What Leopold’s final portfolio reveals about the next wave of AI investing

Every quarter, Wall Street’s biggest investors are forced to show their hand. This time, the 13F filings revealed more than $200 billion in positions spread across the AI stack—from chips and memory to cloud, power, and payments. On the surface, some of the moves look bearish: big funds trimmed Nvidia and other high-flying names. But look closer and a different story emerges: capital is rotating, not leaving. The money is digging deeper into AI infrastructure.

What 13F filings actually tell you

Before diving into specific investors, it helps to understand what a 13F is. Any investment manager with over $100 million in US equities must file a 13F within 45 days of the end of each quarter. It’s a snapshot of their long positions—what they owned at quarter-end—not a real-time trading log.

These filings don’t show short positions, leverage, or derivatives in detail, but they’re still one of the best ways to see where “smart money” is concentrating its bets, especially around big themes like AI.

Inside Leopold’s final portfolio: right thesis, wrong risk

Leopold Aschenbrenner’s portfolio has become a kind of cautionary legend. His firm has since been liquidated, so this 13F is effectively a postmortem of what went wrong—and what he was early on.

His book was extremely concentrated:

• Around 28.5% in SanDisk
• Around 28% in Micron
• About 9.5% in Bloom Energy
• Additional positions in TSMC, Coreweave, and a few “neo cloud” providers

In other words, more than half the portfolio was in AI memory, with leverage on top. Directionally, the bet was clear: AI demand would drive an explosion in memory and infrastructure. That thesis still looks intact. The problem was position sizing and leverage. Even a correct macro view can blow up a portfolio if you’re too concentrated and over-levered.

Why AI memory still matters

Despite volatility in memory stocks, the underlying demand story is strong. Recent earnings from memory giants like SK Hynix showed record profits, even as share prices sold off on fears of over-leverage and a near-term top.

SanDisk, in particular, has become a focal point because of a new product category: high-bandwidth flash (HBF). To understand why this matters, it helps to separate a few memory types:

• DRAM: the standard, general-purpose memory used in most computing
• HBM (high-bandwidth memory): tightly coupled to AI GPUs for training large models
• High-bandwidth flash: a newer category, optimized for inference workloads

Inference—the process of running models in production, not just training them—has become a huge capex sink for hyperscalers like Google and Amazon. High-bandwidth flash is designed to feed these inference workloads efficiently, and SanDisk is both an early mover and a dominant supplier. The company has already booked backlog out to the end of 2027, suggesting demand is not a short-term blip.

The core memory thesis: if AI usage keeps compounding, we don’t yet have enough fabs and supply to meet multi-year demand. Leopold’s directional call on memory may age well; his risk management did not.

Berkshire Hathaway quietly leans into Google

On the other end of the spectrum from high-octane, leveraged bets is Berkshire Hathaway. Now led operationally by Greg Abel, Berkshire moves slowly but with massive size when it’s convinced.

In the latest quarter, Berkshire added roughly $17 billion to Alphabet (Google). That’s a huge swing for a firm known for patience. Why Google, and not pure-play memory or GPU names?

The thesis centers on Google’s unique position across the AI stack:

• Models: Google builds frontier models and research (e.g., Gemini, DeepMind’s legacy work).
• Custom chips: TPUs and other in-house accelerators to train and run those models.
• Cloud: Google Cloud Platform with rapidly improving margins.
• Distribution: Search, Gmail, Workspace, Android, YouTube—products that touch billions of users daily.
• Capex: Google is on track to spend around $250 billion in capex, much of it AI-related, with a long history of strong returns on invested capital.

Owning the full stack means Google can tune its models, chips, and products together, and then push them out over its massive distribution channels. Leadership shakeups in AI (Demis Hassabis stepping aside, Jeff Dean leaving, Sergey Brin re-engaging) look more like a reset than a retreat. For long-term investors, it’s a bet that Google finally starts fully capitalizing on the technology it invented.

