June 24, 2026

The AI Buildout Is Not Just a Chip Trade

My current research map for the long-term AI infrastructure sleeve of my portfolio: power, cooling, memory, grid equipment, chips, agents, and the physical buildout behind the biggest trade I can see.

First of all a heads up, this is a very long form post on how I'm trading this current market. This one is different for me, it is probably the longest email I've written to you all ever, it's about 5,000 words and I tried to be as brief as I could while still getting it all out there. Of course I used AI to help me write this, and for those of you who like to reply to tell me I made spelling errors or run on sentences, hopefully you'll not see that as much here. I could have leaned on AI more to explain all the different technologies and nuances of each but that'd be a 50,000 word document, so forgive me if some of these subjects or acronyms don't make sense immediately.

Feel free to hit me up with your thoughts, ideas or questions...and with that I hope you enjoy this, I had a blast writing it.

Not investment advice. This is research, process, and how I am thinking about one sleeve of my own portfolio. Markets change. Prices matter. Position sizing matters. Risk management matters.

Most of my work is systematic trading.

That is the foundation of my portfolio. Rules. Backtests. Out-of-sample validation. Live monitoring. Multiple strategies stacked together so I am not depending on one idea, one market, or one regime.

That is still the core.

But when you get enough of that running on autopilot, it frees you up to work on the stuff that does not fit neatly into a five-minute chart or a simple signal rule.

This is one of those things.

The AI buildout is the biggest trade I can see right now. Not the easiest trade. Not the cleanest trade. Definitely not a straight line. But the biggest.

And I do not think the market is framing it correctly.

Most people still talk about AI like it is a chatbot cycle, or a software cycle, or maybe a chip cycle if they are a little more market aware.

I think that is too small.

The better frame is that AI is becoming a physical-world industrial buildout.

Power. Cooling. Grid equipment. Transformers. Switchgear. Memory. Optical networking. Construction. Gas turbines. Nuclear. Cybersecurity. Robotics. Lab automation. Defense autonomy. On-device compute.

The question is not just, “which chatbot wins?”

The better question is:

Where does accelerating AI capability collide with scarce physical capacity, regulated access, or locked-in distribution?

That is where I think the opportunity is.

And that is why I am putting together a dedicated research and portfolio sleeve around this.

Not to chase every AI headline.

Not to hide in cash waiting for the “bubble to pop.”

Not to convince myself it is already over because some names look extended.

The point is to stay on the right side of the biggest trade around, while also being prepared for the turbulence, the drama, the sentiment shifts, the positioning washouts, and the moments where something actually changes.

Because something will change.

Tomorrow some new model, chip, energy deal, regulation, construction bottleneck, or scientific breakthrough could change part of the thesis.

That is the whole reason to track it.

Why I care about this so much

I got into artificial intelligence originally back in the early 2000s, tinkering with neural networks.

Then in 2017, when Google released “Attention Is All You Need” and the transformer showed up, I veered in that direction for a bit. I was interested. I knew something was happening there but at the time, I found more success in machine learning than AI.

I had early access to GPT-1 and GPT-2. But I did not know large language models, chatbots, and agents were going to be the thing.

I kept trying to use them for trading or any other ideas I had, but they were not good at it, and there was a lot more juice in there than I was able to squeeze.

Then ChatGPT came out in 2022 and it blew me away.

Not because it was perfect. It was not. It hallucinated. It made mistakes. It sounded confident when it should not have.

But it could do a lot of amazing work I had been struggling with for years. It took my work to another level almost immediately.

The first time I watched it write code for me, I knew this was not going away. It was only going to get better.

One funny thing happened early on.

A lot of the best engineers I knew blew it off. They said it was a joke. It could not do what they could do. It made too many mistakes. It was just autocomplete. All of that.

But one friend of mine, one of the most autistic engineers I knew, was using it constantly.

He called it tab coding.

You would type, the AI would suggest the next code change, and you would hit tab.

Then slowly, all the other engineers started coming around. They took their time. They all had their reasons. But eventually it became ubiquitous.

Since November 2022, I have spent the majority of every day working with these AI tools.

The speed at which this has taken over engineering is astonishing.

