Astera Labs
This is the archived full research. Read the story
The wiring inside an AI computer: chips that carry data between the AI chips in a rack, so a signal still arrives clean at today's speeds.
The scores
Growth3/4 ↑
98.5% growth is organic and PCIe 6 is broad-based, but the guided Q3 step narrows to one lead customer's Scorpio ramp.
Business quality3/4 ↓
Advantaged, not protected: the signal-rebuilding part is held by a physics limit and months of testing; the switch is held by nothing but being first.
Financial position4/4 →
Net cash, no debt of any kind, no financing need, positive free cash flow, and acquisitions paid in cash out of operations.
Execution record3/4 →
Beat its own guidance at every print since the 2024 listing, but ten quarters is too short a record to earn the level above.
Per share capture3/4 →
Diluted count rose 2.33% against 98.5% revenue growth and deals are cash, but 3.0% of the company is written to one customer.
Potential2/4 →
The upside is large and nothing is committed: no volume, term or backlog anywhere, so it rests on somebody else's spending decision.
Risk profile2/4 ↓
Four direct buyers are 82% of sales, their identity is not established by the disclosure, and one foundry partner makes every integrated circuit.
Valuation2/4 ↑
About 84.9 times operating profit, 154% of a three-stage value near $124 at an 8% return: full, beyond the 20% margin.
Evidence quality3/4 →
Everything central is filed and current, but the prize and the content-per-accelerator figure rest on management claims graded rather than verified.
If it works 0.9× · if it fails you keep about 36%.
The Verdict
We rate Astera Labs Unattractive. It sells the chips that carry data between AI chips inside a computer rack, and sales have more than doubled in a year, so the business is solid. But four buyers take most sales, a large customer builds its own switch and the margin is heading down, while the price sits above our value even on the analysts' steep path. That path needs its new switch to sell widely; so far one customer takes it in volume. As a bet it is priced for success: the older chips cover about a third of the price. More volume customers would change the view.
The long caseIf hyperscalers keep building their own accelerator racks, each one needs a switch joining the chips, and Astera's is one of the few sold to all comers.
How we weighed it
The price against our value
Against it
Price at 154% of our value
The price sits above our value on a steep path; the bear case is far below.
The business
In its favour
Solid
A solid business whose sales more than doubled in a year.
Growth against what the price assumes
Cuts both ways
Growth is fast, but whether a second customer adopts the new switch is unanswered.
What could go wrong
Against it
Customers build their own, four buyers take most sales and the margin is heading down.
Against competitors
Against it
Both rivals its own filing names cost less for each dollar of profit.
How its group is behaving
Cuts both ways
AI chip names move together on spending news.
The latest news and report
In its favour
The last report met the sales and margin bars we set.
The evidence under this - 11 claims
- The step depends on one lead platform. Management said it expects to ship Scorpio X to additional customers 'by year-end' and that those programmes are in pre-production or qualification, so the guided quarter is one customer's ramp.
- Astera's own risk list in the same release names the absence of long-term commitments from customers; there is no backlog or minimum purchase anywhere in the filings to convert.
- The company has beaten its own revenue guidance at every print since listing, which cuts both ways: it is evidence the guide is conservative and evidence that the guide is not the ceiling being priced.
- 2026-11-03editedthe guided quarter reported: $584.9M, 4.4% above the top of the range
- Nothing filed carries a content-per-accelerator figure, a unit count or a design-win count. The number exists only on a call and cannot be checked against any statement.
- 'Engaged' is the company's own word and covers pre-production shipment, qualification and evaluation alike; management separately said only the lead customer reaches high-volume production in Q3 2026.
- Content per accelerator is a price times an attach rate, and the attach rate is decided by the platform designer rather than by Astera - inside an NVIDIA NVL72 rack it is zero.
- 2026-11-03editedthe Q3 release carries no customer table or content-per-accelerator figure; the Q3 10-Q and call, not yet read, are where they land, so the shelf life is extended to them
- Annualising one guided quarter of a business growing 40% sequentially flatters the multiple if the ramp continues and overstates it if the ramp is a step rather than a trend. Both readings are printed rather than one chosen.
- The pack's own enterprise value nets only cash and cash equivalents and misses $1,141.5M of marketable securities, so its EV/EBIT of 165.7 is struck on a larger enterprise value than the balance sheet supports. The filed figure is used here and the divergence is recorded in data_notes.
- Measured on non-GAAP operating profit, which is what the sell side models, the same price is about 30 times the annualised guided figure. The gap between the two is stock pay, and which one is right is the question the frame answers rather than assumes.
- 2026-09-25editedre-struck on the 2026-09-24 close after a 33.9% price increase since the thesis; the operating-profit figures are unchanged
- 2026-11-03editedkept as the multiple at the thesis price; C56 re-strikes it on the Q3 quarter
- No product-family revenue is filed (C43), so whether Scorpio becomes the largest family can only be checked against what management says on the next call.
- Four direct customers were 82% of Q2 2026 revenue (C10), and the guided Q3 step is attributed to the lead customer's ramp (C05), so 'additional customers' carries no volume or date beyond year-end.
- Astera files no product-family revenue (C43), so how much of the step is Scorpio X, and how much is Aries, Taurus and the PCIe Scorpio P line also growing, cannot be checked.
- The step is one lead customer's ramp (C05); a single platform can stop as quickly as it started.
- 2026-11-03editedthe Q3 release carries no customer table or content-per-accelerator figure; the Q3 10-Q and call, not yet read, are where they land, so the shelf life is extended to them
The company
An AI computer is a rack of dozens of AI chips passing data to each other constantly. At today's speeds the industry's own standard lets a signal cross only about 500 millimetres of circuit board before it must be rebuilt.
Astera sells the parts that do that: signal rebuilders, cable modules, memory controllers and, newest, switches that join many AI chips together. It matters: in one long Meta training run, 8.4% of the interruptions were traced to switches and cables.
