What if the most durable AI investment isn't the next breakthrough application, but the companies selling the hardware required to run all of them? 👀

We’ll follow the money underneath the AI boom, from memory and storage to custom chips and data centers, and examine where demand is real, where valuations are stretched, and what could break the thesis.

Read through to the end — the framework at the close is the part most busy investors can reuse every week.

"The individual investor should act consistently as an investor and not as a speculator."

— Benjamin Graham

A fixed $500 a month is that consistency: you add through the dull years instead of waiting to speculate on the breakout.

Carpenter Technology Corp. $CRS ( ▼ 0.75% ) closed at $396.84. Five years earlier it was about $34.25. That is a +$362.59 move, or +1,058.66% in total — roughly 63%/yr on average if you held the whole stretch. That pace is extreme. It is not a forecast, and it is a poor default to project forward blindly.

Look for on the chart: the flat stretch into 2024, the 2025–2026 rise toward the $625.98 52-week high, and the pullback to $396.84 (52-week low $233.78) — DCA would have bought more shares early and fewer into the late strength.

Lesson: Late compounding. A large share of this five-year gain showed up in the back half of the window, after years that looked ordinary. Past results never guarantee the future — a 63%/yr average is a historical outlier, not a coupon.

Next Horizon: another verified 5-year chart, same $500/month frame, same honest catch.

Want a cleaner look at this name? Open CRS on Snowball Analytics — price, fundamentals, and history in one place. Context for the chart above, not a buy signal.

Every few weeks, another headline appears suggesting that the AI boom has gone too far.

The technology is moving too quickly. Spending is getting excessive. Data-center construction is becoming difficult to justify. Companies are spending billions before anyone knows exactly how much revenue all this AI will ultimately produce.

Those concerns are worth taking seriously.

But there is a distinction that matters if you are trying to make sense of the AI trade without spending your entire day following every headline: the future of AI applications and the demand for the physical infrastructure underneath them are not exactly the same thing.

Even if AI development becomes more disciplined, the computers already being installed still need memory. They still need storage. They still need networking and optical connections. They still need power and cooling. And the machines already sitting inside data centers can continue generating revenue long after a newer generation of hardware arrives.

That creates a different way to look at the AI investment opportunity.

Instead of trying to predict which AI application becomes the next giant, consider the companies supplying the components that nearly every serious AI system needs.

That brings the focus to Micron Technology (MU), SanDisk (SNDK), Lumentum (LITE), Marvell Technology (MRVL), Cerebras Systems (CBRS), CoreWeave (CRWV), and Nvidia (NVDA).

They are not interchangeable businesses. Some look more attractive because of demand, some because of valuation, and some because their growth expectations may already be reflected in the stock price.

The difference matters.

The most important idea in this entire discussion is simple: AI does not run in the cloud in some abstract sense.

It runs on physical machines inside physical data centers.

And those data centers have limits.

Power has to reach the facility. Cooling systems have to handle increasingly dense racks. Chips have to communicate with memory. Data has to move between machines. Storage has to hold enormous amounts of information. Someone has to own and operate the infrastructure.

That means an AI slowdown does not automatically translate into an immediate collapse in demand for every component.

A significant amount of infrastructure has already been ordered, contracted, financed, or installed.

The source article points to roughly $1.7 trillion of customer commitments across Amazon, Microsoft, and Google-related cloud businesses as evidence of the scale of demand already being built into the system. The important part is not the headline number itself. It is what those commitments require before they can become revenue: data centers, electricity, cooling, networking, memory, storage, and processors.

There is also a capacity problem.

The newest Nvidia systems require substantially more power and cooling than older generations. Many existing data centers simply cannot accommodate the newest racks without major infrastructure upgrades.

That creates an unusual situation: older chips can remain useful because replacing them is not as simple as unplugging one server and plugging in another.

In other words, the AI infrastructure cycle is not necessarily a single race where yesterday's equipment becomes worthless as soon as tomorrow's chip arrives.

It can become a layered system.

That distinction is important for anyone trying to separate the durable part of the AI opportunity from the hype surrounding individual applications.

Micron Technology $MU ( ▲ 0.0% ) sits in one of the most important parts of that infrastructure.

AI models require enormous amounts of high-performance memory, particularly as models become larger and AI systems handle longer and more complicated workloads.

That changes the economics of memory.

According to the article, Micron's operating margin had increased dramatically as demand tightened the market, moving from roughly 11% of sales two years earlier to about 80% in its latest quarter.

That kind of improvement is exactly what investors should expect when supply is constrained and demand remains strong.

But it also creates the biggest question: how long can those economics last?

Memory has historically been a cyclical business. When supply catches up with demand, pricing can change quickly. So the important event for MU is not simply another strong quarter. It is whether the shortage remains tight enough to support the current profitability.