This is when I started taking individual investments in different companies that I thought would be a big part of this.

And it is not just engineering anymore.

It is a main part of every day work in research, finance, operations, customer support, data analysis, cybersecurity, content, legal work, education, and soon the physical world through robotics and lab automation.

My big thesis early on was that companies that control their own chip design, and in some cases their own chips, would have an advantage. That is still playing out now, especially with Google and TPUs, Apple and its own silicon, and the hyperscalers moving deeper into custom silicon.

I wrote a big piece about TPUs in 2024 before it became something most investors were talking about, because I had been tinkering with TensorFlow and I understood what an ASIC like a TPU could do, what it could not do, and how it compared to a GPU or CPU.

That came from years of stacking and racking bare metal boxes all night long in freezing data centers, building systems that had to handle millions of users, watching what broke, watching what scaled, and watching what looked good in theory but failed in the real world.

That background matters here.

Because this is no longer just about model demos.

It is about whether you can get enough power, cooling, chips, memory, networking, land, permits, labor, and capital into the same place fast enough.

It's about how well you can access all that and do something important with it like creating LLM's that the government shuts down, biological research, physics, math and millions of other important uses.

That is the trade now, and I'm about to take you deep down that rabbit hole...

The core thesis

The market still wants AI to be a clean software story.

It is not.

It is an industrial buildout.

The models matter. The chips matter. But the bottleneck is moving from “who has GPUs?” to “who can build AI factories?”

Because what comes out the other end of the AI factory matters.

An AI factory is not just a room full of NVIDIA chips.

It needs:

  • Power
  • Grid interconnects
  • Transformers
  • Switchgear
  • Backup generation
  • Cooling and heat rejection
  • High-density racks
  • HBM and advanced packaging
  • Optical networking
  • Fiber
  • Land
  • Water
  • Permits
  • Construction labor
  • Commissioning speed
  • Cybersecurity
  • Monitoring
  • Physical operations

That is why I think the interesting opportunities are not only in the obvious AI names.

Some of them are obvious for a reason. NVIDIA is obvious for a reason. Microsoft, Google, Amazon, Meta, Broadcom, TSMC, ASML, AMD — these are all real parts of the story.

But the asymmetric work is usually not in staring at the same chart everyone else is staring at.

It is in asking what every additional AI factory needs.

And when you ask that, the trade gets much wider.

Electrical gear. Power distribution. Cooling. Mechanical contractors. Engineering and construction firms. Gas turbines. Existing nuclear. Memory. Storage. Optical networking. Copper. Cybersecurity. Identity. Agent governance. Lab equipment. Robotics.

And we're already well into that move.

The bottleneck moved from GPUs to power, cooling, and construction

The first phase of the AI trade was easy to understand.

Models needed GPUs. NVIDIA had the GPUs. Everyone bought NVIDIA.

That was correct.

But now the bottleneck is broadening.

Goldman Sachs has estimated data center power demand could grow about 160% by 2030. Data centers are expected to move from roughly 1–2% of global electricity demand to something more like 3–4%.

Morgan Stanley has framed the same problem as an electricity-market buildout, with global electricity demand rising by more than 1 trillion kilowatt-hours per year through 2030 and data centers contributing a meaningful share of that incremental growth.

The exact numbers will move around. Forecasts always do.

But the direction is clear.

AI demand is becoming electricity demand.

And electricity is not software.

You cannot deploy a substation from GitHub.

You cannot solve a transformer shortage with a prompt.

You cannot talk a utility into giving you firm power tomorrow because your model benchmark went up.

This is where the physical world starts to matter again.

Large transformers, switchgear, breakers, substations, high-voltage equipment, and the people who know how to install and commission all of it are becoming critical bottlenecks.

Some of this equipment can have lead times measured in years.

That changes the investment map.

It pushes attention toward companies like Eaton, Schneider Electric, ABB, Siemens, GE Vernova, Hubbell, Powell Industries, nVent, Vertiv, Quanta Services, EMCOR, Comfort Systems, MYR Group, MasTec, and other less glamorous names that touch the electrical and mechanical guts of the AI buildout.

This is the part of the trade that does not always sound exciting at first.

