Astera Labs, Inc. Common Stock ALAB

Key data

The business

Astera Labs sells the connective tissue inside an AI server. When a GPU talks to a CPU, to memory, or to another GPU across a rack, the electrical signal degrades over copper as data rates climb to PCIe Gen5 and Gen6. Astera's core product, the Aries retimer, is a small chip that cleans and restores that signal so the link holds. Around it sit Scorpio fabric switches for stitching GPU clusters together, Taurus for Ethernet cabling, and Leo for CXL memory pooling, all managed by one software layer called COSMOS. The customers are the hyperscalers and the AI accelerator platforms, and Astera's content rides inside the reference designs those platforms ship. The moat is placement: once a retimer is designed into a hyperscaler's board and validated with its software, it is expensive and slow to rip out mid-platform. The thing a data-center engineer actually holds is a postage-stamp chip that costs a few dollars and decides whether a million-dollar rack boots clean.

The numbers

The shape is a business that inflected hard in 2025 and is still accelerating. Revenue grew 104.5% year over year in the June quarter and 27.3% sequentially, the fastest sequential step in over a year, which for a component maker means the platforms it is designed into are ramping into volume, not just sampling.

The reported bottom line is the number to distrust. Net income lurches quarter to quarter on taxes, not operations: the June quarter carried a roughly $50M tax benefit tied to a valuation-allowance position and excess benefits on equity compensation, the same tax dynamic flagged a quarter ago, and that watch-item held rather than washing out, flattering net income by about a third. Operating income is the honest read, and it climbed to a record without any tax help. On earnings taxed at a normal rate the multiple is far higher than the headline 153x.

Across full years the inflection is unmistakable: three straight years of losses through 2024, then a swing to real profit in 2025 as revenue more than doubled, and first-half 2026 already near the whole of last year. Gross margin sits near 75% but slipped to 73.3% in June as the mix shifted toward hardware-heavier products, and that matters because a connectivity supplier's entire case rests on staying silicon-margin, not box-margin. Free cash flow ran about $277M over the last year against no debt and roughly $7 a share in cash. What the market does not fully price, in my read, is how fast a large customer could integrate or dual-source the retimer, and the print that settles it is the next platform's design-win disclosure.

Management

The founders run it and their pay is not the tell, their selling is. Co-founder CEO Jitendra Mohan took $2.2M in 2025, about 1% of net income and almost all cash, a restrained package after the $50.9M founder stock grants that landed at the 2024 IPO. Against that, insiders sold about $123M over the last twelve months across 99 transactions with zero open-market buys, including co-founder and president Sanjay Gajendra selling $6.0M in August and director Manuel Alba selling $27M on a single September day. That is only 0.23% of market cap and much of post-lockup founder selling runs on pre-set 10b5-1 plans, but the filings reviewed this run do not cleanly separate planned from discretionary, so the honest read is heavy one-way selling with the plan detail unresolved. There are no buybacks and no dividend, which is correct at this multiple, and the only capital deployed was a small unnamed acquisition.

How it fails or surprises you

Customer designs it out. Revenue leans on a handful of hyperscalers and accelerator platforms. If one integrates the retimer function onto its own silicon or dual-sources to Broadcom, Marvell, or Credo on the next platform, a large slug of revenue leaves at once, not gradually. The print that reveals it first is a soft sequential guide or a named large-customer share decline in the 10-Q.

Rack content compounds (right tail). As racks move to GB300 and Rubin-class systems, Astera's dollars per rack can multiply beyond retimers into Scorpio fabric switching and CXL memory. If Scorpio ramps into volume, revenue per rack steps up faster than models assume. The print is fabric and switching emerging as a disclosed line growing sequentially, and the market is not paying for it today.

The margin-and-tax mirage. The single fact the bull case explains least is that the 73.3% gross margin dip and the tax-inflated net income arrived together. If mix keeps pressing gross margin toward the low-70s while the tax tailwind reverses, the reported 30% net margin evaporates and true earnings power sits well below the headline. That is the number that would prove the story wrong.

Closing thoughts

This is mostly a distribution the market already prices. At 150 times trailing and about 100 times forward earnings, the tape assumes years of hyperscaler-funded growth and durable socket share, so the edge is not in spotting the growth, it is in judging whether content-per-rack and design-in position survive the next platform transition. Part of that resolves on a named print, the next round of design-win disclosures, and part of it is raw exposure to an AI capex cycle no filing can resolve. My judgment, labeled as judgment: at this price the left tail is the fatter one near term, because any wobble in sequential growth or gross margin re-rates a 100x-plus multiple violently, while the right-tail content expansion is real but slower to show up. What is at risk if a customer in-sources is most of the premium, against an upside that needs several quarters of fabric-switch volume to earn its keep.

The bet is still that the AI data-center builders keep buying Astera's signal chips and keep designing them in as the processors get faster. What breaks it is a large customer moving the function in-house or the capex cycle cooling, and the one pair of numbers that tells you first is sequential revenue growth decelerating while gross margin slides below the low-70s at the same time. If those two turn together for two quarters, the read is wrong regardless of the headline.