The AI Trade Is Getting Bigger
For the past several years, the AI investment story has been dominated by one question:
Who makes the chips?
That question created one of the greatest semiconductor rallies in history and turned Nvidia into the clearest symbol of the AI revolution.
But the next phase may be much bigger than GPUs.
The real opportunity is compute.
AI models are becoming more capable, inference workloads are exploding, enterprises are deploying AI agents, and frontier labs are racing to build increasingly powerful models.
Every one of those developments requires one thing:
More compute.
And compute requires far more than a GPU.
It requires:
Silicon → Memory → Networking → Data Centers → Cooling → Electricity → Grid Infrastructure
This is why I believe the next phase of the AI cycle could create opportunities far beyond traditional semiconductor companies.
The AI infrastructure buildout is becoming an economy-wide capital cycle.
1. The Numbers Are Becoming Almost Difficult to Comprehend
The scale of spending is the first thing investors need to understand.
The four largest U.S. hyperscalers — Amazon, Microsoft, Alphabet and Meta — could collectively spend roughly $725 billion or more on capital expenditures in 2026, with estimates moving toward roughly $1 trillion in 2027.
That is extraordinary.
And perhaps even more important:
Wall Street has repeatedly underestimated the spending cycle.
AI capex expectations have been revised higher again and again.
That tells us something important.
The market may still be thinking about AI infrastructure using yesterday’s assumptions.
The question isn’t simply:
“Will hyperscalers continue spending?”
The better question is:
“How much compute can the global economy absorb before physical infrastructure becomes the limiting factor?”
That is a very different investment question.
2. Nvidia Can Win Even If Nvidia’s Market Share Falls
This is one of the most important concepts for investors to understand.
A common mistake is assuming:
Lower Nvidia market share = bearish for Nvidia.
Not necessarily.
Imagine Nvidia controls 90% of a $100 billion market.
That’s $90 billion.
Now imagine its share falls to 60%, but the market grows to $500 billion.
Nvidia’s revenue becomes $300 billion.
Its market share fell dramatically.
Its revenue still more than tripled.
That’s the dynamic I believe investors need to watch.
Custom ASICs from Google, Amazon, Microsoft and Meta are becoming increasingly important, particularly for predictable, high-volume inference workloads.
That doesn’t necessarily mean GPUs lose.
It means:
The total compute market gets bigger.
And that creates multiple winners.
3. The ASIC Revolution Is Real
Google’s TPUs, Amazon’s Trainium, Microsoft’s Maia and Meta’s MTIA are evidence that hyperscalers don’t want to rely entirely on external GPUs forever.
Why?
Three reasons:
Cost.
Efficiency.
Control.
When a company runs AI workloads at enormous scale, even small improvements in performance per dollar can translate into billions of dollars.
That’s why custom silicon matters.
The likely future isn’t:
GPU OR ASIC.
It is:
GPU + ASIC.
Training complex models and handling constantly changing workloads still favors flexible accelerators.
But high-volume inference can increasingly favor purpose-built silicon.
And that creates another major investment opportunity.
4. Broadcom May Be One of the Biggest “Picks and Shovels” of the ASIC Era
If Nvidia is the dominant GPU infrastructure company, Broadcom is emerging as one of the most important beneficiaries of the custom-silicon revolution.
Google.
Meta.
Microsoft.
OpenAI and other AI developers.
Multiple major technology companies are pursuing customized accelerators.
And Broadcom sits directly in the middle of much of this ecosystem.
That makes $AVGO particularly interesting because its AI opportunity isn’t dependent on winning the GPU market.
It can benefit from the proliferation of AI accelerators themselves.
Marvell is another company worth watching as custom silicon expands.
The larger thesis:
The AI semiconductor market is becoming more diversified, not smaller.
5. Memory Could Become the Next Major Bottleneck
This is one area I believe investors still don’t pay enough attention to.
AI doesn’t just need compute.
It needs enormous amounts of high-bandwidth memory.
Modern AI accelerators increasingly depend on HBM to feed processors with data fast enough to keep them working efficiently.
And the supply chain is concentrated.
That creates an interesting dynamic.
You can have:
- GPUs available
- Data centers under construction
- Customers ready to deploy
…but still have constraints because you don’t have enough HBM.
That puts companies such as:
$MU — Micron
SK Hynix
Samsung
at an increasingly strategic position within the AI supply chain.
The market has spent years talking about GPU shortages.
The next bottleneck may increasingly be:
Memory.
And eventually:
Power.