But every megawatt needs equipment.

Every data center needs electrical balance-of-plant.

Every high-density rack needs power distribution.

Every power contract eventually has to become a real site with real gear installed by real people.

That is where I think a lot of investors are still underweight mentally, even though most of these stocks have already moved.

Liquid cooling is not optional at the high end

Another big shift is rack density.

As systems move toward NVIDIA GB200 NVL72-class racks and beyond, the power density gets intense. We are talking about rack-scale systems that can push around 100kW or more per rack.

At that point, air cooling becomes expensive, inefficient, or just not physically realistic.

Liquid cooling becomes the direction of travel.

This does not mean “cooling is solved.”

That is the mistake.

Even if newer systems can use warmer water, and even if some future designs can be chiller-light or chillerless in certain climates, you still need CDUs, pumps, loops, filtration, leak detection, redundancy, controls, water or glycol systems, heat exchangers, and heat rejection.

You still need someone to design it, install it, maintain it, and fix it when something breaks.

So the beneficiaries are not just the chip companies.

They are also the cooling and thermal infrastructure companies: Vertiv, Modine, Trane, Carrier, Johnson Controls, AAON, Comfort Systems, Watsco, and the private suppliers that make the boring pieces nobody cares about until the rack does not work.

This is the pattern across the whole AI buildout.

The market gets excited about the model.

Then the model needs a chip.

The chip needs HBM.

The rack needs liquid cooling.

The data center needs power.

The power needs transformers and switchgear.

The site needs construction labor.

The cluster needs optical networking.

The agent needs identity and permissioning.

The company needs governance and cybersecurity.

The physical world keeps showing up.

Gas turbines became AI infrastructure

This is one of the stranger parts of the thesis, but it makes sense once you follow the chain.

Hyperscalers and AI labs need firm power now.

Not in 2032.

Now.

Renewables matter. Batteries matter. Nuclear matters. Grid upgrades matter. But when you need reliable 24/7 power at scale, and you need it faster than the grid can provide it, gas turbines and onsite power become part of the conversation.

That pulls in GE Vernova, Siemens Energy, Mitsubishi Heavy, reciprocating engines, temporary generation, midstream gas infrastructure, pipelines, and power producers.

This does not mean gas is some perfect answer.

It creates emissions issues, regulatory issues, local opposition, fuel logistics, and political risk.

But the market does not get to choose the cleanest PowerPoint answer. It has to deal with time.

AI load is immediate.

Power infrastructure is slow.

That mismatch is investable.

It also makes existing firm power more valuable.

Which is why existing nuclear and merchant power matter more in the near term than some of the SMR (Small Modular (Nuclear) Reactors) stories people want to believe.

I am interested in nuclear. I think nuclear matters strategically. I think it matters a lot for the 2030s.

But if someone is telling me a speculative SMR company solves the 2026–2029 AI power problem, I get skeptical.

Existing nuclear, uprates, merchant power, data center PPAs, and companies that already control firm generation are much more relevant right now.

That points toward names like Constellation, Vistra, Talen, NRG, Cameco, Centrus, BWX Technologies, and others in that chain.

The speculative SMR names may have moments. Some may eventually become real.

But the burden of proof is higher.

AI needs power now.

Memory may be more important than people think

Everyone understands GPUs now.

Fewer people understand how much of AI is really a memory bandwidth problem.

Training needs memory.

Inference needs memory.

Long context needs memory.

Agents need memory, logs, retrieval, checkpointing, vector stores, storage, and fast access to data.

HBM is a real bottleneck.

Micron has already reported HBM revenue crossing $1 billion in a fiscal quarter, and demand for data center DRAM has been strong because AI systems are memory hungry.

That is why I do not want to think about the AI trade as just compute.

Compute without memory bandwidth is stranded.

A GPU cluster that cannot move data fast enough is not worth what you paid for it.

So memory, advanced packaging, storage, and networking all become part of the thesis.

Micron, SK Hynix, Samsung, Western Digital, Seagate, Pure Storage, NetApp, and the advanced packaging ecosystem all deserve attention.

Some of these are cyclical businesses. They can hurt you if you treat them like software companies.

But the structural demand is real.

