Silicom
An Israeli maker of custom network cards that slot into other companies' servers and security boxes, now pitching cards for AI and for encryption built to resist quantum computers.
The scores
Growth3/4 ↑
Four quarters of accelerating growth (17% to a guided 66%), but the inflection is under a year old, riding ramping design wins.
Business quality2/4 →
The moat rests mainly on the company's own claim that majors decline this segment, with one repeat-customer example and a real 2023-2024 collapse.
Financial position4/4 →
No debt, ~$74-107M of cash and securities against a $303.4M market cap, funded entirely from operations and past capital raises.
Execution record3/4 ↑
Guidance was beaten and raised twice in 2026, and EPS beat in 6 of 8 quarters - but the pattern is only two quarters old.
Per share capture3/4 →
Share count has shrunk via buybacks with only modest new-share issuance, but the newly effective shelf is a real, flagged stock-sale risk.
Potential2/4 →
AI-inference and the 2026 design wins are all conditional, not firm; the prize itself has no independently measured addressable slice.
Risk profile3/4 →
One serious, disclosed concentration (top-3 customers 28%) plus a single-country (Israel) R&D/HQ base and a disclosed 2024 cyberattack, all survivable and understood.
Valuation2/4 →
EV/Sales of 3.77x TTM is nearly triple the prior-fiscal-year median, ahead of a proven GAAP-profitable core, with AI-inference still a small contributor.
Evidence quality3/4 →
Central claims are mostly primary filings plus independent outside sources (10-Ks, Nvidia, Dell'Oro, the White House), but several forward claims rest on management assertions.
If it works 0.6× · if it fails you keep about 46%.
The Verdict
We rate Silicom unattractive. It designs custom network cards for servers, and the business is solid: after sales halved in 2024, growth has sped up four quarters running. The shares have nearly tripled in a year on a bigger story, cards for running AI, which management expects to bring only a few million dollars this year, with no firm deal and insiders selling. The price is full, not cheap: that story is paid for up front, before it is a business. As a bet it is fragile: even if the new lines work, the shares would be worth less than today; if they fail, a little under half is left.
The long caseAs AI moves from training models to running them, and governments force encryption upgrades against quantum computers, new network and security hardware is needed; Silicom must win real volume.
The evidence under this - 6 claims
- 2026-10-29editedQ3 2026 release (S25) raised full-year guidance to $96-98M and set Q4 at $27-28M; the $93-95M and Q3 $25-26M figures it replaces were reached and passed
The specific dollar figure is management's own guidance with no independent verification of the underlying customer commitments (see C11's binding status).
- 2026-10-29editedQ3 2026 release (S25) read: it describes AI Inference as 'significant long-term potential upside' but gives no revenue figure, so the Q2 guide stands unrefreshed; valid_through rolled to the 20-F
The 20-F's own risk factors state that Design Wins do not guarantee purchase volumes and that ramp timing is outside the company's control; FY2024's 53% revenue collapse (C04) is a prior period where previously expected volumes did not materialize on schedule, so a 'potential'/'anticipated' figure has a demonstrated history of not converting as guided.
Selling shares received from option exercises is ordinary liquidity and does not by itself indicate a negative view of the business; the unchecked Rule 10b5-1 box says the sales were not pre-scheduled under a claimed plan, not why they were made.
- 2026-09-27editedprice drift re-strike: returns and 52-week position re-read at the 2026-09-25 close (+197.4% a year, 1.0% off the high, against +139% and 26.9% off at 2026-09-03); Form 4 XML read for the 10b5-1 flag, which settles the counter-evidence's open point - no sale was under a plan
Revenue growth has genuinely reaccelerated for four consecutive quarters (17% to a guided 66%), which is an observed trend, not only a narrative, and partly justifies a re-rate independent of the AI-inference story.
- 2026-09-27editedprice drift re-strike: EV/Sales re-read at 3.77x TTM (from 2.70x at 2026-09-03) with the vendor EV of C27; the 1.34x median was mislabelled 'three-year' - the pack takes it over six prior fiscal years
The company
A company building a server, a security appliance or a network box needs a card that moves data in and out, often in an odd shape or with a special job such as encryption. The chip giants sell standard cards and, Silicom says, rarely customise them for smaller buyers.
Silicom designs the card around a bought-in or programmable chip, writes its software, and has subcontractors build it. The customer then spends twelve months or more testing it inside its own product before ordering in volume.
The evidence under this - 3 claims
Broadcom's and Marvell's own 10-Ks (C16, C17) describe capability across a very wide range of customers and applications and say nothing about declining small or customized accounts; the claim that majors structurally avoid Silicom's segment rests solely on Silicom's own framing of its competitors' priorities, which a competitor with a different incentive might dispute.
