Skip to content
We Should Believe Jensen’s Claim That Reasoning Increases Demand for Compute by 100x
Nvidia
Nvidia's January quarter and April guidance marked a transition from a company reporting breathtaking upside to one delivering modest upside. This shift supports the narrative that Nvidia's best days are behind it. However, I believe Jensen is on the right track in claiming that reasoning increasingly drives a 100x increase in demand for compute. If true, the company will return to reporting impressive upside in CY26.

Key Takeaways

The April guidance was good but not great, indicating a shift in the pace of upside.
The company redirected investor focus from the first half to the second half of the year. I continue to believe there is room for modest revenue upside throughout the year.
In the long term, the central question is: Should we believe Jensen’s claim that reasoning increases the demand for compute by 100x? The answer is yes.
Despite the emergence of cheaper alternatives to Nvidia GPUs including custom silicon, I believe they will remain the dominant workhorse for the next 3-5 years.
1

April guidance

Nvidia guided April revenue to be about $1B above the Street, implying a 2% guide-up. This compares to whisper numbers, which expected a 4% guide-up. In other words, they missed the whisper number in their guidance.

Separately, gross margin guidance came in at 71% versus the Street’s expectation of 72%. Additionally, reading between the lines suggests gross margins will decline further in the July quarter. To soften the gross margin guidance, the company argued that margins would materially improve in the October and January quarters, returning to the mid-70s, close to the company’s all-time high of around 77%. The rationale for this rollercoaster in gross margins is improving production yields as Blackwell ramps.

Overall, this outlook can be summed up as good, but not great. For Nvidia to build momentum around its shares, the outlook needs to be extraordinary and the company fell short of that hurdle.

2

Navigating 2025

Notably missing from the CFO’s commentary was an update on how long Blackwell is expected to be sold out this year. Following last year’s October quarter, the company guided that Blackwell would remain sold out through this year’s October quarter. Some analysts had anticipated that guidance would extend the sellout throughout CY2025.

Under normal circumstances, the company’s omission of specifics regarding Blackwell demand would be viewed negatively, suggesting demand is not as robust as previously believed. In isolation, the lack of detail was directionally negative, as investors typically assume the worst when a company provides less information. However, a more optimistic interpretation is that the company is shifting away from providing one year demand guidance to focus investors on the bigger picture, namely, its ability to sustain higher revenue growth for longer than most expect.

While we didn’t get a direct answer on how long Blackwell will be sold out, Jensen responded to a question from CNBC’s Jon Fortt by stating that his confidence in Blackwell is even stronger than three months ago for two reasons: first, production has ramped up, and second, demand for Blackwell is “greater than ever.” In theory, given the company’s growth, stating that Blackwell demand is “greater than ever” doesn’t reveal much. That said, overanalyzing his wording can get us tangled up. The reality is that Jensen remains bullish on Blackwell, which bodes well for 2025.

Regarding gross margins rebounding by year-end, the reason for this rebound is that as Blackwell ramps, manufacturing yields and margins improve. However, this recovery is partly contingent on stable China export controls, as China currently accounts for about 15% of sales. If restrictions tighten further, it could negatively impact sales of the high-margin H20 GPU in China. To illustrate how scaled-back the H20 is compared to Nvidia’s top-tier products, Jensen noted on the call that it has just 1/60th the power of the GB200.

3

The need for compute should go to the moon

The central message from the company during this earnings call is that the need for compute has essentially not even started. Jensen made the case that we are transitioning from a world where the most common AI prompt was a basic “one-shot” query – where you receive a simple answer in a second – to a world where models reason. In these “reasoning models,” compute hardware requirements increase exponentially.

The idea is that next-generation models could require hundreds, thousands, or even millions of times more compute than today’s one-shot prompts. That begs the question: Is that realistic?

I believe the answer is yes.

In the future, we will likely need 100x more compute to support reasoning models. One quick way to test this is by measuring how long a model “thinks” as a proxy for how much compute it needs. From my own experience, a low-level reasoning prompt takes between 2–10 minutes. Taking a midpoint of 5 minutes implies 300 seconds of compute time, which is 150–300x longer than a one-shot prompt. This back-of-the-envelope estimate supports Jensen’s bullish outlook.

As for the million-times claim? That’s on the table too. A one-million-fold increase in compute translates a two-second response into 23 days of reasoning. While this may sound aggressive, in a world where AI agents self-prompt around complex problems, it’s feasible that they could “reason” for more than 23 days.

Another way to validate this projection is by looking at what has happened with compute hardware over the past decade as a benchmark for the next decade. Supercomputing performance has increased 50-fold in the past ten years. AI training computation, which is a more aggressive benchmark, has increased roughly 10,000x since 2019.

The big picture: I believe Jensen is on the right track and we are just scratching the surface when it comes to the demand for compute.

4

Nvidia vs. Custom Silicon

Despite the emergence of cheaper alternatives to Nvidia GPUs including custom silicon, I believe they will remain the dominant workhorse for the next 3-5 years.

Its true, an increasing amount will be supplied by more affordable methods, but Nvidia’s high-demand, pole-position status will remain intact. On the earnings call, Jensen addressed concerns about custom silicon, stating that Nvidia is still 2-8x cheaper than custom chips when it comes to the cost of producing tokens. He also noted the large gap between designing a custom chip and bringing it into large-scale production.

While I don’t have a direct read on the cost differences between custom silicon and Nvidia’s chips since public data on these chips is limited, I agree with Jensen on the gap between design and production. Building a chip is much harder than it sounds.

Disclaimer

Subscribe to our newsletter

"*" indicates required fields

This field is for validation purposes and should be left unchanged.
Back To Top