Optical networking is the nervous system

The bigger the clusters get, the more important networking becomes.

If the chips cannot talk to each other fast enough, the system does not perform.

Training clusters need high-speed interconnects.

Inference clusters need networking.

Distributed compute needs networking.

Data centers need fiber, optics, switches, cables, retimers, and all the physical and logical infrastructure that keeps the system moving.

That brings in Arista, Broadcom, Marvell, Ciena, Coherent, Lumentum, Corning, Cisco, Credo, Astera Labs, and others.

This is not as simple as “buy every optical name.”

Some are already priced for a lot. Some are cyclical. Some are lower quality than they look in the moment.

But conceptually, the network fabric is not optional.

AI factories are not just piles of chips.

They are systems.

Google, Apple, and the importance of owning your stack

One of my earlier theses was that companies that control more of their own compute stack would have an advantage.

I still believe that.

Google is the obvious example with TPUs.

TPUs are not magic. They are not better than GPUs at everything. But if you understand what an ASIC is, you understand the appeal. For the right workload, with enough scale, a custom chip can be a major advantage.

Google has models, data, distribution, cloud, YouTube, Android, search, and its own AI accelerators.

That is not a small thing.

Apple is different.

Apple is not trying to win the chatbot leaderboard in the same way OpenAI or Anthropic is.

Apple’s AI strategy is about OS control.

A small on-device model for everyday tasks. Private Cloud Compute for larger requests. Apple silicon. Privacy. Distribution. App intents. iPhone, Mac, iPad, Watch, and the whole ecosystem.

Apple does not need the best chatbot in a browser tab.

It needs AI to become a private default action layer across the devices people already use.

If Apple gets that right, it protects the device cycle and services ecosystem. It also commoditizes a lot of standalone consumer chatbot use cases.

That is why the AI trade is also about distribution.

The best model does not always capture the most value.

Sometimes the default surface wins.

Microsoft has Office, GitHub, Azure, Windows, enterprise identity, and procurement.

Google has search, Workspace, Android, YouTube, TPUs, and cloud.

Meta has open models, consumer distribution, ads, and social surfaces.

Apple has the device and the OS.

Amazon has AWS, logistics, robotics, and massive internal automation.

The model matters.

But distribution, workflow, identity, and data decide where the money goes.

Open-source and local AI will change buying behavior

Open models do not need to beat the frontier models outright to matter.

They just need to be good enough for a lot of work.

DeepSeek, Qwen, Llama, and other open-weight models have made that obvious.

Enterprises are not going to send every sensitive workflow to one frontier API forever.

They will use a hybrid stack.

Frontier models for the hardest tasks.

Open or local models for cost control, privacy, latency, sovereignty, customization, and offline work.

This creates a different hardware demand curve.

AI PCs and workstations matter.

Local inference matters.

RAM matters.

VRAM matters.

Storage matters.

On-device chips matter.

Private AI servers matter.

This pulls in Apple, AMD, Intel, Qualcomm, Arm, Dell, HP, Lenovo, NVIDIA, Micron, Samsung, SK Hynix, Western Digital, Seagate, and others.

I do not think every consumer is going to run frontier-class AI locally.

That is not the point.

The point is that enough work moves local or private that the hardware requirements of normal computing go up.

More memory. More storage. More acceleration. More private infrastructure.

That becomes part of the cycle.

Cybersecurity becomes a chokepoint

AI agents are not just chatbots.

They are going to become users.

Privileged users.

They will have access to files, databases, browsers, terminals, code repos, tickets, cloud accounts, CRMs, payment systems, and internal tools.

That is incredibly useful.

It is also dangerous.

If an agent can do real work, it can also make real mistakes.

If an attacker can manipulate the agent, the attacker may inherit the agent’s permissions.

This creates a new category around agent identity, least-privilege access, tool permissions, telemetry, prompt-injection defense, audit logs, DLP, model risk controls, patch validation, and incident response.

Cyber offense also gets cheaper.

Agents can find vulnerabilities, write exploit code, test patches, automate phishing, scan systems, and run workflows that used to take teams of people.

So the defense side has to upgrade too.

