All Newsletters19 August 2026

“Ai value will not arrive in one lump to be distributed downwards. It will come workflow by workflow in the hands of people who understand the work.. A payroll team in Oslo, a tax team in Lisbon…” Neil Lawrence, DeepMind Professor of Machine Learning, University of Cambridge.

Talking about the 5 layers of AI (being Energy, Chips, Infrastructure, Models, Applications): “They are all important. The one that is the most important, of course, is the Ai Application Layer.” Jensen Huang, Nvidia.

“We are working exactly on the projects that we want to work on and if the workload were to change dramatically, and I don’t mean the algorithms, I actually mean the workload, and that depends on the shape of the market.” Jensen Huang, Nvidia.

“It takes all the running you can do, to keep in the same place.” Red Queen to Alice in Lewis Carroll’s Through the Looking Glass.

“Sometimes, I feel the past and the future pressing so hard… there’s no room for the present.” Evely Waugh.

The Rise of the Model Giants Part 2

This section follows on from our previous Newsletter which can be found here.

The corollary of the Rise of the Model Giants, is the fall of the Hardware King, Nvidia.

How did this happen?

In the early phase of the AI build-out, model labs and hyperscalers mostly consumed NVIDIA in the obvious way: buy the bare metal and take the full CUDA software ecosystem with it. That made sense. NVIDIA offered not just chips, but a mature, integrated stack of libraries, compilers, kernels, networking, developer tools and operational knowledge. For a frontier lab racing to train larger models, that ecosystem was often as valuable as the GPU itself.

Anthropic’s strategy changed that. There are many different types of AI workflow but Anthropic recognised that many large-scale AI workloads are dominated by a relatively narrow set of operations: matrix multiplication, attention, normalisation, activation functions, collective communication and memory movement. A general-purpose NVIDIA GPU remains more flexible and enjoys the richest software ecosystem, but not all of that flexibility is required all the time for transformer training and inference.

Rather than being locked into one hardware platform, Anthropic built the engineering capability to run Claude across NVIDIA GPUs, Google TPUs and AWS Trainium. That does not mean abandoning vendor software stacks, it means investing in the internal layers that make those stacks usable at frontier-model scale: model code, kernels, compiler paths, distributed training systems, sharding strategies, memory management, numerical stability, checkpointing, failure recovery and orchestration across very large clusters.

The hard work sits between the model and the silicon. Anthropic still relies on platform-specific ecosystems such as AWS Neuron for Trainium and Google’s TPU software environment, but it is also doing the specialised engineering needed to make those accelerators productive for its own workloads. In some cases that means writing or tuning low-level kernels or it sometimes means reshaping computation graphs, changing scheduling strategies, adjusting data layouts, or routing different workloads to different hardware backends.

Where specialised accelerators lack some of the flexibility of GPUs, the workload can often be compiled, scheduled, partitioned or adapted so that most of the computation runs efficiently on the cheaper or more available platform. Less common operations may still require custom kernels, compiler workarounds, host-side orchestration, or, where necessary, fallback to other hardware.

The strategic trade-off is therefore clear. A lab can pay more for NVIDIA’s mature, general-purpose ecosystem, or it can spend more internal engineering effort to make TPUs and Trainium work efficiently at scale. Anthropic chose the second path where the economics justify it: not replacing NVIDIA entirely, but reducing dependence on NVIDIA by building a multi-platform compute stack.

Now everyone has caught on:

When did this happen?

Phase 1: NVIDIA GPUs and CUDA became the default substrate for modern AI.

Phase 2: High-level machine-learning frameworks began to abstract parts of CUDA away. Developers increasingly worked in PyTorch, JAX, XLA and other higher-level systems, while the lower layers translated that code into hardware-specific execution.

Phase 3: Google proved that a custom ASIC, if paired with a serious compiler and runtime stack, could beat general-purpose GPUs on important machine-learning workloads. Many people got confused, the lesson was not simply “ASICs are better”, it was that specialised silicon only works when the software stack, compiler, networking fabric and operating model are built around it.

Phase 4: At LLM scale, the decisive unit of competition stopped being the chip alone. It became the full system: chip, compiler, memory hierarchy, interconnect, distributed runtime, serving stack, power availability, procurement and model architecture. Remember when Jensen said NVDA was not a chip company.

Phase 5: Anthropic made this explicit. It says its compute strategy uses Google TPUs, Amazon Trainium and NVIDIA GPUs. That is the important change. The frontier lab is no longer so structurally dependent on one accelerator supplier. It treats compute as a portfolio, and they believe this is to the advantage of the training of the model.

