Markets Confusing? Ask Edgen Search.
Instant answers, zero BS, and trading decisions your future self will thank you for.
By David Hartley | Senior Equity Analyst — Technology (AI / Data Center) — 2026-05-02 Rating: Overweight (Sector) Tickers: $NVDA, $AVGO, $MRVL, $VRT, $CAT, $1879.HK, $GNRC, $ETN, $STX, $POET
If you opened your brokerage app this week and saw Caterpillar hit a new all-time high, Vertiv up 270% over the past year, and a Hong Kong silicon photonics IPO called Lightelligence jump 383% on its first day, you probably had the same question I did: what is going on, and what do these companies have in common?
The short answer is that hyperscalers — Microsoft, Meta, Amazon, Alphabet, Oracle — are about to spend roughly $527 billion on AI infrastructure in 2026. That number was $465 billion when Q3 2025 earnings season started. It moved up another notch this past week. And the wild part isn't that the number is big. The wild part is that most of the spending doesn't go to Nvidia.
This piece is a map. We'll walk through the three layers of the physical AI stack — the chips, the stuff inside the rack, and the stuff outside the rack — and rank 10 stocks by where I'd put fresh money today. Eight of the ten already have full deep-dive articles in our research, so this overview connects them into one picture. If you only read one section, read the layer that matches the part of the stack you find easiest to imagine: a graphics card, a metal rack with fans, or a building with a backup generator out back.
Why this sector matters right now
The story everyone tells is "AI needs chips, so buy Nvidia." That story was right in 2023 and 2024. In 2026 it is incomplete. Hyperscaler capex guidance from this past Q1 reporting cycle showed a clear pattern: Microsoft is now signaling roughly $190 billion in fiscal-year capex, Meta around $145 billion, and Alphabet and Amazon stepping up another notch each. Sum the announced numbers and the consensus has moved from $465 billion at the start of Q3 2025 earnings season to roughly $527 billion as of this week.
Of that $527 billion, only a slice ends up at Nvidia. A bigger chunk than most retail investors realize goes to the boring middle of the stack — power, cooling, switchgear, generators, racks, and the increasingly important "optical interconnect" piece that ties GPUs together. This is why CAT, with a roughly $63 billion construction-equipment backlog, just printed a Q1 strong enough to push the stock up 9% to a new all-time high. It's why Vertiv, which makes the power and cooling that lives inside the data-center rack, is up 270% over the past year. And it's why a Hong Kong-listed silicon photonics startup nobody had heard of a month ago booked the largest first-day pop on the exchange in roughly a decade.
The week's three signals — CAT's $63B backlog, Lightelligence's +383% IPO, and consensus capex moving up $62 billion in six months — all point at the same thing. The money is real, the money is committed, and it's flowing to places beyond the Magnificent Seven.
Layer 1 — Chip and Interconnect (5 stocks)
This is the layer everyone sees: the silicon that does the math. NVDA still anchors it. But the more interesting story is what happens at the edges — custom silicon designed for one customer, and the photonics that ties chips together inside a rack.
1. NVDA — Buy
Nvidia is still the only company shipping a full-stack AI accelerator that scales from a single GPU to a 72-card rack. The H100 → H200 → B200 → Rubin roadmap is the baseline every hyperscaler builds around, and the software moat (CUDA, NCCL, cuDNN) keeps competitors paying a tax to even compete. The only honest debate now isn't "is NVDA the leader" — it's "how much of hyperscaler training spend gets pulled to in-house custom silicon over the next three years." We think the answer is "less than the bears claim, more than the bulls admit," and at current levels that still leaves NVDA as the cleanest core holding in the cluster. For the broader chip-layer landscape, see our semiconductor sector overview. Forecast page: NVDA forecast.
2. AVGO — Buy
Broadcom is the quiet winner of the custom-silicon story. AVGO designs the AI ASICs that Google and Meta actually deploy at scale — not GPUs, but accelerators tuned to one workload at one customer. Each design win is multi-year and multi-billion. AVGO also owns the networking switches (Tomahawk, Jericho) that move tokens between thousands of those chips, which makes it a rare two-leg bet inside the same rack. If you believe hyperscaler in-house silicon eats share from merchant GPUs, you almost have to own AVGO. See our semiconductor sector overview for the full peer landscape.
