Yahoo Finance

On Nvidia's current role in the AI boom

Nvidia is the backbone of the AI revolution. It invented the GPUs that power virtually every major AI system on the planet.

And has evolved into something much more than a chip company: a full-stack, rack-scale AI systems company. And its CUDA software is so deeply embedded across every cloud, AI lab, and frontier model that it’s become the de facto AI infrastructure operating system.

Put simply, you can’t build the future of AI without Nvidia.

Where it stands today

In its most recent quarter (fiscal Q1 2027, reported May 20), Nvidia reported $81.6 billion in revenue, up 85.2% from a year earlier. Its data center business—essentially its AI business—grew about 92% to $75 billion. Free cash flow hit a record $49 billion, and gross profit margin held steady right around 75%.

Meanwhile, management guided next quarter to about $91 billion in revenue. As CEO Jensen Huang put it, demand has “gone parabolic.”

The growth engine is Blackwell, Nvidia’s current generation of AI systems, which is ramping faster than any product in the company’s history.

Nvidia also deepened a major partnership with Anthropic and now runs essentially every frontier AI model—OpenAI’s, Anthropic’s, Google’s, xAI’s (SpaceX), and more. And it’s rewarding shareholders, returning a record $20 billion this quarter while raising its dividend.

The long-term story is stronger than ever.

Where it’s going (and why I'm still bullish)

We’re still early in what Nvidia’s founder and CEO Jensen Huang describes as AI’s third major wave: agentic AI.

First came generative AI (chatbots that answer questions). Then came reasoning AI (systems that think a problem through). Now we’re in the agentic AI era, where “agents” do multi-step work on their own like a tireless digital employee.

Each rung up this ladder demands dramatically more computing power, and Nvidia is the biggest beneficiary every time.

Behind agentic AI sits an even bigger wave: physical AI, where things like autonomous drones, vehicles, and robots operate in the real world.

Management sees at least a $1 trillion revenue opportunity from its current and next-generation systems (Blackwell and Vera Rubin) from 2025 through 2027.

Nvidia’s rack-scale platforms deliver the industry’s lowest cost per token and highest throughput. And its next generation architecture, Vera Rubin, begins shipping in the second half of this year.

Meanwhile, Nvidia’s new Vera CPU—purpose-built for agentic workloads—opens a fresh $200 billion market the company has never tapped before.

Bottom line: I continue to believe Nvidia is on track to become a $10 trillion company before 2030, possibly as early as 2028. The stock will have pullbacks along the way—sometimes sharp ones—but I view them as buying opportunities.

For more analysis on growth stocks, check out my free investing letter here .

Rambus ranks #6 in my Memory Stock Power Rankings

Think of an AI system as a kitchen. The processor (GPU, TPU, XPU, etc.) is the chef. Memory is the pantry.

If the pantry is across the parking lot, the chef spends the day jogging back and forth for ingredients. It doesn’t matter how fast she can chop and cook; the pantry is a major bottleneck.

That’s basically what’s happening inside AI data centers. $50,000 AI chips often sit idle, eating electricity, time, and cash, waiting for the pantry to feed them data.

Rambus $RMBS is like a crew that comes in and installs features that allow ingredients (data) to get from the pantry to the chef faster and more efficiently.

The company does this through two main businesses.

First, it designs and sells physical “helper” chips—tiny, specialized semiconductors that sit on memory modules and help data move faster and more reliably.

Memory modules are essentially plug-and-play packages used in computers and data center servers. They’re basically a circuit board with DRAM chips (from companies like Micron and SK Hynix) and supporting chips (like Rambus’s) that make them work and boost their performance.

Rambus is one of just three companies in the world that makes the “memory interface chipset” servers need for fast and efficient high capacity memory. (Montage Technology and Renesas are the other two.)

Second, Rambus licenses its technology blueprints (called intellectual property or IP) to other companies including Nvidia, Micron, SK Hynix, Samsung, Intel, Broadcom, and Qualcomm—collecting royalties when these companies use its inventions in their products.

