Hyperscalers vs Neoclouds: How I Invest in the AI Buildout

I keep asking myself one question when I look at my portfolio: where does the money in the AI buildout actually land?

Everyone got the first answer right. Buy Nvidia. That trade was obvious by 2023 and it's crowded now. The second answer is less obvious, and it sits one layer down from the model everyone talks about: data centers, memory, optics, and electricity.

I invest from the Philippines through two accounts. Interactive Brokers holds the core positions, the ETFs, and anything listed outside the US. GoTrade is where I open small fractional positions when I want to start a name with a few thousand pesos instead of waiting to afford a whole share. IBKR does the heavy lifting. GoTrade lowers the cost of being curious.

These are my notes from mapping the space, including the parts I got wrong first. Not advice.

Why I stopped thinking of "cloud" as one thing

For years I assumed cloud was cloud. AWS, Azure, Google. Rent a server, done. Then I read a Nutanix breakdown on neoclouds and it reframed the whole category for me.

Traditional cloud was built for variability. Web apps, databases, microservices, thousands of small tenants sharing hardware. Training a frontier model is the opposite workload. It's one enormous job that runs for weeks and needs every GPU talking to every other GPU at full speed, all the time.

Different problem, different machine.

Four differences do the work in the investment case.

• Bare metal beats virtualization for this job. Hyperscalers run multi-tenant hypervisors so one box can be sliced across many customers. Great for isolation. Bad for latency. Neoclouds hand the GPU over directly.
• Networking is the real moat. Training a multi-billion parameter model generates enormous east-west traffic between GPUs. Standard SDN chokes on it. Non-blocking InfiniBand or 800G Ethernet keeps tens of thousands of GPUs in sync without packet loss.
• Pricing is simpler. Hyperscalers stack compute, storage IOPS, API calls, and egress. Neoclouds quote an hourly rate for a cluster that runs continuously. Labs burning 24/7 compute care a lot about that.
• New hardware lands first. Neoclouds only do acceleration, so they often deploy the newest Nvidia architecture weeks before hyperscalers finish allocating supply globally.

The core positions, the boring half

My IBKR account holds the anchors. These are the names I don't check daily and don't trade around.

• Microsoft (MSFT). Azure is where enterprise OpenAI deployments land, plus Copilot monetization across Office.
• Amazon (AMZN). AWS defending share with custom silicon (Trainium, Inferentia) to fix the unit economics on AI workloads.
• Alphabet (GOOGL). TPUs plus Nvidia. GCP keeps winning AI-first startups.
• Nvidia (NVDA). Powers both sides of this fight.
• Broadcom (AVGO). Custom AI ASICs and the switching silicon underneath the fabric.
• Arista (ANET). Ethernet interconnect for scaled-out AI data centers.
• Vertiv (VRT). Liquid cooling and thermal management. Unglamorous. Non-optional.

Nothing clever here. This is the Zone 2 base of the portfolio. Easy miles that compound while I go do something else.

The memory trade I missed early

I got this one wrong for about a year, and it still annoys me.

Everyone stared at the GPU. The actual squeeze was on memory. A Blackwell part is useless without high bandwidth memory stacked next to it, and HBM production sits with three companies. Then the second wave hit. Every cluster needs somewhere to put training sets, checkpoints, inference caches, and vector stores, which pulled NAND and enterprise drives into the same shortage.

Memory used to be the ugliest cycle in tech. Boom, glut, price collapse, repeat. AI bent the demand curve because HBM gets sold on multi-year contracts instead of spot.

I hold Micron as the main position here. It's the cleanest US-listed way to own HBM without opening a Korea account, and it's the one I keep adding to. I also hold a smaller SanDisk position on the storage side, sized as a satellite rather than a core holding. That's the whole memory sleeve. One conviction name, one tag-along.

The honest risk: memory is still a cycle. AI smoothed it, it didn't delete it. If capex pauses, these fall harder and faster than the hyperscalers do. I size them like growth positions, never like anchors.

The growth half is about power, not chips

This is the part that changed how I think about the whole sector.

The bottleneck stopped being GPUs. It's electricity and grid interconnects. You can order racks. You cannot order 500 MW and have it show up next quarter.

The scarce asset isn't the chip anymore. It's the substation, the land, and the signed interconnect agreement.

So the interesting names aren't only the compute providers. They're the companies already sitting on power.

The private pure-plays set the comps. CoreWeave is the benchmark neocloud, Nvidia-backed, built partly on repurposed crypto mining infrastructure. Lambda Labs and Crusoe Energy run low-cost GPU clouds tied to flexible or clean energy.

The public names I can actually buy:

• Nebius (NBIS). Pure-play AI infrastructure, large GPU clusters across Europe and North America, direct Nvidia allocations.
• Applied Digital (APLD). Builds and operates ultra-dense data centers designed for AI from the ground up.
• Galaxy Digital (GLXY). The Helios campus in West Texas. Roughly 2,200 acres and 1.6 GW of approved power, converted from Bitcoin mining to AI and HPC. Phase I delivered 133 MW of critical IT load to CoreWeave under a 15-year lease with about $1.4B in project financing. Phase II adds another 260 MW.
• Hut 8 (HUT). Over 1,000 MW under management across Power, Digital Infrastructure, and Compute. Signed a gigawatt-scale deal with Anthropic and Fluidstack for up to 2,295 MW, with liquid cooling designed for 180 kW racks.
• IREN (IREN) and TeraWulf (WULF). Power interconnects, land rights, and high-voltage substations getting retrofitted for high-density AI.

I started most of these small on GoTrade. Fractional shares, small size, no drama. If the thesis holds I size up through IBKR. If it breaks, I lost lunch money and learned something.