The evidence under this - 4 claims
- Nothing outside the company describes COSMOS at all - no customer, no standards body, no benchmark. It is the least externally corroborated claim in the moat.
- Broadcom put hardware in-network collectives into Tomahawk Ultra and contributed its Scale-Up Ethernet framework to the Open Compute Project, so the capability management presents as differentiating is also being standardised by a rival.
How it makes money
It sells each part once, to the factories and cloud companies building the racks, and keeps 73.3 cents of every sales dollar after the cost of making it. Nothing is under contract: four buyers were 82% of sales in the June quarter, and some are factories buying on somebody else's behalf.
The evidence under this - 3 claims
- The same guidance shape preceded the Q4 2025 print and proved too conservative: gross margin was guided to about 74% for Q1 2026 and printed 76.26%.
- Management also says Scorpio's average margin across use cases is 'broadly in range with our corporate gross margins', which argues the mix effect is smaller than the guide implies.
- 2026-11-03editedthe guided quarter reported: 72.3% GAAP gross margin
- 2026-11-03editedthe Q3 release carries no customer table or content-per-accelerator figure; the Q3 10-Q and call, not yet read, are where they land, so the shelf life is extended to them
Where the money comes from, stream by stream
| Stream | Share of sales (Q2 2026) | How it earns it |
|---|---|---|
| Scorpio smart fabric switches | not disclosed | PCIe and scale-up switch chips joining accelerators inside AI racks; management expects this to be the largest family from Q3 2026. |
| Aries retimers | not disclosed | Chips that rebuild PCIe and CXL signals over board and cable distances; a record quarter in Q2 2026. |
| Taurus smart cable modules | not disclosed | Modules inside active Ethernet cables between switches and servers, sold through cable makers. |
| Leo CXL memory controllers | not disclosed | Controllers that let servers attach extra memory over CXL; volume to two hyperscalers is expected in 2027. |
| Engineering services | 0% | Paid design work for customers, $3.2M in FY2024 and nothing in FY2025. |
Who runs it
The two co-founders still run Astera: Jitendra Mohan has been chief executive since 2017, and Sanjay Gajendra runs products and customers as president. The finance chief, Desmond Lynch, joined from Rambus in March 2026. Insiders own about a tenth of the company, mostly the two founders, and they have sold shares but never bought. The yearly bonus rides on sales and a profit margin that leaves out pay in shares; their shares vest with time, not results, until performance shares start this year.
The evidence under this - 4 claims
- The bonus's margin measure excludes stock-based compensation, which was 16.2% of revenue over the four quarters to June 2026 (C07), so the largest cost the shareholder bears is outside the plan.
- Until PSUs start, the co-founders' equity vests with time alone; nothing in the 2025 plan pays on GAAP profit or per-share results.
Why it wins, and who could beat it
Why customers come back
A part is tested for months against one specific processor in one specific machine. After that nobody swaps it mid-generation, and the next generation starts from the part already there.
Amazon shows it from outside: it took the right to buy Astera shares in 2022, 2023 and February 2026, each vesting only as it keeps buying. And Astera keeps more of each sales dollar than the rivals the filings let us compare.
The evidence under this - 3 claims
The threat
The switch is the part others already make. Inside NVIDIA's own 72-chip rack, NVIDIA's switch joins the chips, so there is nothing to win. Amazon builds its own switch for its own AI chips. Microchip sells the same three jobs, and Credo names Astera as one of its three main rivals.
The evidence under this - 4 claims
The rivals: NVDA, AMZN, MCHP, MCHP, CRDO, MRVL, AVGO.
Who else does this, rival by rival
| Ticker | Rival | Where it competes | How far along | Priced against it |
|---|---|---|---|---|
| NVDA | NVIDIA NVSwitch, inside the GB300 NVL72 rackPriced against it: One segment only | Scorpio X switches, Leo CXL and optics | On sale | One segment only |
| AMZN | AWS NeuronSwitch, in the Trainium3 UltraServerPriced against it: Different product | Scorpio X switches, Leo CXL and optics | On sale | Different product |
| MCHP | Microchip Switchtec PCIe Gen 6 switchesPriced against it: Shown, not counted | Retimers, cable modules and PCIe switches | On sale | Shown, not counted |
| Microchip CXL memory controllers | Scorpio X switches, Leo CXL and optics | On sale | ||
| CRDO | Credo active electrical cables and retimersPriced against it: Yes | Retimers, cable modules and PCIe switches | On sale | Yes |
| MRVL | Marvell TechnologyPriced against it: Yes | - | On sale | Yes |
| AVGO | BroadcomPriced against it: Shown, not counted | - | On sale | Shown, not counted |
How big it could get
How big this could get
The market is AI chips whose maker does not supply the connections. Nobody counts those, so it can be bounded but not sized. The spending behind them can: Amazon alone spent $173.03B on property and equipment in the year to June, up 60.7%.
Astera says its switch is worth more than $1,000 on each AI chip, against $50 to $100 at its 2024 listing, and puts the independent switch market at about $20B a year by 2030.
The evidence under this - 4 claims
- Nothing filed carries a content-per-accelerator figure, a unit count or a design-win count. The number exists only on a call and cannot be checked against any statement.
- 'Engaged' is the company's own word and covers pre-production shipment, qualification and evaluation alike; management separately said only the lead customer reaches high-volume production in Q3 2026.
- Content per accelerator is a price times an attach rate, and the attach rate is decided by the platform designer rather than by Astera - inside an NVIDIA NVL72 rack it is zero.
- 2026-11-03editedthe Q3 release carries no customer table or content-per-accelerator figure; the Q3 10-Q and call, not yet read, are where they land, so the shelf life is extended to them
- It is the seller's own estimate of its own market, published in a results release, with no methodology stated and no third party cited.