That points toward companies like Palo Alto Networks, CrowdStrike, Fortinet, Zscaler, SentinelOne, Cloudflare, Okta, Datadog, Microsoft, Google, Palantir, and others that can sit in the identity/security/governance layer.

This is one of the areas I think investors may still underappreciate because it is less exciting than chips.

But if agents become coworkers, security becomes mandatory infrastructure.

The next step-functions: not just software

Coding is the first obvious labor market being restructured.

That does not mean software engineers vanish.

It means the workflow changes.

The human moves from typing every line to setting goals, reviewing plans, writing tests, checking outputs, controlling permissions, and deciding what is actually worth building.

But coding is not the end.

The next step-functions are likely in areas where digital agents can control real-world or regulated workflows with enough safety and supervision.

Pharma and biotech

AI can help with target discovery, protein and compound generation, assay design, trial design, literature mining, regulatory documentation, and lab workflow planning.

But biology is still the bottleneck.

Molecules still fail in humans.

That is why I am more interested in the combination of AI plus lab automation, instrumentation, validated datasets, and wetlab throughput than in blindly chasing every AI-drug pure play.

Watch the big pharma companies, lab tools, automation, and data platforms: Lilly, Novo Nordisk, Merck, Regeneron, Recursion, Schrödinger, Thermo Fisher, Danaher, Agilent, Illumina, 10x Genomics, Revvity.

Robotics

Language, vision, planning, and code make robotics more practical.

But the near-term win is not a perfect humanoid in every home.

The near-term win is constrained environments: warehouses, factories, inspection, hospitals, logistics, agriculture, cleaning, security, construction prefab, and defense drones.

Amazon’s warehouse robot fleet is the practical model. Not one magic robot. Fleets of machines doing constrained work better over time.

Watch Amazon, Tesla, Symbotic, Teradyne, Rockwell, Zebra, Cognex, ABB, Fanuc, Keyence, Hyundai, Honeywell, Deere, and Intuitive Surgical.

But be careful with humanoid hype. Unit economics, uptime, safety, service, and task generalization matter more than demo videos.

Defense autonomy

AI makes autonomy, targeting, simulation, logistics, cyber, electronic warfare, drones, satellites, and sensors more valuable.

This pulls in defense primes, autonomy companies, rare earths, secure chips, power electronics, and edge compute.

Watch Palantir, AeroVironment, Kratos, L3Harris, Northrop, RTX, Lockheed, General Dynamics, BWX Technologies, MP Materials, Rocket Lab, MDA, Iridium, and other space/defense infrastructure names.

Construction and permitting

AI will not instantly automate construction.

But it can compress design, estimating, procurement, scheduling, permitting packets, change orders, site documentation, safety monitoring, and prefab workflows.

At the same time, AI data centers create demand for the companies that can actually build complex electrical and mechanical projects.

That keeps me interested in Quanta, EMCOR, Comfort Systems, Sterling, MYR Group, MasTec, Fluor, Jacobs, AECOM, Autodesk, Bentley, and Trimble.

Education, finance, legal, insurance, and creative work

Document-heavy and rules-heavy industries are obvious agent targets.

AI can draft, reconcile, summarize, audit, check, route, and explain.

The winners are probably not generic wrappers.

The winners are systems of record and workflow owners that can embed AI into the place where the work already happens.

The losers are companies selling expensive seats for work an agent can compress.

That is one of the big questions for SaaS.

Does AI increase ARPU because the software becomes more valuable?

Or does AI reduce seats because fewer humans need to touch the interface?

The answer will differ by category.

But I do not want to own weak seat-based software just because it added an AI button.

SpaceX and xAI: real edge, not magic

The SpaceX/xAI angle is important, but not in the way people usually talk about it.

The point is not “space data centers next year.”

The point is that speed of physical deployment is becoming a model-capability advantage.

If two companies can get similar chips, the winner may be whoever can marshal land, power, cooling, construction, networking, and commissioning the fastest.

xAI’s Colossus-style buildout made that visible.

Musk companies are unusually good at physical execution speed. They can move capital, construction, suppliers, engineering, and logistics faster than normal organizations.

That matters.

But they cannot repeal physics.

Power is still slow.

Permitting still matters.

