performs a relatively fixed amount of computation when a user completes a task, whereas reasoning models can consume more computing power by generating thousands of internal tokens before producing a final answer.

AI agents increase that demand further because they can continuously search, write code, analyze documents and interact with other systems without waiting for a new human prompt after every step.

NVIDIA’s data center revenue has already reached $89 billion in a single quarter, representing growth of 117% from the previous year and approximately 93% of the company’s total revenue.

This suggests that the growth shown in the chart is not based entirely on distant expectations because the company is already generating enormous revenue from the expansion of AI infrastructure.

The transition from Blackwell to Rubin could further expand the market because NVIDIA says Rubin can reduce inference costs by as much as 10 times and train certain mixture of experts models with four times fewer GPUs.

Although greater efficiency may appear bearish for hardware demand, lower computing costs can make many more AI applications economically viable and cause total usage to grow faster than the cost per task declines.

This is similar to what happened with cloud computing, because cheaper and more accessible computing did not reduce demand but instead enabled companies to create far more software and process substantially more data.

The strongest version of the bull case is therefore not simply that companies will train progressively larger models, but that billions of people and software agents could eventually consume AI inference continuously.

when that happens, every improvement in model intelligence could create new applications, while every reduction in token costs could encourage users to generate more tokens.

Training started the boom but inference and agents could make it much bigger.

We’re already positioned around that shift at Milk Road. If you want to see exactly what we’re buying, check out the link below.

https://t.co/thIhK9ZH4E https://x.com/MilkRoadAI/status/2099954784946290843