Using ASICs is nothing new. It is the growth in Revenues from Anthropic AND the discovery that they have achieved what OAI has achieved with 80% less cost that woke everyone up. Anthropic is proving it is worth accepting the engineering tax (hundreds/thousands of engineer-years for software stacks) to access alternative silicon at scale.

Let’s switch approach and now look at the actual enterprise workflows that AI is being asked to execute and then assess which may be the best chip to use for each workflow, then we get a table that may look something like the below

IF we ask GPT the obvious next question: what is the breakdown of the use of differing semiconductor hardware by type in a Data Centre used by Anthropic for solely Agentic Workflows? We get the following:

One can assess the threat for NVDA which currently has a Market Share of about ~80% AI accelerator / AI training-inference silicon revenue. It was NOT such a big problem when Anthropic was just one player of many BUT it becomes a very big problem if this market becomes winner takes all, and Anthropic wins.

This is why Jensen is very keen to mention any deals NVDA does with Anthropic:

What happens now?

There are many different types of inference demand:

1. Commodity / high-throughput tokens
Cheap tokens, huge volume, latency matters but price per token is low. This is where ASICs, TPUs, Groq-style architectures, custom inference chips, etc. may try to win.

2. Premium / high-ASP tokens
Harder, more valuable inference: reasoning, agents, enterprise workflows, coding, financial analysis, long-context, multi-modal, real-time tool use. Even if the chip produces fewer tokens per second, the value per token is much higher, so customers will still pay for more flexible/expensive Nvidia architecture.

Groq/TPU/Trainium/custom ASICs will win the cheap, predictable, high-volume inference. So NVDA bought Groq.

“Adapt or perish, now as ever, is Nature’s inexorable imperative.” H. G. Wells

But Nvidia remains strongest where inference is complex, dynamic, agentic, memory-heavy, high-value, or needs the full CUDA/software ecosystem. Jensen believes NVDA will always retain this section of the Inference market.

Then NVIDIA deepened its push into the data-centre CPU market with Vera, a processor designed for agentic AI, reinforcement learning and data orchestration. Building on Grace, Vera can operate as a standalone CPU or alongside Rubin GPUs, extending NVIDIA’s reach beyond accelerators and giving it greater control over the entire Data Centre AI-computing stack. Note both the CPU and GPU market are roughly equally sized per annum in value terms.

At this point, I strongly suggest a read of the following article (click the image):

Just to note, NVDA are now also going after the consumer CPU Market, they have announced NVIDIA’s First Consumer ARM PC Chip which they claim is 1.8x faster task completion vs x86 on agentic AI, RL and data processing.

So what else did Jensen do to ensure NVDA’s dominant position in the future? NVIDIA has increasingly acted as both supplier and financial backstop to the AI infrastructure boom, creating what critics describe as circular financing: it sells GPUs to neoclouds, invests in them and, in arrangements such as CoreWeave’s $6.3 billion agreement, commits to purchase residual unsold computing capacity. NVIDIA has now announced partnerships with Apollo, BlackRock, Blackstone, Brookfield, Goldman Sachs and KKR intended to mobilise more than $500 billion of third-party capital for AI infrastructure. The strategy reflects how difficult and expensive it can be for young, highly leveraged neoclouds to secure financing when utilisation and GPU residual values remain uncertain; NVIDIA’s offtake commitments make these projects more bankable while simultaneously supporting demand for its own chips.

At this point, the obvious question is: look, why does Jensen not swallow his pride and go to Dario at Anthropic and do huge five year deal with a tasty discount for Anthropic to announce sole use of NVDA systems? NVDA’s leverage is reinforced by supply: NVIDIA is reported to have secured substantial TSMC fabrication and advanced-packaging capacity several years in advance, giving it privileged access to one of the industry’s principal bottlenecks. Anthropic, meanwhile, appears to have underestimated how quickly demand would outgrow its available infrastructure and is now securing capacity wherever it can, including from third-party providers and neoclouds whose systems are often built on NVIDIA architecture. Compute scarcity therefore keeps pulling Anthropic back into NVIDIA’s ecosystem, despite its efforts to diversify across TPUs and Trainium and despite the well-documented policy and personal friction between Jensen Huang and Dario Amodei. Yes there is an ego problem here between two people who consider themselves geniuses. Does that ever end well?

So then Jensen decided to turn to another genius: Elon Musk. Musk has committed SpaceX, which now includes xAI, to a colossal expansion of computing capacity from roughly 1.4 gigawatts toward as much as 10 gigawatts by the end of 2027 and has said its future AI infrastructure will be built exclusively on NVIDIA’s Vera Rubin platform. This effectively turns one of the world’s most aggressive compute build-outs into captive demand for NVIDIA, giving Huang an anchor customer capable of absorbing a significant share of the company’s future chip production.