Cooling still matters.

Heat rejection still matters.

Gas turbines bring regulatory and environmental friction.

Orbital data centers are an interesting science-fiction-to-reality candidate, especially if Starship changes launch economics. But orbital compute has brutal constraints: thermal rejection, radiation hardening, maintenance, launch cadence, orbital debris, downlink bandwidth, and unit economics.

For 2026–2029, I think the public-market beneficiaries are still mostly terrestrial.

Power. Electrical gear. Cooling. EPC. Memory. Optics. Gas. Grid equipment.

Where I am more skeptical

This is a momentum and thesis sleeve, but that does not mean believing every story.

There are several buckets where I think narrative can outrun reality.

Quantum pure plays

Quantum is real science.

But many public quantum equities can trade like they are much closer to broad commercial utility than they actually are.

I want to see revenue quality, actual customer usage, error correction progress, qubit quality, and evidence that workloads beat classical alternatives economically.

Until then, I treat a lot of it as “too much future, too little current revenue.”

SMR pure plays

I like nuclear.

But AI power demand is here now.

SMRs may matter in the 2030s. They do not solve most near-term data center power needs.

Regulatory risk, fuel, first-of-a-kind construction, cost overruns, and timeline risk are real.

So I would rather prioritize existing nuclear and merchant power before paying wild prices for stories that still need years of proof.

Fuel-cell AI power stories

Fuel cells may win niches.

But data centers need reliability, serviceability, fuel logistics, cost certainty, and scale.

“AI power solved by fuel cells” is too broad for me.

Bitcoin miners pivoting to AI

Some miners may own valuable power and land.

That matters.

But a mining site is not automatically an AI data center.

You need fiber, redundancy, cooling, uptime SLAs, security, enterprise customers, capex discipline, and a totally different operating standard.

Some will pull it off.

Many will sell the story.

Generic AI apps

If the product is “ChatGPT but for X” and there is no proprietary workflow, data, distribution, or system-of-record lock-in, I am skeptical.

Platform owners and open models can crush a lot of that margin.

Weak seat-based SaaS

If AI agents reduce the number of human seats, then software priced only around human seat expansion becomes vulnerable.

Outcome-based pricing, usage-based pricing, workflow ownership, and data ownership matter more.

The baskets I care about

This is the working map I am using.

Not a buy list.

Not a recommendation.

A research map.

1. Physical bottlenecks

Grid equipment, transformers, switchgear, power distribution, cooling, EPC, gas turbines, existing power, copper, and construction labor.

Watching: ETN, Schneider, ABB, Siemens, GEV, HUBB, POWL, NVT, VRT, PWR, EME, FIX, STRL, MYRG, MTZ, MOD, TT, CARR, CEG, VST, TLN, NRG, FCX, SCCO, TECK, BHP, RIO.

2. Compute and platform owners

The obvious but still important layer: accelerators, custom silicon, cloud, distribution, operating systems, and developer ecosystems.

Watching: NVDA, AMD, AVGO, MRVL, ARM, TSM, ASML, MSFT, GOOG, META, AMZN, ORCL, AAPL.

3. Memory, storage, and networking

HBM, DRAM, NAND, storage, optical networking, switching, retimers, fiber, and cluster fabric.

Watching: MU, SK Hynix, Samsung, WDC, STX, PSTG, NTAP, ANET, AVGO, MRVL, CIEN, COHR, LITE, GLW, CRDO, ALAB, CSCO.

4. Security and governance

Agent identity, permissions, telemetry, audit logs, DLP, endpoint, cloud security, SIEM/SOAR, patch validation, model-risk controls.

Watching: PANW, CRWD, FTNT, ZS, S, NET, OKTA, DDOG, MSFT, GOOG, PLTR.

5. Next-step-function industries

Bio and lab automation, robotics, defense autonomy, construction, energy optimization, legal, finance, insurance, education, entertainment, and synthetic media.

This is where I expect the next wave of “that was impossible two years ago” moments.

6. Skepticism bucket

Quantum pure plays, speculative SMRs, fuel-cell AI power stories, weak bitcoin-miner pivots, generic AI wrappers, and seat-based SaaS that assumes human seats only go up.

