Newsletter: Press obsessed Bubbles, NVDA Valuation, NVDA Inference systems, GPU Rental prices, China’s Services surplus

All Newsletters10 April 2026

“Difficulties strengthen the mind, as labor does the body.” Seneca.

“Man’s mind, once stretched by a new idea, never regains its original dimensions.” Oliver Wendell Holmes.

Bubbles

It has come to our attention that Polar Capital Technology Trust has released a memo that specifically refutes that the stock market is in a bubble: 2026: The year AI’s impact becomes impossible to ignore | Polar Capital Technology Trust plc

This is in addition to an excellent presentation they have released, that had this truly Prophetic slide (which we have applauded):

However, apart from the “Chapeau” to the PCT team, our point is that this same presentation has 5 slides called “AI Bubble? We Don’t Think So.” This strongly suggests to us that the PCT team are rather bored of discussing Ai bubbles with investors, as are we.

During a recent Teams conversation we had with Ben and Alastair they made the interesting observation that the mere mention of Ai now triggered a Pavlovian “Bubble” reaction. A final thought from the latest PCT newsletter: “The macroeconomic bull case here rests on Al resulting in greater productivity gains than disruption in the near term, leading to a disinflationary growth reacceleration accompanied by lower interest rates. Encouragingly, industries with higher Al adoption rates are showing slightly larger productivity growth over the past year in official US data, according to Goldman Sachs.”

We are not going to reproduce here the excellent work of PCT, after all our Haters say we just replicate their Portfolio, but we do suggest you review the slides which start at page 28 of their January presentation, as they offer some excellent arguments as to why we were a long way from a Bubble.

We have inserted a couple of slides of our own below on this point:

The Internet Bubble

During the Internet Bubble, Mark used to appear on a fringe Sky TV channel called The Money Channel on a week by week basis:

After appearing numerous times, Mark appeared in his regular slot and told the audience that he felt that the Internet Boom had finally become a Bubble and it was time to cash out. He recounts this event very vividly because one of the anchors was a very attractive woman who he had found very desirable until she hissed at him like a chimera of a hyena and a snake as he was escorted from the studio. His young heart broke, albeit very temporarily. Mark was never asked to return.

My point is that there is a difference between a Boom and a Bubble. Bubbles involve frenzied overvaluation and blind buying. Booms involve the roll out of new technological change to bring about a more efficient and productive future. There is a BOOM now, but there is not a Bubble. Mark saw a Bubble in his twenties, and he sees just a Boom now. It does not normally work that way round! There are constantly pockets of Bubbles in every market, but there is no widespread Ai bubble. Large capitalisation stocks are trading logically aligned with their growth prospects, and they are generating huge cash inflows.

For example, and looking at Nvidia, Bubbles don’t involve EV to EBITDA ratios of <15x for growth of greater than 50 per cent! Capital Overspend Booms are also not the same as Bubbles. We will see in time IF the Capital Spend Boom is a Rational Boom or an Overspend Boom. However, Mark also tired back in the Internet Boom days of listening to the unimaginative tell him that High Street Retail would never be disrupted by Ecommerce. The key point is it is pointless discussing these matters as it is a subjective stance on how the future will outturn. We have our opinion and others have theirs. Just as we all have differing opinions on Religion & Politics. However, looking at our Portfolio today it does not feel like a Bubble to us:

Stratechery, the widely read Technology commentator, describes in some detail what has changed to make this phase of the Era of Ai less likely to be a Bubble.

For that latter point, you may remember our previous Newsletter from November 2025 where we laid out in numbers, on a token basis, how the provision of Ai services by the Hyperscalers was earning an attractive ROIC. Please note that was before the arrival of agentic Ai in Q1 2026 and please remember what I said in my last Newsletter on “removing the bloat” when reading Stratechery again:

His concluding points are key:

Beth Kindig recently published a chart called: “Tech Bubble” Warnings Cost Investors a 55-% Nasdaq-100 Run”:

Her excellent article on this matter can be found here: https://io-fund.com/broad-market/nasdaq-100-returns-vs-tech-bubble-warnings. She makes several great points such as the barriers to entry for Ai companies are far higher than most bubble technologies; the current constraints to supply (power, memory, fabs etc) that will hold back any overbuild; the vast unsated demand for Ai; the power, planning and human electrician constraints; and that Ai is driving significant revenues, profits and ROIC already.