How it makes money
Each design a customer approves, which Silicom calls a design win, turns into repeat orders for as long as the customer's product sells, and Silicom counts over 400 still active. Sales are lumpy because they follow customers' own buying: in 2024 they fell 53.2% as customers worked through stock they had built up.
The evidence under this - 3 claims
- 2026-10-29editedQ3 2026 release (S25): the count rose from 7 to 8 of a targeted 7-9
Where the money comes from, stream by stream
| Stream | Share of sales (H1 2026) | How it earns it |
|---|---|---|
| United States | 82.0% | Network cards, smart cards and edge platforms sold into US cloud, security and telecom customers' own products under multi-year design wins. |
| Europe, including Switzerland | 8.5% | The same design-win products sold to European equipment makers and service providers. |
| Asia-Pacific | 5.6% | The same products sold to customers and contract manufacturers in Asia. |
| Israel | 4.0% | The same products sold to Israeli equipment and security vendors. |
Who runs it
Liron Eizenman has run Silicom since 2022; his father co-founded it and chairs the board, and the finance chief has served since 2005. Insiders own about 5%, mostly the chairman, and several sold in August. Bonuses pay on hitting most of the revenue or profit budget; share awards simply vest with time. Shareholders voted down the boss's share awards twice, and the board granted them anyway.
The evidence under this - 4 claims
The August 2026 open-market sales in C22, among them 20,000 shares by the Active Chairman, came after the table's date, so the insiders' share is now lower than 5.42%.
The thresholds are fractions of a budget the board sets, so a revenue-only year at 80% of plan can pay with an operating loss, as every year since 2023 has been.
The 2026 override added a performance test, which the earlier grants lacked.
Why it wins, and who could beat it
Why customers come back
Once a card has passed a customer's year-long testing, replacing it means repeating that test, at a cost to the customer's own schedule. The evidence is repeat business: a July 2026 win with an existing customer triples its expected 2027 spending with Silicom, to nearly $10M.
Broadcom's and Marvell's own annual reports describe broad chip businesses rather than custom card making, which fits Silicom's claim that few rivals chase its niche. No customer's filing confirms it.
The evidence under this - 5 claims
Customer C's revenue collapsed from $26.8M (2023) to $2.56M (2025) despite presumably being a qualification-locked, designed-in customer - direct evidence that the switching-cost mechanism did not prevent a large revenue loss, even if the cause was the customer's own demand/inventory cycle rather than a supplier switch.
The 20-F's own risk factors state that Design Wins do not guarantee purchase volumes and that ramp timing is outside the company's control; FY2024's 53% revenue collapse (C04) is a prior period where previously expected volumes did not materialize on schedule, so a 'potential'/'anticipated' figure has a demonstrated history of not converting as guided.
The threat
The chip giants sell the same job on their own terms. NVIDIA's BlueField cards take networking, storage and security work off the main processor and ship inside standard servers from Dell, Lenovo and Supermicro. Broadcom and Marvell sell the chips that let a big enough customer build its own card, and the AI prospects Silicom courts could design theirs in-house.
The evidence under this - 4 claims
Broadcom's and Marvell's own 10-Ks (C16, C17) describe capability across a very wide range of customers and applications and say nothing about declining small or customized accounts; the claim that majors structurally avoid Silicom's segment rests solely on Silicom's own framing of its competitors' priorities, which a competitor with a different incentive might dispute.
The rivals: NVDA, AVGO, MRVL, and 1 more that is private, in-house or listed abroad.
Who else does this, rival by rival
| Ticker | Rival | Where it competes | How far along | Priced against it |
|---|---|---|---|---|
| NVDA | Nvidia (BlueField DPU platform)Priced against it: Small part of it | Established networking: server adapters, FPGA Smart Cards and edge | On sale | Small part of it |
| AVGO | BroadcomPriced against it: Small part of it | Established networking: server adapters, FPGA Smart Cards and edge | On sale | Small part of it |
| MRVL | Marvell (OCTEON DPU line)Priced against it: One segment only | Established networking: server adapters, FPGA Smart Cards and edge | On sale | One segment only |
| In-house | The internal engineering teams of Silicom's own AI-inference prospectsPriced against it: In-house | New lines: AI-inference NICs, post-quantum crypto and white-label switching | On sale | In-house |
How big it could get
How big this could get
Dell'Oro, a research firm, says the market for such smart network cards grew from $270M in 2020 to $848M in 2024, 26% a year. Nobody splits out the custom, smaller-customer slice Silicom says the giants leave to it, so its own market is unmeasured.
