Business

Know the Business — How Astera Labs Actually Makes Money

Astera Labs is a single-end-market, single-supplier, fabless connectivity-silicon company whose entire revenue line depends on three hyperscalers letting it sit inside their AI servers. The technology is genuine; the moat is real but narrow; and the financial model — once you strip away non-cash compensation — looks like a software company wearing a chip company's costume.

This page does three things: (1) it dissects how Astera converts a $1,000-per-accelerator dollar-content opportunity into a 39% non-GAAP operating margin; (2) it pressure-tests the moat by walking through where competitors can break in and where they can't; and (3) it lays out the only valuation lens a serious buyer should use on a stock that trades at roughly 31x sales.

1. The thirty-second economic engine

Astera designs four families of connectivity chips, hands the masks to TSMC, ships the finished IC (or a board built around it) to a hyperscaler's contract manufacturer, and gets paid on shipment. There are no licenses, no royalties, no software subscriptions, no recurring services. Every dollar of revenue is a unit of silicon (or a small board carrying it) leaving a warehouse.

What turns this otherwise ordinary product business into a 75% gross margin engine is the design-win lock-in. A retimer that has been co-designed into a hyperscaler's custom AI accelerator board is essentially a single-source part for the 2–3 year life of that platform — qualifying a substitute would require re-spinning the customer's board, re-validating signal integrity, and re-running their internal release. The pricing power that comes from being unswappable for 24 months is what holds gross margin in the high 70s while volume grows 17× over three years.

The business does not monetize through a moat on the chip itself. The chip is a few square millimeters of silicon on TSMC 7nm/5nm. The moat is in being the company the customer trusts enough to put inside a board that will spend $500M of capex in production. That trust is built through the Interop Lab, the COSMOS software stack, and a single-vertical engineering team — none of which is replicable in a few quarters.

2. The four products — what each one is, and who threatens it

For each family: what it sells, what share Astera has, and the named competitor most likely to take revenue away.

Aries is the franchise. Of the four lines, only Aries has a defensible share lead today. Management has reported "millions" of Gen 6 ports shipped already, and industry-research sources triangulate roughly 55% share of the AI-accelerator retimer market. This is the line Broadcom is now attacking. Whether Aries holds 50%+ share through the Gen 6 → Gen 7 transition is the single most important business question for the stock.

Scorpio is the upside. Fabric switching is a structurally larger market than retimers — switches cost 5-10× a retimer ASP and there are more per rack. Scorpio P (Gen 6) is already shipping; Scorpio X (scale-up, 320 lanes) shipped in initial volumes in Q1 FY2026 and is expected to become the largest product line by the end of 2026. This is also the most contested market — Broadcom Atlas is the incumbent, and Marvell bought XConn in early 2026 specifically to compete here.

Leo and optical are convex bets, not earnings drivers. CXL memory pooling is real but adoption is slow. The Microsoft Azure M-series ramp in late 2026 is the first material commercial signal. Optical (NPO and CPO) and UALink-based switches are 2027+ stories. Underwriting any of them in a base case is a mistake.

3. How revenue actually grew 17× in three years — and why the next leg looks different

Two distinct phases sit inside the trajectory:

Phase 1 (1Q23 to 4Q24): the Aries Gen 5 ramp. Revenue grew roughly 8× across eight quarters — from $17.7M to $141.1M — as Aries retimers ramped at NVIDIA H100/H200 platforms and the first custom AI accelerator at the lead hyperscaler. Within that window the FY24 four-quarter ramp alone more than doubled revenue ($65.3M → $141.1M). Almost all of this was a single product scaling at a small number of customers. Gross margin held in the mid-70s (excluding a one-off Q1 FY23 cost spike), meaning the ramp was demand-driven, not price-cut.

Phase 2 (1Q25 to current): product diversification under sustained AI capex. Revenue still doubling YoY, but now driven by Aries Gen 6 ramps, the early-stage Scorpio fabric-switch ramp, and Torus AEC volumes. Management said in Q1 FY2026 that Scorpio will be the largest product line by year-end — replacing Aries as the lead franchise. This is the better kind of growth: less concentration on a single line, but it does mean the growth is dependent on a different set of design wins holding.

The Q1 FY2026 guide implies a 2026 revenue trajectory in the $1.45–1.65B range — roughly a doubling again. That is unusual durability for a hardware ramp; it requires Scorpio to hit a multi-hyperscaler ramp (two new hyperscalers expected to start receiving Scorpio P series in late 2026 per management) and Aries Gen 6 to keep extending.

