AI Infrastructure Investment in 2026: HBM, Data Centers, Power and Liquid Cooling
Beyond NVIDIA: The AI Semiconductor Supply Chain, Power Bottleneck and the Next Investment Opportunity
A three-part analysis of NVIDIA, HBM, advanced packaging, networking, power infrastructure, cooling and the evolving AI investment cycle.
NVIDIA Is Growing Faster, But the Bigger Story Is the Supply Chain
For the past several years, NVIDIA has been at the center of the artificial intelligence investment boom.
The company started the AI revolution from the perspective of computing.
Its GPUs became the engines powering large language models, generative AI applications, AI agents, inference systems, scientific computing, and increasingly sophisticated data-center workloads.
But NVIDIA's latest earnings report suggests that the AI story has entered another stage.
The important question is no longer simply whether NVIDIA can sell more GPUs.
The bigger question is whether the rest of the infrastructure can keep up with the speed at which AI computing is being deployed.
That distinction could become increasingly important for investors.
Because an AI accelerator does not operate alone.
It needs high-bandwidth memory.
It needs advanced packaging.
It needs networking.
It needs servers.
It needs power delivery.
It needs cooling.
It needs a data center.
And ultimately, it needs access to electricity.
This means the next phase of the AI boom could increasingly become a supply-chain story rather than a single-company story.
NVIDIA's latest numbers change the scale of the discussion
NVIDIA reported fiscal second-quarter revenue of $96.22 billion for the quarter ended July 26, 2026.
Data-center revenue reached approximately $89 billion, representing 117% year-over-year growth and accounting for the overwhelming majority of the company's quarterly revenue.
NVIDIA also guided for approximately $108 billion of revenue for its fiscal third quarter.
The significance of these numbers goes beyond another quarterly earnings beat.
NVIDIA is now operating at a scale where quarterly revenue is approaching the level that only a handful of the world's largest technology companies can consistently achieve.
More importantly, the company expects another major step up in revenue.
Reuters reported that NVIDIA expects approximately 70% revenue growth for the fiscal year ending January 2028, a forecast that was significantly above Wall Street's expectations at the time of the announcement.
This is a powerful signal.
AI infrastructure demand has not yet behaved like a conventional technology cycle in which demand rapidly normalizes after the initial wave of adoption.
Instead, the industry is still expanding.
But expansion creates another problem.
The faster the AI industry grows, the more difficult it becomes to expand every part of the supply chain at the same speed.
That is where the next investment story begins.
A GPU Is Only One Piece of the AI System
Imagine a hyperscale data center preparing to deploy tens of thousands of advanced AI accelerators.
The first instinct may be to think about GPU availability.
But the actual deployment process is much more complicated.
The GPU requires high-bandwidth memory.
The GPU and memory must be integrated through sophisticated packaging technologies.
Multiple processors need extremely fast networking.
The servers need power distribution.
The racks need thermal management.
The facility requires cooling infrastructure.
The building requires transformers, switchgear, generators, cables, and electrical systems.
And before any of this happens, the developer needs land, permits, grid access, construction capacity, and financing.
The AI infrastructure chain looks something like this:
AI models create demand for compute.
Compute creates demand for accelerators.
Accelerators create demand for HBM and advanced packaging.
Large AI clusters create demand for networking.
Higher compute density creates demand for power.
Higher power density creates demand for cooling.
More data centers create demand for grid infrastructure.
This is why NVIDIA's earnings can be interpreted as a signal for a much larger industrial ecosystem.
The company sits near the center of the chain, but the chain extends far beyond NVIDIA.
HBM: The Memory Layer Behind the AI Boom
One of the most important links in this ecosystem is high-bandwidth memory, commonly known as HBM.
AI accelerators need to process enormous amounts of data.
As models become more complex and inference workloads become larger, memory bandwidth becomes increasingly important.
This makes HBM fundamentally different from simply adding more conventional memory capacity.
The objective is not just to store more data.
The objective is to move data rapidly enough to keep powerful processors supplied with information.
This is one reason the HBM market has become strategically important.
The major suppliers include SK hynix, Samsung Electronics, and Micron.
According to Counterpoint Research data cited by Reuters, SK hynix held approximately 58% of the global HBM market in the first quarter of 2026, while Samsung Electronics and Micron each held about 21%.
That concentration matters.
When a semiconductor component becomes both technically difficult and strategically important, the suppliers capable of producing it at scale can become critical parts of the broader AI ecosystem.
