30% of 2028 Data-Center Capacity Puts NVIDIA’s Growth at Risk

By Tom Whitfield

About 30% of the U.S. data-center capacity planned for 2028 could be at risk, according to a UBS estimate reported by Reuters. The estimate applies to proposed capacity additions, not capital spending companies have already committed. For NVIDIA, the warning matters because its growth depends on those data centers reaching operation on schedule.

What puts planned capacity at risk?

Power, transmission, labor, cooling, financing and local approvals sit between AI plans and operating capacity. A delay at any link can push revenue out while commitments and payment exposure remain.

Digital Realty has 1.5 gigawatts under construction, and 63% of it is pre-leased. Equinix lists projects only after permitting hurdles clear and power is largely secured. The contrast favors operators with sites that can move from planning into construction.

How committed are NVIDIA’s supply plans?

During NVIDIA’s fiscal second quarter of 2027, Data Center revenue reached $89.023 billion, up 117% year over year. Management guided third-quarter revenue to $108.0 billion plus or minus 2%.

Supply obligations rose to $279.00 billion, mostly tied to memory purchases for Vera Rubin production. Those dollars are committed, whereas the capacity UBS identified remains proposed. The mismatch is the core risk: NVIDIA has committed much of its supply while some buildings meant to house the chips remain plans.

CFO Colette Kress said, “This is a supply-constrained outlook.” Management named data-center availability, power, land and shell capacity, cooling, construction and labor as constraints outside NVIDIA’s own factories.

Broadcom chief executive Hock Tan said, “To deploy an AI data center at scale, land, power and shell are major constraints and dictate timing.” He also cautioned that Broadcom was not asserting that all 30 gigawatts would enter production in 2027 and 2028. The near-term constraint sits outside NVIDIA’s factories, where land, power, shells and labor must arrive on schedule.

How has the market repriced the risk?

The source places the AI selloff against a possible $2.5 trillion Wall Street loss, although its analysis focuses on a separate capacity constraint. Chip, equipment and power shares recorded sharply different monthly moves:

• NVIDIA fell 21.4%, from $237.47 on October 7 to $186.64 on November 6, leaving the stock up 0.31% for the year.
• Broadcom fell 22%, Eaton fell 20.78%, and Constellation Energy fell 16.57%.
• Data-center landlords held up better. Equinix slid 1.85%, while Digital Realty fell 3.47%.

The relative performance appears to reward companies that already own permitted, powered sites. Investors distinguish between proposed AI demand and capacity that has cleared the main development hurdles.

What exposure remains beyond chip sales?

NVIDIA’s guarantee obligations to AI cloud and data-center partners cap at $108.5 billion. Days sales outstanding, which measures how long customers take to pay, rose from 45 days to 60 because of extended payment terms.

The company has invested nearly $50 billion in frontier AI labs. It also expects AI labs it may support with its balance sheet to make up roughly a quarter of its business next year. These links make delivery timing more than a semiconductor issue, because delayed facilities can leave financing and customer commitments outstanding.

How could memory costs change the economics?

Non-GAAP gross margin was 75.0% last quarter, and guidance calls for 74.0% plus or minus 50 basis points. Management expects margins to bottom in the 71% to 72% range in the fourth quarter as memory prices climb.

Kress said memory scarcity is driven in large part by the AI buildout itself. For the fiscal year ending January 2028, analyst revenue estimates range from $506.483 billion to $778.471 billion.

If 30% of planned 2028 capacity shifts later, NVIDIA could retain rising costs and payment risk while some growth moves out. Jensen Huang argues that NVIDIA’s computing platform is transferable and can be reassigned to other customers, but that flexibility needs finished buildings with power ready for hardware. Transferability buys time, but it does not create powered capacity.

What happens next?

NVIDIA’s next earnings report will test whether the $108.0 billion third-quarter guidance held while memory costs climbed. On the power side, Constellation Energy’s chief executive expects a PJM response around November and cleaner co-location rules in the first to second quarter of 2027.

AI demand can stay strong while chip investors wait for concrete, copper and transformers to catch up. The next test is delivery rather than demand: NVIDIA must defend $108.0 billion of quarterly guidance while memory prices rise and operators turn proposed 2028 capacity into powered, permitted sites.