The evidence under this - 2 claims
Broadcom's and Marvell's own 10-Ks (C16, C17) describe capability across a very wide range of customers and applications and say nothing about declining small or customized accounts; the claim that majors structurally avoid Silicom's segment rests solely on Silicom's own framing of its competitors' priorities, which a competitor with a different incentive might dispute.
Who has signed
Four 2026 design wins carry dollar figures: about $3M a year from a European encryption firm, $5M or more a year from each of two security companies, and nearly $10M in 2027 from an existing customer. Every one is described as potential or anticipated, with no binding volume or term.
The evidence under this - 2 claims
The 20-F's own risk factors state that Design Wins do not guarantee purchase volumes and that ramp timing is outside the company's control; FY2024's 53% revenue collapse (C04) is a prior period where previously expected volumes did not materialize on schedule, so a 'potential'/'anticipated' figure has a demonstrated history of not converting as guided.
The results
The numbers
Sales grew 6.6% in the year to Dec 2025, faster than the −53% in the year before.
Operating margin is −20%, down from 13% in the year to Dec 2022.
Running costs take 50¢ of each sales dollar, up from 24¢ in the year to Dec 2021.
Shareholders' money lost 9.4% a year, down from 6.7% in the year to Dec 2021.
It used $3.35M of cash in the latest year, so the $42.1M it holds lasts about 12.6 years at that pace.
Yearly sales growth has ranged from −53% to +20% over the last 5 years, with sales falling in 2 of them.
The share count fell 18% in 4.0 years, and pay in shares costs 4.5% of sales.
All the numbers
Filed figures from the company's 10-K and 10-Q, read 2026-10-04. Fiscal years end in December.
| Dec 2021 | Dec 2022 | Dec 2023 | Dec 2024 | Dec 2025 | |
|---|---|---|---|---|---|
| Size and growth | |||||
| Sales | $128M | $151M | $124M | $58.1M | $61.9M |
| Sales growth | +20% | +17% | −18% | −53% | +6.6% |
| Profit | |||||
| Gross margin | 35% | 35% | 23% | 29% | 31% |
| Operating margin | 10% | 13% | −23% | −23% | −20% |
| Net margin | 8.2% | 12% | −21% | −24% | −19% |
| What growth costs, per sales dollar | |||||
| Research and development | 16% | 14% | 17% | 34% | 32% |
| Sales and marketing | 5.1% | 4.6% | 5.6% | 10% | 11% |
| General and admin | 3.6% | 3.0% | 3.4% | 7.5% | 7.4% |
| Pay in shares | 2.2% | 2.4% | 2.7% | 5.4% | 4.5% |
| Capital spending | 2.0% | 1.4% | 0.9% | 1.6% | 1.9% |
| Cash | |||||
| Cash from operations | $1.08M | −$4.09M | $31.9M | $18.3M | −$2.16M |
| Free cash flow | −$1.51M | −$6.18M | $30.8M | $17.4M | −$3.35M |
| Balance sheet | |||||
| Cash and short-term investments | $37.6M | $34.8M | $54.9M | $72.1M | $42.1M |
| Returns and shares | |||||
| Return on capital | 6.8% | 9.6% | - | - | - |
| Shares (diluted) | 6.97M | 6.80M | 6.70M | 6.02M | 5.71M |
What growth costs
Growth makes Silicom more profitable. Most of its costs are engineers whose pay barely moves with sales, so when sales fell by half the loss ballooned, and as sales recover each new dollar costs less to win. The parts in each product grow with sales, so the gain comes below that line.
The evidence under this - 2 claims
The last report (29 Jul)
The third quarter of 2026 beat the company's own forecast again: sales of $26.7M, up 71% on a year earlier, against a promised $25M to $26M. The full-year forecast rose to $96M to $98M.
On the adjusted measure it made $0.3M, its first profit in this run; on the reported measure it still lost $0.3M. Cash and bonds fell $10M to $45.0M as stock for customers rose. The release gave no AI sales figure, no named customer and no order size.
What it did to the thesisSales and profit moved the right way, which the case already granted. The AI story the price pays for got no new numbers, and the company still lost money on the reported measure, so the view stays negative.
The quarter in detail
Earnings per share $-0.16 against $-0.24 expected.