- It is a merchant figure, so it depends entirely on how much of the accelerator base sits inside racks where the accelerator vendor supplies the fabric - the split no public source establishes. NVIDIA's NVL72 and AWS's NeuronSwitch both sit outside it by construction.
The results
The numbers
Sales grew 115% over the twelve months to Sep 2026, about the same pace as the 115% in the year to Dec 2025.
Operating margin is 26%, up from −25% in the year to Dec 2023.
Running costs take 48¢ of each sales dollar, down from 149¢ in the year to Dec 2022.
Each dollar put into the business earns 24% a year after tax, up from 15% in the year to Dec 2025.
It turned 22% of sales into free cash, $345M over the last twelve months.
Yearly sales growth has ranged from +45% to +242% over the last 3 years.
The share count rose 439% in 3.8 years, and pay in shares costs 15% of sales.
All the numbers
Filed figures from the company's 10-K and 10-Q, read 2026-11-09. Fiscal years end in December.
| Dec 2022 | Dec 2023 | Dec 2024 | Dec 2025 | 12 mo to Sep 2026 | |
|---|---|---|---|---|---|
| Size and growth | |||||
| Sales | $79.9M | $116M | $396M | $853M | $1.56B |
| Sales growth | - | +45% | +242% | +115% | +115% |
| Profit | |||||
| Gross margin | 73% | 69% | 76% | 76% | 74% |
| Operating margin | −75% | −25% | −29% | 20% | 26% |
| Net margin | −73% | −23% | −21% | 26% | 30% |
| What growth costs, per sales dollar | |||||
| Research and development | 92% | 63% | 51% | 36% | 34% |
| Sales and marketing | 31% | 17% | 31% | 9.4% | 6.8% |
| General and admin | 26% | 14% | 24% | 10% | 7.8% |
| Pay in shares | 38% | 9.2% | 59% | 19% | 15% |
| Capital spending | 4.8% | 2.4% | 8.6% | 4.4% | 5.9% |
| Cash | |||||
| Cash from operations | −$35.9M | −$12.7M | $137M | $319M | $437M |
| Free cash flow | −$39.8M | −$15.5M | $102M | $282M | $345M |
| Balance sheet | |||||
| Cash and short-term investments | - | $149M | $914M | $1.19B | $1.38B |
| Returns and shares | |||||
| Return on capital | - | - | - | 15% | 24% |
| Shares (diluted) | 34.2M | 37.1M | 131M | 180M | 184M |
What growth costs
Growth has been making Astera more profitable: research spending keeps rising, but sales have risen far faster, so each new sale costs less to win, and the company turned from losses to profit in 2025. It expects to keep a little less of each sales dollar as switches grow, so further gains must come from holding costs down.
The evidence under this - 2 claims
- Research spending held at about 36 cents of each sales dollar from FY2025 into 2026 rather than falling further, as design centres were added (C37).
- FY2024 is distorted by stock pay of 59 cents per sales dollar when listing-triggered RSUs vested, so the trend is read across FY2023, FY2025 and the trailing year, not year by year.
- Management guides gross margin down to about 72% and names 70% as the long-term target (C06), so further margin gains must come from operating costs.
- The same guidance shape preceded the Q4 2025 print and proved too conservative: gross margin was guided to about 74% for Q1 2026 and printed 76.26%.
- Management also says Scorpio's average margin across use cases is 'broadly in range with our corporate gross margins', which argues the mix effect is smaller than the guide implies.
- 2026-11-03editedthe guided quarter reported: 72.3% GAAP gross margin
The last report (3 Nov)
Quarter ended September 2026, filed figures.
Astera's September quarter was another record: sales of $584.9M, up 49% in three months and 154% in a year, above the $540M to $560M it had promised. Gross margin was 72.3%, in line with the 72% promised.
Scorpio, the switch that drove the jump, became the biggest product family. The company promises $730M to $750M next quarter. What the release does not say is whether a second customer has started buying the switch in volume.
What it did to the thesisIt met the sales and margin bars we set, which supports the growth story. The question the case turns on, whether a second customer buys the new switch, is not answered in the release, so the report left that open.
The quarter in detail
Sales growth from the previous quarter, filed figures.
Earnings per share $1.32 against $1.19 expected.
The evidence under this - 4 claims
- The release names no second Scorpio X platform shipping to revenue and files no product-family revenue, so the step cannot be attributed to the lead customer or to several.
- Four customers took 82% of Q2 revenue and the Q3 10-Q, not yet filed, carries the new table.
- The company has beaten its revenue guide at all of the quarterly prints in the pack's history but one, so the guide is a floor rather than a forecast, and a floor can still be a single customer's ramp.
- The guide gives no split by customer or product.
- The step depends on one lead platform. Management said it expects to ship Scorpio X to additional customers 'by year-end' and that those programmes are in pre-production or qualification, so the guided quarter is one customer's ramp.
- Astera's own risk list in the same release names the absence of long-term commitments from customers; there is no backlog or minimum purchase anywhere in the filings to convert.
- The company has beaten its own revenue guidance at every print since listing, which cuts both ways: it is evidence the guide is conservative and evidence that the guide is not the ceiling being priced.
- 2026-11-03editedthe guided quarter reported: $584.9M, 4.4% above the top of the range
- The same guidance shape preceded the Q4 2025 print and proved too conservative: gross margin was guided to about 74% for Q1 2026 and printed 76.26%.
- Management also says Scorpio's average margin across use cases is 'broadly in range with our corporate gross margins', which argues the mix effect is smaller than the guide implies.
- 2026-11-03editedthe guided quarter reported: 72.3% GAAP gross margin
What management did with the money
Insiders in 2026Q2: 9 purchases and 205 sales, from their Form 4 filings.
The price, and where we disagree
In short. Full. The price is about 154% of what we think the business is worth, at an 8% yearly return.
The shares are down about 36% from the $360.51 of the thesis, while Q3 profit came in well above the year before, so the multiple on last year's profit fell by more than half.