The one position I didn't choose: Ionic Digital (IOND)

Every other name here was a decision. This one arrived in my account.

I'm an Ionic Digital shareholder because of the Celsius Network bankruptcy, not because I ran a screen and liked what I saw. Tens of thousands of former Celsius creditors got the same shares in the same way. We didn't buy in. We got handed equity as part of the settlement and then had to go figure out what we owned.

What we own turns out to be interesting. Ionic Digital is a digital infrastructure and Bitcoin mining company that provides power-ready sites for high-performance computing and AI workloads. It was formed out of the Celsius bankruptcy exit and started trading on Nasdaq under IOND on July 28, 2026.

So it lands in exactly the same bucket as Galaxy and Hut 8: energized sites, substations, and land looking for a higher-value tenant than a hash rate. Same pivot, different origin story.

The position I'm still building: SpaceX (SPCX)

I hold SpaceX and I plan to keep adding for the next two years. Not because of rockets.

SpaceX (SPCX) is increasingly viewed by major Wall Street firms as an AI infrastructure play rather than just an aerospace company, driven by massive scaling in AI compute, high-capacity data center leases, and massive capital expenditure. (Yahoo Finance)

That reframing is the whole reason I own it. SpaceX listed on Nasdaq under SPCX in June 2026 in the largest IPO ever recorded, and the story since then has been less about launch cadence and more about compute. Musk has said he wants to exit 2026 with more than 2 GW of compute and get closer to 10 GW by the end of 2027, built on Nvidia silicon.

Then there's Starmind, the orbital compute effort with Nvidia. Solar power is abundant up there and vacuum solves part of the cooling problem, which are the two constraints strangling terrestrial data centers right now. I have no idea if it works at scale. I'm not underwriting it.

What I'm actually underwriting is simpler. Every gigawatt claim in this post depends on someone solving power, cooling, and capex at a scale nobody has attempted. SpaceX is one of the few organizations with a track record of doing exactly that in a different domain. I'm treating it as a long-horizon position I dollar-cost into through IBKR, not a trade.

The layer I only noticed this year: optics

Once you accept that networking is the moat, the next question is what carries the signal. At 800G and above, copper runs out of road fast. Distance, heat, and power draw all work against it. So the interconnect between racks and between buildings moves to light.

That's optical transceivers, silicon photonics, laser sources, and fiber. Lumentum, Coherent, and Ciena sit here, along with a set of Chinese manufacturers most Western investors have never screened. It's a small, unsexy supply chain sitting directly on top of a multi-decade buildout.

I don't have the domain knowledge to pick individual photonics winners, and I'd rather admit that than fake it. So this is one I hold as a basket.

The ETF route

Not every position needs to be a single stock. Three funds cover this theme from different angles.

Roundhill Neocloud ETF (NCLD)

Launched August 2026 on Nasdaq, 0.65% expense ratio, actively managed. Unlike older cloud ETFs stuffed with SaaS, NCLD targets GPU-as-a-Service platforms, high-density data center operators, liquid cooling, and interconnect makers. Concentrated by design, which cuts both ways.

Roundhill Photonics & Optics ETF (LYTE)

Launched August 6, 2026 on Cboe BZX, 0.65% expense ratio, actively managed with limited turnover from quarterly rebalancing. It holds companies drawing at least half their revenue from optical transceivers, laser sources, silicon photonics ICs, optical interconnect systems, photonic substrates, photodetectors, and fiber infrastructure. Top holdings include Lumentum, Coherent, Eoptolink, Zhongji, and Ciena.

This is my answer to the optics layer. The fund's own framing is that optical connectivity is a secular story tied to the AI infrastructure buildout, and I agree with that framing more than I trust my own stock picking inside it. Details on the Roundhill LYTE page.

iShares AI Infrastructure UCITS ETF

Broader scope: semiconductor capital equipment, data centers, and the utilities powering them. The UCITS structure matters more than people outside the US realize, and it's a real reason my ETF positions live on IBKR rather than anywhere else.

How I actually split it

Core anchors give me cash flow and downstream software exposure. The growth sleeve gives me asymmetric upside on power and density. The ETFs let me own a layer I understand structurally but can't stock-pick inside, which is exactly what LYTE is doing for me on optics.

Key takeaways

• Hyperscalers and neoclouds are different machines, not different brands. Virtualization, networking topology, and pricing model are the tells.
• The bottleneck moved from chips to power. Gigawatt-scale interconnects are the scarce asset now.
• Memory was the trade hiding in plain sight. HBM and NAND ride the same demand curve as the GPU.
• Optics is the next layer down. Copper can't carry 800G far, so light does.
• When you understand a layer structurally but can't pick inside it, buy the basket and stop pretending.
• The market is repricing SpaceX as a compute company, not an aerospace one. That's a bet on execution at gigawatt scale, so I'm entering slowly.
• Two accounts solve two problems. IBKR for size, breadth, and non-US listings. GoTrade for cheap curiosity.

What I'll do next

Track two numbers on the growth names: contracted megawatts and lease duration. Announcements are cheap. Signed multi-year offtake at delivered capacity is not.

Watch HBM contract pricing and NAND spot prices, because memory tells you the cycle is turning before the GPU names do. Keep buying SPCX on a fixed schedule across the next two years and check the compute guidance against delivered capacity, not slides. Keep adding to core positions monthly regardless of price, and keep the neocloud sleeve small enough that a 50% drawdown is annoying instead of fatal. These names move violently in both directions and I'd rather be early and small than right and wiped out.

The AI supercycle is a physical buildout. Concrete, copper, glass, transformers, and cooling. The split between hyperscalers and neoclouds is the most useful thing I've figured out this year, and I expect it to keep paying for a while.