Nvidia (NVDA)
Q4 Earnings
We first covered Nvidia (NVDA) on 30th May 2023 in an “initiating coverage” type report. That report can be found here.
We also did follow-up reports, some of which can be found here, here, and here.
This work convinced us of the investment case for Nvidia. We allocated 2-3% of our portfolio to the stock and it later became nearly 8% of the portfolio.
We had sold some along the way, but it is still our largest holding accounting for 7%. Along the way, Nvidia has become the largest company in the world with a market capitalisation of $ 4.7trn.
We were interested in the Q4 Results reported by the company two weeks ago.
Q4 2025 Results
The headlines were as follows:
Nvidia reported better-than-expected earnings and revenue in Q4.
Next quarter revenue forecast was higher than expected.
This was the 10th or 11th successive quarter in which they beat expectations.
Record quarterly revenue of $68.1bn, up 20% from Q3 and up 73% (y/y). Operating Profit grew by 84.3% (y/y) indicating strong operating leverage in the business. Net Profit grew 91.5% (y/y).
Record quarterly Datacentre revenue of $62.3bn (91% of total revenues, up 22% from Q3 and up 75% (y/y).
Record full-year revenue of $215.9bn, up 65% for the full year. Datacentre revenue for 2025 was $194bn (up 68% (y/y)).
GAAP operating expenses rose by 45% Y/Y This was powered by compensation growth related to more headcount, as well as rising compute infrastructure costs via their explosive growth. This growth rate is high, but it is much less than the growth in total revenue.
Both Gross and Operating Margins grew on a y/y basis
Nvidia reported better-than-expected earnings and revenue for the fiscal fourth quarter.
Next quarter revenue forecast was higher than expected.
Financial Profile
The quarterly revenue of $68bn is 14x larger than just five years ago. Although the growth rate now is lower than it was during the 2024 “take off” phase, it is still an impressive 72%. In the last five years, the revenue CAGR growth rate has been nearly 69%.
In the last five years the Operating Profit CAGR growth rate has been nearly 100%.
Gross Margins and Operating Margins are higher than they were five years ago. Operating cash flow growth has been impressive.
Operating cash flow growth has been impressive. In Q4 2026, operating cash was $36.1bn.
Nvidia’s capital expenditures, though growing, are relatively small thanks to its capital-light business model. Therefore, there is not much difference between Operating Cash Flow and Free Cash Flow. In the most recent quarter Nvidia generated Free Cash flow of $34.9bn.
If we look at annual data, Operating Cash Flow topped $ 100bn for the first time for the fiscal year ending January 2026.
Full-year Free Cash flow was $96bn. Nvidia has become a huge generator of cash.
How has Nvidia deployed its cash?
Part of the free Cash has been used for repurchases of shares, investments (acquisition of equity stakes in other companies, stock repurchases and other financing activities.
For the year, we returned $41bn or 43% of free cash flow to our shareholders in the form of share repurchases and dividends.
Total cash on the balance sheet has also increased significantly and is currently at $62bn.
In the last five years the stock has given a CAGR return of 69% or an absolute return of 12.7X. For investors and employees, these have been life changing returns.
Outlook
The company’s outlook is positive.
Total revenue is expected to be $78 billion, ±2%. This represents a growth of 77% (y/y).
GAAP gross margins are expected to be 74.9% ±50 basis points.
The guidance is even more impressive as it excludes China. The US administration has not given a green light to sell the older H200 chips to China. China is a $50bn per year market, so when and if the market finally opens, Nvidia stands to benefit significantly. It should be noted however that the Chinese government is reported to be nudging their companies to buy chips from local firms such as Huawei.
Excellent results but a poor market reaction.
This was another excellent set of quarterly results from Nvidia. Growth rates continue to be astonishingly high as the AI investment boom continues to defy all superlatives. However, the stock has been volatile and has fallen 8% since the results. Mr Market is not impressed. We need to ask Why Not?
The Bear Case
Once there is a directional trend in a stock price, there are some easy explanations as why the stock has fallen or risen. With respect to Nvidia’s share price decline, a few factors come to mind.
The stock has had a huge run and is now over owned. Most investors have a good allocation to the stock and do not wish to add to their holdings. In fact, it makes sense to sell to take some profits. In the absence of buying, some modest selling pressure can lead to large price declines.
The company is investing large sums of money in some of its customers. It is never a good sign when a company must finance its largest customers. This is a red flag and a reason to sell.
The five large hyperscalers, Amazon, Microsoft, Alphabet, Oracle and Meta account for about 50% of Nvidia’s total revenues. They five have increased their capital expenditure budgets again for the current year to a combined total of about $700bn. Nvidia expect this to reach $1trn by 2030. The bear case is this number cannot increase much more.