The next question is the second customer. The call and the quarterly filing may answer it before February; the Q4 results in mid-February are the formal test.
The evidence under this - 5 claims
- The share count is the thesis's; diluted shares of about 187M would put the value near $41B and every multiple about 8% higher.
- Annualising a guided quarter of a business growing 27% sequentially flatters the multiple if the ramp continues and overstates it if the ramp is a step.
- The release names no second Scorpio X platform shipping to revenue and files no product-family revenue, so the step cannot be attributed to the lead customer or to several.
- Four customers took 82% of Q2 revenue and the Q3 10-Q, not yet filed, carries the new table.
- The same guidance shape preceded the Q4 2025 print and proved too conservative: gross margin was guided to about 74% for Q1 2026 and printed 76.26%.
- Management also says Scorpio's average margin across use cases is 'broadly in range with our corporate gross margins', which argues the mix effect is smaller than the guide implies.
- 2026-11-03editedthe guided quarter reported: 72.3% GAAP gross margin
How we value it
- Method
- Cash the business will produce, in three stages
- Required return
- 8%
- Long-run growth
- 4%
- Moat
- An average moat
| Bear | Base | Bull | |
|---|---|---|---|
| Sales growth, years 1 to 3 | 30%, 19.1%, 21.5% | 60%, 38.1%, 43.1% | 75%, 47.6%, 53.9% |
| Normal operating margin | 16.6% | 20.8% | 22.9% |
| Return on new capital after year ten | 8% | 12% | 15% |
| Value a share | $41.37 | $124 | $212 |
- Price
- $191
- Priced on
- 6 Nov
- Price as a share of the base value
- 154%
- Yearly return the price offers on the base case
- 6.6%
- Risk check: valued at 10%, the rate for a cyclical or indebted business
- $81.25
Where each figure comes from
| Figure | Source | Where exactly |
|---|---|---|
| Sales, the last twelve months | Filed report | Quarterly report (10-Q) for the period to 30 Sep 2026, filed 4 Nov 2026; the four quarters to that date. |
| Operating margin today | Filed report | Quarterly report (10-Q) for the period to 30 Sep 2026, filed 4 Nov 2026; the four quarters to that date. |
| Tax rate | Our method's rule | The normal rate where no rate is filed to read: 25% for a US company, 21% elsewhere. |
| Sales per dollar of new capital | Filed report | Annual report (10-K) for the period to 31 Dec 2025, filed 20 Feb 2026: the year's sales over its operating capital, held to between 0.5 and 6. |
| Shares | Filed report | Quarterly report (10-Q) for the period to 30 Sep 2026, filed 4 Nov 2026; the diluted share count. |
| Cash and investments | Filed report | Quarterly report (10-Q) for the period to 30 Sep 2026, filed 4 Nov 2026; the balance sheet on that date. |
| Debt | Filed report | Quarterly report (10-Q) for the period to 30 Sep 2026, filed 4 Nov 2026; the balance sheet on that date. |
| Sales growth, years 1 to 3 | Analysts' forecasts | FMP's consensus for the years to 31 Dec 2026, 31 Dec 2027, 31 Dec 2028 and 31 Dec 2029: the average of 10 to 20 analysts' revenue estimates, 60%, 39%, 43.6%. FMP publishes how many analysts, not who they are. |
| Normal operating margin | Analysts' forecasts | FMP's consensus operating margin for the year to 31 Dec 2028, the average of 20 analysts' estimates, 48.5%, less stock pay of 15.2% of sales and less the 12.9% the analysts add back on their adjusted basis: 20.4%, which the base case takes as 20.8%. FMP publishes how many analysts, not who they are. |
| Return on new capital after year ten | Our method's rule | Set by our reading of the moat: the 8% rate with none, 12% with an average one, 15% with a strong one; the bear case earns the rate. |
| The bear and bull cases | Our method's rule | Our method's spread: the bear case takes half the growth and a fifth off the margin, the bull a quarter more growth and a tenth more margin. |
| Price | Market price | FMP's closing price on 6 Nov 2026. |
What each case rests on
Bear caseGrowth runs at half the analysts' path, 30%, 19% and 22%, the margin settles a fifth lower at 16.6%, and new capital earns only the 8% rate: the downturn test.
Base caseRevenue grows 60%, 38% and 43% over three years, the analysts' path (C57), held to at most 60% a year, then fades toward 4%, closing 30% of the gap each year; the margin moves to 20.8%, the analysts' year-three figure after stock pay and on the filed basis (12.3% less).
Bull caseGrowth runs a quarter above the analysts' path, 75%, 48% and 54%, the margin reaches 22.9%, and new capital earns 15%.
What moves the base value
| Lower | Higher | |
|---|---|---|
| Growth in years 1 to 3, a quarter lower or higher | $86.66 | $176 |
| The normal margin, 15% lower or higher | $107 | $141 |
| Capital per dollar of growth, a quarter more or less | $122 | $125 |
| Return on new capital after year ten, at the rate or one step up | $95.72 | $135 |
| The rate, a point higher or lower | $98.24 | $167 |
| Long-run growth a point lower | $112 | - |
Value a share as growth and margin change
| Normal margin ↓ · growth → | 50% | 75% | 100% | 125% | 150% |
|---|---|---|---|---|---|
| 16.6% | $50.13 | $71.15 | $101 | $142 | $197 |
| 18.7% | $55.17 | $78.91 | $112 | $159 | $222 |
| 20.8% | $60.22 | $86.66 | $124 | $176 | $246 |
| 22.9% | $65.26 | $94.41 | $136 | $193 | $270 |
| 25% | $70.30 | $102 | $147 | $209 | $294 |
Growth is shown as a share of the base case's. Green: the price is at or under 80% of that value. Red: at or over 120%.