The evidence under this - 5 claims
- 2026-10-29editedQ3 2026 release (S25) raised full-year guidance to $96-98M and set Q4 at $27-28M; the $93-95M and Q3 $25-26M figures it replaces were reached and passed
- 2026-10-29editedQ3 2026 release (S25): GAAP loss narrowed to $0.3M and non-GAAP turned to a $0.3M profit; GAAP loss persists
- 2026-10-29editedQ3 2026 balance sheet (S25): the same $107M of working capital plus securities, but a tenth of it moved from cash into inventory
The specific dollar figure is management's own guidance with no independent verification of the underlying customer commitments (see C11's binding status).
- 2026-10-29editedQ3 2026 release (S25) read: it describes AI Inference as 'significant long-term potential upside' but gives no revenue figure, so the Q2 guide stands unrefreshed; valid_through rolled to the 20-F
What management did with the money
The balance sheet is sound: $107M of working capital and securities, about $19 a share, and no debt. Buybacks cut the share count by about 15% between 2023 and 2025. But in June a registration allowing new share sales took effect, and the chief executive said any use would likely be to fund AI.
Insiders in 2026Q2: 8 purchases and 18 sales, from their Form 4 filings.
The evidence under this - 3 claims
- 2026-10-29editedQ3 2026 balance sheet (S25): the same $107M of working capital plus securities, but a tenth of it moved from cash into inventory
- 2026-10-04edited6,738,706 and 5,706,142 are ordinary shares outstanding (net of treasury shares) from the 20-F's statement of changes in equity, not a diluted count; the label was wrong and the figures stand (S01)
- 2026-10-29editedQ3 2026 release (S25) read for a raise: paid-in capital rose $3.9M since December, consistent with option exercises and share pay, and the release announces no offering
The price, and where we disagree
In short. Full. The price pays for the AI story before it is a business: the recovery is real, but nearly triple the usual price for each dollar of sales is not yet earned.
The shares are back at their high of the past year, up 14.8% in the last week with no company news or filing behind the move. In August the chief executive, the finance chief and two directors sold, and none of those sales was under a pre-arranged trading plan.
The next answer comes on 29 October. Sales of $25M to $26M are forecast; what matters more is whether management names a customer, a volume and a term for its AI and switching orders, instead of the guarded words of the last call.
The evidence under this - 3 claims
Selling shares received from option exercises is ordinary liquidity and does not by itself indicate a negative view of the business; the unchecked Rule 10b5-1 box says the sales were not pre-scheduled under a claimed plan, not why they were made.
- 2026-09-27editedprice drift re-strike: returns and 52-week position re-read at the 2026-09-25 close (+197.4% a year, 1.0% off the high, against +139% and 26.9% off at 2026-09-03); Form 4 XML read for the 10b5-1 flag, which settles the counter-evidence's open point - no sale was under a plan
- 2026-10-29editedQ3 2026 release (S25) raised full-year guidance to $96-98M and set Q4 at $27-28M; the $93-95M and Q3 $25-26M figures it replaces were reached and passed
The 20-F's own risk factors state that Design Wins do not guarantee purchase volumes and that ramp timing is outside the company's control; FY2024's 53% revenue collapse (C04) is a prior period where previously expected volumes did not materialize on schedule, so a 'potential'/'anticipated' figure has a demonstrated history of not converting as guided.
How we read the multiple
Counting cash and debt, buyers pay about $3.80 for each $1 of Silicom's yearly sales, about 2.8 times the $1.30 of a typical past year. It has lost money on its official accounts every quarter.
So the extra can only be paying for the new businesses: AI, encryption and network switching, together under a tenth of this year's sales.
Against its own history. Nearly triple its own normal for each dollar of sales, while the profit that once justified a lower price has not come back.
The evidence under this - 3 claims
Revenue growth has genuinely reaccelerated for four consecutive quarters (17% to a guided 66%), which is an observed trend, not only a narrative, and partly justifies a re-rate independent of the AI-inference story.
- 2026-09-27editedprice drift re-strike: EV/Sales re-read at 3.77x TTM (from 2.70x at 2026-09-03) with the vendor EV of C27; the 1.34x median was mislabelled 'three-year' - the pack takes it over six prior fiscal years
- 2026-10-29editedQ3 2026 release (S25): GAAP loss narrowed to $0.3M and non-GAAP turned to a $0.3M profit; GAAP loss persists
The specific dollar figure is management's own guidance with no independent verification of the underlying customer commitments (see C11's binding status).
- 2026-10-29editedQ3 2026 release (S25) read: it describes AI Inference as 'significant long-term potential upside' but gives no revenue figure, so the Q2 guide stands unrefreshed; valid_through rolled to the 20-F
Priced against its rivals
None of the listed rivals above sells the same thing to the same buyers as a large part of what it does. NVIDIA and Broadcom do too, but as a small part of much bigger companies, so their price says little about this one. A price comparison needs at least two, so the price is judged against its own history instead.
