If NVIDIA’s net margin stays at 50% forever, return on equity (ROE) is 100%, the terminal growth rate is 4%, and the investor’s required rate of return is 9% (pretty standard in the US), NVIDIA would need to generate approximately $620 billion in revenue. Compared to about $250 billion now.
Though that’s not a completely impossible thought if trillion-level investments actually materialize.
But as @Pohjolan_Eka challenged there, one can have many opinions on the final sustainable level for the time being.
Up until now, the market has constantly underestimated how massive this investment cycle will become, which is visible in the “staircase-like” behavior of NVIDIA’s forecasts and the fact that, for example, memory stocks only started skyrocketing about a year ago.
This discussion can be approached from the perspective of technological sustainability, but also from an economic perspective. As I wrote above, the numbers are so large that the whole world will soon have to start digging deep into its pockets if the current trend continues. No wonder the stock market isn’t quite buying Jensen’s comment about $3–4 trillion in investments, and NVIDIA isn’t yet priced on the assumption that its normalized revenue would be over a trillion a year.
Spending time in the corridors of large corporations, I see a quite different sentiment. The demand isn’t about the goal of AGI. No peak in AI spending has been passed, and it is still a matter of investment levels. Whether enough is being invested compared to others (competitors) is the first (and sometimes only) question, followed by questions about top/bottom-line impacts, and only much later, somewhere at the lower coder level, might someone ask if the same impact could be achieved more cheaply. There is no demand ceiling when management is constantly worried that someone might disrupt their business by putting more firepower into AI. This phenomenon sometimes leads to “tokenmaxxing” or HR-style employee surveys on how and how much you use AI in your work. If a gem is found in the AI suggestion box, the money for expensive GPU capacity is easily found in IT and R&D budgets. 3-5% of revenue is no problem when disruption is both hoped for and feared. Granted, small companies lack the same enthusiasm for investments, and large corporations do calculate and prioritize project profitability, but large companies have plenty of leverage and the benefits on paper look good enough—at least compared to traditional IT and R&D investments.
By the time a taxi driver or a manager says that enough tokens have been consumed, there is enough AI, and the same results can be achieved with cheaper home computer capacity, then I will believe that hardware investments based on NVIDIA GPUs will decline. Much smaller and cheaper computing capacity than the latest iPhone is also sufficient for calls and messaging. There may well be an AI bubble elsewhere, just as there was in the phone business, and not everyone survived the smartphone transition.
I can’t say whether it will show up first in individual data center companies as capacity failing to sell, or in Nvidia having to lower prices to move all its production. Both lead to the same result, just via slightly different routes. Recently, SpaceX already built too much data center capacity for itself, but luckily they managed to lease it to Anthropic with a 90-day notice period. If contracts across the field are that short, it’s going to get ugly if supply exceeds demand in the near future, considering massive investments have been made and capacity needs to be leased out even at a loss.
This happens constantly on the hardware side: a “good enough” level is reached, and then the purchasing pressure created by consumers disappears from the market entirely. Examples include CD-ROM burning speeds or broadband connection speeds. In the early stages, when no product on the market was good enough to meet needs, people would come from the next house over to admire a one-megabit cable connection. But nowadays, almost no one wants or needs a gigabit connection, so higher speeds are no longer meaningful to consumers in any way.
People used to buy separate sound cards for computers because there were such massive differences, but today no one cares what sound card is in a computer because they are all good enough for the average user. When specs are good enough, price becomes the primary deciding factor, and only a small minority wants those SOTA (State of the Art) products.
This dynamic doesn’t seem to be accepted as an axiom in the AI discussion. If you have an IQ 50 AI and someone else develops an IQ 70 AI, that’s a huge leap in usability. But if you have an IQ 150 AI and a competitor develops an IQ 250 AI, it might not matter at all except to a very small group of users, and people won’t be willing to pay for that hyper-intelligent AI when a super-intelligent one is available at half the price.
Claude and OpenAI are examples of commodity AI companies that have no genuine product-based competitive advantage besides their brand. They are still relevant right now because some users are willing to pay extra for performance, but we are very close to—or have already passed—the point where any competing “generic” (Pirkka) AI would suffice for most users. Claude’s alleged 1 trillion valuation is completely incomprehensible in this context, unless their tools are automatically bundled into every Windows OS like Word and Excel.
I agree with you, and this is typical in the phase of adopting the first AI solutions, where management is in an “AI psychosis” and limitless AI capex is sought with great haste, either due to fear of disruption or imagining massive productivity gains once AI solutions are forced as widely as possible across the organization. There’s also a lot of that classic corporate R&D maneuvering where a project important to one’s own organization is externally dressed up in AI clothes to hustle the funding from management for the desired investment.
The next phase is the shift to AI opex. Those previously acquired, hastily cobbled-together AI “spaghetti” solutions start breaking down and generating high maintenance costs; the feared disruption never came, and productivity gains didn’t materialize as desired, no matter how much people tried to massage the statistics. However, token costs run continuously, so they begin to be optimized to reduce the number of tokens used, and new investments start being viewed with a critical eye to avoid repeating the mistakes of the first phase.