Let us consider the bear case in more detail. The big 5 hyperscalers are already using most of their operating cash flow to fund capital expenditure and some have resorted to issuing significant amounts of debt. Oracle, for example, was already highly leveraged even before the AI investment binge.
The hyperscalers need these to start generating much higher revenues and cash from the AI investments. The hyperscalers are likely to have overestimated the cash likely to be generated due to AI investments. When this realisation hits, they will cut Nvidia orders drastically and the latter’s revenue will decline.
A slightly different argument is current GPU demand is driven by the need to train Large Language Models (LLMs). Once the bulk of LLM training, focus will switch to inference, as large numbers of knowledge workers and consumers increasingly deploy the trained models. Inference will be an increasingly dominant source of AI revenue and revenue expansion for the hyperscalers.
However, for inference, the hyperscalers will not use just GPUs. Instead, they will mostly use Application Specific Integrated Circuits (ASICs). They are developing their own ASICs which go under names like Trainium (Amazon), TPU (Alphabet) and Maia (Microsoft). In this case, AI will grow as inference demand increases exponentially, hyperscalers will invest and expand to meet this demand but much of the chip expenditure will be on in-house designed chips, rather than GPUs produced by Nvidia or AMD. Therefore, Nvidia’s revenue will not continue to grow at current rates. Its growth is likely to slow down or even reverse.
These are some of the (well-known) bearish arguments that Nvidia sceptics have advanced.
Nvidia’s (new) bullish case
In the post-earnings conference call, Nvidia laid out a new vision of where this fast-changing market is going. The CEO Jensen Huang argued we are undergoing a historic change and entering a new era.
The interesting question is not about revenues in the next few quarters. The key is whether the $700bn of hyperscalers’ annual capital expenditure represents the peak of a hardware cycle or the first phase of a permanent economic shift to higher expenditures. Jensen Huang made a strong case in support of the latter view.
…compute equals revenues now in this new world. In a lot of ways, that’s the reason why we say it’s a new industrial revolution. There are new factories, new infrastructure being built, producing tokens is going to be the future of computing, which I believe, and I think largely the industry believes, then we’re going to be building out this capacity from this point forward and continue to expand from here.
In other words, the current capital expenditure is not a one off: in the future we will see even more investment in AI compute capacity. Nvidia is confident about this is because they argue we are at an agentic AI inflection point. This is not a new idea, but Nvidia is giving it much greater emphasis now.
Agentic AI means we are entering the era of Generative Software which is a fundamentally different way of doing software. Traditionally, the way software was done is like music recording. The software programme (and recorded music) are compiled and pre-recorded once and effectively set in stone. The user can access this an infinite number of times at zero marginal cost.
“The way we used to do software was pre-recorded. Everything was captured a priori. We pre-compile the software, we pre-write the content, we pre-record the videos.
In the age of AI, things are, and will be done, very differently: after each user makes a request, software is created anew each time by AI agents. The software is not pre-recorded. This is generative software AI, and it requires a lot more computing capacity.
Now everything is generative in real time... Just as a computer has a lot more computation capability than a DVD player, artificial intelligence needs a lot more computing capability than the way we used to do software in the past.”
One key factor is the release in February 2026 of some new AI tools and plugins by OpenAI and Anthropic, which can used with ChatGPT and Claude.
Generative software is fundamentally different. When Claude Code takes a coding task, it doesn’t retrieve a stored answer, it reasons. It breaks the problem apart, writes code, tests it, iterates, sometimes spawns sub-agents working in parallel. A single request might generate tens of thousands of tokens over minutes or hours, each requiring real-time computation. The gap between serving a static page and running a multi-step reasoning chain isn’t 2x or 10x. It’s closer to 1,000x.
The move from pre-recorded software to generative software is a profound change and it changes the nature of compute. Compute is not capital expenditure anymore; it is working capital. A material which is needed by the hyperscalers to offer their clients and which generates cash for them.
Once those tokens are monetized, once each token represents a unit of economic output, compute is no longer a cost centre but becomes a revenue engine.
As the costs of using AI declines, the cost of generating tokens falls and many new applications will become economically viable, and this will lead to greater demand. We cannot at this juncture foresee the nature of the new work that technological advance will lead to. We did not foresee how the internet would lead to streaming music, social media, search engines, effective AI and hundreds of other things.
The release of tools such as Claude Code, OpenAI Codex and Claude Cowork to users means all kinds of users, not just software developers, can now perform new tasks and complete old tasks much faster than was the case before.
Whenever the cost of computing has fallen in the last six decades, it has led to sustained increase I demand as many new use cases became feasible. This means hyperscalers will see demand for compute and they will invest in it to create revenue
“Compute directly translates to intelligence and revenue growth. Tokens are profitable.”