How we read the multiple
Astera's reported profit is lifted by tax benefits, so this thesis values it on operating profit after the shares paid to staff.
Those figures use the mid-September price. At the $229.67 close on 2 November, buyers pay about 95 times the operating profit of the last four quarters, or about 37 times if the promised December quarter lasted a year.
Neither is the true figure for a business growing this fast. The price assumes September is a floor, not a peak.
Against its own history. Cheaper than its own normal on sales and dearer on what it owns, but that normal includes years when it lost money.
The evidence under this - 5 claims
- Annualising one guided quarter of a business growing 40% sequentially flatters the multiple if the ramp continues and overstates it if the ramp is a step rather than a trend. Both readings are printed rather than one chosen.
- The pack's own enterprise value nets only cash and cash equivalents and misses $1,141.5M of marketable securities, so its EV/EBIT of 165.7 is struck on a larger enterprise value than the balance sheet supports. The filed figure is used here and the divergence is recorded in data_notes.
- Measured on non-GAAP operating profit, which is what the sell side models, the same price is about 30 times the annualised guided figure. The gap between the two is stock pay, and which one is right is the question the frame answers rather than assumes.
- 2026-09-25editedre-struck on the 2026-09-24 close after a 33.9% price increase since the thesis; the operating-profit figures are unchanged
- 2026-11-03editedkept as the multiple at the thesis price; C56 re-strikes it on the Q3 quarter
- The share count is the thesis's; diluted shares of about 187M would put the value near $41B and every multiple about 8% higher.
- Annualising a guided quarter of a business growing 27% sequentially flatters the multiple if the ramp continues and overstates it if the ramp is a step.
- The step depends on one lead platform. Management said it expects to ship Scorpio X to additional customers 'by year-end' and that those programmes are in pre-production or qualification, so the guided quarter is one customer's ramp.
- Astera's own risk list in the same release names the absence of long-term commitments from customers; there is no backlog or minimum purchase anywhere in the filings to convert.
- The company has beaten its own revenue guidance at every print since listing, which cuts both ways: it is evidence the guide is conservative and evidence that the guide is not the ceiling being priced.
- 2026-11-03editedthe guided quarter reported: $584.9M, 4.4% above the top of the range
Priced against its rivals
The competitors its own annual report names, each confirmed against the rival's filing. Measured on company value for each $1 of operating profit, cheapest first:
| Ticker | Company | company value for each $1 of operating profit |
|---|---|---|
| CRDO | Credo Technology Group Holding | 59.0× |
| MRVL | Marvell Technology | 128.4× |
| ALAB | Astera Labs | 165.7× |
| AVGO | Broadcom not in the median | 38.7× |
| MCHP | Microchip Technology Incorporated not in the median | 55.5× |
The middle of the other 2 is 93.7. Broadcom Inc., Microchip Technology Incorporated compete with it in one product line inside a much bigger company, so their price says little about this business and they are not counted in the middle. Astera Labs, Inc. is dearest of 3.
Rivals' figures come from a data provider, so read the spread of the group rather than any single pair.
Both rivals its own filing names cost less for each dollar of operating profit, so the extra paid here is this company's own.
Also checked and set aside: Rambus (a substitute, not the same product).
Private or foreign rivals that cannot be priced: Montage Technology Co., Ltd., Parade Technologies, Ltd.
The bet: what it pays if it works
The bet is on three young lines: the Scorpio X switch joining AI chips in a rack, in volume at one customer, and memory and optical chips due in volume in 2027.
The older chips, valued like the listed rivals, cover about a third of today's price. If the switch doubles its first pace and the other two each reach Astera's 2023 size, the whole is still worth less than today: an expensive bet.
If all three stall, shareholders keep the older chips and the cash.
| Scorpio X switches, Leo CXL and optics operating profit in two to three years | Accelerators fitted with Scorpio X a year needed | Valued like its peers | Like the cheapest |
|---|---|---|---|
| $210.00M | about 630.0K at 23% kept | 0.8× | 0.6× |
| $415.00M (the big win) | about 1.5M at 23% kept | 1.2× | 0.9× |
| $647.00M | about 2.6M at 23% kept | 1.7× | 1.2× |
To pay 2× (a Fair bet) it would need about $1.26B of Scorpio X switches, Leo CXL and optics operating profit (about 5.3M accelerators fitted with Scorpio X a year at 23% kept); to pay 5× (a Cheap bet), about $3.57B of Scorpio X switches, Leo CXL and optics operating profit (about 15.4M accelerators fitted with Scorpio X a year at 23% kept), at the same multiple.
Each row is Scorpio X switches, Leo CXL and optics alone, plus the rest of the company (Retimers, cable modules and PCIe switches) counted at about $16.13B.
The steps it has to take
- Done Scorpio X in volume production at the lead customer
- In progress Scorpio the largest product family by revenue due 2026-11
- Not started Scorpio X shipping to additional customers due 2026-12
- Not started No single customer above a quarter of revenue due 2027-12
- Not started Leo CXL in volume at two US hyperscalers due 2027-12
- Not started First optical product in volume production due 2027-12
- Not started Scorpio X at about $1.2B a year due 2028-06
The detail behind the label
The evidence under this - 9 claims
- The step depends on one lead platform. Management said it expects to ship Scorpio X to additional customers 'by year-end' and that those programmes are in pre-production or qualification, so the guided quarter is one customer's ramp.
- Astera's own risk list in the same release names the absence of long-term commitments from customers; there is no backlog or minimum purchase anywhere in the filings to convert.
- The company has beaten its own revenue guidance at every print since listing, which cuts both ways: it is evidence the guide is conservative and evidence that the guide is not the ceiling being priced.
- 2026-11-03editedthe guided quarter reported: $584.9M, 4.4% above the top of the range
- Nothing filed carries a content-per-accelerator figure, a unit count or a design-win count. The number exists only on a call and cannot be checked against any statement.