HBM4 and HBM4E Raise the Technical Bar
The industry is now moving toward increasingly advanced generations of HBM.
HBM4 is expected to become an important memory technology for next-generation AI accelerators.
HBM4E represents another step toward higher bandwidth and improved efficiency.
But the important investment point is not simply that the next generation is faster.
The manufacturing process also becomes more difficult.
Higher layer counts.
More complex stacking.
Higher bandwidth requirements.
More demanding thermal characteristics.
Tighter integration with advanced packaging.
All of these factors increase manufacturing complexity.
SK hynix said in June 2026 that it had shipped samples of a 12-layer HBM4E product to major customers and said the product could achieve speeds of up to 16 gigabits per second per pin, with more than 20% better power efficiency than previous products.
The company also announced plans to significantly expand wafer capacity over the coming years.
This illustrates the central problem of the AI semiconductor industry.
Demand is increasing quickly. But supply cannot be created instantly.
A new semiconductor fab takes years to plan, finance, construct, equip, qualify, and ramp.
That time difference is becoming one of the defining characteristics of the AI infrastructure cycle.
Advanced Packaging May Be Just as Important as the Chip
Another part of the AI supply chain that deserves much more attention is advanced packaging.
For conventional chips, packaging was often treated as a final manufacturing step.
For modern AI accelerators, that view is becoming outdated.
Advanced packaging has become part of the architecture itself.
Technologies such as TSMC's CoWoS and SoIC allow high-performance compute dies and memory components to be integrated into increasingly sophisticated systems.
TSMC has been expanding CoWoS and SoIC capacity as AI demand increases. Industry reports have repeatedly identified advanced packaging capacity as an important constraint for AI accelerator production.
This creates another important investment connection.
A company can have an excellent AI processor design.
It can have sufficient wafer capacity.
It can have customers ready to purchase the product.
But if advanced packaging capacity is insufficient, the final system cannot be produced at the required scale.
This is a classic bottleneck.
And bottlenecks can move.
At one point, the industry worried primarily about GPU wafers.
Then attention moved toward HBM.
Now advanced packaging has become another critical layer.
TSMC and ASE Illustrate Two Different Parts of the Packaging Story
TSMC is particularly important because it combines advanced semiconductor manufacturing with advanced packaging capabilities.
Its CoWoS technology is closely associated with the integration of AI accelerators and HBM.
But TSMC is not the only company relevant to the packaging ecosystem.
OSAT companies such as ASE Technology also play an important role in semiconductor assembly, packaging, and testing.
This distinction matters because the AI semiconductor supply chain is becoming more specialized.
A future AI system may involve:
A chip designer.
A leading-edge foundry.
An HBM supplier.
An advanced packaging provider.
A networking supplier.
A server manufacturer.
A power-system provider.
A cooling company.
A data-center operator.
The final AI product is therefore the result of a network rather than a single manufacturing process.
That is why investors looking beyond NVIDIA should study the entire chain.
NVIDIA's $96.22 billion quarterly revenue and $89 billion data-center business demonstrate that AI infrastructure demand remains exceptionally strong. But the next constraint may not be the GPU itself. HBM, advanced packaging, networking, electricity, and cooling are becoming equally important parts of the AI build-out.
From HBM to Networking, Power and Cooling
If Part 1 explains why the AI semiconductor story is expanding beyond NVIDIA, Part 2 explains where that expansion is happening.
The most interesting development is that AI demand is moving into industries that historically had little connection with semiconductor investing.
Networking.
Electrical equipment.
Power generation.
Cooling.
Construction.
Transformers.
Grid infrastructure.
These industries are becoming increasingly important because AI computing is becoming physically demanding.
Networking Is Becoming a Core AI Component
A massive AI cluster is not simply a collection of GPUs.
Thousands or tens of thousands of accelerators must communicate with one another.
That requires extremely fast networking.
This is where companies such as Broadcom and Marvell become relevant.
Broadcom is involved in networking silicon, connectivity, and custom AI accelerators.
The company's own AI infrastructure materials describe its business as covering both AI networking and custom accelerators.
Broadcom is therefore interesting because it represents a different side of the AI semiconductor market.
NVIDIA sells a vertically integrated AI computing platform.
Broadcom can benefit from the growing demand for customized infrastructure and networking.
This distinction becomes particularly important as hyperscale customers attempt to optimize their infrastructure for specific workloads.
NVIDIA Versus Custom Silicon
This creates one of the most important competitive questions in AI.