The September 30 earnings report was identified in the source as an important checkpoint for that question.

The stock's price trend remained strong, but valuation had become more demanding relative to Micron's own history. That creates a reasonable distinction between liking the business and liking the price.

A great company can still become an expensive stock.

Tip: With MU, watch memory pricing, margins, and supply conditions together. Strong demand is valuable, but unusually high profitability can also attract new supply.

SanDisk $SNDK ( ▲ 0.59% ) represents another piece of the memory equation.

While high-performance memory helps AI systems work with information quickly, storage holds the enormous volume of data those systems need to access.

SanDisk became an independent public company after being separated from Western Digital in early 2025. That gives investors a much shorter public-market history to analyze than they have with established semiconductor companies.

But the financial improvement has been significant.

The source describes SanDisk as barely breaking even a year earlier, compared with an operating margin of approximately 78% in its latest quarter.

The stock was also trading at less than eight times expected earnings for the following year, while its PEG ratio was around 0.3 in the analysis.

That combination makes SNDK interesting because investors were not being asked to pay an enormous earnings multiple for a business experiencing rapid improvement.

The limitation is that the company does not have a decade-long public history to provide the same valuation context available for older semiconductor businesses.

That means there is less evidence for determining whether the current earnings profile is sustainable.

For a busy investor, that is the number worth remembering. The opportunity is not simply “AI needs storage.” It is whether SanDisk can turn the current supply environment into sustainable earnings without eventually being hit by the normal cyclicality of the storage industry.

Lumentum $LITE ( ▼ 0.23% ) is another example of a company benefiting from AI infrastructure while demonstrating why valuation cannot be ignored.

Lumentum produces optical components, including lasers that help convert electrical signals into light so data can travel quickly through fiber connections.

As AI systems become more powerful, communication between processors, memory, and other parts of a data center becomes increasingly important.

The company has made a dramatic financial improvement. The source notes that Lumentum moved from losing money on each dollar of sales two years earlier to generating roughly 27 cents of operating profit per dollar.

Revenue growth has also been substantial.

But that is precisely where the investment question becomes more difficult.

A company can execute extremely well and still have a stock price that gets ahead of the underlying business.

The article cites Lumentum trading at roughly 195 times its previous year's profit compared with a historical valuation closer to 70 times. The PEG ratio looked much more reasonable because earnings were growing quickly, but the valuation based on current profits remained demanding.

That is an important distinction.

The business does not necessarily need to disappoint for the stock to struggle. Sometimes expectations simply become too high.

Tip: LITE demonstrates why strong revenue growth is only half the equation. Always ask how much future growth is already embedded in the stock price.

Marvell Technology $MRVL ( ▲ 0.36% ) operates in another critical part of the AI infrastructure chain.

The company designs custom chips used by major cloud providers that want specialized processors instead of relying exclusively on off-the-shelf Nvidia products.

That puts Marvell in a potentially attractive position as large technology companies increasingly develop their own AI infrastructure.

One particularly significant arrangement involves Google. The source states that Google obtained the right to acquire approximately 6.7% of Marvell while committing to roughly $120 billion of chip purchases through 2033.

That is a remarkable amount of potential business.

But there is a catch.

When investors know that a company has years of demand already lined up, the stock can begin reflecting those future earnings long before the revenue arrives.

That is the concern with MRVL.

The source places the stock at roughly 80 times previous-year earnings compared with a historical valuation around 33 times. The underlying business may be benefiting from a genuine long-term trend, but the market can still overestimate how much of that growth belongs to shareholders at today's price.

The lesson is particularly useful when evaluating AI infrastructure stocks: visibility is valuable, but visibility is not the same thing as undervaluation.

Cerebras Systems $CBRS ( ▼ 8.87% ) takes a different approach.

The company develops specialized AI computing systems built around extremely large processors. Rather than selling only individual components, Cerebras can provide complete computing systems that customers use under multi-year arrangements.

The business has generated attention because of the size of its contracted backlog.

The source cites approximately $25 billion in signed orders compared with less than $1 billion of expected sales for the year.

That creates an enormous potential runway.

But it also creates a question that should never be ignored: will those orders convert into actual revenue on schedule?

Cerebras reported a large quarterly loss, although the article points out that much of the reported loss was related to stock-based compensation rather than an equivalent cash outflow. The company also had approximately $8.6 billion in cash according to the source.

That gives it substantial financial resources, but cash does not eliminate execution risk.

Cerebras has also been public for only a short time. With limited public-market history, there is less information available to establish what a normal valuation should look like.

That makes the backlog especially important.