The bet: what it pays if it works
This bet is on the three new lines only: chips for AI inference, post-quantum encryption cards and white-label switches. The rest of Silicom is counted separately, at its own normal price for sales it already guides to.
Even if all three new lines reach the sizes management has named, the company would be worth less than today, which makes this an expensive bet. If they come to nothing, the old business and the cash still carry much of today's price.
| New lines: AI-inference NICs, post-quantum crypto and white-label switching revenue in two to three years | Design wins worth $5M a year needed | At the stated multiple |
|---|---|---|
| $12.00M | about 2 at 100% kept | 0.6× |
| $18.00M (the big win) | about 4 at 100% kept | 0.6× |
| $50.00M (market forecast) | about 10 at 100% kept | 1.0× |
Analysts size Smart NICs near $5B, post-quantum at $0.5-1.9B and front-end data-centre switching near $20B a year; shares of 0.4%, 2% and 0.1% give Silicom about $50M of new-line revenue in 2029. On that forecast, $50.00M pays 1.0× from today's value: an Expensive bet on the market case. New lines: AI-inference NICs, post-quantum crypto and white-label switching would be worth about $150.00M, 0.9× what the price pays for it.
To pay 2× (a Fair bet) it would need about $147.73M of New lines: AI-inference NICs, post-quantum crypto and white-label switching revenue (about 30 design wins worth $5M a year); to pay 5× (a Cheap bet), about $429.62M of New lines: AI-inference NICs, post-quantum crypto and white-label switching revenue (about 86 design wins worth $5M a year), at the same multiple.
Each row is New lines: AI-inference NICs, post-quantum crypto and white-label switching alone, plus the rest of the company (Established networking: server adapters, FPGA Smart Cards and edge) counted at about $120.60M.
The steps it has to take
- Done First AI-inference production order received
- Not started Q3 2026 print discloses AI-inference or switching orders with a counterparty, volume and term due 2026-10-29
- In progress First white-label-switching production orders, expected before year-end due 2026-12
- In progress FY2026 AI-inference revenue lands at the guided $3-4M due 2027-02
- Not started FY2027 guidance puts AI inference near $10M due 2027-02
- Not started Post-quantum win deploying at its anticipated $3M a year due 2027-12
The detail behind the label
The evidence under this - 5 claims
The specific dollar figure is management's own guidance with no independent verification of the underlying customer commitments (see C11's binding status).
- 2026-10-29editedQ3 2026 release (S25) read: it describes AI Inference as 'significant long-term potential upside' but gives no revenue figure, so the Q2 guide stands unrefreshed; valid_through rolled to the 20-F
The 20-F's own risk factors state that Design Wins do not guarantee purchase volumes and that ramp timing is outside the company's control; FY2024's 53% revenue collapse (C04) is a prior period where previously expected volumes did not materialize on schedule, so a 'potential'/'anticipated' figure has a demonstrated history of not converting as guided.
- 2026-10-29editedQ3 2026 release (S25) raised full-year guidance to $96-98M and set Q4 at $27-28M; the $93-95M and Q3 $25-26M figures it replaces were reached and passed
Revenue growth has genuinely reaccelerated for four consecutive quarters (17% to a guided 66%), which is an observed trend, not only a narrative, and partly justifies a re-rate independent of the AI-inference story.
- 2026-09-27editedprice drift re-strike: EV/Sales re-read at 3.77x TTM (from 2.70x at 2026-09-03) with the vendor EV of C27; the 1.34x median was mislabelled 'three-year' - the pack takes it over six prior fiscal years
- 2026-10-29editedQ3 2026 balance sheet (S25): the same $107M of working capital plus securities, but a tenth of it moved from cash into inventory
Where we disagree with the market
A cheap, recovering hardware business with explosive AI upside on top: a turnaround plus a free bet.
The price already sits far higher on sales, ahead of the evidence: AI is about 4% of revenue, no win is firm, and insiders sold.
The gap shows only when the call is read against the press release. The release calls AI a game-changer; on the call the number is $3M to $4M, and the order details are withheld for competitive reasons.
We would be early rather than wrong if 2026's design wins turn into firm orders at their stated size through 2027.
What analysts say, and the evidence
The evidence under this - 3 claims
The specific dollar figure is management's own guidance with no independent verification of the underlying customer commitments (see C11's binding status).
