Jensen Huang did not hedge. Speaking to reporters on the sidelines of an AI summit convened by King Charles III at Dumfries House in Scotland on September 17, Huang said: “I expect Nvidia to sell twice as many chips this next year as we do this year.” That is a unit-volume forecast, not a revenue guide, and it lands at a peculiar moment: Nvidia’s August 26 outlook called for about 70% revenue growth in fiscal 2028, and CFO Colette Kress repeatedly framed that outlook as supply constrained, saying customer forecasts point to growth doubling next year.
So Huang is effectively saying the constraint gets solved. Bulls and bears on the desk agree that is the right question. They disagree violently on the answer.
The Bull Case
The optimists point to visibility. Huang has argued Nvidia has far greater supply chain visibility now than a year ago, when a single lab was driving the buildout. Today, he said, multiple frontier labs are scaling in parallel alongside an open-model ecosystem and physical AI coming online.
The demand base is genuinely broader. On the August 26 earnings call, Kress told analysts that customer forecasts “point to our growth doubling next year,” while the ~70% outlook reflects supply constraints rather than any demand ceiling. On that reading, Huang’s doubling is not a stretch target. It is what happens when the packaging and memory log-jam clears.
The packaging side has a credible path. Industry estimates have put TSMC’s CoWoS capacity around ~80,000 wafers per month in 2026, with projections toward roughly 120,000 to 140,000 by end-2026 and around 170,000 by end-2027. Optical interconnect suppliers, notably Coherent and Lumentum, stand to benefit as higher GPU shipment volumes push data rates and optical module adoption upward across both scale-out and scale-up architectures.
The Bear Case
The skeptics start with memory, and the numbers there are uncomfortable. During Nvidia’s fiscal second-quarter earnings call on August 26, Kress said the company is witnessing “extreme pricing conditions in memory,” with price increases already exceeding internal forecasts and heading higher into next year. Nvidia’s own SEC filing for that quarter said its supply and capacity commitments jumped to $279 billion from $119 billion the prior quarter, and the company said the increase was primarily related to the procurement of memory. A company locking in that kind of forward spend is not signaling that scarcity resolves quickly.
The margin trajectory says the same thing. Nvidia guided gross margin to about 74% in fiscal Q3, bottoming in fiscal Q4 at 71% to 72%, and then settling at 72% to 73% in fiscal 2028 as price increases take effect. The key point is that Nvidia is explicitly modeling a margin trough before any recovery, which is consistent with elevated memory input costs persisting into fiscal 2028.
Then there is the packaging ceiling. Even with TSMC’s aggressive expansion, investors continue to model CoWoS as a binding variable, because advanced packaging, bonding, and memory integration capacity can lag demand even when wafer starts are available. If that holds, packaging constraints remain a gating factor through 2027, not transistor scarcity.
What Investors Are Missing
The unit-versus-revenue distinction cuts both ways, and most of the discussion misses that. Nvidia sells far more than data center GPUs. It also sells networking switches, optical and networking silicon, and client and embedded chips, including the custom Nvidia Tegra processor widely reported to be inside Nintendo’s Switch 2. A Blackwell rack system and a laptop GPU each count as one chip sold. They do not count as the same dollars. A doubling in units could therefore coexist with revenue growth well below 100%, which is precisely what the formal guidance implies.
The cleaner read-through from Huang’s comment may belong to the infrastructure layer rather than Nvidia itself. Goldman Sachs Research projected US data center power demand rising from 41 gigawatts in 2026 to 66 gigawatts in 2027. Every chip shipped into a rack needs power delivery and cooling that does not depend on which architecture wins.
Stocks to Watch
Nvidia (NVDA) owns the upside if the unit doubling materialises, but the margin trough arrives first. The company reported $96.2 billion in fiscal Q2 revenue, up 106% year over year, with data center at $89.0 billion. The stock has already priced considerable optimism.
TSMC (TSM) is a key physical gating factor via CoWoS. CoWoS capacity is growing fast, but it is still widely modeled as tight versus demand into 2027. If the doubling happens, TSMC’s advanced packaging revenue compounds with it.
Broadcom (AVGO) benefits whether or not Nvidia executes. On its fiscal Q3 2026 earnings call, Broadcom said it expects fiscal 2026 AI revenues of roughly $58 billion and said it has secured supply to support AI semiconductor revenues of approximately $115 billion in fiscal 2027. Custom silicon demand is largely independent of Nvidia’s unit count.
Vertiv (VRT) is the constraint no one tracks on chip-focused desks. Data center power density is not a cycle. Every new generation of AI silicon requires more power and more cooling, making Vertiv’s addressable market essentially co-extensive with the AI infrastructure buildout itself, with no dependency on picking which chip architecture wins.
Marvell (MRVL) and Coherent (COHR) sit at the optical interconnect layer that scales directly with GPU shipment volumes, as rising per-chip data rates push optical module adoption deeper into the cluster architecture.