Different companies move at different clock speeds depending on the organization’s experience and management’s vision, but the assumption that investment faucets would remain permanently open for traditional mature-stage companies that heavily optimize operational cost-efficiency would, in my opinion, be a rather bold assumption. Yet these companies make up the largest share of the economy.
I agree in principle, but sometimes technological development can also stop because a certain technology reaches a dead end due to, for example, the laws of physics, as with CDs. If it were possible to manufacture an optical storage system that combined the benefits of CD/DVD/Blu-ray (archival stability and physical protection against overwriting) with modern speed and capacity, it would surely be a bestseller, but apparently, it is simply too difficult or impossible. Broadband connections, on the other hand, are a very good example of where “good enough” suffices for almost everything, so demand gradually fades.
No doubt these thoughts exist in many large corporations, but I would like to see a CIO/CTO state this out loud in the mainstream media. Usually, the opposite is said; the companies you mentioned (UPM and Outokumpu) both report on their use of AI in their annual reports.
My own experience is that “AI spaghetti” is accepted onto the roadmap much more willingly than traditional “IT spaghetti.” AI OpEx costs are perceived as simpler and more flexible than normal software licenses and service maintenance. Personally, I see a future scenario where AI essentially takes over IT and R&D budgets (5-8%). It is significantly easier to build your own AI agents or deploy the AI capabilities of existing products than it is to justify the implementation of yet another piece of software/hardware and related services (spaghetti) in a large corporation. An example of this is SAP Joule, which in part supports the growth of NVIDIA’s hardware and software business in large enterprises.
Some CD-ROMs are a curiosity-level application that affected a few, whereas the advent of AI is a paradigm shift that will eventually affect almost everyone. Even if your examples regarding CDs or internet speeds were accurate, it is far too early to draw a direct analogy from them to the future of AI.
You wouldn’t say something like that out loud in any annual report, because then you’d seem old-fashioned. At the turn of the decade, there was a very similar enthusiasm for the metaverse, and traditional large corporations made highly questionable investments in virtual reality centers of excellence and hardware. The Zeitgeist drives a lot of executives’ technology investment priorities, but while you can momentarily fill downtown Helsinki with electric scooters if you’re investing with other people’s money, at some point the business has to start sustaining itself profitably through consumers’ wallets or B2B customers from the traditional economy. For the hardware investments driving NVIDIA’s stock to be at a sustainable level, players like OpenAI and Anthropic would need to reach profitability, and in my opinion, there are no signs of that yet. Now, of course, as they attempt to go public, quarterly results might be momentarily manipulated so that such a narrative can be constructed.
Non-deterministic spaghetti is a hell to maintain and generates far more wasted costs in the form of misused tokens, so the maintenance costs of AI software are much higher than those of traditional deterministic spaghetti IT. There just isn’t enough wide-scale experience available on this yet.
I would recommend trying to ask that AI what the technological significance of CD-ROMs was for society and how widely they were used everywhere in people’s lives, including the business world
Below is an article about how Nvidia is investing heavily in technology where data is transmitted using light instead of electricity.
The idea is to reduce the energy consumption of AI infrastructure and alleviate bottlenecks in data centers, among other things, as traditional data transfer relying on copper and electricity is already starting to limit scaling. This type of technology also interests other tech giants, but widespread adoption is still in its early stages because large-scale production of such complex technology is challenging to implement.
Key Points
Nvidia has committed at least $6.5 billion to companies developing photonics technology since March this year.
Photonics is considered to be a more efficient way to transfer data than the current standard process of using more costly electricity running on copper, which is thought to be a major blocker to the rollout of AI.
“The amount of silicon photonics technology capacity that we need is substantially higher than the world has today,” Nvidia CEO Jensen Huang said at GTC in March.
Alright, the Vera Rubin production and distribution ramp-up has now officially started! The world’s first VR200 NVL72 delivered by Dell to Coreweave has already passed L11 diagnostics - i.e., 72 GPUs, 36 CPUs, and NVLink are working flawlessly.
So, the H2/26 schedule communicated in March seems to be holding up well, and we are even slightly ahead of it. As I see it, the markets aren’t expecting much VR sales even for Q3. Now it seems like it’s coming - possibly quite a bit, even though I understand the H-series is still generating sales in the billions.
The price of such a rack has been speculated to be 7 million per unit.
There probably isn’t anything majorly new in this story. It explains how China is making a serious effort to break away from Nvidia.
Companies operating in the country are developing their own chips for cars, AI, and computing, because Nvidia is expensive and an uncertain choice from a geopolitical perspective. The article states that these things won’t happen overnight, but the direction is clear: Chinese models and devices are increasingly being built to run on the country’s own chips.
Key Points
Companies in China are increasingly developing alternatives to Nvidia chips.
That’s the case even for less-advanced Nvidia semiconductors used in driver-assist systems.
Newer Chinese AI models are also expanding compatibility to homegrown chips.