- 2026-10-29editedQ3 2026 release (S25) read: it describes AI Inference as 'significant long-term potential upside' but gives no revenue figure, so the Q2 guide stands unrefreshed; valid_through rolled to the 20-F
The 20-F's own risk factors state that Design Wins do not guarantee purchase volumes and that ramp timing is outside the company's control; FY2024's 53% revenue collapse (C04) is a prior period where previously expected volumes did not materialize on schedule, so a 'potential'/'anticipated' figure has a demonstrated history of not converting as guided.
Selling shares received from option exercises is ordinary liquidity and does not by itself indicate a negative view of the business; the unchecked Rule 10b5-1 box says the sales were not pre-scheduled under a claimed plan, not why they were made.
- 2026-09-27editedprice drift re-strike: returns and 52-week position re-read at the 2026-09-25 close (+197.4% a year, 1.0% off the high, against +139% and 26.9% off at 2026-09-03); Form 4 XML read for the 10b5-1 flag, which settles the counter-evidence's open point - no sale was under a plan
What could go wrong, and how we would know
The risks
- Wins that never become orders. Every 2026 win is described as potential. If they slip, this year's growth has nothing firm behind it going into 2027.
- Still losing money. Research spending exceeded the profit left after making the cards in 2025, and every quarter through June 2026 was a loss.
- Customers who vanish. One customer went from $26.8M of sales in 2023 to $2.56M in 2025. The top three are 28% of sales.
The evidence under this - 4 claims
The 20-F's own risk factors state that Design Wins do not guarantee purchase volumes and that ramp timing is outside the company's control; FY2024's 53% revenue collapse (C04) is a prior period where previously expected volumes did not materialize on schedule, so a 'potential'/'anticipated' figure has a demonstrated history of not converting as guided.
- 2026-10-29editedQ3 2026 release (S25): GAAP loss narrowed to $0.3M and non-GAAP turned to a $0.3M profit; GAAP loss persists
What we still do not know
- Who placed the first AI production order, how large it is and for how long. The chief executive would not say.
- How big Silicom's niche really is. Nobody measures the custom slice of the market separately.
The evidence under this - 3 claims
The specific dollar figure is management's own guidance with no independent verification of the underlying customer commitments (see C11's binding status).
- 2026-10-29editedQ3 2026 release (S25) read: it describes AI Inference as 'significant long-term potential upside' but gives no revenue figure, so the Q2 guide stands unrefreshed; valid_through rolled to the 20-F
What would change our view
- Unattractive→AttractiveThe inference line becomes material, at a normal multiple
It needs all of:
- ○AI sales reach about $10M in next year's forecast, up from $3M to $4M, with seven to nine new design wins a year.checked at the next results, due in Feb 2027
- ○Reported operating losses end with the promised return to quarterly profit in the second half of 2026.checked at the next results, due in Feb 2027
- ○The company's value falls to about 1.34 times last year's sales, its own normal in earlier years, from 3.77 at the last close.about ≈$18.76 today, 64% below the last close
- Reopens the researchThe funding question is settled
It needs:
- ○Silicom pays for the AI and switching push without selling new shares before those orders are firm.watched as it happens
We would also be wrong if
- The year's design wins fall short of seven to nine, with no reason given.
- A big customer drops away again and full-year growth falls short of the 2026 forecast.
What is coming
- 29 October 2026 Q3 2026 results: Revenue against the $25-26M guide, whether gross margin holds near 30%, and whether AI-inference or white-label orders are disclosed as firm.
Every dated event ahead
- 29 October 2026 Q3 2026 results: Revenue against the $25-26M guide, whether gross margin holds near 30%, and whether AI-inference or white-label orders are disclosed as firm.
- January 2027 Q4 and full-year 2026 results: Whether 2026 lands in the $93-95M guide, which decides the CEO's $83M RSU test, and whether R&D per sales dollar keeps falling.
- April 2027 FY2026 annual report on Form 20-F: New insider holdings after the August sales, whether buybacks resumed or the shelf was drawn, and any customer concentration shift.
- June 2027 2027 Annual General Meeting: The Active Chairman's re-election and the next pay votes, after shareholders rejected executive grants in 2024 and 2026.
Since we wrote this
- 29 OctEarningsSales up 71% and a non-GAAP profit, but no AI numbersRevenue beat the guide, full-year guidance was raised and non-GAAP profit came early; AI inference still has no revenue figure, so Unattractive and Medium hold.
- 5 OctThesis reviewThe bet now names what it ridesThe owner asked that every bet name the macro trend behind it, after BWXT's bet was missing its chief executive's case for nuclear over gas.
- 4 OctThesis reviewOwner's checklist: revenue streams, cost of growth, management, catalystsThe owner's checklist found the thesis silent on who runs Silicom, what growth costs, where the money comes from and what is dated ahead; this pass fills those four blocks from the 20-F and 6-Ks.
