Newsletter on Portfolio Changes

All Newsletters10 February 2026

Prometheus is a figure from Greek mythology, he was a Titan who stole fire from the gods and gave it to humans, and was punished by Zeus for it.

Prometheus syndrome is our era’s version of the myth: where Man punishes fellow Humans for trying to steal intelligence from Humanity and give it to Machines.

“If you don’t have a crisis, make one.” Michael Dell.

“Anyone who does not believe in miracles is not a realist.” David Ben Gurion.

Factsheet Commentary

For our full Factsheet, link here.

The Portfolio

The portfolio has changed materially and in this Newsletter we will explain why.

Let’s discuss Microsoft first:

We have tweeted throughout the last 12 months that Charles Lamanna was failing to make traction with Co-pilot at Microsoft and they were losing the agent/co-pilot wars, which Anthropic is clearly winning.

There is now a Capex War between the Hyperscalers as they rush to satisfy unmet demand and they can see the attractive future returns from being the predominant multi model platform provider. Yes they will earn great returns in the future. We have laid out in previous emails how the hyperscalers are making attractive margins and ROICs. Please see: Circular Financing, no ROIC on Ai and other such nonsense and this article from Citywire: Demystifying the Economics of Ai. Nonetheless, it is our belief that the Market views these names as a “Show Me” story whilst Capex growth > Revenue growth. So, in the meantime, it was best to reposition.

The history of ROIC:

The Journos read this article re GOOG and tell you that this DC investment will collapse GOOG’s ROIC. But GOOG has been investing hard in DC Capex for years….and its ROIC has rocketed upwards! (Check the stats below). So who do you believe GOOG management with a decades of experience OR some Journo?

The really good news is that the agents are finally coming.

We would strongly suggest you watch this video that we curated/edited of Claude Cowork in action.

The New Positions

It is worth reading this Citywire article too: Look out for Optics.

We have switched back into purely Hardware Ai plays from holding a material stake in a Hyperscaler (MSFT), but we have widened the number of Hardware positions owned mainly to include Optics.

This may help in your understanding of the new names in the portfolio. Here is how a single packet of data may move through a data center.

1. The Traffic Controller: Arista Networks

The journey begins at the “Top of Rack” switch. In a cluster of thousands of GPUs, data needs to move between processors with zero latency. This is the domain of Arista Networks.

Think of Arista as the air traffic controller for the data center. Its high-speed Ethernet switches determine exactly where data packets go. Unlike older switches designed for email and web browsing, Arista’s switches are optimized for AI workloads, which move in massive “elephant flows” that can easily clog a standard network.

2. The Speed of Light: Coherent and Lumentum

Here is the engineering bottleneck: Electricity is too slow and generates too much heat to move data between these powerful switches and servers at the speeds AI requires. The data must be converted into light.

This happens in a device the size of a pack of gum called an optical transceiver.

Coherent Corp. manufactures the transceiver module that plugs directly into the switch. It is the physical vessel that translates electrical signals into optical ones.
Inside that Coherent module, the actual beam of light is often generated by a laser chip manufactured by Lumentum.

In this symbiotic hand-off, Lumentum provides the raw photonics (the laser), and Coherent packages it into a module (the transceiver) that may plug into Arista’s hardware.

3. The Thermodynamics: Vertiv

While the data is moving at the speed of light, the physical hardware is generating heat at the speed of a toaster oven. A rack of Nvidia’s Blackwell chips consumes over 100 kilowatts of power which is ten times the density of a standard server rack from five years ago.

Traditional air conditioning is insufficient for this density. Vertiv provides the liquid cooling distribution units (CDUs) that pump coolant directly to the cold plates sitting on top of the chips. Vertiv also manages the high-voltage power distribution, ensuring the massive surge of electricity needed to run the Arista switches and optical modules doesn’t trip the facility’s breakers.

4. The Long Haul: Ciena

Finally, the AI model is often too large to reside in a single location. It must be synchronized with data stored in a facility miles away.

The signal leaves the local Arista network and is handed off to an optical transport system built by Ciena.

