2026-05-15 — views $ALAB · Astera Labs · Aries / Leo / Taurus / Scorpio
Astera Labs (ALAB) — the first pure-play public bet on AI rack interconnect IP
ALAB ships four product lines that wire the AI rack — PCIe retimers, CXL memory controllers, Smart Cable Modules, and the new Scorpio Gen6 fabric switch. Picks-and-shovels for hyperscale buildouts.
Astera Labs (NASDAQ: ALAB) is the first US public-market pure-play on AI rack interconnect IP — the silicon that wires GPUs to each other, to memory, and to the network inside a hyperscale AI cluster. Founded 2017, IPO’d March 2024, and now selling into all five major US hyperscalers as well as multiple NVDA reference platforms.
For builders, the company is less interesting as a stock pick than as a category-defining read: the picks-and-shovels for scale-up matter as much as the GPU itself once cluster sizes exceed what a single rack can hold.
What ALAB actually ships — four product families
| Line | Function | Where it sits |
|---|---|---|
| Aries | PCIe retimers (Gen 3 → Gen 6) | CPU↔GPU, CPU↔NVMe inside a server |
| Leo | CXL Memory Connectivity Controllers | Pooled memory expansion via CXL |
| Taurus | Smart Cable Modules (electrical + active optical) | Top-of-rack scale-out cables |
| Scorpio | Smart Fabric Switches (PCIe Gen 6) | Rack-level GPU-to-GPU fabric |
Aries is the historical revenue engine — every modern PCIe Gen 5 / Gen 6 server design that crosses ~30cm of trace needs a retimer to keep the signal clean, and ALAB is the volume incumbent. Leo is the CXL play, smaller in revenue but a leading indicator for memory disaggregation. Taurus and Scorpio are the AI rack plays: Taurus wires GPUs to NICs and top-of-rack switches over electrical cable; Scorpio is the new fabric switch that allows >8 GPUs to share a PCIe Gen 6 fabric within a rack — directly competing with NVDA’s NVLink switch on the open standard side.
Why Scorpio matters more than the other three combined
NVLink is closed; PCIe is open. For hyperscalers building non-NVDA clusters (AMD MI-series, custom silicon, future Cerebras / Groq racks), PCIe Gen 6 is the native scale-up fabric. Scorpio is the first commercial-grade switch shipping on the open-standard PCIe Gen 6 stack. If hyperscalers want a credible alternative to a fully NVDA-locked rack, they need Scorpio (or its successor) to work.
The bull case: Scorpio shipped in volume in 2025 with NVDA’s Blackwell platform tier and continues into Rubin in 2026. The bear case: NVDA-only racks dominate the AI buildout through 2027, and the open-standard fabric volume stays a smaller slice than projected.
Picks-and-shovels framing
For builders thinking about durability of the AI infra layer:
- The compute layer is hyper-competitive (NVDA vs AMD vs Cerebras vs Groq vs all the custom silicon). Margin compression is the long-run story for compute.
- The interconnect IP layer is structurally less competitive. Aries-class retimers are a duopoly (Astera Labs + Marvell). Smart Cable Modules and CXL controllers are even more concentrated. Switching costs are high — once a server platform validates a specific retimer, swapping vendors costs more in re-validation than in chip price.
This is the picks-and-shovels version of the same gold-rush — compute vendors fight each other, but everyone needs the IP that connects the compute.
Practitioner note
For builders ordering or specifying infrastructure:
- If your roadmap involves non-NVDA GPUs, ALAB Scorpio is on the short list of dependencies. Validate sample availability and platform reference designs before committing to a non-NVLink scale-up architecture.
- PCIe Gen 6 timing is the variable to track. Gen 6 platforms started shipping in 2025; volume ramp is 2026. Servers ordered today for 2027 delivery will be Gen 6 native — and Aries Gen 6 retimers will be in the BOM whether the buyer notices or not.
- For non-hyperscale buyers, the ALAB product names show up indirectly via OEM platforms (Dell, HPE, Supermicro reference designs). You don’t order ALAB silicon directly; you order a platform that contains it. The spec sheets tell you which.
- CXL is the watch-list product, not the revenue driver. If pooled-memory CXL deployments mature (the persistent 2024-2026 promise), Leo becomes the second growth leg. If they don’t, Leo stays a niche line and the bull case rests on Scorpio.
The under-considered angle: ALAB and CRDO are complementary, not competitive. Astera Labs lives inside the server box (PCIe domain); Credo lives between racks (Ethernet / 800G+ domain). A hyperscaler ordering both is the norm, not the exception. Confusing them — or assuming one will eat the other — misses the structural distinction in where they sit on the cluster topology.