§ I
Where I am standing
My name is Lars Gross. I am an AI engineer based in Zurich. I do not run a fund. I do not manage other people's money. I am the sole manager of this book. What I do professionally is build production AI systems: transformer inference stacks, retrieval pipelines, evaluation harnesses, the unglamorous plumbing that turns a research artifact into something that survives a Tuesday at 3pm under real load.
The tools I build for work are the same tools I use to research positions. I make investment decisions using privately developed AI analysis systems that process technical research, chip roadmaps, patent filings, inference benchmarks, and power infrastructure data to surface asymmetric opportunities. These are not dashboards with canned signals. They are purpose-built systems that let me evaluate the same engineering substrate I work with every day.
That vantage point matters, because almost every analyst writing about “AI exposure” today is reading the same earnings transcripts, the same sell-side desks, the same Stratechery paragraph. They are pattern-matching to the dot-com era or to consumer internet. I am reading the kernel commits, the CUDA release notes, the MLPerf submissions, the latency budgets that decide which model is actually cost-viable at scale.
That gap between what people say about AI and what the engineering substrate is actually doing is where the mispricing sits.
§ II
The thesis, in one paragraph
We are in the early years of the largest capital-deployment cycle of my lifetime: an industrial build-out of compute, energy, and data infrastructure to support models that will become economically indispensable across every white-collar function. Markets understand this directionally. They consistently mis-estimate its magnitude, its durability, and who actually captures the rent. The book is concentrated where the engineering reality and the market narrative are furthest apart.
The market has priced a narrative of AI progress. It has not priced the engineering reality of AGI. These are different things. Most participants — fund managers, analysts, commentators — are working from a mental model that treats transformers as a smarter search engine and the next model release as an iPhone iteration. Anyone who has read the actual research, run the actual benchmarks, or built against the actual APIs in the last eighteen months knows that this model is wrong by an order of magnitude. Progress looks incremental from the outside. It looks compounding from the inside. That asymmetry of perception — what practitioners in the field call situational awareness — is the structural edge this book is built on.
§ III
Five convictions that drive position sizing
1. Inference, not training, is the durable demand.
Training spend is lumpy, headline-driving, and finite per model generation. Inference is the recurring revenue layer: every query, every agent loop, every embedding refresh. Every capable model that gets deployed creates a permanent floor of inference demand that compounds with adoption. Capex narratives that focus on training H100 counts miss this entirely. The book is biased toward names that capture inference economics: memory bandwidth, networking fabric, power-efficient accelerators, and the dark-fibre and cooling layers that make inference at scale physically possible.
2. The bottleneck moves. Be honest about where it is.
In 2022 the bottleneck was GPUs. In 2024 it was networking and HBM. In 2025-2026 it is power and grid interconnect. Each phase produced its own asymmetric winner that the consensus only priced after the move. The discipline is not to be permanently attached to one bottleneck name: it is to know which physical constraint is currently binding and rotate weight toward the names that resolve it. I update this view in writing, quarterly, in the letters.
3. Software margins are not safe yet.
The most contrarian view I hold is that a meaningful slice of the application-software layer is at structural risk over the next five years. Not because demand collapses: because the model layer commoditises features that used to be a moat, and because new entrants ship with a fraction of the headcount. I size software exposure carefully, and I am willing to be short specific names where the engineering thesis is clear.
4. AGI is not priced in. Neither is what it breaks.
The market has priced “AI as a feature.” It has not priced AGI as a regime. These are different bets with different magnitude. Systems capable of autonomously executing multi-week knowledge-work projects — plausible within this decade on current scaling trajectories — would re-rate labour, capital allocation, and the relative value of every business that monetises cognitive output. That is not in the consensus discounted cash flow. The infrastructure required to run those systems is also not fully priced: the power, the chips, the networking, the cooling. The book is long that infrastructure, sized for an outcome that most models still treat as a tail risk rather than a base case.
The corollary is that the hype layer — consumer AI applications, wrapper startups, “AI-powered” SaaS priced at forty times revenue — is overvalued relative to what AGI actually does to application software. When the model layer commoditises, most of the application moat disappears. I avoid that trade, and in selective cases I am on the other side of it.
5. Finance is the next vertical to be structurally repriced.
Financial institutions are beginning to run AI inference at scale for credit decisioning, trading, and compliance workflows. Banks and fintech platforms that own the inference layer rather than outsource it will see structural cost advantages and durable moat extensions over the next several years. This is not a pure AI trade. It is a financial-sector trade with an AI catalyst. The book holds selective exposure to institutions where the AI thesis has a clear, verifiable engineering component: where the internal systems are being built, not bought off a vendor shelf, and where the economics of that choice are not yet reflected in the multiple.
§ IV
How I size, when I sell, what I avoid
Concentration over diversification.
The book holds roughly ten to twenty positions at any time. Top five typically account for more than half of capital. Diversification is the price you pay for not having a view. I have views.
Position-size by conviction, not by market cap.
A small-cap with a verifiable technical moat and asymmetric upside deserves more weight than a mega-cap consensus name where the entire street agrees with me. The biggest mistake an individual investor can make is to hold mega-caps at index weight and call it a strategy.
Options and leverage are tools, not strategies.
I use long-dated calls on specific binary catalysts where the engineering evidence is strongly asymmetric and the implied volatility under-prices the catalyst. I use leveraged instruments selectively for short-duration trades around clear technical events. Every such position is disclosed in the register with strike, expiry, and rationale. The default instrument is the underlying stock.
What I avoid.
Pre-revenue AI consumer apps. Anything whose moat depends on an OpenAI API rate limit. Pure-play training-compute names priced for permanent peak demand. Crypto narratives dressed in AI clothing. Sell-side consensus mega-cap longs at index weight: I either have a real view or I don't own them.
§ V
What I owe the people reading this
Transparency, on a delay short enough to be useful and long enough to not be reckless. Every open position is on the public register. Every closed position stays on the register: including the losers, with the actual realised return, not the sanitised version. Theses are written before the position is opened and dated. When I am wrong, the post-mortem is published.
I do not promise alpha. I do not promise a smooth ride. Concentrated portfolios can and will draw down thirty percent or more. I promise only that what you see here is what I am actually doing with my own capital, and that the reasoning will be written down before the outcome is known.
§ VI
A note on what this is not
This site is not investment advice. It is not a solicitation. It is not a fund. I am not a registered adviser in any jurisdiction. I run a personal book and I write about it publicly because the discipline of writing is what keeps the reasoning honest.
If at some point in the future I do raise outside capital, it will be on terms that are clearly disclosed and to qualified investors only. People who have been reading the letters will know first.
Lars Gross · Zurich · Since 2023