6. The Real AI Bottleneck May Be Electricity
This is where the AI investment thesis gets much bigger.
You can manufacture more GPUs.
You can manufacture more servers.
You can build more data centers.
But you can’t simply manufacture electricity overnight.
That’s the problem.
The IEA expects data-center electricity demand to rise dramatically through 2030.
AI is turning data centers into some of the largest new electricity consumers being added to the grid.
And AI workloads can be extremely power intensive.
That creates a massive second-order investment opportunity.
The AI trade is moving from:
Silicon → Electricity.
7. Welcome to the “Silicon-to-Substation” Trade
This is the framework I would use to understand the next stage of AI infrastructure.
Layer 1 — Compute
$NVDA
$AMD
The accelerators.
Layer 2 — Memory
$MU
$SNDK
SK Hynix
Samsung
The high-bandwidth memory required to feed the accelerators.
Layer 3 — Networking
$AVGO
$ANET
$CRDO
$MRVL
Moving enormous quantities of data between GPUs, servers and data centers.
Layer 4 — Data Centers
$NBIS
$CRWV
$IREN
$CIFR
$APLD
The physical locations where compute actually lives.
Layer 5 — Power
$VST
$CEG
$GEV
Generating and supplying the electricity.
Layer 6 — Grid Infrastructure
$ETN
$PWR
Transformers, switchgear, transmission and electrical infrastructure.
This is the part of the AI trade that I believe could surprise investors.
The further you move down the stack, the less crowded the narrative becomes.
8. The Energization Gap
Here’s a concept I would watch extremely closely through 2027–2030.
Announced capacity is not the same as operational capacity.
A company can announce:
“We are building a 1 GW AI data center.”
That doesn’t mean 1 GW of AI compute is operating.
You need:
- Land
- Permits
- Construction
- Servers
- GPUs/ASICs
- HBM
- Networking
- Cooling
- Transformers
- Grid connection
- Generation capacity
And increasingly:
Power availability.
This creates what I call the Energization Gap.
There could be a significant amount of AI compute that is physically built but cannot immediately operate at full capacity because the electricity infrastructure isn’t ready.
That distinction could become extremely important for investors.
9. Nuclear and Natural Gas Become AI Infrastructure
This is where the AI story intersects with traditional energy.
If electricity demand accelerates faster than renewable generation and transmission can be built, the industry will need reliable baseload and dispatchable power.
That puts nuclear and natural gas directly into the AI conversation.
Companies such as:
$CEG
$VST
$GEV
are therefore not simply traditional energy or industrial businesses.
They can increasingly be viewed as AI infrastructure companies.
That doesn’t mean every utility automatically becomes an AI winner.
Valuation still matters.
Execution still matters.
Regulation matters.
But the structural connection is becoming impossible to ignore.
10. Data Centers Become the New “Factories”
There is another way to think about this.
The industrial revolution built factories.
The internet built data centers.
The AI revolution is building AI factories.
But these factories consume enormous amounts of electricity and require specialized infrastructure.
That makes companies involved in data-center development potentially very valuable.
The NeoCloud ecosystem is particularly interesting because these companies can provide GPU infrastructure to customers that don’t want to build everything themselves.
That is why names such as:
$NBIS
$CRWV
$IREN
$CIFR
$APLD
have become part of the broader compute conversation.
But this is also where investors need to be careful.
These companies can have enormous growth potential.
They can also have enormous capital requirements.
Growth without financing discipline can destroy shareholder value.
11. The Biggest Risk: AI’s Circular Financing
This is the part of the thesis that should not be ignored.
Nvidia invests in AI companies.
AI companies purchase compute.
Cloud companies build data centers.
Data-center companies buy Nvidia hardware.
Cloud providers finance infrastructure.
And some of those same companies become customers, investors or suppliers of each other.
This creates a fascinating feedback loop.
When demand keeps accelerating:
The loop becomes a growth engine.
But if AI demand slows dramatically:
The same loop can amplify the downside.
That is why I don’t think the correct question is:
“Is AI a bubble?”
The better question is:
“Which companies have real cash-flow-producing demand, and which companies depend primarily on the next round of financing?”
That distinction could become critical during the next bear market.
12. China Creates Another Long-Term Variable
The AI compute race isn’t happening only in the United States.
China is aggressively developing domestic alternatives across:
- AI accelerators
- HBM
- semiconductor manufacturing
- data centers
- AI models
Huawei has become increasingly important in China’s AI-chip ecosystem, while CXMT is working to improve domestic HBM capabilities.
The U.S. still maintains significant advantages in advanced AI compute and semiconductor technology.
