Astera Labs

Astera Labs (ALAB) trades on the Nasdaq. It does not sell GPUs and does not rent cloud capacity — it sells the wiring between them. Aries retimers clean up degrading signals carrying PCIe and CXL traffic over the length of a server board, and Scorpio fabric switches route those signals across a rack of dozens of wired-together GPUs. As racks grow from eight GPUs toward Nvidia’s 72-GPU NVL72 class and beyond, every added GPU needs more of this silicon, not less — a toll collected inside the rack rather than a bet on any single chip architecture winning outright.

The company started in 2017 in Santa Clara, when three Texas Instruments veterans quit to build connectivity chips out of a co-founder’s garage, betting that data-center interconnect hadn’t kept pace with the GPU clusters forming around it. See The AI Capex Supercycle for the spending cycle Astera sits inside, and Nvidia and CoreWeave for the chipmaker and the neocloud on either side of the socket it fills.

Almost no independent coverage treats Astera Labs as anything but an IPO-week headline. A company whose revenue has grown roughly sevenfold since its debut, while the record on it goes quiet, is exactly the gap this report exists to close.

TEXXR’s corpus catches Astera Labs early, before AI infrastructure was the story. A $50 million Series C led by Fidelity in September 2021 valued the fabless startup at $950 million, and a $150 million Series D that November more than tripled that to $3.15 billion — both well ahead of the public AI-infrastructure boom. The IPO filing, in February 2024, disclosed 2023 revenue of $115.8 million, up 44.9% year over year, with the net loss narrowing to $26.3 million. By March 8, the company said it would seek up to $534 million; by March 20 it priced above range, at $36 a share, for roughly $713 million at a $5.5 billion valuation. TechCrunch called it a referendum on how much investors wanted in on AI, a full year before CoreWeave ran its own version of the same test. The market’s answer arrived fast: shares jumped 72% on debut day, closing at $62.03 for a market value near $9.46 billion.

What TEXXR’s corpus does not carry is everything since. Revenue kept compounding — reaching $396.3 million in 2024, then $852.5 million in 2025, with net income turning positive at $219.1 million — as the product line widened past retimers. Scorpio P, the industry’s first PCIe 6 fabric switch, reached volume production in 2025, and by that September PCIe Gen6 products already exceeded 20% of revenue, with the initial attach running mostly to Nvidia’s NVL72 racks and custom accelerator systems. Nvidia extended the relationship again that May, naming Astera a collaborator on NVLink Fusion — connectivity for third-party accelerators that want into Nvidia’s own scale-up fabric, with revenue not expected until fiscal 2027. One customer relationship went further than a design win: in February 2026, Amazon took a warrant for shares — up to 3.26 million of them — vesting against as much as $6.5 billion in future purchases of its switches, retimers, and optical products, the same pay-to-be-a-customer structure recurring across the capex cycle’s financing loops. By Q1 2026, revenue reached $308.4 million, up 93% year over year, and CEO Jitendra Mohan told investors Scorpio X — the newer scale-up switch — was shipping initial volumes, ramping through the second half of 2026, alongside a CXL memory controller running in private beta with Microsoft Azure. Benzinga’s shorthand for the customer base — Nvidia, AMD, Amazon, and Google — names exactly the four relationships that do not appear as a single edge anywhere in TEXXR’s graph (see The Graph, below). None of it — the Amazon warrant, the Azure beta, the NVLink Fusion deal — has produced a TEXXR article or edge.

The dependency on Nvidia and AMD’s own roadmaps cuts both ways. On AMD’s side, the MI400 “Helios” rack due in the second half of 2026 will use Broadcom switches instead of native UALink silicon for scale-up connectivity, because neither Astera Labs nor Marvell will have UALink switches ready in time — a design win Astera does not yet hold on the rack architecture built as AMD’s answer to NVL72.

TEXXR’s knowledge graph carries exactly nine hyperedges touching Astera Labs across its full history — eight financial, one a single product-launch mention, zero partnership and zero acquisition edges. Querying the graph directly for edges between Astera Labs and Nvidia returns zero. Between Astera Labs and AMD: zero. Between Astera Labs and Amazon: zero. That is the entire record of a company whose 2025 10-K disclosed one customer above 70% of revenue, and its top three customers accounting for roughly 86% — the graph has no trace of the relationship that is, by the company’s own filing, most of its income statement.

TEXXR’s own volume comparison sharpens the gap further. Marvell, competing in the identical PCIe/CXL/optical connectivity market, carries at least 32 tagged articles in the corpus, including one Jensen Huang endorsement that sent its stock up 32.52% in a single session. Astera Labs, arguably the purer-play name in the same niche, carries six — clustered around its IPO week, with nothing since.

Positioning: how much of ALAB is sold short, on the record brokers file twice a month. It lags by design — read it as where the crowd stood, not where it stands.

Order flow is signed before it becomes revenue — the earliest of the three views. From the capital-flow ledger.

• Whether a partnership edge ever appears. Nine edges, all financial or launch, none connecting Astera to the customers that generate its revenue — a first partnership edge would mark the record catching up to the filing.
• The Scorpio X ramp. Management’s own bet is that the scale-up switch, ramping through the second half of 2026, becomes a bigger revenue line than the retimer business that built the company.
• The UALink-to-Ethernet gap on AMD’s Helios rack. AMD is shipping its 2026 flagship rack on Broadcom’s Ethernet silicon because Astera’s own UALink switches aren’t ready — a concrete near-term miss on the design win that matters most outside Nvidia.
• NVLink Fusion revenue, not expected before fiscal 2027. A bet that non-Nvidia accelerators will want into Nvidia’s own fabric, tracked here as a specific date rather than a slogan.
• The Amazon warrant’s vesting schedule. Warrant shares vest against actual purchases, up to $6.5 billion through 2033 — a forward revenue signal that either materializes in filings or doesn’t.
• Customer concentration in the next 10-Q. One customer above 70% of revenue, three above 86%, is the number to hold against every future quarter — not the design-win press releases layered on top of it.

Astera Labs lives or dies by the same constraint as these names — small equal-weighted groups, one constraint each. Not investment products: see all categories.