Nvidia FQ2 25 – Full steam ahead

Beat but no real raise.

• Excellent results and guidance were met with disappointment from the fickle market which wanted to see yet another beat and raise and so while the AI boom remains very much on track, its failure to get even bigger has allowed a sliver of doubt to creep in.
• FQ2 25 revenues / EPS were $30.0bn / $0.68 nicely ahead of expectations of $28.7bn / $0.65.
• Nvidia is one of the few companies in AI that is working properly when it comes to financial performance.
• Revenues were up 122%, gross margins expanded YoY, and expenses grew by 48% allowing the benefits of operating leverage to accrue to shareholders with diluted EPS growing by 168% YoY.
• Furthermore, almost all of net income flows into cash from operations from which the only big draws on the cash balance come from returning cash to shareholders.
• This is how a mature, well-run company should perform and Nvidia stands in stark contrast to other AI companies like OpenAI, which is burning cash like there is no tomorrow and raising even more money despite having billions in revenue.
• There is no question in my mind about which end of the AI industry I would want to invest in if I had to take a position in it directly.
• Despite this exemplary financial performance, there were a couple of hiccups.
  • First guidance, which fell short of the most lofty expectations.
  • FQ3 25 will see revenues of $31.9bn – $33.2bn ($32.5 +/- 2%) which is above the current average estimate of $31.7bn but below the top end of the range.
  • The market was clearly hoping for another blowout where expectations were pushed up yet again and the shares fell by 6.9% in after-hours trading as a result.
  • I think that Nvidia has much better visibility than it is implying, and I would not be surprised to see this guidance beaten with revenues of around $34bn in Q2 25.
  • This is because Nvidia has more demand than it can handle right now meaning that sales are determined by how much capacity it has booked at TSMC and not by customer orders.
  • The problem is that the market was already baking this in which is why the shares sold off when Nvidia failed to raise guidance meaningfully.
  • Second, Blackwell delay which in my opinion is not a big problem.
  • Nvidia has had to make minor design changes to Blackwell to get the yield to where it needs to be, and this has caused the schedule to slip by several weeks.
  • In the grand scheme of things, this is irrelevant as Blackwell has no competition and customers will be more than happy to wait for their turn to pour cash into Nvidia’s coffers.
  • However, it may have dented sentiment somewhat but I think it will quickly be forgotten once volume shipments commence in FQ4.
• The net result is that everyone is still building AI like crazy and there are no immediate signs of demand slackening.
• However, the law of large numbers means that growth is going to slow markedly over the next 12 months and even Nvidia admitted this stating that data centre would grow “quite significantly next year”.
• To me, this means something like 30% to 40% YoY in fiscal 2026 revenue which is much lower than fiscal 2024 which is forecast at 102% YoY but it is still excellent performance at this scale.
• This is where the problems begin because consensus is already expecting 39% YoY in revenue growth in fiscal 2026 which will lead some to wonder whether this figure is indeed too high.
• However, in the grand scheme of things, the valuation is not that demanding because Nvidia is so good at ensuring that the benefits of its efforts flow to the shareholder.
• This is why the shares are currently trading on 30.3x 2026 PER which is broadly in line with the big digital ecosystems even while Nvidia is growing revenues significantly faster.
• This is why the shares fell by 7% rather than 30% underlining that Nvidia remains one of the lowest-risk ways to invest directly in the AI bubble.

• Generally, I remain nervous about AI valuations and the prospect of the generative AI bubble bursting, which is why I continue to prefer the adjacencies of inference at the edge and nuclear power to run all of these data centres.