You should read the article in full but a few quotes of note:

“TSMC, the world’s largest contract chipmaker and critical supplier for AI logic chips, has stated that its advanced-node capacity is roughly three times short of what AI demand requires, even amid ongoing expansion efforts. This reflects a real gap between what customers want and what TSMC can physically produce.”

“There is also more evidence that cloud is now late cycle, with the market being extremely saturated with more than 30,800 SaaS companies worldwide, each competing with one another for wallet share as production differentiation narrows. Market saturation preceded the disruptive fears from AI-based solutions automating workflows.”

Marc Andreessen, the highly successful venture investor, has also held forth on the wealth destructive effects of listening to the Bubble Luddites:

Conclusion: This section has conclusively answered the question: “Mark don’t you think Valuations are stretched, and we are in a Bubble?”

What to avoid during the Ai Boom

Turning back to the excellent Beth Kindig, within the previously mentioned article, she also calls the end of the Software 1.0 Computing cycle and presents the following chart as proof:

Please note these names above were some of the darling stocks of many Growth portfolios but not a single one can be found in the MNL LN portfolio. Whilst we are on that subject of what to avoid, we are going to lay out which companies and sectors we do see as being destructed by this Capital Expenditure Boom. The opening months of 2026 have witnessed a dramatic leap forward in artificial intelligence, driven largely by the arrival of highly autonomous AI agents like Claude Code. This breakthrough has caused a stark split in Tech sector performance. Markets are aggressively questioning the long-term survival of traditional, per-user SaaS pricing models now that AI agents can execute tasks independently.

Furthermore, this upheaval is no longer confined to software companies; it is actively spilling over into publishing, finance, and healthcare. Moving forward, we anticipate the next major waves of disruption will heavily impact banking, insurance, retail, manufacturing and commercial real estate.

Software as a Service. We dealt with this point in our last Newsletter, here is an excerpt from Ben Thompson:

“In the shorter-term, however, the real risk I see for software companies is the fact that while they can write infinite software thanks to AI, so can every other software company. I suspect this will completely upend the relatively neat and infinitely siloed SaaS ecosystem that has been Silicon Valley’s bread-and-butter for the last decade: identify a business function, leverage open source to write a SaaS app that addresses that function, hire a sales team, do some cohort analysis, IPO, and tell yourself that you were changing the world.

The problem now, however, is that while businesses may not give up on software, they don’t necessarily want to buy more — if anything, they need to cut their spending so they have more money for their own tokens. That means the growth story for all of these companies is in serious question — the industry-wide re-rating seems completely justified to me — which means the most optimal application of that new AI coding capability will be to start attacking adjacencies, justifying both your existence and also presenting the opportunity to raise prices. In other words, for the last decade the SaaS story has been about growing the pie: the next decade is going to be about fighting for it, and the model makers will be the arms dealers.”

Advertisers. May I suggest you read these transcripts from this excellent podcast by Patrick O’Shaughnessy’s “Invest like the Best”.

“What would you be worried about if you were one of these fairly monopolistic owners of a massive ad network like the ones we’ve discussed? We’ve got their Uber and Amazon in the mix, DoorDash, Facebook, Google. If you were there running their ads businesses, what would scare you?

“Consumer behaviour changes, where they don’t open up the apps anymore, but they use agentic interfaces, they use AI interfaces which are not owned by my company or this company to do their transactions. If you assume that a big percentage of things are repeat, then could you put those repeat things on autopilot through an agent and you never open the app? So you lose opportunities then advertise, and you lose the relation with the customer over time, because the customer starts trusting the AI agent.”

“(with Ai) These things have been brought together. It’s the dream of any advertising person and these are complex multi-phase searches. That’s the other beautiful thing.
You search, but each of the queries is like a search, and then you search again and you’re just building up searches. At Google, you typically search and then you lose the person because they go off and click and you don’t hear. These are natural language queries right for amazing, amazing targeting.”

Microsoft Co-pilot. From last July through late January, the percentage of Co-pilot subscribers who use the product as a primary option decreased from 18.8% to 11.5%, according to a survey of more than 150,000 respondents in the U.S. by market research firm Recon Analytics. This happened while the share of paid users who choose Google’s Gemini as the first option increased from 12.8% to 15.7%. Those surveyed who switched from Co-pilot said they found better quality elsewhere, with some citing poor user experience and restrictive usage limits. Workers who have access to subscriptions for Co-pilot, ChatGPT and Gemini choose ChatGPT and Gemini at a higher rate than Co-pilot, the Recon data shows. In recent weeks, a new AI product from Anthropic called Claude Co-work has drawn praise for its ability to work in and across 365 applications in the ways Co-pilot users have found difficult.

