Nvidia’s $91 Billion Quarter: I Cannot Tell If This Is the Peak or the Beginning

Nvidia just reported earnings that make me uncomfortable. Not because the numbers are bad. Because they are so good that I cannot tell whether I am looking at a generational company or the peak of a cycle that is about to correct. The headline: $58.3 billion in revenue for the quarter ending April 2026, up from $18.8 billion in the same quarter last year. A guidance of $91 billion for next quarter. A gross margin north of 70%. These are numbers that belong to a software company, not a company that ships physical silicon on substrates manufactured by TSMC.

The numbers that matter and the one that worries me

Data center revenue was $51.2 billion, which is 88% of total. That concentration is the number I keep staring at. Nvidia is essentially a one-product company at this scale: Hopper and Blackwell GPUs sold to hyperscalers and sovereign AI programs. Gaming, automotive, professional visualization, they still exist in the earnings release but they are rounding errors against the data center business.

The B200 and GB200 ramp is driving the guidance. Nvidia said the Blackwell transition is the fastest product ramp in company history, which given the CoWoS packaging constraints at TSMC I have been tracking, means either the packaging capacity expanded faster than expected or Nvidia is pulling forward orders from later quarters to hit the number. I do not know which, and that uncertainty matters.

The $91 billion guidance implies roughly 45% sequential growth. Wall Street expected around $85 billion. The beat-and-raise cycle continues, and at some point you have to ask whether this is sustainable execution or whether every quarter is borrowing from the next one. I do not have an answer. I have been wrong about Nvidia’s ceiling three times in the last two years, so take my skepticism with the appropriate grain of salt.

Who is actually buying all of this

The customer concentration question is the one nobody on the earnings call asked directly enough. Microsoft, Google, Amazon, and Meta collectively represent the majority of Nvidia’s data center revenue. The sovereign AI programs, where governments like Saudi Arabia, UAE, India, and France are building national AI compute infrastructure, are growing fast but still represent a smaller share. The xAI Colossus deployment I covered in the SpaceX IPO deep dive consumed roughly $18 billion in GPU purchases alone.

Here is what I keep thinking about from my FDI seat: the sovereign AI buildouts are the growth vector that could sustain this beyond the hyperscaler capex cycle. When a country decides it needs its own AI infrastructure for national security and economic competitiveness, the purchasing decision is political, not purely economic. France is doing this through Mistral and government-backed compute investments. The UAE is doing it through G42 and Cerebras. India just announced a national AI mission with substantial GPU procurement. These buyers are less price-sensitive and less likely to switch to AMD or custom silicon because the switching cost includes retraining entire national technical workforces on a different software stack.

The margin question that nobody talks about honestly

A 71% gross margin on hardware is abnormal. Intel’s gross margins in its peak were around 60%. TSMC runs about 55%. AMD’s are closer to 50%. Nvidia at 71% means either the product has no real competition (which is partially true because of CUDA lock-in), or the pricing power is temporary and will compress as AMD’s MI300X gains share and custom silicon from Google (TPUs), Amazon (Trainium), and Microsoft (Maia) matures.

I lean toward compression over the next 12-18 months, but not collapse. The CUDA ecosystem is a genuine moat, not because the hardware is unbeatable, but because the switching cost for enterprises that have built their entire inference pipeline on CUDA is measured in engineering years, not dollars. Cerebras has OpenAI as a customer for inference-specific workloads where the architecture advantage is large enough to justify the switch. AMD is picking up share at the margins. But for the vast majority of enterprise AI workloads, CUDA is the default and will remain so through at least 2027.

What the stock buyback tells you

Nvidia authorized $50 billion in additional stock buybacks, bringing the total authorization to about $100 billion. The company also raised its quarterly dividend, though at $0.01 per share the dividend is symbolic rather than material. The buyback is the interesting signal: it tells you management believes the stock is either fairly priced or underpriced at current levels, and it gives them a mechanism to return cash without committing to a dividend yield that constrains future capital allocation. For a company generating this much free cash flow, it is the rational choice. Whether it is the right signal for shareholders who might prefer that capital going into R&D or acquisitions is a separate question I am not going to pretend to answer definitively.

The part that connects to everything else I cover

Nvidia’s earnings are the financial expression of everything happening across the AI infrastructure stack. The SK Hynix HBM profits exist because Nvidia orders are driving HBM demand. The Chicago data center boom exists because hyperscalers need somewhere to put these GPUs. The RTX Spark announcement at Computex exists because Nvidia wants to extend the architecture from data center to laptop before someone else owns the edge. Every piece of the stack I track on this site connects back to this earnings report.

But I want to be careful about treating Nvidia’s dominance as permanent. I have covered enough tech cycles to know that the company everyone thinks is unbeatable usually is, right up until it is not. Intel was unbeatable in 2015. Qualcomm was unbeatable in mobile in 2012. Nvidia’s position is stronger than either of those, because the CUDA ecosystem is a deeper moat than anything Intel or Qualcomm built. But moats get crossed. The question is when, not whether.

For now, the numbers say Nvidia is executing at a level that has no precedent in semiconductor history. $91 billion in guidance for a single quarter. I am going to write that number one more time because it still does not feel real: ninety-one billion dollars. From a company that was doing $27 billion per year as recently as 2023. Whatever happens next, this is the quarter that future business school cases will point to as the moment AI infrastructure spending went from “significant” to “the largest capital allocation event in technology history.” I just cannot tell you yet whether those cases will be filed under “visionary execution” or “peak of cycle.” Ask me in eighteen months.