Ciena’s role is “Data Center Interconnect” (DCI). It takes the light signals generated by the Lumentum/Coherent components and amplifies them for long-distance travel over fiber optic cables. Ciena ensures that the latency between two buildings is so low that the AI model “thinks” the two buildings are actually one giant room.

Ok, its simplistic but it gives you an idea. The key to widening the diversity of hardware names is not to bear too much “future architecture change” risk. If you are a third line supplier to the huge racks primarily designed by Nvidia or others, you can easily get designed out of their future. So there are plenty of hardware names running hard at the moment but we would rather reposition into the relatively lower and medium risk names (noting that all Tech is High Risk).

Great so the overweight position in Microsoft has gone and the portfolio from position number 2 onwards is more balanced, so what about Nvidia?

NVIDIA

2026 will be the year that the Market starts to see who will win the War for the Inference Chips. Most people assume that NVDA will maintain its market share in GPUs. As a reminder GPUs are the flexible workhorse that is commonly used to train models and undertake inference on models when flexibility in the hardware is required.

Whereas, TPUs are ASICs that are specialised for inference at scale. TPUs have to be set up in advance, while GPUs are more flexible on the fly.

Why TPUs were created (the original problem)

Around 2013–2015, Google faced a crisis:

Search ranking
Ads targeting
YouTube recommendations
Gmail spam filtering

All relied on:

Neural networks
Real-time inference
Billions of requests per day

CPUs were too slow
GPUs were too power-hungry and expensive

So Google designed TPU v1 (2015) for inference only.

Even modern LLMs:

Still rely on repeatable tensor ops
Even if the content is dynamic
The math is fixed

So TPUs remain ideal.

A TPU (Tensor Processing Unit) needs you to decide ahead of time:

what data will flow through it,
how it’s configured,
and what exact operations (kernels) it will run.

All of that must be pre-planned before the computation starts. A GPU, on the other hand, can decide and launch its operations while the program is running. It’s more adaptable and can handle changing tasks without as much upfront planning.

Analogy:

A TPU is like a factory assembly line that must be fully designed before it starts running.
A GPU is like a workshop where tools can be picked up and used as needed during the job.

So the core idea is: TPUs trade flexibility for efficiency, GPUs trade some efficiency for flexibility.

TPUs use HBM as its memory facilitator but it is important to note that there are also SRAM based inference chips.

1️⃣ HBM-based inference chips

(capacity & flexibility first)

Examples

GPUs (NVIDIA A100 / H100 / B200)
Google TPUs (v4 / v5)
AWS Trainium / Inferentia
NVDA Rubin CSX

Memory traits

Large off-chip HBM
Can handle very large models
Dynamic memory access
Kernel scheduling at runtime

Pros

Run huge models (70B, 100B+)
Flexible workloads
Strong ecosystem support

Cons

Higher latency
Less predictable timing
Power and cost overhead from HBM

2️⃣ SRAM-based inference chips

(latency & determinism first)

Examples

Groq LPU (GroqChip-1)
Cerebras (largely SRAM-centric, though architecturally different)
Some edge AI accelerators

Memory traits

Mostly or entirely on-chip SRAM
Static, pre-planned memory access
No cache misses, no page faults

Pros

Ultra-low latency
Fully deterministic performance
Very high tokens/sec per watt

Cons

Model size constrained by SRAM
Less flexible
Requires compilation upfront

So actually using Claude we have updated and restructured our Nvidia model and we have reverse engineered the model from a bottom up enterprise workflow approach.

Here are our cohorts of workflow types:

Each of these workflows is best served by a different type of semiconductor, which creates the following model. Note that Agents and Robots are complex, and require GPUs and we believe that as Ai gets more intelligent it will undertake more complex processes.

You will see that our Price Target, even with loss of Market Share and Reducing Margins, is >$500 for 2030. So NO we are not going to reduce our Nvidia position unless the facts change or the price hits $500 early.

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.

We have released a new presentation for Q1 2026 which updates on our thinking for the year ahead. Please do review this presentation here.