- 27 SepPrice moveShares up 32% on no company news; still UnattractiveThe shares are back at their high and cost nearly three times their usual price per dollar of sales. Nothing the company filed moved; the August insider sales were not pre-planned.
- 3 SepThesis writtenUnattractive, medium convictionThe core business is recovering for real, but the AI story behind a 139% rise is about 4% of this year's sales, and no 2026 win is firm.
Also checked 3 times with nothing new to act on.
Sources 25
Sources - 25
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 filing2026-04-28business overview, competition, risk factors, MD&A, income statement/balance sheet, segment note, customer concentration, geographic salesbacksC04C05C07C08C09C12C14C21C23C24C27C29C30C36C38C39C40C41C42C43C44
- S02primarycompany filing2026-07-29Q2/H1 2026 GAAP and non-GAAP results, guidance, balance-sheet summary, management commentarybacksC01C02C03C06C31
- S03primarytranscript2026-07-29SILC Q2 2026 earnings call transcriptprepared remarks and Q&A; the Q&A disclosed AI-inference dollar guidance, shelf-filing rationale, gross-margin outlook and design-win detail
For AI inference, the number to consider for 2026 is roughly $3 million to $4 million. That is the range we expect for this year. ... This was our third post-quantum cryptography design win ... We expect to scale towards an anticipated annual deployment of around $3 million. ... a win with a $5 million per year potential with a Tier 1 global security leader. ... This win triples our expected business with this customer to nearly $10 million in 2027.
backsC10C11C12C13 - S04primarycompany filing2026-06-30confirmation of the first AI-inference production order milestoneread, but no claim rests on it
- S05secondaryvendor data2026-09-03SILC research pack (FMP/IBKR/FINRA via research.py)price/returns, insider Form 4 activity, FINRA short interest, EPS-beat history, vendor quarterly/annual fundamentals (advisory, cross-checked against S01)backsC26C31
- S06primaryregulator2026-09-03SEC EDGAR filing index for SILC (CIK 0000916793)confirms an EFFECT filing (registration statement effectiveness) dated 2026-06-30backsC13
- S07primarynews2026-09-03News digest batch for SILC (news_digest.py)five kept headlines (all company-issued PRNewswire releases) with measured next-session price/volume reactions
Silicom Secures $5M+/Yr Design Win with Tier-1 Cloud-Based Cyber Security Customer for 5G-Enabled Edge Networking Device
backsC26 - S08primarycompany filing2025-12-18Broadcom Inc. 10-K AVGOBroadcom's own description of its business, confirming an immaterial-perimeter mismatch with Silicom's product linebacksC16
- S09primarycompany filing2026-03-11Marvell's own description of its business, confirming an immaterial-perimeter mismatch with Silicom's product linebacksC17
- S10primaryindustry data2026-09-03independent Smart NIC market-size series and analyst framing of Nvidia's competitive threat; the article's own original publish date is not stated on the page, so 'date' is set to the date it was retrieved
the Smart NIC market will grow at a 26% compound annual growth rate, from $270M in 2020 to $848M by 2024
backsC19 - S11primarycompany release2026-09-03Nvidia's own description of BlueField DPU capability and its OEM-partner distribution model; the release's own original publish date was not confirmed in the fetched excerpt, so 'date' is set to the date it was retrieved
offload critical networking, storage and security tasks from CPUs
backsC18 - S12primarygovernment data2026-06-22the dated federal PQC migration mandate behind the machine.current driver
transition all HVAs and high impact systems to use PQC for key establishment by December 31, 2030 ... transition all HVAs and high impact systems to use PQC for digital signatures by December 31, 2031 ... require covered contractors to comply by December 31, 2030, with NIST's FIPS, including all applicable FIPS incorporating PQC compliant algorithms. Executive Order 14412, signed June 22, 2026.