Rambus basically sells chips and IP that serve as the connective tissue that makes fast memory work.

How Rambus fits the AI story

Several ways. I’ll hit on two.

For example, in Part II of this series I explained how the base die—the control chip at the bottom of the memory stack—moved from a DRAM process to a logic process with HBM4, and that this turns the high bandwidth memory needed for AI into custom silicon.

Rambus built its next generation HBM controllers (and its HBM controller IP) explicitly for this custom-HBM world.

A second way: Most folks think about AI as GPUs. But as inference and agentic AI workloads scale, the systems around the GPUs get busier—orchestration, data management, real-time execution, moving work between steps. That’s CPU work.

More CPUs means more memory modules, and each module comes with a Rambus chipset.

Meanwhile, the upgrades to the new module formats, MRDIMM and SOCAMM2, which I talked about in Part I , means more Rambus content per module.

Why Rambus ranks #6

Rambus scores very high on strategic positioning and management execution. It also has a durable moat. But it scores low on growth because of the nature of what it sells.

What I mean is that Rambus is a unit- or volume-driven business. It makes more money when more memory ships. It doesn’t really benefit from the pricing supercycle that defines this sector right now.

Plus, the same shortage that’s allowing the memory makers like Micron and SK Hynix to charge insane prices is constraining total unit volumes of the conventional modules that use Rambus’s interface chips.

I should also note that Rambus and the other two companies I mentioned earlier that make memory interface chipsets (Montage Technology and Renesas) are currently the subject of a DOJ investigation for collusion on pricing. So the limited pricing power we just talked about could become even more limited.

See my other names in my rankings here .

Situational Awareness (2026) meets LTCM (1998): Lessons for Long-Term Investors (Part I)

Recently, the market provided an important structural case study regarding systemic risk and capital concentration. Situational Awareness, a hedge fund managed by Leopold Aschenbrenner, underwent a significant deleveraging event that resulted in a rapid portfolio unwinding across public compute equities.

Prior to the unwinding, fund communications signaled strong conviction in the thesis, encouraging fresh capital inflows into the vehicle.

Until recently, the fund had drawn widespread market attention following an extraordinary performance sequence in 2025, expanding its assets under management from 225 million dollars to over 20 billion dollars in an exceptionally short timeframe.

The fundamental thesis behind Situational Awareness originated from a 165-page research document published by Aschenbrenner following his academic work at Columbia University and subsequent role on OpenAI’s superalignment team.

As documented by media reports, his tenure at OpenAI concluded following internal reviews regarding security protocols. Shortly thereafter, he authored the essay "Situational Awareness," outlining a core macroeconomic view: Artificial General Intelligence (AGI) would materialize by 2027, rendering physical compute infrastructure assets deeply undervalued.

By September 2024, the thesis attracted capital from prominent technology figures—including Stripe co-founders Patrick and John Collison, former GitHub CEO Nat Friedman, and Pioneer founder Daniel Gross—launching the fund with 225 million dollars.

Unlike established institutions built over decades, the vehicle expanded rapidly without navigating prior full market cycles. According to Financial Times reporting, while the firm grew to 20 staff members, portfolio risk management structures remained remarkably unconstrained.

Fund terms included a 25 million dollar minimum investment threshold and a 2-year lock-up period. Crucially, operating without traditional regulatory caps on position sizing allowed the fund to allocate over 40 percent of its total equity into a single enterprise.

While high concentration can be justified in businesses with durable moats and non-cyclical cash flows, applying high concentration to hyper-cyclical infrastructure sectors introduces profound downside risks.

When Explosive Returns Mask Underlying Leverage

The strategy initially delivered substantial paper gains. After generating a 2,065 percent return in 2025, the fund reported an additional 439 percent gain in early 2026, bringing total AUM to roughly 20 billion dollars.

The core portfolio holding was Nebius $NBIS , accounting for approximately 40 percent of assets. The remaining balance was distributed across infrastructure suppliers such as CoreWeave $CRWV , Micron $MU , SanDisk $SNDK , and Bloom Energy $BE, along with digital infrastructure operators like Iren, Core Scientific, and Applied Digital. This structure represented a single, unified macro bet on persistent global compute scarcity.