- 'Engaged' is the company's own word and covers pre-production shipment, qualification and evaluation alike; management separately said only the lead customer reaches high-volume production in Q3 2026.
- Content per accelerator is a price times an attach rate, and the attach rate is decided by the platform designer rather than by Astera - inside an NVIDIA NVL72 rack it is zero.
- 2026-11-03editedthe Q3 release carries no customer table or content-per-accelerator figure; the Q3 10-Q and call, not yet read, are where they land, so the shelf life is extended to them
- No product-family revenue is filed (C43), so whether Scorpio becomes the largest family can only be checked against what management says on the next call.
- Four direct customers were 82% of Q2 2026 revenue (C10), and the guided Q3 step is attributed to the lead customer's ramp (C05), so 'additional customers' carries no volume or date beyond year-end.
- Astera files no product-family revenue (C43), so how much of the step is Scorpio X, and how much is Aries, Taurus and the PCIe Scorpio P line also growing, cannot be checked.
- The step is one lead customer's ramp (C05); a single platform can stop as quickly as it started.
- 2026-11-03editedthe Q3 release carries no customer table or content-per-accelerator figure; the Q3 10-Q and call, not yet read, are where they land, so the shelf life is extended to them
- No Leo or optical revenue, volume or term is disclosed; management says most CXL designs are in qualification or early stage this year.
- The chief executive greeted the CXL traction with 'it was about time', after years of investment with little to show; the optical market figure comes with no source.
- Astera's FY2023 revenue was retimers at the start of the AI build-out, a different product in a different market; nothing says Leo or optics will follow the same path.
- Management states no revenue figure for either line, and the optical products are not yet in production (C40, C51).
- Annualising one guided quarter of a business growing 40% sequentially flatters the multiple if the ramp continues and overstates it if the ramp is a step rather than a trend. Both readings are printed rather than one chosen.
- The pack's own enterprise value nets only cash and cash equivalents and misses $1,141.5M of marketable securities, so its EV/EBIT of 165.7 is struck on a larger enterprise value than the balance sheet supports. The filed figure is used here and the divergence is recorded in data_notes.
- Measured on non-GAAP operating profit, which is what the sell side models, the same price is about 30 times the annualised guided figure. The gap between the two is stock pay, and which one is right is the question the frame answers rather than assumes.
- 2026-09-25editedre-struck on the 2026-09-24 close after a 33.9% price increase since the thesis; the operating-profit figures are unchanged
- 2026-11-03editedkept as the multiple at the thesis price; C56 re-strikes it on the Q3 quarter
- 2026-11-03editedSeptember balance sheet from the Q3 release
Where we disagree with the market
The switch is simply an upgrade: more sold on each AI chip, a bigger market, the same company doing better.
The switch is the part rivals want most. It is about to be the largest line and is the least protected.
Analysts are 73.1% positive and none says sell. On the June-quarter call every question about the switch asked how fast or how much; none asked who else could supply it.
The difference is protection. The older signal chips are protected by a limit written into an industry standard; the switch only by being first. We are wrong if it reaches volume at a second and third named customer in 2027 with the margin near 72%.
What analysts say, and the evidence
The evidence under this - 6 claims
- Nothing filed carries a content-per-accelerator figure, a unit count or a design-win count. The number exists only on a call and cannot be checked against any statement.
- 'Engaged' is the company's own word and covers pre-production shipment, qualification and evaluation alike; management separately said only the lead customer reaches high-volume production in Q3 2026.
- Content per accelerator is a price times an attach rate, and the attach rate is decided by the platform designer rather than by Astera - inside an NVIDIA NVL72 rack it is zero.
- 2026-11-03editedthe Q3 release carries no customer table or content-per-accelerator figure; the Q3 10-Q and call, not yet read, are where they land, so the shelf life is extended to them
- The same guidance shape preceded the Q4 2025 print and proved too conservative: gross margin was guided to about 74% for Q1 2026 and printed 76.26%.
- Management also says Scorpio's average margin across use cases is 'broadly in range with our corporate gross margins', which argues the mix effect is smaller than the guide implies.
- 2026-11-03editedthe guided quarter reported: 72.3% GAAP gross margin
What could go wrong, and how we would know
The risks
- Customers build their own. Amazon already builds its own switch for its own AI chips, and says so publicly.
- Four buyers, 82% of sales. Some are factories buying for somebody else, so the percentages may not show where real demand sits.
- A margin heading down. Promised at about 72% for September against 73.3% in June, with 70% named as the long-term level.
- Insiders only sell. No open-market purchase against 508 sales, and both founders sold large blocks on four quarterly dates by choice, not under a plan.
The evidence under this - 5 claims
- 2026-11-03editedthe Q3 release carries no customer table or content-per-accelerator figure; the Q3 10-Q and call, not yet read, are where they land, so the shelf life is extended to them
- The same guidance shape preceded the Q4 2025 print and proved too conservative: gross margin was guided to about 74% for Q1 2026 and printed 76.26%.
- Management also says Scorpio's average margin across use cases is 'broadly in range with our corporate gross margins', which argues the mix effect is smaller than the guide implies.
- 2026-11-03editedthe guided quarter reported: 72.3% GAAP gross margin
What we still do not know
- Whether the more than ten customers management says are engaged on the switch become named platforms shipping it, or the ramp stays one customer's.
- How much it sells into China: 30.1% of 2025 sales by billing address, while management says China will stay in single digits.
The evidence under this - 3 claims
- Nothing filed carries a content-per-accelerator figure, a unit count or a design-win count. The number exists only on a call and cannot be checked against any statement.
- 'Engaged' is the company's own word and covers pre-production shipment, qualification and evaluation alike; management separately said only the lead customer reaches high-volume production in Q3 2026.
- Content per accelerator is a price times an attach rate, and the attach rate is decided by the platform designer rather than by Astera - inside an NVIDIA NVL72 rack it is zero.