Will hyperscalers continue relying primarily on NVIDIA?
Or will they increasingly build their own accelerators?
The answer is likely to be more complicated than either extreme.
NVIDIA has a major advantage in software.
CUDA and its broader software ecosystem have created a significant developer and infrastructure advantage.
Once an organization builds software, models, libraries, workflows, and engineering expertise around a particular platform, switching becomes expensive.
This creates a form of ecosystem lock-in.
But hyperscalers have another incentive.
Cost.
If a company operates millions of workloads, even a relatively small reduction in cost per inference can become financially significant.
There is also a control issue.
Designing a custom accelerator allows a company to optimize hardware around its own workloads.
Google has developed TPU technology.
Amazon has developed its own AI accelerator programs.
Microsoft has invested in custom silicon.
Other major technology companies are also exploring customized AI hardware.
Broadcom has become an important partner in this custom-silicon ecosystem.
In October 2025, OpenAI and Broadcom announced a collaboration to co-develop systems involving AI accelerators and Ethernet networking, with deployment targeted to begin in the second half of 2026 and continue through 2029.
OpenAI also announced a separate inference-chip collaboration with Broadcom in 2026.
These developments do not mean NVIDIA is losing its leadership.
Instead, they demonstrate that the AI semiconductor market is becoming more diversified.
NVIDIA may remain dominant in general-purpose accelerated computing while custom silicon takes a larger role in highly optimized workloads.
That would actually expand the total semiconductor opportunity.
The Physical Bottleneck: Electricity
The AI semiconductor story eventually reaches one unavoidable question.
Where does the electricity come from?
This is where the AI boom begins to look less like a software revolution and more like an industrial infrastructure cycle.
Morgan Stanley Research estimated that U.S. data-center electricity demand could reach 74 gigawatts by 2028 and projected a potential shortfall of approximately 49 gigawatts in available power access.
The exact figures will depend on construction schedules, grid expansion, generation capacity, and the actual power requirements of future AI systems.
But the broader message is clear.
Power availability can become a constraint on AI expansion.
And power infrastructure cannot be built at semiconductor speed.
A new AI accelerator generation can arrive relatively quickly.
A transmission line cannot.
A new transformer cannot necessarily be delivered immediately.
A new substation cannot be built overnight.
A nuclear power project can take many years.
Permitting can add additional delays.
AI demand moves at software speed.
Physical infrastructure moves at industrial speed.
That difference may determine the pace of the next stage of AI deployment.
Transformers Are an Invisible AI Bottleneck
The transformer is not an exciting AI technology.
It does not run a language model.
It does not generate images.
It does not appear in AI benchmarks.
But without transformers, large data centers cannot receive and distribute electricity efficiently.
This creates an unusual investment situation.
The most important bottleneck in a technology boom can sometimes be a component that has almost nothing to do with the technology itself.
The same principle applies to switchgear, cables, substations, backup generators, and power-management equipment.
This is why companies such as Eaton and Schneider Electric are increasingly relevant when analyzing the AI infrastructure cycle.
They operate in electrical distribution and power-management markets that can benefit from the expansion of data centers and electrification.
The AI boom therefore creates a bridge between semiconductor investing and industrial investing.
Generac: A Simple Example of AI Demand Reaching the Factory Floor
Generac provides an interesting case study.
The company is traditionally associated with generators.
At first glance, that may appear far removed from artificial intelligence.
But data centers require reliable backup power.
As AI facilities become larger and more power-intensive, the importance of resilient electrical systems increases.
Reuters reported in August 2026 that Generac was investing approximately $250 million to expand production for data-center demand and had a reported backlog of around $1.6 billion.
This is exactly the type of second-order effect investors should watch.
NVIDIA sells an AI accelerator.
That accelerator increases demand for AI servers.
AI servers increase demand for data centers.
Data centers increase demand for electricity.
Electricity requirements increase demand for backup power.
A company selling generators can therefore benefit indirectly from AI.
The connection is not obvious until the entire chain is examined.
Cooling: The Other Side of the Power Problem
Every watt consumed by an AI accelerator eventually becomes heat.
The more computing power packed into a rack, the greater the thermal challenge.
This is why liquid cooling has become increasingly important.
Companies such as Vertiv and nVent are positioned within the broader electrical and thermal-management ecosystem supporting data centers.
Vertiv has extensive exposure to data-center power and cooling infrastructure.
nVent operates in electrical connection and protection technologies, including solutions relevant to data-center infrastructure.