The story was best explained by the blog Kristal Lens which is written by Krista.ai. We will borrow extensively from their work in this piece.
We are seeing the change from pre-recorded software to generative software as shown below:
Source: Kristal.ai
Generative software requires up to 1000X more compute load than traditional pre-recorded software according to Nvidia.
Source: Kristal.ai
The AI inflection point has been reached because Anthropic and Open AI has released tools which will allow millions of workers to do their work much more efficiently.
Anthropic’s Claude Cowork agent platform is revolutionary and has opened floodgates for enterprise AI adoption.
ChatGPT moment of agentic AI has arrived. With partnerships spanning Anthropic, Meta, OpenAI, and xAI, NVIDIA is deployed across every cloud, and with our ability to build full stack AI infrastructure from the ground up or support them in the cloud, we’re uniquely positioned to partner with frontier model builders at every stage: training, inference, and AI factory scale-out.
Anthropic’s annualised run rate of revenues has doubled in the last three months even though it is capacity constrained. Given this, it is almost certain to invest more to increase capacity and grow more.
Kristal.ai also highlight the increasing importance of Networking in Nvidia’s revenues.
Source: Kristal.ai
Nvidia realised early that networking was going to be more important as the hyperscalers would need to connect their computers more effectively to improve efficiency and performance. In 2021, they acquired Mellanox, a Networking specialist.
Networking has gone from 8% to 18% of Data Center revenue in one year, growing 267% (y/y) in Q4. Full-year networking exceeded $31 billion, 10x what it was in FY21, the first year after the Mellanox acquisition.
We’re also now the largest networking company in the world. If you look at Ethernet, we came into the Ethernet market about a couple of years ago, into Ethernet switching, and I think that we’re probably the largest Ethernet networking company in the world today,. Spectrum-X Ethernet has been a home run for us.
In the giant datacentres, effective networking of all the computers and networking of the datacentres to each other can be unlock significant efficiency gains.
You know, the difference of it when you built a $10bn or $20bn AI factory, the difference of 10%, you know, and it could be easily 20% on the effectiveness and the utilization of your network for your data center, that translates to real money.
This reflects a profound change in what Nvidia delivers. They do not just deliver chips but whole computer networks. Thousands of racks of computers which are expertly “wired up” together to optimise performance.
If Nvidia was just a chip provider, it would be easy to switch to AMD, if the latter offered a cheaper or more efficient new chip. However, switching away is much more difficult if Nvidia provides complex computing systems at scale.
The NVL72 is a single rack-scale computer. Nine NVLink switch nodes, each with two custom chips, connecting 72 GPUs through proprietary fabric with liquid cooling engineered for a specific thermal envelope.
You cannot substitute a Broadcom switch or an AMD GPU into this system. The components are fused into an architecture. By controlling the networking fabric and the software stack (CUDA) as well as supplying the most advanced GPUs, NVIDIA is ensuring that the cost-per-token remains lower on its platform. By locking in the customer and increasing the switching cost, Nvidia is widening its moat.
Source: Kristal.ai
Switching is not an option due to the cost
Source: Kristal.ai
As Kristal Lens notes “once you’ve built your data center around NVL72 racks, plumbed the liquid cooling, provisioned the power density, deployed the NVLink fabric, you are physically locked in for the useful life of that infrastructure. Five to seven years. Switching isn’t a procurement decision; it’s a construction project.”
Evidence for this can be found in the rise of Networking Revenue. Customers are increasingly paying Nvidia to provide large scale solutions; to network large numbers of computers so they efficiently act as one giant computer.
Two views on Nvidia
So, we have two contrasting views on Nvidia.
The bear view says the AI capital expansion is at or nearing a cyclical high. Once people have the AI compute capacity they need, they will stop investing and Nvidia will see a rapid decline in orders.
The Nvidia view argues the opposite. Newly released tools enable many workers (not just developers) to become much more efficient. These developments will increase the speed of exiting work and make many new types of work feasible for the first time.
This work will generate many more tokens and people will be willing to pay for them. The hyperscalers (and others) will be happy to invest more in AI datacentres as they can see a relatively short and certain payback period.
The world is shifting to Generative AI which will require much more compute. Nvidia is gearing up to capture more of this demand and to provide not just chips but whole racks of thousands of networked and configured computers.
Valuation
NVIDIA’s growth, though impressive, is supply constrained. They cannot supply all the demand they have. That means quarter-to-quarter noise matters less than understanding how long this cycle can run and what the business looks like when demand normalizes.
Contrary to what you read in the press, Nvidia’s stock is not very expensive. Given the current price of $185.7 per share and the two-year forward EPS of $10.8, the two-year forward multiple is just17.2X. This seems quite reasonable value for a fast-growing stock with an ROE of 100% and a ROCE of 97.2%. These are elevated and will fall back and normalise soon.