3. MRVL — Buy
Marvell is the second-place player in custom AI silicon, but second place in this market is still a great place to be. MRVL designs Trainium for AWS and Maia for Microsoft — two of the four largest AI workloads on Earth. The data center segment now represents more than three-quarters of company revenue, and the optical-interconnect business inside it is where the most interesting margin expansion sits. The risk is concentration: if one of those two anchor customers walks, the model breaks. The reward is a company that has effectively become a custom-silicon design house with two of the biggest customers in tech. Full thesis: MRVL stock analysis — AI custom silicon 2026.
4. Lightelligence (1879.HK) — Hold
Lightelligence is the contrarian-interesting name on this list. The company makes co-packaged optics, or CPO — silicon photonics that move data between AI chips at light speed instead of through copper traces that bottleneck once you stack enough GPUs. The IPO debut on the Hong Kong exchange popped 383% on day one, the largest first-day move in roughly a decade. It is also the cleanest pure-play public bet on CPO commercialization. Why Hold and not Buy? The valuation now bakes in nearly perfect execution from a company that has yet to prove it can ship at hyperscaler volumes. The stock will trade like a 90% pure beta on AI sentiment for the next 6-12 months. We'd rather wait for either a real customer ramp or a 25-30% pullback. Full breakdown: why Lightelligence (1879.HK) surged 383% on IPO debut.
10. POET — Sell (failure-case lesson)
POET Technologies is on the same technology curve as Lightelligence — silicon photonics, CPO. Same physics, same end customer pool, same theoretical TAM. And yet POET dropped 47% in a single session this past week after Marvell publicly walked away from a development engagement that the market had been treating as a path to commercialization. The lesson here isn't "photonics is bad." The lesson is that within an emerging technology category, the gap between the company that lands the anchor customer (Lightelligence with its sponsor relationships) and the company that loses the anchor customer (POET) is a 5x stock price difference inside one week. If you're going to play CPO, play the proven one. Full diagnostic: why POET Technologies crashed on Marvell cancellation.
The Lightelligence-versus-POET pair, sitting next to each other on the same technology, in the same week, is the cleanest case study in this sector this year. Same tech, opposite outcomes — and the gap is who got the contract, not whose technology was better.
Layer 2 — In-Rack Power and Cooling (2 stocks)
Once the chips exist, you have to feed them and keep them from melting. A Blackwell rack pulls roughly 120 kilowatts. The previous generation was about 30. That four-times jump in density is why power distribution and liquid cooling went from a forgotten line item in the data-center budget to one of the fastest-growing segments in the entire industrial complex. This is the layer between the building's substation and the back of the GPU — and it is where the most boring-looking stocks have produced the most spectacular returns.
5. VRT — Buy
Vertiv is the cleanest pure-play. The company makes the rack-level power distribution units, the busways, the in-row cooling, and the liquid-cooled coolant distribution units that every modern AI rack needs. Roughly 70% of revenue is now AI-data-center exposed, and the order book backlog stretches deeper than the company has visibility for in any prior cycle. The stock is up 270% over the past year, which is the obvious problem — but the thesis isn't "is it cheap on trailing earnings," it's "is the secular ramp going to continue for another three to five years." We think it does. Full thesis after the rally: Vertiv stock thesis after 270% rally. Forecast page: VRT forecast.
7. ETN — Buy
Eaton is the diversified version of the same trade. ETN sits one level upstream from Vertiv — switchgear, transformers, the medium-voltage equipment that takes utility power and turns it into something a data-center floor can use. Roughly 15-20% of company revenue is now AI-data-center linked, and that share is climbing every quarter. ETN trades at a more reasonable multiple than VRT and gives you exposure to the same tailwind without the pure-play volatility, but you also get less torque per dollar of capex. Think of it as the IBM-of-electrical-infrastructure trade — slower compounding, lower drawdowns. We don't have an Edgen deep-dive on ETN yet; that's on the roster.
Layer 3 — Outside-Rack Backup Power and Storage (3 stocks)
The third layer sits outside the white space. It's the diesel generators that keep a data center running when the grid blinks, the hard drives where models park their training data, and the trucks and turbines that get the campus built in the first place. This is the layer Wall Street keeps under-pricing because it doesn't sound like an "AI stock." It is.