The investment thesis is not that every data center will suddenly switch completely to liquid cooling.
The more reasonable thesis is that increasing rack density creates pressure for more sophisticated thermal-management solutions.
Air cooling can remain useful.
Liquid cooling can become increasingly important in high-density environments.
Hybrid systems may also remain common.
The final architecture will depend on rack density, workload, facility design, water availability, energy efficiency, and cost.
That nuance matters.
AI infrastructure is not a single technology.
It is an engineering optimization problem.
The Supply-Demand Mismatch Is the Real Story
The most important idea across this entire article can be summarized in one sentence.
AI demand is accelerating faster than physical infrastructure can be built.
A hyperscaler can decide today to increase AI capital spending.
But the resulting infrastructure may take years to become operational.
Land must be acquired.
Permits must be approved.
Power must be secured.
Transformers must be ordered.
Substations must be built.
Buildings must be constructed.
Cooling systems must be installed.
Servers must be delivered.
Chips must be packaged.
HBM must be available.
The final system must then be tested and commissioned.
This creates a long chain of dependencies.
A delay at any one point can delay the entire project.
That means investors should increasingly pay attention to lead times and infrastructure schedules rather than only quarterly chip shipments.
The AI boom is spreading into networking, electrical equipment, generators, cooling, construction, and grid infrastructure. The central mismatch is simple: AI demand can increase quickly, but electricity, transformers, transmission lines, substations, permits, and data centers take much longer to build.
The Bull Case, the Bear Case and the Next AI Bottleneck
The strongest investment story is rarely the one with only positive data.
The AI semiconductor ecosystem has enormous growth potential.
But it also faces increasingly complicated risks.
The current environment should therefore be viewed as a combination of structural opportunity and rising execution risk.
The Bull Case: AI Infrastructure Has Not Reached Its Peak
The strongest bullish argument is visible in NVIDIA's latest numbers.
Quarterly revenue reached $96.22 billion.
Data-center revenue reached $89 billion.
The company expects approximately $108 billion of revenue in the following quarter.
And NVIDIA is projecting approximately 70% revenue growth for the fiscal year ending January 2028.
Those numbers are difficult to reconcile with the idea that AI investment is already collapsing.
The demand is real.
Hyperscalers continue to build.
AI laboratories continue to train and deploy models.
Enterprises are experimenting with AI agents and inference.
Sovereign customers are investing in domestic AI infrastructure.
Industrial companies are beginning to use AI for automation and optimization.
The customer base is therefore expanding.
That is one of the strongest arguments supporting the long-term AI infrastructure thesis.
The Bull Case: AI Is Moving From Training to Inference
There is another structural development investors should monitor.
The AI industry is gradually shifting from a world dominated by model training toward a world where inference becomes increasingly important.
Training requires enormous computing resources to create models.
Inference happens every time users interact with those models.
If AI agents become embedded into search, software, customer service, robotics, healthcare, financial services, logistics, and industrial systems, inference workloads could become enormous.
This creates demand for efficient computing.
And inference can create opportunities for both general-purpose GPUs and specialized accelerators.
This is one reason the custom-silicon market deserves attention alongside NVIDIA.
The future may not be NVIDIA versus custom chips.
It may be NVIDIA plus custom chips plus CPUs plus specialized accelerators, all serving different workloads.
That would make the semiconductor market larger rather than smaller.
The Bear Case: AI Capital Spending Could Become More Selective
The biggest risk is not necessarily that AI disappears.
The bigger risk is that the growth rate slows.
Technology companies are committing enormous amounts of capital to AI infrastructure.
Morgan Stanley estimated that four major U.S. technology companies could spend roughly $630 billion on data centers and AI chips during 2026.
Even if AI remains highly valuable, investors eventually need to ask whether these investments generate sufficient economic returns.
If AI revenue grows slower than infrastructure spending, management teams could become more selective.
That could lead to a change in the investment cycle.
Instead of asking:
How many GPUs can we buy?
Companies may start asking:
How much revenue does each GPU generate?
How efficiently can each rack operate?
How much electricity does each workload consume?
Can custom silicon reduce cost?
Can inference be optimized?
This would not necessarily end the AI boom.
It would make the boom more economically disciplined.
The Bear Case: NVIDIA's Margins Could Face Pressure
One of the most important developments in the latest earnings report is not just revenue growth.
It is the pressure created by rising component costs.