If contracts convert according to expectations, the investment case becomes easier to understand. If significant orders are delayed or canceled, the valuation could look very different.

Tip: With CBRS, signed backlog deserves close attention, but signed orders should not automatically be treated as guaranteed future earnings.

CoreWeave $CRWV ( ▲ 1.39% ) sits closer to the finished infrastructure itself.

Rather than simply manufacturing components, CoreWeave operates computing infrastructure and rents access to Nvidia-powered systems to customers that need AI capacity but cannot build enough data-center capacity themselves.

The growth has been extraordinary.

The source says quarterly revenue has grown from roughly $400 million two years earlier to about $2.6 billion, while customers have contracted for approximately 4.2 gigawatts of power compared with roughly 1.5 gigawatts currently operating.

That illustrates both the opportunity and the problem.

Demand is clearly large, but satisfying that demand requires enormous capital spending.

CoreWeave has been borrowing to build infrastructure ahead of customer demand, and the source cites approximately $13.7 billion of cash used over the preceding 12 months against $7.6 billion of sales.

That is not a minor detail.

When infrastructure companies grow this quickly, investors need to watch the relationship between revenue growth, capital expenditures, debt, and free cash flow.

A company can have enormous demand and still produce disappointing shareholder returns if the cost of satisfying that demand becomes too high.

CRWV therefore represents one of the clearest examples of why AI demand alone is not enough. The economics of delivering that demand matter just as much.

Then there is Nvidia $NVDA ( ▲ 0.52% ).

Nvidia remains at the center of the AI hardware ecosystem because its processors are powering a huge portion of the industry's current computing demand.

The scale of the financial improvement is difficult to ignore. According to the source, quarterly revenue has risen from approximately $30 billion two years earlier to around $96 billion, while the company now generates an operating margin of roughly 66%.

At the same time, demand continues to exceed what Nvidia can immediately supply.

The valuation argument is where things become interesting.

The source cites Nvidia at roughly 28 times previous-year earnings compared with a five-year historical average around 52 times, with a PEG ratio of approximately 0.35 at the time of the analysis.

That does not mean NVDA is risk-free.

The company has also guaranteed up to $18 billion of certain customer data-center leases, creating an exposure that deserves attention if customers experience financial trouble.

And at a multitrillion-dollar market capitalization, the question of future scale is legitimate. Nvidia no longer needs to prove that AI is commercially important. The challenge is continuing to grow at a rate that can justify the expectations attached to the company.

Still, the underlying numbers explain why Nvidia remains central to the infrastructure discussion.

The investment case is no longer simply about selling chips. It is about maintaining a position at the center of a computing ecosystem where demand for processing power continues to expand.

Tip: With NVDA, the question is no longer whether AI needs Nvidia's technology. The bigger question is how much of the industry's future growth can continue flowing through Nvidia at today's scale.

The most useful takeaway from these companies is that the AI investment story does not have to depend on predicting which chatbot, AI agent, or application ultimately wins.

There is another layer underneath it.

AI needs memory.

It needs storage.

It needs optical connections.

It needs specialized processors.

It needs finished computing systems.

It needs data centers, electricity, cooling, and networking.

That creates a chain of businesses that can benefit even if the ultimate winners among AI applications remain unknown.

But that does not make every AI infrastructure stock automatically attractive.

MU faces the normal cyclicality of memory.

SNDK has limited public-market history.

LITE's business has improved dramatically, but its valuation can become demanding.

MRVL has significant future commitments already reflected in its price.

CBRS has a massive backlog but limited trading history and execution risk.

CRWV has enormous demand but also enormous capital requirements.

NVDA has exceptional financial performance, but its sheer size means the company must continue delivering extraordinary growth to justify its valuation.

That is the distinction worth carrying forward.

The AI story can be right while an individual AI stock is still overpriced.

For someone who does not have time to chase every market headline, that is a much more useful framework than simply deciding whether AI is “over” or “the next bubble.”

Watch the physical constraints.

Watch the orders.

Watch margins.

Watch capital spending.

Watch free cash flow.

And most importantly, watch whether the growth investors are paying for actually arrives.

If the enormous backlog of AI-related commitments begins shrinking materially, the thesis changes. If those commitments continue turning into data-center demand, computing workloads, memory consumption, storage requirements, optical connections, and chip orders, the physical infrastructure story remains much harder to dismiss.

The market can argue about AI all day.

The machines still have to run.

Want a simple next step for the $500 habit?

Our Fast Track Blueprint walks through a trusted online broker setup and a clear portfolio starting path.

Thanks for reading. This format is built to be fast to open, clear to understand, and useful enough to act on — without pretending past returns continue forever.

Disclaimer: This newsletter is for informational purposes only and is not financial advice. Consult a qualified advisor before investing.