Berkshire’s move also rhymes with a broader pattern in the filings: hyperscalers—especially Google and Amazon—are the quiet winners of this AI cycle.

Hyperscalers as the AI core: Google and Amazon

Across many of the top funds, two names kept showing up as big winners: Alphabet and Amazon. The logic is straightforward:

• They control massive cloud platforms (GCP and AWS).
• They’re spending eye-watering amounts on AI capex.
• They already have proven, diversified revenue and strong margins.
• They can monetize AI through cloud, ads, SaaS, and internal productivity.

Amazon, in particular, is becoming one of the most important AI exposure vehicles in public markets. Through AWS, it sells compute and storage to AI startups and enterprises. It has a strategic stake in Anthropic. And it’s rolling out its own AI chips like Trainium and Inferentia to improve economics and reduce dependence on Nvidia.

When you combine that with CEO Andy Jassy’s background running AWS, you get a leader who deeply understands cloud infrastructure economics. His public commentary and AWS numbers have made many investors even more bullish on AI’s long-term impact on Amazon’s margins and growth.

If you’re building your own AI portfolio and want a simple, diversified starting point, hyperscalers are often the first layer people consider. For a more structured approach, you might find frameworks like the one in this 1-hour-a-day AI portfolio guide helpful.

The forward-looking crowd: Nvidia, SpaceX, and the infrastructure edge

Some investors are willing to go further out on the risk curve to capture more upside from AI infrastructure. Two names that stand out in this camp are Brad Gerstner and Gavin Baker.

Their portfolios are heavily skewed toward the hardware and infrastructure that actually makes AI possible:

• Nvidia: still a core holding, with multi-billion-dollar positions.
• SpaceX: a huge private stake (over $4 billion in Gavin Baker’s case), reflecting a bet on Starlink, data centers, and AI in space and at the edge.
• TSMC and ARM: the foundational chip manufacturers and IP providers.
• Neo clouds and specialized GPU clouds like Coreweave: nimble providers focused on AI workloads.

These investors are not just buying the obvious GPU names; they’re going deeper into the components and “plumbing” that will be required as AI scales.

Power and optics: the next AI bottlenecks

One of the most interesting themes emerging from these filings is the focus on power and optics. Once you accept that AI demand is likely to keep growing, two hard constraints appear: electricity and data movement.

Today, most data inside data centers moves over copper cables using electrical signals. That’s fine at small scale, but as you pack tens or hundreds of thousands of GPUs into clusters, copper becomes inefficient, hot, and expensive.

This is where photonics—using light to move data—comes in. Companies like Coherent and Astera Labs are building the optical interconnects and components that let GPUs talk to each other faster, over longer distances, using less power.

The trade looks something like this:

• More GPUs → more data to move between chips and racks.
• Copper hits physical and economic limits.
• Optical interconnects step in as the scalable solution.
• The suppliers of those optics and materials become critical chokepoints.

Gavin Baker has been especially vocal and early on this theme, taking sizable positions in optics and interconnect companies. Year-to-date, some of these names (like Coherent) have already seen strong performance, but the thesis is multi-year: if AI clusters keep scaling, optics becomes non-optional.

On the power side, investors are circling everything from traditional utilities to nuclear, renewables, and even more exotic solutions—like the jet-engine-style generators reportedly being deployed behind some data centers. The core idea: AI is not just a software story; it’s a physical infrastructure story, and power is the ultimate bottleneck.

Nvidia’s own portfolio: betting on its ecosystem

Even Nvidia, the central winner of the AI GPU boom, now looks more like an ecosystem investor than a pure chip vendor. Its own 13F shows large holdings in:

• Intel: around $30 billion, reflecting a belief in the importance of CPUs to orchestrate massive GPU clusters and AI agents.
• SpaceX: a big private stake, aligned with the expectation that SpaceX will be a major Nvidia GPU customer.
• Coreweave and other neo clouds: partners that buy Nvidia GPUs at scale and resell them as cloud capacity.
• Coherent, Nokia, Synopsys, and other infrastructure and tooling providers.

