ALAB

The short version. Astera Labs has gone from $116M of revenue in 2023 to a company guiding $540-560M for a single quarter (Q3 2026), on the back of AI server build-outs. We think its business is worth about $149 a share today. Three things decide whether that is too high or too low:

1. How long the step-up lasts after 2027. We hold Astera's share of AI accelerator spending (0.47% in 2026) through 2027 and let it rise to 0.65% by 2031 as new switch customers, a new open standard for linking accelerators (UALink) and optical products arrive. That gives revenue of $3.5B in 2027 and $5.8B in 2031.
2. One customer. In 2025 one end customer bought more than 70% of what Astera sold (10-K 2025). The filings do not name it. If its switch program slips or is shared with a second supplier, 2027 can be a fifth lower, worth -$28 a share.
3. How much of the profit owners keep after stock pay. Stock-based pay is about half of Astera's operating cost (48% in the Q3 2026 guide). We charge it in full as a cost, and count the awards already granted in the share count once.

Avoid: at $261.24 Astera trades 75% above our $149 fair value and close to our optimistic case ($288). The price needs revenue growth after 2027 about 19 points a year faster than ours, or the Street's own revenue and margins plus a much lower required return than we use (section 5). Our blind value, committed before we saw the price, was $138; the model audit found that our cost assumption fell behind our own Q3 forecast and that three tax assets were missing, together +$11 (section 9).

Terms. The central case is our best estimate; the pessimistic and optimistic cases are coherent alternative outcomes we weight 25% each. Non-GAAP figures are the company's own, before stock pay and a few other items; GAAP figures follow US accounting rules. EBITA is operating profit before amortization of intangible assets bought in acquisitions. Accelerator spend is the revenue of the AI chip makers (NVIDIA's data center line plus Broadcom's AI line), our sector note's measure of AI hardware spending.

1. What Astera sells and what limits it

Astera sells four families of connectivity products, all used inside AI servers and racks (10-Q Q2 2026; calls):

• Aries retimers clean up PCIe signals so that accelerators, processors and network cards can sit farther apart. This was the founding product.
• Scorpio switches route traffic among many chips. The P-Series links accelerators to processors and network cards ("scale-out"); the X-Series links accelerators to each other inside one rack ("scale-up"). Management says Scorpio becomes the largest product family from Q3 2026 and that X-Series alone carries $1,000 or more of content per accelerator (management claim, Q2 2026 call 2026-08-04 and Citi conference 2026-09-09).
• Taurus Ethernet cable modules and Leo CXL memory controllers, smaller today; Leo is expected at two large cloud customers in 2027 (management claim, Q2 2026 call).

Astera designs chips and buys manufacturing from one foundry (10-K 2025). Gross margin was 73.7% (non-GAAP) in Q2 2026 and is guided to about 72% in Q3, falling toward management's long-term 70% target as switches and modules, which carry lower margins, grow in the mix (management claim, Q2 2026 call).

Who pays. Three layers, which the filings keep separate:

The filings do not say whether Amazon is the end customer above 70%; we do not assume it. Billing addresses are China 39%, Singapore 29% and Taiwan 27% of Q2 2026 revenue, which are the contract manufacturers' locations; management says Chinese end demand is in the single digits (management claim, Q2 2026 call).

What limits it. (a) The pace of AI hardware spending at a handful of cloud companies. (b) Broadcom, which sells PCIe switches and retimers and is many times larger; management names it as the main switch competitor (Citi 2026-09-09). (c) Accelerator makers building the link in-house (NVIDIA's NVLink is the model for scale-up). (d) A single foundry.

2. Recent results against the bar

Astera has beaten the midpoint of its own revenue guidance in each of the last eight quarters, by +4.2% to +16.0% (median +9.6%), and none came in below the midpoint (company releases, Q3 2024 to Q2 2026; preview.txt). Q2 2026 revenue was $392.4M against a $355-365M guide; non-GAAP gross margin 73.7% against about 73%; non-GAAP operating margin 39.1% (release 2026-08-04). Operating cost has run above its guide in six of the eight quarters, by up to +10.4%, so the profit beats are smaller than the revenue beats.

