UBS elaborates on the prosperous prospects of AI infrastructure: NVIDIA holds a trillion-dollar revenue opportunity, and data center revenue is expected to double again?

NVIDIA not only delivered several earnings reports that exceeded expectations, but its subsequent growth prospects may surpass market imagination.

On Tuesday, UBS analyst Timothy Arcuri stated in a client report that NVIDIA holds "tens of gigawatts" of AI infrastructure projects, conservatively estimated to be worth over $1 trillion. If this prediction comes true, NVIDIA's data center revenue is expected to reach $400 billion annually within 2-3 years, nearly double the current market expectations.

Last week, UBS analysts Steven Fisher, Amit Mehrotra, and others pointed out that the construction boom of AI data centers is not expected to manifest in the real economy or provide structural tailwinds until the second quarter of 2026.

Analysts believe this view indicates that the prosperity of AI infrastructure has surpassed the cyclical concept and shifted towards an exponential infrastructure expansion model. The construction of AI data centers in the digital age can be likened to the "New Deal" of the 1930s, except that it is not building highways and dams, but GPUs and megawatts of power. It is reshaping America's infrastructure, but the dominant players are large tech companies.

Additionally, UBS also analyzed that during the conference call, the GB200 shipment data was more about reassuring investors that the rack issues have been resolved, while NVLink is driving explosive growth in network business, and it is expected that improvements in Blackwell's profitability and cost reductions will push gross margins back to around 75% by the end of fiscal year 2026.

Trillion-Dollar Potential Projects: An Overlooked Growth Engine

According to UBS analysis, the "tens of gigawatts" potential AI infrastructure projects mentioned by NVIDIA in the latest earnings call, based on a conservative estimate of 20 gigawatts and the company's pricing range of $40-50 billion per gigawatt, present a total revenue opportunity of at least $1 trillion.

More critically, the timeline is key. UBS believes these projects will be launched within 2-3 years, meaning NVIDIA could achieve approximately $400 billion in "visibility" data center revenue annually—this figure is nearly double UBS's expectation for its data center revenue in fiscal year 2026 ($233 billion).

UBS analyst Timothy Arcuri pointed out: this is clearly exciting as Crusoe alone, a digital infrastructure developer, has about 20GW of project pipeline, and this is just one of many developers. This finding should alleviate investors' concerns about the sustainability of growth.

The Truth Behind GB200 Shipment Data

The GB200 rack data provided by NVIDIA during the earnings call has left investors puzzled. The company stated that major hyperscale customers are deploying nearly 1,000 NVL72 racks per week on average, equivalent to 72,000 Blackwell GPUs, and are expected to further increase capacity this quarter.

Taken literally, this means that each hyperscale customer's GPU utilization rate is approaching 1 million per quarter—far exceeding most consensus expectations. **However, UBS believes that NVIDIA is not conveying revenue "utilization rates," but rather trying to reassure investors that the GB200 rack issues have been resolved, and a large number of racks are flowing from ODM and OEM manufacturers to customers **

UBS stated, "We will not overinterpret these numbers, as we believe the company is more interested in explaining the situation rather than providing quantitative data."

NVLink Drives Explosive Growth in Network Business

NVIDIA's network revenue grew to approximately $5 billion in the first fiscal quarter, a quarter-over-quarter increase of 64%, with $1 billion coming from a significant increase in NVLink revenue. UBS believes this is almost entirely related to the increase in shipments of the GB200 NVL72 rack-scale system, each of which contains 72 GPUs in the NVLink domain, representing a significant upgrade compared to the HGX system's maximum of 8 GPUs.

As NVIDIA recognizes NVLink revenue separately for the NVL72 system, network revenue is expected to more closely track NVL72 rack shipments, although there may be a slight lag.

Gaming Business Recovery: Channel Restocking Rather Than AI Reallocation

The significant improvement in gaming revenue in the first fiscal quarter (a nearly 50% quarter-over-quarter increase, far exceeding expectations) has raised investor concerns that the RTX 50 series graphics cards are being diverted to other markets.

UBS believes any such diversion is extremely limited for several reasons: the supply of Blackwell-based RTX GPUs in the gaming channel remains too constrained; the RTX 50 series GPUs are based on PCIe and do not support NVLink for expansion; NVIDIA initially prioritized data center applications, somewhat limiting the supply of Blackwell to the gaming channel. The growth in the first fiscal quarter was primarily driven by channel restocking after severe supply shortages.

Clear Path to Gross Margin Recovery

The overall improvement in Blackwell's profitability and declining costs remain the main drivers for gross margins to return to around 75% by the end of fiscal year 2026. UBS believes the GB300 will play an important role, with NVIDIA potentially recognizing a small amount of revenue in the second fiscal quarter, with significant volume expected in the third fiscal quarter.

In the long term, value-based pricing remains a key function of NVIDIA's gross margins, involving a combination of hardware and software, with Dynamo, released at the GTC conference, being a typical example that can enhance inference speed on NVIDIA hardware by over 30 times.

Overall, NVIDIA's financial report and UBS's visibility into the multi-year AI infrastructure boom indicate that this boom is more about exponential infrastructure expansion rather than a product cycle. UBS analysts Steven Fisher, Amit Mehrotra, and others pointed out last week that the construction boom of AI data centers is expected to manifest in the real economy and provide structural momentum only by the second quarter of 2026