Nvidia: Jan-25 beat, Apr-25 guidance beat, the company’s tone hasn’t changed (demand for Blackwell is “extraordinary“, gross margins will improve late FY26)
but CEO is clearly saying that demand is growing faster than expected.
CEO is clearly saying that demand is growing faster than expected and has visibility into 2025-26 revenues “we have a fairly good line of sight of the amount of capital investment that data centers are building out”
Models are getting better and using up to 100x more computing power (ie chain of thought and reinforcement learning). The AI market is today Consumer, tomorrow Agentic, later on Physical.
The stock is not expensive at 30x FY26 EPS and 24x FY27 EPS. Get over it.
Demand is growing faster than expected for 2 reasons:
Models are evolving. AI is evolving beyond perception and generative AI into reasoning (chain of thought and reinforcement learning techniques) and that can require 100x more computing. 3 areas of model development:
Pre-training is still growing
Post-training, reinforcement learning, fine-tuning, model distillation require orders of magnitude more compute than Pre-training
Inference time scaling and reasoning where a single query demands 100 times more compute
The number of use-cases / applications is growing:
“We've really only tapped consumer AI, search and some amount of consumer generative AI, advertising, recommenders. The next wave is coming, agentic AI for enterprise, physical AI for robotics, and sovereign AI”.
Nvidia is losing share for Inference?
Conventional wisdom says that Nvidia’s GPU are great for Training but Inference will be implemented on proprietary ASICS (the internal chips of AWS, Google, Microsoft). Nvidia says otherwise:
“Our inference demand is accelerating"
'Blackwell was architected for reasoning AI inference"
"Blackwell has great demand for inference"
"The vast majority of our compute today is actually inference and Blackwell takes all of that to a new level. We designed Blackwell with the idea of reasoning models in mind”.
Internal ASICS are a lot cheaper than Nvidia GPU?
Your Tech analyst should have explained this to you, I mean it’s plain & simple: if you are a big cloud CSP, the price of Nvidia’s GPU or Nvidia’s very high margins don’t matter to you (but keep investors flabbergasted). Total Cost of Ownership (TCO) is what matters. This includes the purchasing cost of the Hopper or Blackwell system but the running cost (ie electricity) is a lot more important. For a computing chip, this means the number of computations per Watt or TOPS / Watt. CEO said:
“We're driven to a 200x reduction in inference cost in just the last two years. We delivered the lowest TCO and the highest ROI”
“if our performance per watt is anywhere from 2x to 4x to 8x (better than competitors or better than before), which is not unusual, it translates directly to revenues. And so if you have a gigawatt data center, if the performance or the throughput in that gigawatt data center is 4 times or 8 times higher, your revenues for that gigawatt data center is 8 times higher”. “the token throughput of our architecture being so incredibly fast is just incredibly valuable”.
I can actually show it to you:
Some hints that Nvidia has a very good visibility into 2025-26
“we have a fairly good line of sight of the amount of capital investment that data centers are building out”. But they won’t quantify.
Shall we worry about lower gross margins?
GM was 79% a year ago, now 73.5% and next quarter guidance 71%. That’s because industrializing Blackwell was 1) very complex and expensive 2) ran into bugs for 3-6 months, as we know.
“Blackwell is a customizable AI infrastructure with several different types of NVIDIA-built chips, multiple networking options, for air- and liquid-cooled data center” “We have some 350 plants manufacturing the 1.5 million components that go into each one of the Blackwell racks, Grace Blackwell racks. It's extremely complicated.”
So the impact of that is that “as Blackwell ramps, we expect gross margins to be in the low 70s.”
But then “when fully ramped, we have many opportunities to improve the cost and gross margin. We'll improve and return to the mid-70s late this fiscal year”.
Which means the transition to Hopper to Blackwell was difficult, and more difficult than expected. But the next transition will be easier “Blackwell and Blackwell Ultra, the system architecture is exactly the same. It's a lot harder going from Hopper to Blackwell because we went from an NVLink 8 system to a NVLink 72-based system. So the chassis, the architecture of the system, the hardware, the power delivery, all of that had to change. This was quite a challenging transition. But the next transition will slot right in“.
Many investors worried that we couldn’t see Blackwell revenues coming thru in Hon Hai or Quanta’s revenues.
Now we have numbers:
4Q25 (Jan-25) Data center Compute revenues was US$ 32.6bn.
In Q4, Blackwell sales exceeded our expectations. We delivered $11.0 billion of Blackwell architecture revenue in the fourth quarter of fiscal 2025.
Given server assembly time, delivery and commissioning time, we should see the supply chain revenues move from Feb-Mar-25.
4Q25 (quarter end Jan-25) beat by 6%
1Q26 guidance (quarter end Apr-25) beat by 10%
Higher revenues than Consensus forecast, but lower margins
Consensus expectations and valuations
1Q26 starts with 65% revenue growth YoY. Consensus has 51% for FY26. This means that by end-FY26 Consensus is thinking about 45% revenue growth – and a further steep slowdown into FY27 to 22% growth.
I think that FY26 revenue growth at ~50% is ok, and that gels with what the big pieces in the supply chain are talking about (ie TSMC, SK Hynix). But FY27 at 22% is too low. It is early days but again based on the supply chain expectations, 30-35% is a lot more likely.
Valuations are just fine, slightly below Trailing average but that doesn’t make the stock “attractive”. The stock is “fairly valued” as in at its average PE multiple. It’s never been cheap.