Subsequent analysis indicates that this performance was driven primarily by balance-sheet leverage rather than non-correlated alpha.

The portfolio operated at approximately 4x leverage relative to net assets, financed through prime brokerage facilities at major investment banks including Goldman Sachs, JPMorgan, and Bank of America. High directional leverage effectively masked the underlying concentration risk during the upward trend.

From a portfolio construction standpoint, applying 4x leverage to volatile, cyclical assets violates standard position-sizing frameworks, such as the Kelly Criterion (where optimal stake equals winning probability multiplied by expected odds minus 1, divided by odds minus 1). Over-allocating leverage beyond mathematical thresholds significantly increases the probability of an automated margin event during periods of market stress.

The Catalyst: Market Shifts and Prime Brokerage Unwinding

The catalyst occurred as financial media reported that Meta $META was evaluating options to monetize excess compute capacity in the open market. This development directly impacted key portfolio holdings, given Meta’s primary role as a major customer for CoreWeave (representing a 21 billion dollar commitment) and Nebius (representing a 27 billion dollar commitment).

Following these reports, market pricing shifted rapidly. CoreWeave declined 38 percent while Nebius dropped 46 percent through July 2026.

Infrastructure providers rely heavily on compute scarcity to maintain elevated margins. When supply expands, operating margins adjust quickly. For a highly leveraged vehicle, price declines trigger immediate collateral adjustments from prime brokers.

At the end of July, collateral requirements surpassed available liquid buffers, prompting prime brokers to execute orderly portfolio liquidations to protect their credit exposure. On July 29, an estimated 16 billion dollar block of shares was unwound in the market to reduce gross exposure and settle outstanding financing lines.

Institutional market participants, including Citadel, provided liquidity during the liquidation event at distressed valuation levels.

Citadel operates with significant gross leverage, but pairs it with centralized risk controls, tight stop-loss parameters, and broad asset diversification—systems refined over multiple market cycles since 2008. As the market rebounded the following day, Nebius rose 46 percent, CoreWeave gained 36 percent, and SanDisk advanced 24 percent, resulting in substantial capital gains for liquidity providers.

In subsequent updates to limited partners, Aschenbrenner’s fund communications highlighted that Situational Awareness remained up 80 percent Year-To-Date in 2026, despite a 67 percent interim drawdown for the month.

This YTD figure was driven by a mark-to-market valuation on a private, illiquid holding in Anthropic. Because prime brokers do not extend leverage against non-public shares, this private holding remained unencumbered during the public equity unwinding. However, capital deployed by 2026 incoming investors was concentrated heavily in the public equity sleeve, experiencing severe capital impairment while remaining bound to the 2-year lock-up structure.

In Part 2, we will draw direct parallels between this unwinding event and the 1998 collapse of Long-Term Capital Management (LTCM), extracting critical risk-management lessons for long-term investors.

Uncovering these structural dynamics requires looking far beyond the generalized noise of conventional platforms. While the rigor of my analysis can be reviewed on my Yahoo Finance Community profile, Quality Investments substack offers an even higher level of depth for investors focused on the long-term drivers of a company—drivers that are frequently overlooked by short-term, simplistic analysis. This is where the true edge of a sophisticated investor lies. Quality Investments is not a platform for casual financial entertainment; it is a serious private research space built for long-term investors, featuring institutional-grade theses, complete valuation models, and portfolio tracking.

Quality Investments | Mario Silva Arteta | Substack

Disclaimer: This post is for informational and educational purposes only and does not constitute financial, investment, or legal advice. The opinions expressed above are solely those of the author based on publicly available data. Always conduct your own independent due diligence.

How Everpure fits the AI story

Everpure $P (formerly known as Pure Storage) sells all-flash storage systems to large enterprises.

While that description is accurate, it’s also almost completely useless because it makes the company sound like a simple hardware “box” vendor. And Everpure’s never really behaved like one of those. Because how the company builds its flash storage systems is unique.

