Silicom AI Inference Orders Put Small-Cap SILC in the Buildout's Path

A small networking vendor in a very large market

Companies spending heavily on AI infrastructure are starting to push more of that money toward inference, and some of it is reaching small suppliers that were never part of the GPU trade. Silicom Ltd., an Israeli maker of network adapters and data-infrastructure hardware, is a clear example. An August analysis on an investor platform argued that the market values the company cheaply on its existing business alone, so its AI inference work comes almost free. The piece cited a market value of about $250 million, about $55 million in net cash and a $180 price target, which it said was roughly 330% above the share price at the time.7

That argument rests on two things: a core business that is clearly growing again, and an inference opportunity that is still mostly unproven. The coverage supports the first more firmly than the second.

What actually happened

The inference story started on May 5. Silicom said an unnamed "AI infrastructure challenger" had chosen its inference-specific hardware for a proof of concept aimed mainly at Tier-1 hyperscalers. The customer placed initial orders for first-half delivery ahead of trials planned for the second half of 2026.1 Silicom said that if the trials succeed, the customer expects a first full-scale deployment of tens of thousands of units. Each unit would sell for several thousand dollars.9 The company also said it had orders from two AI compute contenders for two separate inference products, plus a development order for a third.2 The stock rose 11.7% the next session.2

On June 30, Silicom announced its first production order for an inference product, with delivery in 2026. It said this lifted expected 2026 inference revenue into the "multi-million-dollar range."5 Shares jumped about 15.7% that day.10 On the second-quarter call, management narrowed that figure to roughly $3 million to $4 million for the year and said it expects further growth in 2027.20 Executives also said they had customized an AI network interface card for a leading inference ASIC and infrastructure vendor. The first evaluation unit has shipped, and initial deliveries are being prepared under a purchase order.17

Silicom explains its pitch in technical terms. It combines adaptable FPGAs with off-the-shelf high-performance networking chips. The company says this helps AI hardware developers get past the "latency wall" and reduces the "hardware lottery" risk of committing to fixed silicon too early.1

The core business is doing the heavy lifting

Silicom's AI headlines get the attention, but its current numbers come from older product lines. Second-quarter revenue rose 59% to $23.8 million, from $15.0 million a year earlier.13 That beat the company's own $20 million to $21 million guidance.17 It also far exceeded a consensus estimate of about $17.4 million.18 Year-over-year growth has sped up from 17% in the fourth quarter of 2025 to 33% in the first quarter and 59% in the second.18 Management raised full-year guidance to $93 million to $95 million, up from $82 million to $83 million.18 It also guided third-quarter revenue to $25 million to $26 million.20

The CEO said the growth came from FPGA products, standard and acceleration adapters, and edge systems. He said switching, post-quantum cryptography and AI inference were not yet meaningful contributors.17 That matters for the valuation case. If the inference business produced nothing, the company would still be on track for more than 50% growth this year.12

Profitability is improving but not there yet. The GAAP net loss for the quarter was $2.1 million, or $0.37 per share, and the non-GAAP loss was $0.9 million.13 Gross margin was 30.4%, near the top of the company's 27%–32% target range.16 Management now expects quarterly non-GAAP profitability in the second half of 2026, earlier than it first planned.13

Where the reporting diverges

Some of the coverage disagrees on details, and some figures are plainly wrong.

Operating expense growth. One transcript says operating expenses rose "only 6.16%" year over year. Another version of the same remarks, and a separate summary, say 16%.161817 The reported figures, $7.2 million rising to $8.3 million, work out to about 15%, so the 16% version is right. Operating leverage is still real, just less dramatic than the lower number suggests.

The switching deal. One transcript summary says a white-label switching win could be worth "$5 billion per year."16 Silicom's own May announcement described a $5 million-per-year design win with a Tier-1 cybersecurity company for a new white-label switch family.12 The billion-dollar figure is an error.

The investor reaction. Coverage of the Q2 results was overwhelmingly positive. Even so, the stock fell 4.37% in premarket trading as investors weighed the continuing losses against the better outlook.18 A filing analysis of the June order was more cautious than the press releases. It noted that Silicom gave no baseline revenue figures, margin data or customer-concentration details, so the order's real financial impact could not be measured.4

The bull-case math. The August investment thesis estimated a $80 billion–$100 billion addressable market for inference. It argued that one customer alone could add $64 million in revenue by 2027, against a 2026 base of about $95 million.7 That roughly matches what a full deployment of tens of thousands of multi-thousand-dollar units would imply. But it depends entirely on a single unnamed customer completing a hyperscaler trial that only starts in the second half of 2026.9 The same thesis lists slowing core momentum, weaker AI infrastructure demand and AI optimism already in the share price as key risks.7

Why the Nvidia backdrop matters

Silicom's timing makes more sense next to what Nvidia has done this year. In May, Silicom's CEO said the industry is moving from capital-heavy training buildouts to large global inference deployments. He argued that general-purpose GPUs are no longer the most scalable route for pure inference, as buyers look for lower total cost per token.9

Nvidia's own product moves support the view that inference needs specialized hardware. At GTC 2026, the company folded Groq's SRAM-based language processing unit into its Vera Rubin platform as the Groq 3 LPX rack, with 256 LPUs per rack, sold alongside the Vera Rubin NVL72.34 The technology came from a $20 billion licensing and talent deal struck in December 2025.35 Rubin CPX, a GPU Nvidia announced in 2025 for long-context inference, was put on hold. A spokesperson said the focus had moved to the Groq-based platform.3438 In August, Nvidia said LPX was in full production, with Nebius as the first cloud customer.40 Executives also said providers can charge more for low-latency tokens.37

Nvidia's move has two implications for small suppliers. On one hand, it confirms that buyers want purpose-built inference hardware and that latency is a problem worth paying to fix. On the other, the biggest supplier in the market is now competing directly for that business. The "challengers" Silicom serves have to win share against an Nvidia stack that is getting more specialized.

The cost pressure behind the shift

Hardware prices explain why hyperscalers are testing alternatives. One mid-2026 pricing survey put a new H100 at $25,000 to $40,000 and an 8-GPU H200 DGX system at roughly $400,000 to $500,000.21 Another estimated eight-GPU HGX B300 servers at $550,000 to $750,000 as of September.22 AI-optimized facilities reportedly cost $20 million or more per megawatt, compared with $7 million to $12 million for conventional builds.21 One estimate found that GPU hardware accounts for only about 35% of five-year cost of ownership. Power, cooling, networking and staffing make up the rest.29

The spending is enormous. Estimates put 2026 capital spending by the largest hyperscalers at about $725 billion to $800 billion, depending on which companies are included.2926 Silicom's full-year inference guidance of $3 million to $4 million is tiny by comparison. That is why the upside looks large on paper, and also why it is fragile.

The reading

The best-supported case for Silicom is about its core business, not AI. Revenue growth is accelerating, results beat guidance, margins are holding near the top of their range, and the balance sheet shows $107 million in working capital and marketable securities with no debt.17 That is already a turnaround story.

The inference business is better viewed as an option than a forecast. The orders are real and the production order shows customers will pay. But the large scenarios depend on unnamed challengers winning hyperscaler trials while Nvidia expands its own inference lineup. The second-half proof of concept will decide whether the $64 million scenario becomes real revenue. Until then, investors are buying a growing networking company that may also benefit from the inference buildout.