These can still trade.

They can still squeeze.

They can still have moments.

But I want a higher burden of proof.

The indicators I am tracking

For this sleeve, I do not just want a list of stocks.

I want a dashboard of what would tell me the thesis is strengthening, weakening, or changing.

The main things I am watching:

  1. Data center power-demand forecasts from Goldman, IEA, EPRI, Berkeley Lab, utilities, and hyperscalers.
  2. Utility interconnect queues and transformer/switchgear lead times.
  3. Gas turbine backlog and delivery slots.
  4. HBM lead times, memory pricing, and HBM4/HBM4E transition.
  5. Rack density and liquid cooling adoption.
  6. Optical networking demand inside AI clusters.
  7. Frontier model progress, especially agent task horizon.
  8. Coding-agent benchmarks, but with cost, reliability, and wall-clock time included.
  9. Open model gap versus frontier models.
  10. AI PC and local AI usage, not just shipments.
  11. Cyber incidents involving AI agents or prompt-injection/tool misuse.
  12. Regulation around cyber-capable and bio-capable models.
  13. Apple developer adoption of App Intents and Private Cloud Compute.
  14. SaaS pricing shifts from seat-based to usage or outcome-based.
  15. Lab automation throughput and AI-designed molecule progress.
  16. Robotics deployments with real uptime and utilization.
  17. Quantum revenue quality versus valuation.
  18. Copper inventories, mine supply, and grid capex revisions.
  19. SpaceX Starship cadence and cost curve for any orbital-compute thesis.
  20. Defense procurement shifts toward autonomy, drones, and AI-enabled systems.

That is the work.

Not “AI good, buy AI stocks.”

That is not a strategy.

The work is tracking where the thesis is actually being confirmed or denied.

How this fits into my portfolio

This is not replacing systematic trading.

It is another sleeve.

The best way to reduce volatility and improve risk-adjusted returns is to stack strategies together.

One strategy can be short-term and systematic.

Another can be medium-term trend.

Another can be market neutral.

Another can be long-term thematic momentum.

This AI buildout sleeve is the long-term, thesis-driven, momentum-driven, future portfolio.

It is still a momentum strategy in the broad sense.

I want to be where capital, earnings, fundamentals, sentiment, and price are all moving in the same direction.

But unlike a pure quant signal, this sleeve needs research.

It needs context.

It needs an understanding of what is actually changing in the real world.

It needs a line in the sand for when the thesis is wrong.

It needs a way to separate turbulence from thesis break.

That distinction matters.

A 20% drawdown in a leader after a huge run is not automatically a thesis break.

A major change in capex, power economics, model economics, regulation, or competitive structure might be.

I want a process that helps me stay in the big winners while they are still big winners, without pretending that every dip is meaningless and every headline is bullish.

That is the balance.

Stay with the trade.

Respect the turbulence.

Update when the facts change.

Why I am turning this into a research product

I am going to be doing this work anyway.

I need it for my own portfolio.

I want the research map, the watchlists, the thesis updates, the portfolio construction, the buy/sell discipline, the turbulence monitor, and the “what changed this week?” process.

The goal is simple:

Stay with the long-term AI buildout trade without either chasing every shiny object or hiding in cash waiting for the bubble to pop.

It is a research and portfolio sleeve for the biggest long-term buildout I can see.

The point is to be early enough to matter, disciplined enough to survive, and flexible enough to update when the facts change.

By tomorrow some part of this report could be obsolete.

I expect that.

That is the point.

I want to be close enough to the work to know when something changes, whether it is real, and what it means for the portfolio.

Bottom line

The world does not just need smarter models.

It needs more electricity.

More memory bandwidth.

More cooling.

More network fabric.

More transformers.

More switchgear.

More construction.

More validated data.

More secure agents.

More physical deployment speed.

That is why I think the AI buildout is bigger than the chip trade.

The chip trade was chapter one.

The next chapters are physical, industrial, regulated, and messy.

That is where the mispricings usually show up.

And that is where I am spending my time.

Source list / starting references

Written by Chris Dover at Pollinate Trading. Signals and strategy verified live on Collective2.

More research →See the live track record →