- 2026-11-03editedthe Q3 release carries no customer table or content-per-accelerator figure; the Q3 10-Q and call, not yet read, are where they land, so the shelf life is extended to them
- The filed table says 30.1% and the statement says single digits. Both can be right - the table is billing address for contract manufacturers and the statement is end demand - but nothing filed reconciles them, so the reader has only management's word for which number describes the exposure that matters.
- The billing-address share rose from 18.3% to 30.1% in one year, which is a large move in whichever thing it measures.
What would change our view
- Unattractive→NeutralThe price falls back within 20% of fair value
It needs:
- ○The share price falls to about $149, back within 20% of what we think the business is worth.about ≈$147.13 today, 45% below the last close
- Unattractive→AttractiveThe price falls 20% under fair value
It needs:
- ○The share price falls to about $99.16, 80% of what we think the business is worth.about ≈$95.47 today, 64% below the last close
We would also be wrong if
- The switch is still one customer's product at the end of 2027.
- The rest of the business settles below 25% growth a year.
What is coming
- 3 November 2026 Q3 2026 results: Revenue against the $540M to $560M guide, gross margin against 72%, whether Scorpio became the largest family as promised, and whether a second Scorpio X platform is named.
Every dated event ahead
- 3 November 2026 Q3 2026 results: Revenue against the $540M to $560M guide, gross margin against 72%, whether Scorpio became the largest family as promised, and whether a second Scorpio X platform is named.
- December 2026 Scorpio X-Series shipping to additional customers: Whether the switch reaches buyers beyond the lead customer by year-end as promised; a slip leaves the guided step resting on one platform.
- February 2027 Q4 2026 results and the FY2026 10-K: The full-year ramp, the 2027 outlook, and whether the 10-K's customer table still shows four buyers taking four-fifths of sales.
- April 2027 2027 proxy statement: The first disclosure of what the new performance shares pay on; GAAP profit or per-share goals would close the gap the bonus's non-GAAP margin leaves.
Since we wrote this
- 9 NovThesis reviewUnattractive: priced at 135% of a value near $141Re-valued in three stages at an 8% return, the method now drawn on every thesis: the price sits beyond the 20% margin of safety.
- 4 NovEarningsAstera's call: a second hyperscaler is shipping its switch in pre-production. View heldQ3 2026 results.
- 3 NovEarningsRevenue beat and margin held; second platform unconfirmedSales of $584.9M cleared the $560M bar and margin was 72.3%, but no second Scorpio X platform is named until the call and 10-Q are read.
- 11 OctThesis reviewScorpio X, Leo CXL and optics priced as one betThe owner's rule of 2026-10-11: a thesis carries one bet, so its early lines fold into one business.
- 5 OctThesis reviewThe long case now sits under the verdictAfter the five-name pilot, the owner asked for the long case on every thesis.
- 4 OctThesis reviewOwner's checklist: revenue streams, cost of growth, management, catalystsThe owner's checklist found the method silent on who runs the company, what growth costs, where the money comes from and what is dated ahead; this pass fills those four blocks from the proxy and filings.
- 25 SepPrice moveShares up 33.9% with no new profit behind it. View heldThe rise moved the company further from the multiple that would make it attractive, not closer. Both routes still wait on the November results.
- 16 SepThesis writtenNeutral, medium convictionA real, fast-growing business whose growth now comes from the switch that NVIDIA sells inside its own racks and Amazon builds for itself.
Also checked 4 times with nothing new to act on.
Sources 23
Sources - 23
Every claim above cites one of these, or states the reasoning that derived it. The quote is frozen at the date it was read: a link can rot, and an agent fetch returns a reading of a page rather than the page. Backs counts the whole claim ledger, so it can name a claim this page does not surface.
- S01primarycompany filing2026-02-20the product families and COSMOS; the named competitor list; the customer and geographic concentration tables with the billing-address basis; the single-foundry supply note; the Amazon warrant history in Note 10; stock-based compensation of $160.0M; R&D of $304.0M; headcount of 756; the aiXscale acquisition and the February 2026 asset purchasebacksC01C07C10C12C14C15C33C37C40C43C48C49C52C53
- S02primarycompany filing2026-08-05the income statement, balance sheet and cash flow for the quarter and half; the income tax note and the valuation allowance reasoning; the Q2 2026 customer and receivable concentration tables with the manufacturing-partner caveat; the 2026 warrant terms and its $280.0M maximum fair value; warrants contra revenue; the single reportable segmentbacksC01C04C07C08C09C10C11C14C15C16C17C28C31C34C38C42C43C48C49
- S03primarycompany release2026-08-04the Q3 2026 GAAP and non-GAAP guidance including revenue, gross margin, operating expenses, tax rate, per-share range and diluted share count; the GAAP to non-GAAP reconciliation; the balance sheet and cash flow statements; the risk list naming the absence of long-term commitments from customersbacksC01C03C04C05C06C07C08C16C17C27C28C37C38C42
- S04primarytranscript2026-08-04PCIe 6.0 above 50% of revenue against a third in Q1; content per accelerator above $1,000 for Scorpio X; more than ten customers engaged on Scorpio X; the 70% long-term gross margin target; the $20B merchant scale-up switching figure and the $4B Taurus figure; the optical roadmap dates; the statement that China stays single digits as a share of revenue; COSMOS in the orchestration path. Twenty-four analyst turns from twelve analysts - the questions are a separate origin from the answers, and none of them asked who else could supply the fabricbacksC03C05C06C13C25C26C36C40C41C44C51C53
- S05primarycompany release2026-02-10the full-year 2025 GAAP and non-GAAP results; the Q1 2026 guidance of $286M to $297M at about 74% gross margin, which is the guide the February repricing turned on; the $20B by 2030 merchant scale-up switching estimate; the Leo CXL deployment in Microsoft Azure M-series virtual machines; the Israel design centrebacksC26C27C34C37C38C41
- S06secondaryvendor data2026-09-16Research pack for ALAB, frozen at content/ALAB/research-pack.jsoncontent/ALAB/research-pack.jsonthe last close, market capitalisation, 20-day range and 252-day levels; the price move decomposition over 252 and 126 sessions; the analyst rating series and price target drift; short interest from FINRA; the earnings history with filed GAAP per-share figures beside adjusted ones; theme co-movement; the filed competitor cohort multiples; the geographic split re-keyed from the filer's XBRL; headcount and revenue per employeebacksC12C27C28C29C31C35C38
- S07primaryindustry data2026-09-16the reach limit that makes the retimer socket exist, stated by the body writing the interface rather than by anybody selling into it
Common Electrical I/O - 224G-MR This project will develop and produce an implementation agreement for a Medium Reach electrical interface operating @144 to 232Gbps signaling over up to 500 mm of PCB and up to one connector. Common Electrical I/O - 224G-LR This project will develop and produce an implementation agreement for a Long Reach electrical backplane interface operating @144 to 232 Gbps signaling over up to 1000 mm of backplane and up to two connectors.