As SemiAnalysis recently observed: “Coding was once the most valuable work of all… Coding is now a beachhead in terms of the disruption that agentic information processing has, and the larger 15 trillion-dollar information work economy is now at risk. There are 1b+ information workers, or roughly 1/3rd of the global 3.6 billion workforce.”

The big issue with Co-pilot is that it does not have connectors into external third party applications like Claude does OR if it does, they have daft limitations that render them non functional. This closed garden approach will not serve MSFT well.

By mastering software development, AI systems such as Claude are now actively participating in the creation of subsequent, more advanced generations of AI, a process researchers refer to as an “intelligence explosion.” OpenAI has confirmed that GPT-5.3 Codex was instrumental in its own development, debugging its training and managing deployments. This recursive improvement cycle suggests that the timeline for AI to achieve superhuman capabilities across various domains is compressing dramatically.

We warned MSFT many times they were being left behind but they treated us as if we were something they had just trodden in:

…and their Reply…

Crowding Out. Another side effect of this Boom, that we all better prepare for, is investment “crowding out”. I keep getting told by UK investors that the Hyperscalers will not continue to fund their Data Centre expansions which will cost multi trillions over the next 5 years. Yes multi trillions! But we disagree, YES they will fund it.

And if they have cash flow limitations then they will raise debt finance just as Oracle raised its $25bn bond issue, with ease. However, the whole ongoing PE and Private Credit collapse will not help which will mean that other projects outside AI will suffer capital investment constraints and less investment.

Conclusion: The disruption risk for your legacy portfolio is now very high. Are you sure the over-valuations are not actually in those disrupted stocks, not the Ai stocks? If Humans are also likely to be disrupted by Ai, don’t you think hedging your future with a portfolio of potentially disruptive, Ai stocks may be a good idea? How about selling all your Private Credit or PE funds?

Nvidia

Transcript for Jensen Huang: NVIDIA – The $4 Trillion Company & the AI Revolution | Lex Fridman Podcast #494 which can be found at https://lexfridman.com/jensen-huang-transcript

Will NVIDIA be worth $10 trillion?

Jensen Huang(01:24:45) I think that NVIDIA’s growth is extremely likely, and in my mind, inevitable. And let me explain why. We’re the largest computer company in history. That alone should beg the question, why? And the reason of course… Two reasons. First, two foundational technical reasons. The first reason is that computing went from being a retrieval-based, file retrieval system. Almost everything is a file… We pre-write something, we pre-record something. You know, we draw something, we put it on the web, we put it in a file. And we use a recommender system, some smart filter, to figure out what to retrieve for you. And so we were a pre-recording, human pre-recording, and file retrieving system. That’s what a computer is, largely.

Jensen Huang(01:25:39) To now, AI computers are contextually aware, which means that it has to process and generate tokens in real time. So we went from a retrieval-based computing system to a generative-based computing system. We’re gonna need a lot more processing in this new world than in the old world. We need a lot of storage in the old world. We need a lot of computation in this new world. And so that’s the first part of it. We fundamentally changed computing and the way how computing is done.

Jensen Huang(01:27:04) The second idea is computers, because it was a storage system, it was largely a warehouse. We’re now building factories. Warehouses don’t make much money. Factories directly correlates with the company’s revenues. And so, the computer did two things. Not only did it change the way it did it, its purpose in the world changed. It’s no longer a computer, it’s a factory. It’s a factory, it’s used for generation of revenues. We’re now seeing not only is this factory generating products, commodities that people want to consume, we’re seeing that the commodities are so interesting, so valuable to so many different audiences that the tokens are starting to segment, like iPhones. You have free tokens, you have premium tokens, and you have several tokens in the middle.

Jensen Huang(01:28:10) And so intelligence, as it turns out, you know, it’s a scalable product. There’s extremely high intelligence products, tokens that you could… that are used for specialized things, people be willing to pay. You know, the idea that somebody’s willing to pay $1000 per million tokens is just around the corner. It’s not if, it’s only when. And so, so now we’re seeing that the commodity that this factory makes is actually valuable, and is revenue generating and profit generating. Now the question is how many of these factories does the world need? How many tokens does the world need? And how much is society willing to pay for these tokens? And what would happen to the world’s economy if the productivity were to improve so substantially? What would happen…

Jensen Huang(01:29:08) Are we, are we gonna discover new drugs, new products, new services? And so when you take these things in combination, I am absolutely certain that the world’s GDP is going to accelerate in growth. I’m absolutely certain the percentage of that GDP that will be used for computation will be 100 times more than the past—mm-hmm—because it’s no longer a storage unit. It’s a product generation unit. And so when you look at it in that context and then you back into what is NVIDIA’s, what does NVIDIA sh—what does NVIDIA do and how much of that new economics, new industry would we have to benefit t—to address, I think we’re gonna be a lot, lot bigger.