backsC20 - S13primarycompany filing2026-03-16the audited FY2025 financial statements and ICFR management report filed six weeks before the 20-F; carries no FY2026 guidance (Exhibit 99.2, the MD&A, was searched for 'guidance'/'outlook'/'2026' and returned no match) - guidance for FY2026 was first disclosed later, in the Q1 and Q2 2026 earnings releasesbacksC32C39
- S14secondaryvendor data2026-09-27SILC research pack (FMP/IBKR/FINRA via research.py), refreshed at the 2026-09-25 closeprice and returns at the 2026-09-25 close, 52-week range, vendor market cap and enterprise value, TTM EV/Sales against the median of prior fiscal years, FINRA short interest settled 2026-09-15, insider rosterbacksC22C25C27
- S15primarycompany filing2026-08-28the Rule 10b5-1 checkbox on each open-market sale: <aff10b5One>0</aff10b5One> on all four August filings, and none of the 15 open-market sales in the trailing 365 days flagged as under a planbacksC22
- S16primaryindustry data2024-02-01an independent analyst's 2028 size for the Smart NIC market and its split in words between hyperscale custom designs and the open-vendor rest of the market; read through ir_fetch
we predict the Smart NIC market to exceed $5 billion by 2028. ... The hyperscale cloud service providers will drive rapid development in server connectivity, including increasing the adoption of customized Smart NICs and DPUs ... The rest of the market, which lags the hyperscale market in the deployment of these technologies, will still be an attractive market for the open vendors offering off the shelf solutions
backsC33 - S17primaryindustry data2024-08-01the same firm's raised outlook six months later, and that the raise came from AI back-end connectivity; delloro.com answered 307 to ir_fetch and press_fetch, so the wire copy of the firm's own release was read through ir_fetch
we project the Ethernet Adapter and Smart NIC market to exceed $16 billion by 2028. We raised our market outlook significantly because of strong requirements for back-end Ethernet server connectivity to support the scale out of AI server clusters.
backsC33 - S18primaryindustry data2024-06-30ABI Research chart data - Post-Quantum Cryptography Revenue (from PT-2994, PQC Readiness in HSM Markets, 2Q 2024) ABI Researchan independent analyst's PQC solutions market to 2028 and its timing against the 2030 migration deadlines; the page dates the research 2Q 2024 without a day, so 'date' is the quarter's end; read through ir_fetch
The Total Addressable Market (TAM) for Post-Quantum Cryptography (PQC) solutions is forecast at US$246 million by the end of 2024, and doubling by 2028, reaching US$530 million. ... Market uptake will accelerate significantly between 2026 and 2028, as organizations increase spending on PQC solutions in order to align with national migration policy requirements for PQC readiness by 2030.
backsC34 - S19primaryindustry data2024-09-16MarketsandMarkets press release - Post-Quantum Cryptography (PQC) Market worth $1,887.9 million by 2029, PR Newswire MarketsandMarketsa second, much higher PQC market forecast to 2029, recorded as the conflicting figure; read through ir_fetch
The Post-Quantum Cryptography (PQC) Market size is estimated to grow from USD 302.5 million in 2024 to USD 1,887.9 million by 2029 at a Compound Annual Growth Rate (CAGR) of 44.2% during the forecast period
backsC34 - S20primaryindustry data2024-07-23an independent five-year size for front-end data-centre switching and the open-software (disaggregation) penetration it expects outside the hyperscalers; read through ir_fetch
Ethernet data center switches deployed in non-accelerated infrastructure, or "Front-End Networks", are forecast to generate more than $100B in sales over the next five years. ... SONiC (Software for Open Networking in the Cloud) adoption is expected to accelerate, achieving a 10-20 percent penetration rate in Tier 2/3 Cloud service providers and large enterprises by 2028.
backsC35 - S21primarycompany filing2026-02-27the scale of a white-label switch rival Silicom's 20-F names: HPS revenue of $5.0 billion in 2025, up 81%, driven by networking switch programs with hyperscalers; read through filing_fetchbacksC36
- S22primarycompany filing2026-06-25which resolutions passed (Orbach re-elected to 2029, both Eizenman salary rises), the shareholders' rejection of RSU grants to the CEO (38,333) and the Active Chairman (42,000), and the board's override of the CEO grant with a forfeiture condition unless FY2026 revenue is at least $83 millionbacksC42C44
- S23primarycompany filing2026-09-28sales by geography and the ultimate customers above 10% for H1 2026, expense lines as a percentage of sales against H1 2025, the treasury-share count unchanged at 2,087,132 and 5,747,611 shares outstanding at June 30, 2026, cash used in operations of $16.2 million on an inventory buildbacksC38C39C43
- S24primarycompany filing2026-10-01the date of the Q3 2026 release (October 29, before the market opens) and the 9:00am ET callbacksC45
- S25primarycompany filing2026-10-29Q3 2026 revenue, gross profit, GAAP and non-GAAP results, the unaudited balance sheet at September 30, 2026, the Design Win count and the raised 2026 guidance
revenues for the fourth quarter are expected to reach $27 - $28 million, representing 66% year-over-year growth at the upper end of the guidance. As a result, full-year 2026 revenue projections are being raised once again to $96 - $98 million
backsC02C03C06C10C12C13C46