The Q3 2026 guide is the largest step Astera has given: revenue $540-560M, 40% above Q2 (earlier guides stepped 8-27%), non-GAAP gross margin about 72%, operating cost $156-160M, non-GAAP diluted EPS $1.16-1.21 and GAAP EPS $0.87-0.92 (release 2026-08-04).

3. Our forecast and value

The model is model.py on the shared method (output.txt is its saved run); every input, with its source, is in assumptions.md.

• 2026 is the reported first half ($700.8M) plus our Q3 ($603M) and Q4 ($727M) from the guidance record (section 7).
• 2027 holds Astera's 2026 share of accelerator spend (0.47%) while that spend grows 72% on our sector note's proposed path (semis-ai 5.3, 2026-10-15).
• 2028-2031 lift that share to 0.65%. Our sector note has accelerator spend growing only 8%, 0%, 4% and 6% in those years, so most of our growth is Astera taking share: new switch customers, UALink switches in 2027, optical links from late 2027 and CXL at two cloud customers (management claims, Q2 2026 call). In 2030 our central is 21% of the $25B market management says Astera can serve (Q4 2025 call); Astera had about 34% of its 2025 market.
• Margins follow management: gross margin at the 70% target from 2028 (we do not let it sit above the target), and operating cost at our Q3 forecast's 27.2% of revenue (the guide's $158M plus Astera's usual overrun, on our $603M) with no further leverage. Management has given no operating-margin target that we found.
• Stock pay is charged as a cost: 48% of non-GAAP operating cost in 2027, falling to 40% by 2031 (51% in 2025, 48% in the Q3 guide). The awards already granted (restricted units, performance units and options, 11.2M shares) are in the 186.4M share count, so the $680.9M of expense still to be booked on them (10-Q note 10) is taken out of future charges ($151M, $312M, $161M and $57M in H2 2026-2029) rather than counted twice.
• Amazon's warrants: the 1.69M shares vested or probable are in the count. The 3.89M not yet vested are left out because their cost already lowers revenue (about 2 points of gross margin, Q1 2026 call); counting them as well is worth -$3.05.
• Cash and tax assets: $1,253M of cash and securities, a $60M tax refund due (10-Q), the federal tax losses ($270M, worth $57M at 21%) and $109M of federal research credits (10-K 2025), less $7M of acquisition holdbacks and $150M kept for operations, is $1,323M ($7.09 a share). Leases are not deducted because rent stays in the operating cost. No debt, no dividend, no buyback; free cash is kept and earns the risk-free rate.

Cost of equity. Astera listed in March 2024, so its own share-price history is about two and a half years long and noisy (raw beta 3.69). We use the median of five connectivity and accelerator peers instead (raw 2.22, adjusted toward 1 to 1.815; FMP 2026-10-23), giving 12.91%. On Astera's own adjusted beta (2.80) the cost of equity is 17.1% and the value $118 (-$31); on the US semiconductor industry's beta of 1.52 (Damodaran, January 2026, secondhand via the advocate review) it is about $163.

The exit. The five-year exit is the average of two estimates that disagree sharply: letting Astera's high returns fade toward its cost of capital over 15 years gives $137; the multiple peers have historically traded on gives $397. They are far apart because peers' historical multiples (36.5x operating profit before acquisition amortization) imply a much lower required return, or much longer high growth, than our 12.91% with returns fading to 25%. We take the method's average rather than pick one. The peer multiple is one change from the method's default: the default divides peers' value by GAAP operating profit, which for Broadcom and Rambus is after billions in acquisition amortization, while Astera's operating profit has almost none. On a like basis the multiple is 36.5x, not 43.3x (anchor_ebita.txt, SEC filings); using the default would add $18.