Jensen Huang(01:29:58) And then the rest of it, to me, is: is it possible for NVIDIA to be a, you know, $3 trillion revenue company in the near future? The answer is, of course, yes. And the reason for that is because it’s not limited by any physical limits. There’s nothing that I see that says, you know, gosh $3 trillion is not possible. And as it turns out, NVIDIA’s supply chain is—the burden is shared by 200 companies. And the fact that we scale out on the backs of, with the partnership of this ecosystem, the question is: do we have the energy to do so? And surely we will have the energy to do so. And so all of these things combined, that number is just a number, you know?

Jensen Huang(01:31:42) Now, NVIDIA is not in the market share business. Almost everything that I just talked about don’t exist. That’s the part that’s hard. You know, if NVIDIA was a $10 billion company trying to take NVIDIA’s share, then it’s easy to see for shareholders that, oh, yeah, if they could just take 10% share, they could be this much larger. But it’s hard for people to imagine how large we could be because there’s nobody I could take share from. You know? And so I think that that’s one of the challenges for the world is the imagination of the future. But I got plenty of time, and I’ll keep reasoning about it, and I’ll keep talking about it, and every single GTC will become more and more real.

Jensen Huang(01:32:27) You know, and then more and more people will talk about it, and one of these days, you know, we’ll get there. But I’m 100% we’ll get there.

Conclusion: This section answers the question “But Mark don’t you think NVDA is overvalued?” No, very clearly we don’t, as we have regularly published a Price Target of $500! We think Jensen makes some very good points in the podcast transcript shown above. If we are wrong on that Price Target of $500, we are minded to consider that we are wrong by being too pessimistic. Beth Kindig has published a very convincing article as follows: Nvidia’s stock path to $20trn Yes that’s right $20trn! We suggest you click the link and read it.

NVDA GTC: we were in the Era of Inference, but now we are in the Era of Agentic Inference

Readers will know that we have often said that the Era of Ai would only get into full flow once we saw functional agents. Hence, all this talk of Ai being split into two segments: Training and Inference, and NVDA was just a Training Chip company, was tosh & nonsense. To illustrate that point, we broke down the chip market into 5 segments and used these for our NVDA Price Target model which we have published in a number of recent Newsletters. Please find small illustrative segment of model:

In 2026, the industry has realized that “Agentic AI” (autonomous systems that reason, plan, and use tools) requires a different hardware profile than simple chatbots required. A chatbot might like to undertake inference on a TPU but Agents require a “Specialized Agentic ASIC” which isn’t just one chip; it is typically a heterogeneous system designed to solve the three bottlenecks of an agent: Context Memory, Sequential Reasoning, and Action Latency.

At GTC, NVIDIA announced the new “Vera Rubin” Trinity, NVIDIA has moved away from the “single GPU” model toward a three-chip architecture specifically for agents.

• • Vera CPU (The Orchestrator): This is NVIDIA’s first “Agentic-first” CPU. Unlike traditional server CPUs, it is built with specialized “Logic-Ahead” branch predictors to handle the sequential tool-calling and API-heavy logic of agents without stalling the GPU.

• • Groq 3 LPU (The Scribe): Following the 2025 technology licensing/integration deals, NVIDIA now includes Language Processing Units (LPUs) in its racks. These are ASICs designed for 1,000+ tokens/sec generation, allowing an agent to “think” out loud and communicate with other agents instantly.

• • BlueField-4 DPU (The Context Memory): This is the most “specialized” agentic ASIC. It manages the KV Cache (the “memory CAPACITY” of the agent’s history). In 2026, agents have million-token contexts; the DPU offloads this memory from the GPU so the agent doesn’t “forget” its task mid-workflow.

