3 AI Infrastructure Stocks Institutions Love but Most Retail Investors Never Heard Of

The interconnect layer beneath AI clusters is the current bottleneck. GPUs are useless if they cannot exchange gradients and tokens at line rate, and the cables, switches, retimers, and optical modules that carry that traffic have become the constraint the industry is racing to relieve. One data point frames the pressure: (NASDAQ:ALAB | ALAB Price Prediction) said PCIe 6 products already made up more than 50% of its total revenue in Q2, up from one third in Q1, evidence that the connectivity generation shift is happening in quarters, not years.

The three names below sit at that layer or feed off it. All three are mid-to-large caps, but they are higher-volatility growth stocks with concentrated customer bases, appropriate only for a small, size-controlled allocation, not a substitute for mega-cap AI exposure.

A note on the headline framing: We do not have current 13F disclosures in hand to name specific institutional holders here. Treat the “institutional” angle as a description of the shareholder base these growth-stage AI infrastructure names typically attract, not a claim tied to a specific filer.

Nebius Group: The AI-Only Cloud Buying the Interconnect

Nebius Group (NASDAQ:NBIS) runs a full-stack AI cloud that rents NVIDIA GPU capacity, storage, and managed services to labs and enterprises training and serving models. In plain English, it is a specialized cloud built only for AI workloads, competing with the hyperscalers rather than trying to be one. It is also one of the largest end-buyers of the interconnect gear the other two names in this piece sell.

The Q2 2026 numbers are the tool-verified data point. Revenue was $582.3 million, up 454% year over year, with the Nebius AI Cloud segment alone at $574.9 million, up 514%. Remaining performance obligations, the multi-year contracted backlog, stood at $37.49 billion. Management reaffirmed FY2026 guidance of $3 billion to $3.4 billion in revenue and $7 billion to $9 billion in annualized run-rate revenue, with an adjusted EBITDA margin of approximately 40%, alongside FY2026 CapEx of $20 billion to $25 billion. CEO Arkady Volozh said on the August 12 call, “We could sell today our entire 2027 capacity on these terms if we wanted to. But we are not doing this.”

Bull case: A pure-play AI cloud with hyperscaler-scale demand, a $27 billion five-year Meta agreement, and a $2 billion NVIDIA strategic equity investment behind it. Risk: it is a capital furnace. H1 2026 CapEx was $8.13 billion, three customers were 24%, 21%, and 14% of revenue, and there is $8.5 billion of convertible debt plus $12.1 billion in uncommenced lease obligations. The stock is up nearly 135% year to date (YTD) through midday Sept. 15, but it also gave back 21.47% over the trailing month. Volatility here is a structural feature of the setup.

Astera Labs: Traffic Control Inside the Rack

Astera Labs designs the connectivity silicon that lets thousands of GPUs, CPUs, and memory devices inside an AI server rack talk to each other at full speed without corrupting data. Its four product families do specific jobs. Aries retimers boost the electrical signal on a PCIe link before it degrades. Taurus does the same for Ethernet and ships as smart cable modules. Leo controls CXL-attached memory. Scorpio is the switch, effectively the traffic controller that routes data among many accelerators inside a scale-up system.

Q2 fiscal 2026, filed August 4, showed revenue of $392.4 million, up 104.5% year over year and up 27% sequentially, beating consensus of $360.81 million. Non-GAAP diluted EPS was $0.80 versus a $0.69 estimate, the eighth consecutive EPS beat. Non-GAAP operating margin was 39.1%. Q3 guidance calls for revenue of $540 million to $560 million and non-GAAP EPS of $1.16 to $1.21. Management expects Scorpio to become the largest product family in Q3, one quarter ahead of plan. CEO Jitendra Mohan said content per accelerator is scaling: “We expect the content opportunity for Scorpio X Series solutions alone to grow well beyond $1,000 per XPU in future generations of AI platforms.” Management cited a merchant scale-up switching market projected to reach $20 billion by 2030 (management-stated figure).

Bull case: the company is graduating from a signal-conditioning niche into a full AI fabric provider as Scorpio ramps across more than 10 customers. Risk: hyperscaler customer concentration, cyclical semiconductor exposure, gross margin variability as the switching mix grows and a rich valuation at a P/E of roughly 203. The stock is up 45% YTD but down 9% over the past week and nearly 19% over the past month. This is size-appropriate for a small allocation only.

Credo Technology: Copper, Optical, and Everything Between

Credo (NASDAQ:CRDO) supplies the cables and chips that carry data at 800G and 1.6T speeds inside and between AI racks. Active electrical cables (AECs) are copper cables with a small chip inside that regenerates the signal, replacing laser-based optical transceivers for short GPU-to-switch runs at lower power and higher reliability. Optical DSPs and silicon photonics chips do the same job in glass for longer distances. Retimers clean up PCIe and Ethernet signals. Zero-flap optics adds a software layer that predicts and mitigates link failures before a cluster goes down.

Q1 fiscal 2027, filed Sept. 1, delivered revenue of $479 million, up 114.7% year over year, with non-GAAP EPS of $1.20 versus a $1.17 estimate, the eighth consecutive EPS beat. Non-GAAP gross margin was 68% and non-GAAP operating margin was 48.2%. Q2 revenue guidance is $525 million to $535 million. Management guided fiscal 2027 to more than $600 million of optical revenue and more than 85% year-over-year total revenue growth. CEO Bill Brennan framed the thesis this way: “As AI infrastructure scales, our job is to provide connectivity that works reliably, uses less power, provides visibility into the network, and keeps expensive processors operating at high utilization.” Credo was also added to the Bloomberg 500 Index following the September reconstitution announced August 26, 2026.

Bull case: a vertically integrated portfolio spanning copper AECs, optical DSPs, silicon photonics, retimers, and diagnostic software, all riding the same AI data-center buildout. Risk: severe customer concentration, with the top four customers at 33%, 28%, 13% and 10% of Q1 revenue. Acquisition-related GAAP gross margin compression, elevated share-based comp of $87.98 million in the quarter, and a tariff regime management itself called fluid all belong in the risk column. The stock is up only 5.89% YTD and is down more than 46% over the past month after a sharp reset, so the momentum framing has reset, while the fundamental story remains intact.

What to Watch Next

The through-line is that GPU spending is meaningless without the interconnect fabric underneath. Nebius is the customer building at gigawatt scale, and Astera Labs and Credo are the silicon and cable suppliers who get paid every time a rack is populated. This is the same pattern behind our free report on seven AI infrastructure names that aren’t chipmakers, the power, cooling and networking companies quietly feeding the buildout. All three trade with the volatility of growth-stage semiconductor and cloud names, well above hyperscaler-level composure. The next catalysts are the Scorpio X production ramp at Astera, the fiscal 2027 optical revenue report at Credo, and Nebius’s connected-power update against its 800 megawatts to 1 gigawatt year-end 2026 target.

Contact [email protected] for any questions or corrections.