Sensitivities (central fair value $149, one change at a time; output.txt):

The simulated median in odds.json ($126) is 16% below the central. A quarter of the weight sits on a bear-leaning world whose median ($71) is further below the central than the bull-leaning world's ($186) is above it, and inside the base world the ranges for 2027 revenue and later growth reach further down than up (base-world median $139).

4. Market expectations and how the price reacted

Weekly closes (Robinhood, regular session): $112-120 through March 2026; a rise to $417 in the week of 2026-06-15 (intraday high $499 in the week of 2026-06-29); $292-334 through August and September; $342 in the week of 2026-10-05 after an intraday $402; $286 in the week of 2026-10-12. On 2026-10-22 the official close was $257.82, and at 10:18 ET on 2026-10-23 the last trade was $261.24 (SOXX $476.71). The company's market value at the price is about $45B (Alpha's data, 2026-10-23).

Running our model backwards (odds.json), the price is justified by revenue growth in 2028-2031 about 19 points a year faster than ours, which reaches about $10.8B in 2031: 1.2% of our sector note's accelerator spend, against 0.47% today and our optimistic case's 1.1%; in 2030 it is $8.4B, about 33% of management's $25B served market (Astera had about 34% of its 2025 market). No other single driver gets there inside its bracket.

Alpha, our momentum model, ranks Astera 13th of about 700 on 2026-10-23: growth A, momentum A, quality A, expectations B, valuation E (670th). It ranks the business's growth and the stock's trend; we value the business. They differ because the price has run far ahead of the value we find, which Alpha's own valuation grade also says.

Timing, a quick chart read: after a fivefold rise from the April low ($98) to the late-June high ($499), the stock has made lower highs ($402 in early October) and is back near the August-September lows of $251-266; below them the late-July low of $245 is the next level. Grade D: a broken uptrend sitting on support. Timing is beside the view, never an input to it.

5. Where we differ from the Street

The Street is below us for 2027-2028 and well above us after it. Its averages (FMP, 2026-10-23) are revenue of $1.9B, $3.1B, $4.0B, $6.4B and $7.1B for 2026-2030 (ours $2.03B, $3.49B, $4.37B, $4.80B, $5.28B) and non-GAAP EPS of $3.99, $6.49, $8.26, $14.60 and $16.80. The 2030 revenue figure averages 15 analysts (low $6.0B, high $8.4B; secondhand, as relayed by the advocate review): our $5.28B is below every one of them. Run through our model (street_check.py, written by the model auditor):

So about $68 of the gap to the Street is forecast (revenue after 2028 and cost leverage we do not credit) and about $104 is how the exit is valued: our method averages the peer multiple with a 15-year fade at 12.91%, and the Street's targets do not. On the method's averaged exit the Street's path reaches its own target only at a 8.0% cost of equity. In 2027-2028 the Street's implied margins are at or below ours; the disagreement is about 2029-2031.

What would settle it: 2029-2030 revenue, which shows first in how many customers take Scorpio X switches during 2027 and in the concentration disclosures; and the cost ratio over the next two quarters (25% of revenue with revenue above $3.5B would move us to about $164).

6. Why the other side could be right

The pessimistic case ($47) is a world where AI spending pauses from 2028, as in our sector note's pessimistic shape, and the lead customer shares its switch program with a second supplier: 2027 growth of 35%, then -10% and -15%, gross margin 66% under price pressure, cost cut but stock pay still near half of it. Nothing in the filings rules this out: the concentration is extreme, Broadcom is a capable incumbent, and a single program's timing has already driven a 40% quarterly step.

The optimistic case ($288) has switches broadening to several cloud customers, UALink and optical arriving on time, and operating cost falling to 25% of revenue (a 45% operating margin, the level Marvell targets for fiscal 2031). Revenue reaches $9.7B in 2031, about 30% of management's served market. Astera's record of beating its own guide every quarter is the best evidence for it.