4. The operating leverage story — and the SBC asterisk

The single biggest swing factor in the FY2025 P&L was not revenue — it was the collapse of SBC as a percentage of revenue as IPO-vesting overhangs ran off. Look at the GAAP vs non-GAAP gap closing:

Three lessons sit inside that chart:

The gross margin is the bedrock and it has not moved. From the first $80M of revenue to the most recent $850M, gross margin has not left a 69–76% band. The 70bp compression in FY2025 came entirely from a richer mix of hardware modules (cable assemblies carry the IC plus PCB content, diluting per-IC margin while raising absolute dollars). Management has been transparent that gross margin runs in the 73–76% band; Q2 FY2026 guide is 73% because of a one-time 200bp non-cash customer-warrant impact. There is no meaningful sign of pricing erosion.

Operating leverage is real, but it's a non-GAAP story. From FY2024 to FY2025, revenue grew 115% while non-GAAP opex grew at roughly half that pace — non-GAAP operating margin jumped from 30% to 39%. Q1 FY2026 R&D spend was $96M, up sharply from the FY2025 quarterly average, and yet non-GAAP operating margin was 36%, still above the FY2025 average. The fabless model is doing what fabless models are supposed to do at scale.

SBC is enormous but the worst is behind us. SBC was 59% of revenue in FY2024 (including $89M of IPO-triggered vesting), 19% in FY2025, and is now running in the mid-teens. That is still 2–3× a normal mature semiconductor, but in a range where buyback policy could plausibly offset dilution — which is what investors need before treating non-GAAP figures as the underlying earnings power. Diluted share count grew 37% in FY2025 (largely from IPO-related vesting); FY2026 dilution should be far smaller.

The right earnings frame for this business today is non-GAAP, while continuing to discount FCF for ongoing SBC at the 15–20% of revenue running rate. Non-GAAP net income of $331M in FY2025 vs. FCF of $282M and SBC of $160M is roughly the right starting frame; conservative SBC-adjusted FCF lands around $120M, and SBC-adjusted FCF yield is well under 1% at current price levels.

5. The customer book — the single biggest risk in plain English

86% of FY2025 revenue came from the top three end customers, and industry research triangulates the largest customer at ~70% of revenue. That largest customer is widely understood (though not disclosed by name in the filings) to be NVIDIA, with the rest of the top three being hyperscalers building custom AI accelerators (the leading candidates are AWS, Microsoft, and Google for Trainium, Maia, and TPU programs respectively).

Why it's bearable:

Design wins are 24-month commitments. A customer cannot just walk away — once a chip is qualified into a board, switching is expensive and risky. The realistic worst case in any given 12-month window is "loses next-generation slot", not "loses all current revenue tomorrow."

The customer base is concentrated but the right concentration. Each of the top three is also a large and growing spender. This is not the apparel industry's "you lose Walmart and you're done." Every dollar of capex Amazon, Microsoft, and NVIDIA add is partially a tailwind for ALAB.

Why it's still dangerous:

The hyperscalers are designing custom silicon for adjacent functions and could in principle internalize PCIe retimers if economics warranted. Today none has, because the volumes don't pay the design cost. That math could change if any one of them grew their internal AI accelerator volumes far beyond current levels.

A single Gen 7 retimer design loss at the largest customer would be a multi-quarter revenue event. Industry research suggests Broadcom's October 2025 Gen 6 retimer launch is already being evaluated for Gen 7 slots. The next 18 months of Gen 7 design wins are the single most important variable for the medium-term thesis.

6. Where the moat actually lives — the four pieces, ranked

Calling a business "high-quality" without naming the mechanism of the moat is the most common mistake in semis. For ALAB the moat is real but narrow: strong inside the Aries franchise, partial inside Scorpio, and barely present in Torus or Leo.

The honest summary: this is a medium-strength, generation-bounded moat. Strong within a given PCIe generation (24-month lock-in is real). Weaker across generations (every 24 months the design competition resets). Essentially absent in Torus (commodity AEC modules where Credo competes head-to-head) and untested in Leo (CXL adoption itself is the limiting factor).

A reasonable model: assume Aries Gen 6 produces $700M+ of cumulative revenue through 2027 with high certainty, then handicap the Gen 7 transition probabilistically based on the Broadcom competitive picture as it develops over the next 12 months.

7. Peer reality check — pure plays vs. diversified incumbents

This is the comparison that most matters for valuation. The lens below is economic comparability — which of these companies has economics actually similar to ALAB's, and which ones don't.

CRDO is the only legitimate full comparison. Same fabless model, same hyperscaler customer base, similar single-product-cycle exposure, similar revenue scale, similar growth rate. CRDO actually grew faster in its latest fiscal year (206% YoY) but has a less defended retimer franchise and more Ethernet AEC dependence. CRDO is the right starting point for valuation; ALAB should trade at a modest premium for the Aries franchise and Scorpio optionality, offset by higher customer concentration.