How does this all end? God knows, but you can see why we have to watch every deal announced by SPCX, NVDA, OAI and ANTH very carefully. Imagine how pleased Jensen will be if Elon’s models start to compete with Anthropic’s?

The volatility of the Technology market

Several weeks ago, UBS raised Micron’s target price from $535 to $1,625, helping drive a 19% one-day share-price jump and pushing Micron above $1 trillion market cap. UBS cited AI-driven memory demand, tighter DRAM/NAND supply, and longer-term supply agreements with partially fixed pricing as reasons for the huge re-rating. UBS’s Micron upgrade was not a normal target-price tweak; it was a full-scale capitulation to a changed earnings regime. The bank lifted its price target from $535 to $1,625! The core argument is that Micron is no longer just a highly cyclical DRAM/NAND producer, but a strategic supplier into an undersupplied AI infrastructure market, with long-term agreements, tighter capacity and rising pricing power giving investors visibility into materially higher earnings and free cash flow. The more interesting point is how common these enormous “catch-up” upgrades have become across the AI supply chain and how asleep at the wheel the IB analysts are. Micron, Sandisk, Seagate, Western Digital, GE Vernova, Nvidia and others have all seen analysts repeatedly chase the share price higher after the market had already recognised the inflection. That pattern suggests the investment banks are increasingly reactive rather than anticipatory: models are being rebuilt after the fact, targets are being revised in huge jumps rather than steady increments, and “Street-high” calls often arrive only after the stock has already made the move. In this world, many are asking is broker research now useless!

My key point is if the consensus is SO far from the Reality are you really surprised this is causing volatility? Do you think that is a positive skew volatility and do you see that positive skew volatility is an opportunity?

“Is the high volatility in a fund the fault of the Fund Manager if the underlying Market the fund is investing in is volatile?” It is interesting to note that psychologists claim that this question is a psychometric IQ-discrimination test. Apparently, around 90 per cent of value investors answer: Yes. It would seem that the roughness of the North Sea is perceived to be the fault of the fishermen that fish it!

“To be resilient, prepare, don’t predict.” Carissa Veliz, University of Oxford.

“He was gifted with the instinct for self-preservation that passes for wisdom among the rich.” Evelyn Waugh

The Summer Season

The Era of AI loves a market drama, especially one that gives risk-averse investors permission to panic first and think later. We had the DeepSeek moment, when the market briefly decided that cheaper models meant the end of the AI capex cycle. Then came the tariff scare, when a short-term policy shock was extrapolated by the confused into a massive structural earnings crisis.

This summer has taken the genre to new heights. Meta’s reported plan to rent out spare compute was immediately translated into proof that the entire sector had overbuilt even while other providers remain capacity-constrained. DeepSeek’s proposed ASIC became fresh evidence that NVIDIA was finished. Then came Leopold Aschenbrenner’s July margin crash. A failure of position sizing and leverage was naturally presented as a failure of the whole AI thesis itself.

The model discourse has been no saner. Every new open-weight release produces another confident declaration that “open source will eat the frontier models”, as though cheaper inference and downloadable weights instantly eliminate the advantages of frontier research, training scale, proprietary data, compliance, distribution and full-stack products. Kimi K3 is reportedly impressive in model architecture, but again is an example of distillation. Naturally, this has allowed both tribes to claim victory. The open-source evangelists say it proves the closed-model moat is dead; the frontier labs say it proves open models can compete only by copying them.

None of this means the risks are imaginary. Capex funding IS causing crowding out, custom silicon will take share (we have dealt with that point above), open-weight models will pressure pricing, and excessive leverage remains an excellent way to turn a sound long-term thesis into a short-term catastrophe. But the market repeatedly confuses marginal evidence with a definitive regime change.

So consider this:

“The AI market is probably less than 3% penetrated. And so everybody’s going to grow.” Ben Horowitz: The Fight Over Open Source AI.

Historically, if you want to estimate where a TAM is going for a market you look at the early, fast adopter then you look at the median spend and you apply a fraction of the early, fast adopter spend. So currently the median spend is $11.38/$7450 = 0.15%.

The mature cloud median is generally 8–15% of P99, with a cross-industry midpoint around 12–13% → ~$1,120 so the striking result is that AI’s current 0.15% ratio is roughly 100 times less converged than a mature usage-based technology market.

Interestingly, the complete SaaS benchmark comes from Zylo’s 2026 index was $9,455 per employee annually, or about $788 monthly. That includes hundreds of established applications across productivity, CRM, security, finance, HR and other functions. Could we see a future where all this and more is spent on intelligence?