6. CAT — Buy
Caterpillar is the one nobody saw coming. CAT's Energy & Transportation segment makes the large diesel and natural-gas reciprocating engines that hyperscalers install as backup power for their campuses. Q1 2026 came in strong enough to push the stock up 9% to a new all-time high, and management raised the construction-equipment backlog to roughly $63 billion — a record. A meaningful chunk of that backlog is now AI-data-center-linked, including the gensets, the campus earthmoving, and the transmission infrastructure to deliver power to the site. The market historically valued CAT as a late-cycle industrial. The new version is part-industrial, part-AI-infrastructure pick-and-shovel. Full breakdown: why Caterpillar stock jumped 9% on Q1 2026 AI data-center power. Forecast page: CAT forecast.
8. GNRC — Buy
Generac is the small-cap version of the genset trade. GNRC has historically been a residential-backup-power story, but the company has been moving aggressively into the commercial and data-center segments and recently won hyperscaler vendor qualification — the gating moment where a hyperscaler approves a vendor for actual orders. The thesis is that the hyperscaler revenue mix grows from a low single-digit share today to a meaningfully larger share over the next 24-36 months, with both higher growth and better margins than the legacy residential business. Higher beta, higher upside, more execution risk than CAT. Full thesis: Generac AI data-center backup-power thesis.
9. STX — Buy
Seagate is the storage layer. AI training generates obscene quantities of data — checkpoints, embeddings, model versions, output logs — and most of it parks on hard drives, not on the more expensive solid-state. STX's Mozaic 3+ HAMR-based drives are reportedly sold out through 2027, which gives Seagate something hard-drive companies historically never had: pricing power. This is the rare trade where a commodity-coded business gets to print growth-stock economics for a couple of years. The risk is the eventual oversupply cycle, but that's a 2028 problem, not a 2026 problem. Full thesis: Seagate STX AI HDD sold-out Mozaic margin thesis.
The ranked table
| Rank | Ticker | Layer | Rating | 12-Month View | Key Thesis | Key Risk |
|---|---|---|---|---|---|---|
| 1 | NVDA | Chip | Buy | Core holding | Full-stack accelerator + CUDA software moat; hyperscaler training still anchored on Blackwell/Rubin | In-house custom silicon (AVGO/MRVL design wins) eating training share faster than expected |
| 2 | AVGO | Chip / Custom Silicon | Buy | Core holding | Custom AI ASICs for Google + Meta + others; networking switch leader inside the rack | Customer concentration; one design loss = real revenue gap |
| 3 | MRVL | Chip / Custom Silicon | Buy | Buy on dips | Trainium for AWS, Maia for Microsoft; data center >75% of revenue mix | Two-customer concentration; either anchor walking breaks the story |
| 4 | 1879.HK Lightelligence | Chip / Photonics | Hold | Wait for pullback | Cleanest pure-play CPO commercialization; +383% IPO debut | Valuation; volume execution unproven; HK micro-float volatility |
| 5 | VRT | In-Rack | Buy | Core holding | Pure-play AI rack power + liquid cooling; ~70% AI-DC revenue | Multiple compression after 270% rally if capex pause |
| 6 | CAT | Outside-Rack | Buy | Core holding | $63B backlog, AI gensets + campus build re-rating; ATH on Q1 strength | Late-cycle industrial cyclicality returns; data-center share still <20% |
| 7 | ETN | In-Rack (diversified) | Buy | Slow compounder | Switchgear + transformer leader; ~15-20% AI-DC revenue mix | Less torque per capex dollar than VRT; multi-year ramp |
| 8 | GNRC | Outside-Rack (small-cap) | Buy | Higher-beta name | Hyperscaler vendor qualification; mix shift from residential to commercial | Execution risk on the mix shift; lumpy hyperscaler order timing |
| 9 | STX | Storage | Buy | 2-year window | Mozaic 3+ HAMR drives sold out through 2027 | Oversupply cycle returns 2028+; commodity nature reasserts |
| 10 | POET | Chip / Photonics (failure case) | Sell | Avoid | Marvell walked away from anchor engagement; -47% in one session | Customer pipeline now thin; story stock without a story |
What could break the sector
I have to put a stake in the ground on what would actually invalidate this sector-level Overweight call. Three named risks, in order of how much I worry about each.
The first risk is a hyperscaler capex pause. Every name on this list, including NVDA, gets a meaningful share of revenue from Microsoft, Meta, Amazon, Alphabet, and Oracle. If any two of those five materially cut 2027 capex guidance — say, one quarter of slowing Azure or AWS growth, or a board-level decision to reset the spend pace — every name on this list takes a 15-30% multiple compression. This isn't a tail risk, it's the central scenario the bear case prices in. It's also the reason I'd keep position sizing reasonable rather than running concentrated.