RMBS is a margin reference, not a peer. Rambus's 79.6% gross margin is the licensing-heavy model — they collect royalties on memory-interface IP that lives in DDR5 buffer chips. Their cost structure is permanently lower than ALAB's because they don't ship product. Useful for understanding the upper bound of IP-style gross margin in this space, but don't anchor ALAB valuation to RMBS multiples.

MRVL and AVGO are the realistic acquirers, not comps. Both have higher revenue scale, lower gross margins, and the strategic motivation to roll up connectivity. A reasonable counterfactual to the public story is "Astera trades at a premium because the long-run end state is a strategic sale at 20-30× sales."

MCHP is a misdirection. The 10-K names Microchip as a Leo competitor, but Microchip's 7% YoY growth, 10% non-GAAP operating margin, and analog-cycle exposure make it an irrelevant comp for ALAB. Treat it as cycle context, not valuation context.

8. The actual valuation lens — what to underwrite, what not to

At the 6/18/26 close of $417, Astera trades at a ~$75B market cap and ~$74B EV (cash and marketable securities of $1.19B, no debt). That works out to roughly ~86× FY2025 sales ($852.5M), about ~74× trailing-twelve-month sales (TTM ≈ $1.0B through Q1 FY26), and about ~48× FY2026 consensus sales (~$1.5B implied by the Q2 guide × 4). The valuation_summary table below shows ALAB at ~32× EV / Sales benchmarked off the FY2025 year-end EV (~$27.1B); that is the lens for like-for-like peer comparison, not the spot multiple a buyer pays at $417. These are software-company multiples, not semi-company multiples.

The right framework has three anchors:

Anchor 1: EV / forward sales vs CRDO. ALAB at ~30-32× trailing sales vs CRDO at ~24× is a 25-30% premium. Justify with the Aries franchise lead + Scorpio fabric-switch optionality + larger absolute cash balance, offset by higher customer concentration. The right zone is 1.0-1.3× the CRDO multiple. Anything outside that zone deserves a specific reason.

Anchor 2: SBC-adjusted FCF yield walk. Today's SBC-adjusted FCF yield is well under 1% — unambiguously expensive on a current-earnings basis. The yield would need to converge to 1.5-2% on consensus 2027 estimates (≈$2.0-2.5B revenue, 40%+ non-GAAP op margin, $700-900M FCF, SBC normalizing to 12-15% of revenue) to earn its multiple on a present-value basis. The investor is paying for the convergence path, not for present earnings.

Anchor 3: Stress test against a 25% hyperscaler-capex pause. The right downside scenario is not "all hyperscalers pull back 50%" — it is "AI capex grows at 15% in 2027 instead of 35%". In that world, ALAB revenue still grows but at 30-40% rather than 60-80%, and the multiple compresses sharply (likely toward 12-15× sales). The stock could lose 50% in a single quarter without the business deteriorating in any structural sense. This is the asymmetric tail risk that justifies sizing the position carefully even for buyers who believe in the long-term story.

9. The handful of metrics that actually move the stock

Only six numbers predict the stock's behavior over 6–18 month windows. The rest is noise.

10. The investor's question — what is this business worth, and who should own it?

The business is high-quality, the customer book is not. Genuine technology, real moat inside the Aries franchise, fabless economics that look like software at scale, a balance sheet with more than $1.2B of cash and no debt. Counter-weighted by the most concentrated end-customer book of any major semiconductor name (86% top-3, ~70% top-1) and a Gen 7 competitive transition that will be litigated over the next 12-24 months by Broadcom's October 2025 entry. A high-quality business with a high-variance customer book deserves a premium multiple but not the highest premium multiple in semis.

The right comparison is CRDO, not the diversified incumbents. Both are pure-plays sized into a structurally large but narrow market. ALAB should trade at 1.0-1.3× CRDO's EV/sales multiple, with the premium tied specifically to (a) the Aries Gen 6 share lead, (b) the dollar-content per accelerator story ($1,000+ and rising per management) and (c) the Scorpio fabric-switch optionality. Anything above 1.4× CRDO is paying for things that have not yet been earned.

The right owner is an investor underwriting an AI-capex multi-year cycle, not a quality compounder. This is not yet a long-duration cash-flow story; SBC-adjusted FCF yield is below 1% and even consensus 2027 estimates leave the yield under 2%. It is a growth story whose terminal value depends entirely on the AI capex cycle remaining structural and on the company continuing to execute generation transitions cleanly. For an investor who is comfortable underwriting the AI capex tailwind through 2028 as the foundational view, ALAB is the most direct pure-play available. For an investor who is unsure whether AI capex compounds at 30% or 15% in 2027, this is the last semi name to own — it is the most sensitive to that single macro variable.

The stock will be expensive on every multiple anyone calculates. The question is whether the rate of compounding inside the business outruns the rate of multiple compression. The single biggest source of asymmetric downside is customer concentration, and that risk does not show up in any of the financial statements until the quarter it actually arrives.