The conclusions are simple. Will you let one volatile month weaken your commitment to the huge opportunities ahead created by AI’s inevitable diffusion through the economy? Are you sure today’s noise will still matter in three months, let alone three years? Rapid growth and structural change will inevitably produce volatility; it cannot be eliminated. When the upside is asymmetric, is volatility a weakness or simply the price of participation? Will any month matter in the course of an Era? Do you see more exciting investment opportunities anywhere else?

The task is not to predict every turn, it is to size exposure to the opportunity intelligently, preserve the ability to stay invested, and allow positive skew to compound over time.

“We’re still in the early days, yet I already believe AI will be the most profound shift of our lifetimes.” Sundar Pichai, Google and Alphabet CEO.

Marvell Technology (MRVL) — Evercore TMT Global Conference
Speakers: Willem Meintjes (CFO/EVP), Ashish Saran (SVP Investor Relations)

Saran: “Our interconnect business in particular is absolutely on fire. I mean that thing’s growing at 70% plus for this year.”

Saran: “These XPUs are no longer monolithic single-chip devices, right? These are basically multiple compute die, HBM stacks. You need high-speed SerDes, you need die-to-die interfaces, you need custom HPM interfaces, you need much more optimized custom SRAM.”

Saran: “The rate of acceleration is increasing. The complexity is going up. Again, scale-up is a great example of it. Scale-up networking outside of one player didn’t really exist at this point in time.”

Saran: “You’re hearing about agentic AI today, all it’s doing is driving a lot more traffic. You need to disaggregate your memory at this point. You need a lot more traffic between XPUs. All of that is again back to networking IP.”

Saran: “The TAM numbers are probably going to keep floating up… I just think that the sheer complexity of the network, right, with agentic, with mixture of experts, with all of these different inferencing models is just a lot more than what we all thought, even just barely a year or two years back.”

AMD (AMD) — BofA Global Technology Conference

Speakers: Jean Hu (CFO), Matt Ramsay (VP Investor Relations)

Hu: “The biggest change during the last few months is really the rise and inflection of agentic AI… It’s not about answering questions anymore. It’s about orchestration. It’s about database access, and a lot of tool execution, and all of those require significant CPU performance.”

Hu: “What we’re seeing is very significant, and the incremental demand for our CPU platforms. That has been really exciting.”

Hu: “We had record CPU performance. CPU business grew more than 50%. We guided the Q2 CPU performance is going to go up year-over-year 70%.”

Hu: “Very large CapEx planning, very long planning cycle. They actually plan way ahead. 2027, we feel really good about both the visibility of the demand side and the visibility of the supply side.”

CPU TAM — Raised From $60B → $120B+

Hu: “Last November, when we had our Financial Analyst Day, we actually outlined how we think about this opportunity… we believe the TAM is going to grow 18% CAGR to $60 billion in 2030. Remember, the general purpose of the CPU used to be just a single-digit growth.”

Hu: “When we had our earnings call, we actually updated our market forecast to more than $120 billion. But frankly, it’s still very early… The pace and the speed of agentic AI adoption has been tremendous.”

Hu: “It’s evolving so fast and more and more complex, you should expect for the agentic AI portion of the market because you have so many diverse workload so complex, the ASP will continue to increase because increasingly, you need a very high core count CPUs — high-performance CPUs.”

Hu: “Agentic AI — you are seeing agentic AI server rack sit in between the traditional servers and the GPUs, and those racks are handling all those different workload… That market, we think, by whatever the $120 billion or $200 billion market opportunity is, the majority of that large market.”

Lam Research (LRCX) — BofA Global Technology Conference, 02 June 2026
Speaker: Doug Bettinger (EVP & CFO)

Bettinger: “I’m a pretty conservative guy, generally speaking. I’m talking as optimistically as you’ve ever heard me if you’ve listened to me talk for a while, because I’ve not seen it as rich as it is right now.”

Memory LTAs Underpinning Visibility

Bettinger (on memory long-term agreements): “That’s probably a part of what’s behind these very rich, robust, longer-term conversations that I was describing earlier in our talk… the confidence in these conversations is very, very high. And this is probably certainly part of it.”

Bettinger: “As long as everybody is as profitable as they are, they’re going to invest and they are investing.”

Capital Return

Bettinger: “Our plans are to return 85% of free cash flow to shareholders. In the last several years, we’ve returned more than that… The last three years we’ve grown it [the dividend] annually 15%. We’ll grow the dividend again this year.”

Key Manchester & London Updates

Please remember that all Fund based news is now posted to: https://www.linkedin.com/company/mnl-ln

Twitter will only see posts on AI, Technology and the global Economy.