NVIDIA has indicated that memory shortages are affecting supply and putting pressure on margins. Reuters reported that memory constraints were limiting supply and contributing to cost pressure.
This creates an interesting contradiction.
The stronger AI demand becomes, the more components NVIDIA needs.
But if HBM and other critical components become more expensive, some of the economics can move toward suppliers.
This is a normal feature of constrained supply chains.
The bottleneck captures value.
That means strong NVIDIA demand can simultaneously be bullish for NVIDIA and bullish for certain suppliers.
But eventually, new capacity can change the balance.
That is why investors should distinguish between:
Demand growth.
Capacity growth.
Pricing power.
Gross margins.
And return on invested capital.
The Bear Case: Advanced Packaging Could Limit Growth
The same issue applies to packaging.
If CoWoS and related advanced packaging technologies cannot expand quickly enough, accelerator shipments can be constrained even if GPU wafer production increases.
TSMC has been expanding advanced packaging capacity in response to AI demand, but capacity expansion itself requires significant capital investment and time.
This is another example of the speed mismatch.
AI models can be updated quickly.
Chip architectures can evolve quickly.
But factories and packaging facilities take years to build.
The result is a supply chain that can remain tight even when companies are spending aggressively to expand capacity.
The Competitive Threat: NVIDIA's Software Advantage Versus Custom Silicon
The NVIDIA competition deserves a deeper examination.
NVIDIA's strongest advantage is not simply GPU performance.
It is the ecosystem around the GPU.
CUDA.
Libraries.
Developer tools.
AI frameworks.
Optimized software.
Networking.
Systems.
Customers have invested years building around this environment.
That creates switching costs.
But custom silicon has a different advantage.
It can be designed specifically for a company's own workloads.
A hyperscaler can potentially optimize performance, power consumption, and cost for its own applications.
This creates a strategic trade-off.
NVIDIA: flexibility, mature software, and a broad ecosystem.
Custom silicon: optimization, control, and potentially lower cost for specific workloads.
Neither side has to completely eliminate the other.
In fact, the likely outcome could be coexistence.
NVIDIA may continue dominating broad accelerated computing while custom chips take an increasing share of predictable, high-volume workloads.
This is particularly important for companies such as Broadcom and Marvell.
Their opportunity is not necessarily dependent on defeating NVIDIA.
It can come from helping hyperscalers build alternative or complementary computing architectures.
China Remains a Major Risk
Geopolitics is another major variable.
NVIDIA's latest guidance does not assume meaningful data-center computing revenue from China.
The company continues to face uncertainty around U.S. export restrictions and the ability to sell advanced AI processors into the Chinese market.
This creates two different risks.
The first is near-term revenue loss.
The second is the acceleration of alternative AI semiconductor ecosystems.
If Chinese customers cannot access the most advanced U.S. accelerators, they have a stronger incentive to develop domestic alternatives.
That could eventually create more competition across the global AI semiconductor market.
Therefore, export controls should not be viewed only as a quarterly NVIDIA issue.
They could influence the architecture of the global AI chip market for years.
What Investors Should Watch From Here
The next phase of the AI semiconductor cycle will require a broader set of indicators.
The first is NVIDIA data-center growth.
If data-center revenue continues expanding rapidly, the broader infrastructure thesis remains strong.
The second is HBM.
Watch capacity, qualification, pricing, technology transitions, and supplier market share.
The third is advanced packaging.
CoWoS and related technologies are becoming increasingly important to the ability to turn AI chip designs into finished products.
The fourth is networking.
As AI clusters become larger, the communication layer becomes more important.
Broadcom and Marvell are therefore companies worth watching within this part of the ecosystem.
The fifth is electricity.
Data-center demand can only grow as quickly as power becomes available.
Morgan Stanley's estimate of a potential U.S. power-access shortfall illustrates the scale of this challenge.
The sixth is cooling.
The growth of high-density AI racks could increase demand for advanced thermal-management systems.
Companies such as Vertiv and nVent provide examples of businesses positioned within this infrastructure layer.
The seventh is electrical infrastructure.
Eaton and Schneider Electric are relevant because the AI data-center build-out requires power distribution and management equipment.
The eighth is backup power.
Generac provides an example of how AI demand can reach traditional industrial businesses.
Finally, investors should monitor hyperscaler capital expenditure.
Microsoft.
Amazon.
Alphabet.
Meta.
And other major technology companies remain among the most important sources of AI infrastructure demand.