Against our central from below: our growth after 2027 is mostly share gain, and share gains in a customer base this concentrated come in lumps. From above, the advocate review's Buy world: Credo, a smaller connectivity peer, already runs non-GAAP operating cost at 19.9% of revenue ($95.2M on $479.0M, release 2026-09-01), while we hold Astera at 27.2% through 2031; and every one of the 15 analysts covering 2030 is above our revenue. With 2028-2030 growth of +40%, +25% and +15% (2030 at 28% of management's served market) and Credo-like cost, our own model gives $259; at the industry beta $283. That world needs no change in the exit. We do not adopt it: cost at a peer's ratio is not Astera's evidence, and the extra growth rests on management's market claim and on Scorpio X customers not yet named in any filing. The one fact that would flip the call: 2030 revenue on course for about $6.1B or more (25% of the served market), shown during 2027 by Scorpio X in production at several customers beyond the lead.

7. Next results: Q3 2026, on 2026-11-03 after the market close (confirmed by the company on 2026-10-06)

Each forecast starts from the company's guide and adds Astera's median gap to its guide over the last eight quarters (preview.py). The ranges are wider than that record alone, because our earlier ranges held only 12 of 29 outcomes, and because this quarter is the largest step Astera has guided, driven by one product's first volume quarter.

What a surprise is worth: our model carries Q4 2026 revenue at $727M. A Q3 that changes our 2027 view matters more than the quarter itself: each $100M of 2027 revenue above our $3,493M, kept through 2031, is worth about +$3.90 a share.

8. Dates to watch, and what would change the view

• 2026-11-03, Q3 2026 results. Revenue at or below the $550M midpoint, or a Q4 guide below $600M, would be the first break in the guidance record and would point toward the pessimistic 2027.
• 2027, UALink switches and CXL ramps (management claims). A slip moves 2028 growth toward our pessimistic path.
• Customer concentration in the 2026 10-K (February 2027). The top end customer's share falling below 60% with revenue still growing would cut the main risk.
• Gross margin below 70% before 2028 would mean the mix shift is faster or pricing weaker than management's target allows.

9. What changed from the blind value, and the review

The blind value (commit 66154e9, pushed before we looked at the price) was $138, pessimistic $46, optimistic $287. After the price, two fresh reviewers worked on draft cf57c4537: a model auditor (the reviewer role) and an opposite-side advocate (the architect role), both Claude Opus 5.5, run one after the other without seeing each other's findings. Each numerical finding was rerun here; review_bridge.py reproduces the accepted changes step by step.

Net: $138 to $149 (+$10.87), the call unchanged (Avoid). The two changes that moved the value were found by the auditor on internal consistency and filings, and both raise the value. Flags answered: 3 (value 43% below the price): no risk is charged twice beyond finding 5; the gap is growth after 2028, cost leverage and the averaged exit. 4 (fade $137 against outside view $397): the 12.91% discount rate inside the fade; the fade equals the outside view only at about 6%. 7 (own beta $118): disclosed; the market's gap is not a beta gap, as our central reaches the price only at a 6.75% cost of equity (advocate). 17 (0.46x the discounted Street target): section 5. Not checked by either reviewer: the sector note's accelerator path, the outside-view EV data, and the Citi conference transcript beyond press summaries.

10. Sources

Primary, all saved in data/ and masked for prices before reading: 10-K 2025 (filed 2026-02-20); 10-Q Q2 2026 (filed 2026-08-05); earnings releases Q1 2024 to Q2 2026; 8-Ks of 2024-12-16, 2025-10-22, 2026-01-20, 2026-02-10 (two) and 2026-06-04. Secondhand: call transcripts Q4 2025 (2026-02-10), Q1 2026 (2026-05-05), Q2 2026 (2026-08-04) and the Citi conference (2026-09-09), Quartr via TradingView. FMP statements and peers' betas, fetched 2026-10-23; SEC XBRL amortization for AVGO, MRVL, NVDA, CRDO and RMBS, fetched 2026-10-23. Sector note docs/research/sectors/semis-ai/note.md (5.3, 2026-10-15).