This creates a kind of circular economy: Nvidia invests in companies that, in turn, buy Nvidia hardware and build services on top of it. For outside investors, one simple heuristic emerges: if you’re not sure where to look for AI infrastructure exposure, studying Nvidia’s cap table and partnerships is a reasonable starting point.

Hedging the AI bet: QQQ puts and risk management

Not every AI bull is all-in. Some of the more sophisticated portfolios show aggressive AI exposure paired with broad market hedges. For example, Gavin Baker holds a large position in QQQ put options.

In plain English, QQQ puts are insurance against a decline in the Nasdaq-100 index. If the overall tech market falls, those puts gain value, offsetting some of the losses. That lets him stay heavily exposed to specific AI infrastructure names while reducing the risk of a broad market drawdown wiping everything out.

It’s a sharp contrast to Leopold’s approach. The lesson is simple: even if you’re convinced AI is the future, risk management—position sizing, diversification, and hedging—still matters.

Payments and AI: why Visa and Mastercard are still in play

One of the more surprising 13F themes came from Bill Ackman, who added positions in Visa, Mastercard, S&P Global, and Netflix, while trimming Amazon. At first glance, this looks counterintuitive. After all, many AI-native companies are trying to build new payment rails and reduce the fees charged by card networks.

But there’s another way to see it. Visa and Mastercard already sit in the middle of global commerce. They have deep integrations with merchants, banks, and corporate finance systems. As AI agents start to make and authorize payments on behalf of users and businesses, the easiest path might be to build on existing rails rather than replace them overnight.

Stripe’s acquisition of OpenRouter and other moves in the AI routing and payments space hint at a future where “plumbing AI” is as important as building models. In that world, incumbents with distribution and trust may have more leverage than it first appears.

Where consensus is forming—and where it isn’t

When you zoom out across all the filings, a few consensus themes emerge:

Most-owned or most-added names:

• Alphabet (Google): the standout consensus winner, with many funds increasing positions.
• Amazon: a top holding for multiple major investors (including Tiger Global and David Tepper).
• TSMC: the essential chip manufacturer behind almost every advanced AI processor.
• SpaceX: a favored private-market bet on AI, connectivity, and data centers.

Emerging but less crowded themes:

• Power and energy: from utilities to nuclear and bespoke generation for data centers.
• Memory: still a strong structural story despite short-term volatility and the shadow of Leopold’s blowup.
• Optics and interconnects: the “wiring” of AI data centers, from Coherent to Astera Labs.
• Payments and routers: Stripe, card networks, and AI model routers as the interface between AI and money flows.

Interestingly, Nvidia itself was not as heavily owned in these filings as you might expect, outside of a few concentrated believers. That may reflect concerns about valuation after its massive run, or simply a preference for owning the broader ecosystem and downstream beneficiaries.

If you’re trying to time AI markets or wondering whether the selloff in AI names is over, it’s worth pairing these insights with broader market context like the analysis in this breakdown of the AI stock selloff.

What this means if you’re building or investing in AI

These 13F filings aren’t a playbook to copy trade, but they do highlight how the most informed capital is thinking about AI:

• The AI story is shifting from just “models and GPUs” to a full-stack infrastructure story: memory, power, optics, payments, and cloud.
• Hyperscalers like Google and Amazon are becoming the default way to own that stack in public markets.
• More aggressive investors are pushing into the edges: neo clouds, optics, power generation, and private names like SpaceX.
• Risk management separates survivors from cautionary tales. Leverage and concentration can kill even a correct thesis.

Whether you’re an investor or a builder, the key question remains the same: will demand for AI models, tools, and agents be higher 10 years from now than it is today? If your answer is yes, then the infrastructure that feeds that demand—chips, memory, power, connectivity, and financial rails—is where the long-term game is being played.