If their capital spending continues rising, the supply chain receives a powerful demand signal.
If spending begins to slow, the entire ecosystem may eventually feel the effect.
The Most Important Question Is No Longer "Who Makes the Best AI Chip?"
That question was appropriate during the first phase of the AI boom.
Today, the question is becoming more complicated.
Who makes the memory?
Who provides the packaging?
Who connects the systems?
Who supplies the electricity?
Who builds the substation?
Who manufactures the transformer?
Who cools the rack?
Who provides backup power?
Who builds the data center?
Who finances the infrastructure?
And who ultimately earns enough revenue from AI to justify all of this spending?
Those questions define the next stage of the AI investment cycle.
The AI Infrastructure Chain Is Becoming a New Industrial Cycle
This may be the most important conclusion.
The AI boom is no longer purely a semiconductor story.
It is becoming a combination of:
Semiconductors.
Memory.
Advanced packaging.
Networking.
Electrical equipment.
Power generation.
Grid infrastructure.
Cooling.
Construction.
Industrial automation.
Data centers.
The beneficiaries of this cycle may therefore come from industries that investors traditionally considered unrelated to technology.
Generac is an example.
Eaton is another.
Schneider Electric is another.
Vertiv is another.
TSMC represents the manufacturing and advanced-packaging layer.
ASE represents the outsourced semiconductor packaging and testing ecosystem.
Broadcom and Marvell represent networking and custom silicon.
SK hynix, Samsung Electronics, and Micron represent the HBM layer.
NVIDIA remains the central accelerator platform.
But none of these companies operates in isolation.
They are pieces of the same infrastructure build-out.
The Real Bottleneck Could Keep Moving
There is one final concept investors should remember.
A supply chain does not have a permanent bottleneck.
The bottleneck moves.
If GPU supply increases, HBM may become constrained.
If HBM capacity increases, packaging may become constrained.
If packaging expands, networking may become the limiting factor.
If chips and networking become available, power may become the constraint.
If power becomes available, cooling may become the constraint.
If power and cooling are solved, permitting or construction may become the constraint.
This is why the AI infrastructure market is so difficult to analyze using only one company or one financial metric.
The bottleneck can move from silicon to memory.
From memory to packaging.
From packaging to power.
From power to cooling.
From cooling to construction.
And from construction back to economics.
Final Conclusion: Look Beyond NVIDIA
NVIDIA's latest earnings provide a powerful confirmation that AI infrastructure demand remains strong.
Revenue of $96.22 billion.
Data-center revenue of $89 billion.
Third-quarter guidance of $108 billion.
A projected 70% revenue growth rate for the following fiscal year.
These numbers demonstrate the scale of the current AI cycle.
But the bigger investment story may be what happens around NVIDIA.
HBM manufacturers must increase capacity.
TSMC and other packaging providers must expand advanced packaging.
Networking companies must support increasingly large AI clusters.
Electrical companies must build the infrastructure needed to distribute power.
Cooling companies must manage increasing rack densities.
Generators and backup-power systems must protect critical data centers.
Utilities must provide more electricity.
Transmission systems must expand.
And data-center developers must navigate increasingly complicated permitting and construction requirements.
That creates enormous opportunity.
It also creates enormous risk.
AI spending could slow.
Custom accelerators could gain market share.
Memory prices could normalize.
Advanced packaging capacity could expand faster than demand.
Power shortages could delay data centers.
China restrictions could reshape global semiconductor competition.
And the financial returns from AI infrastructure could eventually come under greater scrutiny.
The most rational way to view the AI semiconductor market is therefore neither unlimited optimism nor an immediate bubble thesis.
It is a supply-chain story.
The technology is advancing rapidly.
Capital spending is enormous.
Demand remains strong.
But the physical world moves more slowly.
That difference between digital speed and physical speed may be one of the most important investment themes of the next several years.
For investors, the lesson is simple.
Do not look only at the company selling the AI chip.
Look at every company required to make that chip useful.
The next major AI investment opportunity may not come from the most obvious part of the AI stack.
It may come from the bottleneck.
And the bottleneck is likely to keep moving.
Research Note
This article is intended for informational and educational purposes. The companies and technologies mentioned above are discussed as examples within the AI semiconductor and infrastructure supply chain and should not be interpreted as individual investment recommendations. Semiconductor demand, capital expenditure, technology transitions, power availability, export controls, competition and valuation can change rapidly.
댓글
댓글 쓰기