But the direction is important.
China is moving toward:
Compute self-sufficiency.
That means export controls may remain a structural factor in the semiconductor industry for years.
For investors, China exposure may increasingly deserve a permanent geopolitical discount rather than simply a temporary regulatory discount.
13. The Most Important Variable: Inference
Here’s the part of the AI story that could change the economics of the entire industry.
Training gets the headlines.
Inference could become the enormous recurring workload.
Once an AI model is trained, every time someone asks it a question, generates an image, writes code, uses an AI agent or interacts with an autonomous system, compute is consumed.
And unlike training, inference happens continuously.
Think about the difference between:
Building the factory
and
Running the factory every day.
As AI becomes embedded into:
- Search
- Software
- Robotics
- Autonomous vehicles
- Healthcare
- Finance
- Customer service
- Enterprise applications
- Personal assistants
the amount of inference could explode.
That could fundamentally change which hardware wins.
14. The Next AI Winners May Not Be the Companies Everyone Is Watching Today
This is my biggest takeaway.
The first phase of AI created obvious winners:
$NVDA
$AMD
$AVGO
$MU
The next phase could create winners in less obvious areas:
Power
Grid infrastructure
Transformers
Cooling
Optical networking
Nuclear
Natural gas
Data centers
Memory
Custom silicon
AI inference
This is why I don’t view AI as a single-sector trade.
I view it as a multi-year capital-expenditure cycle.
15. My AI Compute Watchlist
If I were building a long-term AI infrastructure watchlist, I’d organize it by layer rather than simply picking the “best AI stocks.”
COMPUTE
$NVDA
$AMD
CUSTOM SILICON
$AVGO
$MRVL
MEMORY
$MU
$SNDK
NETWORKING / OPTICAL
$ANET
$CRDO
$MRVL
DATA CENTERS / NEOCLOUD
$NBIS
$CRWV
$IREN
$CIFR
$APLD
POWER
$VST
$CEG
$GEV
GRID
$ETN
$PWR
The important thing isn’t owning every name.
It’s understanding where the bottleneck is moving.
16. What Could Break the Thesis?
No investment thesis is complete without the bear case.
There are several.
1. AI capex slowdown
If hyperscalers suddenly reduce spending, the entire ecosystem could re-rate.
2. Model efficiency
If models become dramatically more efficient, required compute per task could fall.
3. Custom silicon
ASICs could reduce demand for expensive general-purpose GPUs.
4. Power constraints
Ironically, the biggest constraint on AI growth could also limit the revenue opportunity.
5. Financing stress
Highly leveraged NeoCloud and infrastructure companies could struggle if capital markets tighten.
6. Valuation
Even an excellent company can be a bad investment at an excessive valuation.
7. Geopolitics
China/U.S. semiconductor restrictions could disrupt supply chains and markets.
17. What I’m Watching
Over the next several years, I would focus less on headlines and more on leading indicators.
Watch #1 — Hyperscaler Capex
Are $AMZN, $MSFT, $GOOG and $META increasing or decreasing spending guidance?
Watch #2 — HBM Pricing
Memory shortages can tell us how intense AI infrastructure demand really is.
Watch #3 — Power Interconnections
How quickly can new AI data centers actually receive electricity?
Watch #4 — ASIC Adoption
How quickly are Google, Amazon, Meta and Microsoft replacing external GPUs with internal accelerators?
Watch #5 — Inference Growth
How rapidly is AI usage translating into recurring compute demand?
Watch #6 — Financing
Are NeoCloud companies generating enough cash flow to fund expansion?
Watch #7 — China
How quickly are Huawei and domestic semiconductor companies closing the technology gap?
The Bottom Line
The AI revolution isn’t ending with GPUs.
It is expanding beyond them.
The next stage of this cycle is about building the physical infrastructure required to turn artificial intelligence into an always-on global utility.
That means trillions of dollars potentially flowing through an ecosystem that stretches from:
GPU → HBM → Networking → Data Center → Cooling → Electricity → Grid
And that is why I believe investors should stop thinking about AI as simply a technology sector.
AI is becoming an infrastructure cycle.
The biggest question for the next decade may not be:
“Who has the smartest AI?”
It may be:
“Who has enough compute to run it?”
And behind that question sits an even bigger one:
“Who has enough power to run the compute?”
That’s the investment chain I will be watching:
Silicon → Memory → Network → Compute → Data Center → Power → Grid.
The AI race is becoming a race for compute capacity.
And we may still be very early in the infrastructure buildout.