The second risk is in-house photonics from Nvidia. NVDA has been quietly investing in its own optical interconnect roadmap. If at GTC 2026 or 2027 NVDA announces a meaningful in-house CPO product that obsoletes the merchant photonics market, both Lightelligence and the merchant transceiver names get repriced lower in days. This is specifically why we rate Lightelligence Hold rather than Buy despite the IPO momentum.
The third risk is customer concentration at one or two hyperscalers. AVGO, MRVL, and to a lesser extent VRT, have meaningful single-customer revenue exposure. The hyperscalers know this and use it in negotiation. Any single design loss or roadmap slip — particularly one that lands at AVGO or MRVL — would crater that name even in a healthy capex environment.
I do not think any of these three breaks in 2026. I think one of them probably matters in 2027 or 2028. That's the calendar I'd plan around.
Bottom line
Sector rating: Overweight. The math is straightforward. Hyperscalers will spend roughly $527 billion in 2026 on infrastructure that physically has to exist somewhere. The chips are one layer of three, and the other two layers — what goes inside the rack, what sits outside it — have less competition, less crowded multiples, and longer-duration order books than the headline names.
Top three picks if I had to size new positions today: NVDA for the core chip exposure, VRT for the cleanest in-rack pure-play, and CAT for the outside-rack story Wall Street is still pricing as a cyclical industrial. Avoid POET — the contract walk-away changed the story and the path forward is unclear.
The Lightelligence-versus-POET pair is the cleanest reminder of how this sector actually works. Same technology, same week, opposite outcomes. The companies that win in AI infrastructure are the ones with the anchor customer. Position your portfolio around the ones that already have one.
Frequently Asked Questions
What does "AI physical infrastructure" actually mean as a sector?
AI physical infrastructure is the real-world equipment that has to exist for AI software to run, and it covers three layers: the chips that do the math (NVDA, AVGO, MRVL, optical photonics), the in-rack power and cooling that feeds and cools those chips (VRT, ETN), and the outside-rack systems that supply backup power and store training data (CAT, GNRC, STX). When hyperscalers announce capex numbers, this is where the money physically goes.
Why is $527 billion the number everyone is using?
That figure is the sum of disclosed 2026 calendar-year capex guidance from the five largest U.S. hyperscalers — Microsoft, Meta, Alphabet, Amazon, and Oracle — based on their most recent earnings calls through Q1 2026. The number was roughly $465 billion at the start of the Q3 2025 reporting cycle and stepped up another $62 billion as guidance was raised this past quarter. Most independent sell-side desks now publish $520-540 billion as the consensus 2026 figure.
If I can only buy one stock from this list, which should it be?
For most retail investors with limited positions, NVDA remains the cleanest single name because it captures the broadest slice of total AI infrastructure spending and has the deepest software moat. If you specifically want exposure to the under-priced layers most retail investors miss, VRT (in-rack pure-play) or CAT (outside-rack with industrial diversification) would be the next two I'd consider. The single biggest mistake is owning ten names equally without thinking about which layer each is exposed to.
Why is POET rated Sell when it's the same technology as Lightelligence?
The technologies overlap, but the businesses don't. Lightelligence has anchor sponsor relationships and is moving toward commercialization; POET lost its biggest development engagement when Marvell publicly walked away earlier this week, and there is no announced replacement customer. In emerging-technology categories, the company that lands the anchor customer survives and the company that loses one usually does not. The price gap reflects that, not a difference in physics.
Will hyperscalers actually spend $527 billion or is this an analyst fantasy?
The number is grounded in management guidance from the actual companies that have to write the checks, not analyst extrapolation. Microsoft, Meta, Alphabet, and Amazon all raised 2026 capex guidance at their Q1 calls; Oracle disclosed multi-year backlog tied to AI capacity contracts. The actual risk isn't that the planned spend doesn't happen, it's that the spend pace gets stretched out 6-12 months if AI demand or token economics disappoint. That stretching would compress multiples in the names on this list, but it wouldn't invalidate the cluster.
Disclaimer: This article is for informational purposes only and does not constitute investment advice. Past performance does not guarantee future results. Always consult a licensed financial advisor and conduct your own research before making investment decisions.
Your money person, finally.
Try Ed free. No credit card. No commitment.