Let’s start with how a normal enterprise flash array gets built…

You basically take a box, fill it with commodity solid-state drives, and run software on top to manage those SSDs. Each drive works like its own separate little kingdom, with its own controller doing translation, wear-leveling, error correction, and garbage collection for its own flash. One SSD has no idea what any other SSD in the system is doing. And each one must keep a chunk of its capacity as a reserve for itself because it has to handle its own worst case alone.

Everpure does things differently. Instead of buying full SSDs, it buys just the raw NAND flash that goes in them, puts it in its own DirectFlash Modules, and manages it all centrally with its Purity software. So the array has one brain that knows everything going on instead of a few dozen brains that know nothing outside their own little kingdom.

Because data management happens at the system level rather than the drive level, the array needs far less spare capacity held in reserve. And because the software sees every cell in the system, it can balance wear across the whole array rather than each drive individually.

CFO Tarek Robbiati summed up the benefits of Everpure’s approach: “We buy the NAND from the same providers that provide SSDs… and we extract more—about 30% to 40% more performance from our flash modules than they do through their own SSDs.”

Same flash, same suppliers, and up to 40% more performance, because of software. That’s Everpure.

The company also throws out the traditional four-to-five-year “rip and replace” upgrade cycle/model. Customers simply subscribe to its Evergreen storage-as-a-service and the hardware refreshes underneath them while they buy capacity as needed rather than trying to guess what they’ll need years ahead of time.

So Everpure sells flash storage that’s denser, faster, and more efficient than the same flash would be in anybody else’s box, wrapped in a subscription service that removes the worst thing about owning enterprise storage.

How Everpure fits the AI story

There are at least three AI angles to Everpure’s story. I’m going to focus on the two I think matter most.

First AI angle: Agents

As you know, every AI system runs on data. If that data is slow, scattered, disorganized, duplicated, or mismanaged, the AI underperforms regardless of how good the model is.

Think about a company deploying an internal AI assistant that needs to search contracts, customer records, support tickets, company policies, product specs, and financial statements. All that data must be reachable, fast, organized, and well managed. Also, the assistant should be able to see exactly what it’s permitted to see, nothing more, nothing less.

Agents raise the stakes big time. A chatbot basically retrieves something and answers. An agent reads documents, checks databases, updates records, opens tickets, and routes information across multiple systems, continuously, with permissions checked at every step. The more agents an enterprise runs, the more work lands on whatever system stores and governs its data—enhancing the value of Everpure’s higher performing systems.

Everpure’s not building agents or selling the chips that run them. It sits underneath, providing the well managed data layer agents need to operate.

Second AI angle: Hyperscalers

In its core business, Everpure buys the NAND flash to build its systems and eats whatever price increases it can’t (or doesn’t want to) pass on to customers. That eats into margins when NAND prices are high, like today.

When the company works with hyperscalers, the model becomes much higher margin. The customer supplies its own NAND. Everpure provides the DirectFlash software, the controllers, and the integration. The Master Supply Agreements that cover these deals carry 75% to 85% gross margins. That’s well above the core range, and it strips out most of the commodity risk.

On August 10, Everpure announced its second top-five hyperscaler storage design win. The stock jumped about 10% that day and hit an all-time high a few days later.

Why Everpure ranks #7 in my Memory Stock Rankings

In the first quarter of fiscal 2027 (ended May 3rd), Everpure’s revenue grew 35% year-over-year to $1.05 billion, and non-GAAP operating profit nearly doubled to $159 million. Remaining performance obligations (backlog/contracted future revenue) reached $3.8 billion, up 41% from the same period last year.

Management then raised full-year revenue guidance by 300 basis points to $4.41 billion to $4.51 billion, which would reflect annual growth of about 22% at the midpoint of that range.

The most important numbers from the quarter: CFO Robbiati said about one-third of year-over-year revenue growth came from pricing and pull-forward. Two-thirds came from volume and customer wins.