We extended the analysis with a DCF valuation making relatively conservative growth and margin assumptions (relative to the analysts’ consensus) and utilises a weighted average cost of capital (WACC) of 9.5%. This suggests the fair value of the stock is about $193 per share which is a 5% premium to the current price.
Our best guess currently that the stock should generate a minimum CAGR of at least 11% over the medium term. The risk to this forecast is skewed mostly to the upside. If the expectations of analysts and of the company prove to be accurate, the stock could well generate a CAGR return of 15% to 18%. In this case the stock will double in 4-5 years.
Conclusions
We have presented two very different views of the prospects for Nvidia.
In the last three decades, Nvidia has grown very strongly and has shown itself as able to read technology trends very well and has shown a remarkable capacity to innovate, design and supply rapidly changing market trends.
We invested in this company three years ago and we have achieved a very satisfactory result. There is a risk that this pleasing experience means we are a little biased in our bullish view of the stock.
However, having considered this, we give the management the benefit of the doubt and we will add 1.5% to our holding so that it now account for about 8.5% of our portfolio
Annexe 1
A summary of some other issues discussed by management in the post earnings conference call.
Nvidia’s co-operation agreement with Groq.
We noted above that Nvidia acquired Mellanox when they saw how important networking was likely to become. The same applies to Inference. They have signed a co-operation agreement with Groq has shaken up the market with an ultra-fast Language Processing Unit (LPU) designed specifically for AI inference. Founded by Jonathan Ross, a former Google TPU engineer, Groq’s technology is built to make AI interactions feel instantaneous and natural.
We recently entered into a non-exclusive licensing agreement with Groq for its low latency inference technology; and welcomed a team of brilliant engineers to NVIDIA. As we did with Mellanox, we will extend NVIDIA’s architecture with Groq’s innovations to enable new levels of AI infrastructure, performance, and value.
Groq’s LPUs are already being rolled out in various sectors, including Tesla’s vehicles, where they’re powering the Grok AI chatbot. The company’s tech is also being used in datacentres and has even caught the attention of the Pentagon for potential use in classified systems
The AI Ecosystem
Nvidia has been making a lot of investments in smaller companies including those which are, or are likely to be, its customers. They are in the middle of an AI ecosystem.
Some of the strategic investments that you’ve made into Anthropic and potentially OpenAI, CoreWeave as well, but also partners, Intel, Nokia, Synopsys. You know, you’re clearly at the center of everything. At the core of everything NVIDIA is our ecosystem. That’s what everybody loves about our business, the richness of our ecosystem.
Just about every startup in the world is working on NVIDIA’s platform.
We’re in every cloud; we’re in every on-prem datacentre.
We’re all over the world’s edge and robotic systems.
Thousands of AI natives are built on top of NVIDIA. We want to take the great opportunity that we have as we’re in the beginning of this new computing era, this new computing platform shift, to put everybody on NVIDIA. Everything is already built on CUDA, and so it’s, we’re starting from a really terrific starting point.
As we build out the entire AI ecosystem, whether it’s in AI for language or physical AI, or AI physics, or biology, or robotics, or manufacturing, we want all of these ecosystems to be built on top of NVIDIA.
Whether it’s in enterprise or in manufacturing, industrial or science or robotics, each one of these ecosystems have different stacks, and we want to make sure that we continue to invest into our ecosystem.
Chip Depreciation Reduction
One problem is chips depreciate very fast. Their useful economic life maybe as low as three years. Buyers must write down their value to zero in three years. However, Nvidia claims work on their CUDA software stack is extending the useful life of the older chips
All of our GPUs are architecturally compatible, which means that when I’m working on optimizing models today, for Blackwell, all of that work and all that dedication to optimizing software stacks and new models also benefit Hopper and also benefit Ampere. It’s the reason why A100 continues to feel fresh and continues to stay performant, years after we’ve deployed it into the world.
Architecture compatibility allows us to do that. It allows us to invest enormously in software engineering and optimization, knowing that our entire install base in the cloud, on-prem, everywhere, from generations of architectures, of GPUs, will all benefit.
We’ll continue to do that, and allows us to extend the useful life, allows us to have innovation, flexibility, and velocity, which translates to performance, and very importantly, performance per dollar and performance per watt for our customers.
Networking
A key development in their networking offering has been the NVLink Switch.
NVLink Switch has enabled us to deliver generationally 50 times more performance per watt. It’s just an incredible leap, and it’s sensible. NVLink Switch is a great invention. It was hard to do. Performance per dollar, 35 times. The leap in inference is incredible. It’s very important to realize that inference equals revenues now for our customers.
As noted above Networking is an integral part of Nvidia’s overall offering to their clients.