Following the GTC 2026 “Vera Rubin” keynote earlier this week, NVIDIA has effectively pivoted from being a “chip company” to becoming the primary architect of Systems for the Agentic Era. While competitors like Google (with TPU v7 “Ironwood”) and AWS (Trainium 3) have excellent specialized silicon, NVIDIA’s Vera Rubin platform is the only one that vertically integrates the entire “Agentic Loop” into a single, coherent rack-scale supercomputer that solves the three specific “physics” problems of autonomous agents:

1. The Logic Bottleneck (Solved by Vera CPU)

Traditional CPUs were the “waiting room” for GPUs. The new Vera CPU, with its 88 custom Olympus cores, is designed specifically for agentic orchestration.

The Moat: Most agentic workflows spend 30-40% of their time on “logic” (using tools, running code, or calling APIs). Vera handles this 50% faster than general-purpose server CPUs, ensuring the GPUs aren’t sitting idle while an agent “decides” what to do next.

2. The Speed-of-Thought Bottleneck (Solved by Groq 3 LPU)

The biggest surprise of 2026 was NVIDIA’s $20 billion acquisition of Groq. At GTC 2026, Jensen Huang revealed that Groq 3 LPUs (Language Processing Units) are now integrated directly into the Vera Rubin racks.

While the Rubin GPU handles the massive “Prefill” (reading millions of tokens of history), the Groq LPU handles the “Decode” (writing the text). This allows agents to generate 1,000+ tokens per second, making multi-agent “debates” or complex coding tasks feel instantaneous.

3. The Memory Access Bottleneck (Solved by NVLink 6)

Agents require massive “KV Caches”—the short-term memory of their current conversation. Even if the DPU manages the memory perfectly, the system “stutters” if moving that data between the Vera CPU (the logic), the Groq LPU (the writer), and the Rubin GPU (the thinker) takes too long.

NVIDIA’s NVLink 6 provides 1.8 TB/s of coherent bandwidth between the CPU, GPU, and LPU.
Competitors using standard PCIe 6.0 struggle here; their agents “stutter” when moving data between the “thinker” (CPU) and the “writer” (GPU). In the NVIDIA stack, they effectively share the same brain.

Conclusion: All this tosh & nonsense about how NVDA is not an inference chip company is wrong! So please note this section directly answers any questions like: “Mark dont you feel that NVDA will become irrelevant as Google and AMZN build their own ASIC architecture?”

Token Growth Explosion and GPU rental prices

Semianalysis have written an excellent article on the explosive growth in Tokens and how this is driving up GPU rental prices that can be found here: Pricing is Only Going One Way For Now, and ROIC Follows. Some key quotes are shown below:

“Taken together, these factors point to a clear conclusion: GPU rental pricing is more likely to continue rising than falling.”

“The dynamic is self-reinforcing. As Neoclouds see supply tighten and prices rise, they move to secure more hardware ahead of further price increases, which only tightens supply and pushes pricing higher still. This echoes the 2023–2024 GPU shortage, when tight supply allowed OEMs to push through outsized margin expansion and drove a sharp spike in server pricing, though we think the server market is mature enough this time around that this may not repeat itself.”

“As we argued in a recent note to our institutional clients, the re-acceleration in GPU rental pricing improves Neocloud ROIC by expanding margins on already-deployed capital. At the same time, higher rental rates extend the economic useful life of existing GPUs, meaning invested capital generates cash flows for longer before requiring reinvestment.”

Again, there has been much Luddite tosh written by the Press about how GPUs would be worthless just a year after being bought, yet now we see their rental prices increasing.

Chinese Ai Models

There is a very interesting recent development in the use of Chinese Models for Agents. This explosion in growth is being driven by Lobster agents using inexpensive per token, Chinese models. This growth should NOT be ignored and could have profound global consequences in the future. See the growth of Chinese model use below:

No, the story here is NOT that China wins and the USA loses. This is a very lazy conclusion. The USA will win too, from its Token exports.

China is using subsidies provided by the Chinese Govt to re-architecture the cost base of its enterprises lower through the use of Lobster agents. This will shift down the cost of its manufacturing base just as the European manufacturing base is having its cost base shifted upwards from energy costs and minimum wage regulations. The first consequence of this movement is that this is probably the final nail in the coffin for the European Manufacturing base.

The second consequence of this change is that China is going to become a powerhouse of exporting Services as well as Goods. Those Services will be tokens. Those exported agent tokens will hollow out the Employment market in Europe (not the USA which will boom) which will lead to social unrest, and the long term destruction of the Liberal Democratic EU block. This has huge Macro Economic & Geopolitical consequences for the decade again. We will be talking more on this matter through tweets and newsletters: