Nvidia's 70% Growth Guidance Comes With Its Own Asterisks

Nvidia told Wall Street this month that it expects roughly 70% revenue growth in its next fiscal year, a figure Jensen Huang repeated at the Goldman Sachs Communacopia + Technology Conference in San Francisco on September 10. The number itself was not new. Nvidia's chief executive and finance chief first issued it as guidance on the company's Q2 fiscal 2027 earnings call on August 26, alongside a quarterly report of $96.22 billion in revenue against a Wall Street estimate of $92.07 billion. What matters for how that guidance should be read is less the size of the number than the format Nvidia chose to deliver it in, and the specific words its executives used to qualify it.

Nvidia has historically confined itself to single-quarter guidance. The August 26 forecast marked the company's first year-ahead revenue projection, implying roughly $680 billion to $700 billion in fiscal 2028 revenue against about $400 billion expected this year. Wall Street's own pre-guidance models had penciled in growth closer to 44%, and consensus estimates reportedly moved only to around 63% after Nvidia's disclosure, still short of the company's figure. Fortune and Yahoo Finance have both framed the switch to annual guidance as a deliberate response to skepticism about an AI spending bubble, an unusual case of a company changing its own disclosure practice to argue with the market rather than simply report to it.

The load-bearing qualifier came from how Huang and Chief Financial Officer Colette Kress described the ceiling on that number. Huang told the Goldman audience that Nvidia's demand is running well ahead of 70%, and that supply is what allows the company to commit to that figure with confidence. Kress has indicated that absent manufacturing and component constraints, year-over-year growth could more than double. That framing matters because it converts the headline growth rate from a forecast of what customers will buy into a statement of what Nvidia's own production lines can build, a distinction that shifts where the risk in the number actually sits. Huang pointed to a concrete data point behind that claim: month-over-month order growth of 27% for systems built around the GB200 NVL72 configuration, which pairs 36 Grace CPUs with 72 Blackwell GPUs.

That supply ceiling is not free. In the same earnings disclosure that raised the growth guidance, Nvidia cut its gross-margin outlook to a 71% to 72% trough. Kress tied the reduction to elevated pricing for high-bandwidth memory, the component in short supply across the industry precisely because of the AI buildout Nvidia's own chips are driving. The growth guidance and the margin cut were issued in the same breath, on the same call, describing the same underlying scarcity from opposite sides of Nvidia's income statement.

The 70% figure also excludes an entire market Nvidia once counted on. The company has said Chinese regulators have restricted licensed sales of its H200 chip and that shipped volumes have come in far below what licenses would allow, and Nvidia has responded by excluding China data-center revenue from its outlook entirely. China was once described internally as tens of billions of dollars in potential annual opportunity. A forecast built on a near-zero contribution from that market is a different kind of commitment than one that assumes access will be restored, and any shift in U.S. or Chinese export policy would move the 70% figure without any change in demand elsewhere.

The commercial relationships Nvidia is citing as evidence for its growth are also the relationships drawing the most scrutiny. Days before Huang's Goldman appearance, Amazon Web Services expanded its GPU commitment to Nvidia by 2 million units, covering Blackwell Ultra and upcoming Rubin chips, pushing its total committed order past 3 million GPUs for delivery in 2027 and 2028; the same agreement includes a 100,000-GPU AI factory built for the U.S. government on secure AWS infrastructure. Amazon's stock traded lower on the news, a sign that investors are less convinced than Nvidia's own guidance suggests that open-ended hyperscaler capex commitments are unambiguously good news. Investor Michael Burry has described Nvidia's pattern of investing directly into the companies that then buy its chips, including a $100 billion compute-guarantee arrangement with OpenAI, as circular spending, comparing it to Enron and pointing to Nvidia's five-year credit-default-swap spreads roughly doubling in two months as a market-priced warning sign. Kress addressed the criticism directly on the August call: "We recognize the scale of this support, and we know some will call this circular financing. We see it differently." The International Monetary Fund and the Bank for International Settlements have separately flagged circular financing arrangements across the AI sector as a systemic risk worth monitoring, without naming Nvidia specifically.

Nvidia's own roadmap is the mechanism by which it says it will actually deliver the number. The company has committed to an annual architecture cadence: the Rubin platform, paired with a new Vera CPU, ships in the second half of 2026 with roughly double the inference performance of the current Blackwell generation, followed by Rubin Ultra in the second half of 2027 at roughly fourteen times Blackwell's performance. That cadence is what supply-constrained guidance depends on holding to schedule. Nvidia still commands an estimated 80% share of the AI-accelerator market, but hyperscaler customers are increasingly building their own chips for in-house use, including Google's Ironwood, Amazon's Trainium 3, Microsoft's Maia 200 and Meta's MTIA, with bookings for that custom silicon growing at roughly a 44.6% annual rate against about 16% for merchant GPUs generally. Those chips do not compete for Nvidia's external customers directly, since hyperscalers build them for their own data centers rather than selling them, but they do shrink the slice of hyperscaler capital spending that Nvidia's guidance can assume is up for grabs. Advanced Micro Devices' MI400, built on TSMC's 2-nanometer process and tied to a multi-gigawatt compute deal with OpenAI, is due in the second half of 2026 as the most direct external challenger to Nvidia's merchant-GPU business.

Nvidia crossed a $5 trillion market capitalization in the days surrounding Huang's Goldman remarks, the first company to do so. The 70% figure sits underneath that valuation as both an operational forecast and, implicitly, the number the market needs to keep believing in for the valuation to hold.