backsC02 - S08primaryindustry data2026-09-16UALink Consortium - Specifications page UALink Consortiumthe size of a scale-up domain as a standards body defines it, which bounds how many accelerators one merchant fabric can address
The UALink 1.0 Specification enables 200G per lane scale-up connection for up to 1,024 accelerators within an AI computing pod, delivering the open standard interconnect for next-generation AI cluster performance.
backsC21 - S09primaryother2026-09-16that the largest accelerator vendor supplies the scale-up fabric inside its own rack, in its own words - the exclusion the prize turns on
NVIDIA Mission Control powers every aspect of NVIDIA GB300 NVL72 AI factory operations, from orchestrating workloads across the 72-GPU NVLink domain to integration with facilities.
backsC18 - S10primarycompany release2025-05-18the route by which Astera silicon enters an NVIDIA rack at all, named by NVIDIA rather than by Astera - both the threat and the company's answer to it
MediaTek, Marvell, Alchip Technologies, Astera Labs, Synopsys and Cadence are among the first to adopt NVLink Fusion, enabling custom silicon scale-up to meet the requirements of demanding workloads for model training and agentic AI inference.
backsC22 - S11primaryother2026-09-16Amazon Web Services - AWS Trainium product page Amazon Web Servicesthat the concentrated buyer builds the scale-up switch itself - the customer_can_build evidence, published by the party with no incentive to flatter a supplier
Trainium3 Next Generation UltraServers scale up to 144 Trainium chips, with NeuronSwitch providing an all-to-all fabric that doubles interchip interconnect bandwidth over Trainium2 UltraServers.
backsC19 - S12primaryindustry data2024-07-31what interconnect failure costs the buyer, counted by the buyer: 419 unexpected interruptions in 54 days on 16,384 GPUs, 35 of them network switch or cable
Llama 3 405B is trained on up to 16K H100 GPUs, each running at 700W TDP with 80GB HBM3 … the synchronous nature of training makes it less fault-tolerant-a single GPU failure may require a restart of the entire job … During a 54-day snapshot period of pre-training, we experienced a total of 466 job interruptions. Of these, 47 were planned interruptions … The remaining 419 were unexpected interruptions, which are classified in Table 5 … Network Switch/Cable Network 35 8.4%
backsC23 - S13primarycompany filing2026-06-15a direct rival naming Astera Labs as one of only three principal competitors, and its own description of selling the same products to hyperscalers, neoclouds, OEMs, ODMs and contract manufacturersbacksC39
- S14primarycompany filing2026-07-31purchases of property and equipment of $173.03B over the trailing twelve months against $107.66B a year earlier - a filed base for the build-out, from the customer that holds Astera's warrantsbacksC24
- S15primarycompany filing2026-05-21a named competitor's own description of shipping PCIe Gen 3 to Gen 6 switches and retimers, and CXL and PCIe memory controllers - all three of Astera's product jobs, from the rival's sidebacksC20
- S16primarycompany filing2026-09-16508 open-market sales and no open-market purchases; the plan flag on each, putting 77.4% of shares sold under a plan adopted earlier; and the four quarterly discretionary blocks of 90,000 to 95,000 shares by each co-founder, which the aggregate could not showbacksC32
- S17primaryvendor data2026-09-16Daily closing price series for ALAB (FMP historical-price-eod)data/cache/fmpthe session-by-session reaction to each of the last four prints, which is what the prior-repricing search needed and what the pack's news reaction block only carries for recent itemsbacksC27C34
- S18primarycompany filing2026-03-11Marvell naming Astera Labs inside a twenty-four-name competitor list, and its own self-description as a fabless data-infrastructure supplier whose data centre end market is 74% of revenueread, but no claim rests on it
- S19primarycompany filing2026-04-23beneficial ownership as of 1 March 2026; the directors' and officers' biographies and tenure; the CD&A: the 2025 bonus metrics, weights and payout, the time-based RSU design, the co-founders' 2024 pre-IPO grants and the plan to add PSUs from 2026; the outstanding equity awards tablebacksC45C46C47
- S20primarycompany filing2026-02-10Michael Tate's retirement as chief financial officer effective 2 March 2026 and his advisory role to 1 September 2026; Desmond Lynch's appointment and his offer letter, including a PSU award vesting on performance milestones set by the boardbacksC46C47
- S21primarycompany release2026-11-03Q3 2026 revenue, GAAP and non-GAAP margins, the Q4 guide, the September balance sheet and the product announcementsbacksC05C06C17C54C55C56
- S22primarycompany filing2026-11-09trailing revenue, operating profit, stock pay, tax, debt, cash and the diluted share countbacksC58
- S23secondaryvendor data2026-11-09the analysts' average revenue and operating profit for the next three fiscal years, and the closebacksC57C58