That matters because the obvious bear case for any storage company right now is that revenue growth is simply a function of temporarily higher flash prices and customers panic-buying ahead of the next price hike. If two-thirds of Everpure’s growth is coming from volume and customer wins (a proxy for market share gains) that quashes the obvious bear case.

I alluded to one reason why Everpure isn’t ranked higher at the end of the prior section when I mentioned the stock hitting an all-time high this month—valuation jumped even faster than the business’s performance.

Everpure currently trades around 40X forward earnings. While I think it does deserve a premium for things like recurring subscription economics, high customer retention, a real software moat, and legit optionality in the burgeoning hyperscaler business, 40 times forward earnings is still a bit steep given where the business is today.

What’s more, Everpure is on the wrong side of the shortage. It buys NAND flash. So rising NAND prices are a cost. And the company has explicitly chosen to pass through less of that cost to customers than competitors. The hyperscaler side of the business can fix this, but only for the portion of revenue that’s from these customers, which is basically a rounding error today.

If you liked this, consider joining my Grow or Die investing letter here .

This is what separates Seagate...

Seagate Technology $STX makes high-capacity HDDs. About 90% of the bits it ships go into data centers.

What separates Seagate is its strategic choice to focus on advancing areal density above all else.

Remember, areal density is basically how many bits you can pack onto a platter, the magnetic spinning disk that stores the data.

There are two ways to ship more storage. You can build more drives. Or you can add capacity to each drive you build.

Seagate’s obsessed with the second path. And you can see what that looks like in the numbers. In the June quarter the company shipped 218 exabytes, which reflected year-over-year growth of 34%. But its drive unit output was essentially flat.

So same number of drives out the door and a third more storage sold.

That’s why Seagate’s gross margins have expanded from the low 30s in mid-2024 to the low 50s in mid-2026. And its operating margins have expanded from the high teens to the low 40s over the same timeframe.

The tech that delivers this operating leverage is called Mozaic, Seagate’s implementation of HAMR, the laser-heating technique I talked about in the Western Digital write up.

How Seagate fits the AI story

As AI workloads move from training to inference to agentic applications and physical AI, data isn’t just growing, it’s compounding. Exponentially more data gets generated and must be retained—for historical context, for compliance, for future reuse. So storage’s role in AI is only growing.

But Seagate’s done something smarter than simply playing to the fact that more storage is needed as AI advances. It wrote a white paper with SK Hynix that reframes how HDDs can be used in the context of AI.

Inference and agentic workloads depend on longer and longer context windows. The model needs to remember what happened across a long conversation or a multi-step task. To do this it builds a KV cache, basically like a set of detailed notes it can refer to for everything it’s already figured out so it doesn’t have to recompute everything at every step of the conversation or workflow.

Typically, KV cache lives in HBM, then spills over to flash SSDs when it gets too big, and hard drives have nothing to do with it.

Seagate’s paper argues you can extend KV cache across HBM, SSDs, and hard drives to retain far more context than you could otherwise afford and avoid recomputing data that’s already been generated.

The payoff here isn’t more storage, it’s freeing up GPU capacity for revenue-generating work—because you’re reducing the compute needed for recomputing context.

So HDDs stop being just the endpoint of data’s journey and become a way to make a GPU cluster more productive.

As far as I know we haven’t seen this sort of setup actually implemented yet, but the argument for it is bolstered by the fact that the world’s top memory company, SK Hynix, which has no particular incentive to talk up HDDs, co-authored the paper. So it’s something to watch.

Why Seagate ranks #5 in my Memory Stock Power Rankings

Seagate scored extremely high on execution, and very high on my other criteria except valuation.

Growth for example: June-quarter revenue hit a record $3.63 billion, up 48.5% year-over-year and about 17% sequentially. That’s the fastest sequential growth since 2012. And management is guiding for 56% year-over-year revenue growth for the September quarter.

At around $850 per share, Seagate trades near 24 times forward earnings. That’s not bad. But it is higher than the five-year average of just under 20.

If it were about 10% cheaper, Seagate would probably move up to #4 on my list. Sign up to Grow or Die to get the rest of my list when it publishes. Go here .