7 Best Small-Cap AI Infrastructure Stocks — October 2026

AI-infrastructure demand did not cool through the summer of 2026 — it accelerated. The four largest hyperscalers (Amazon, Microsoft, Alphabet, Meta) now guide combined calendar-2026 capital expenditure to roughly $700–730 billion, nearly double 2025's ~$410 billion, and Amazon raised its own plan to about $220 billion on July 30 while explicitly citing higher memory prices — AWS grew 37% year over year in Q2, its fastest rate since 2021. Memory has become the cycle's defining bottleneck: all three merchant DRAM makers describe 2026 HBM output as effectively sold out, and TrendForce projects supply-demand gaps of 4.9% in DRAM, 4.2% in NAND and 5.1% in HBM — the widest since 2011. In optical interconnect, 1.6T transceivers were commercialized this year, and the co-packaged-optics transition has begun in parallel.

The small-cap layer is where those constraints turn into actual revenue lines: the optical DSPs inside 800G modules, the interconnect IP inside AI SoCs, the persistent memory in storage accelerators, the power silicon in AI servers, the test handlers that qualify HBM. This is the successor to our April 2026 AI-infrastructure list, rebuilt score-first from our 2,200+ ticker universe on the September 6 snapshot. Four names return (Aeluma, Arteris, MaxLinear, Everspin); Ouster (lidar, not datacenter infrastructure), inTest (revenue declining) and M-tron (primarily defense RF) drop out; GSI Technology, Alpha & Omega and Cohu join as compute-in-memory, AI-server-power and test/HBM-inspection sub-segments. Five of the seven sit under $2 billion in market cap — MaxLinear ($5.25 billion) returns for continuity with the April list, and Cohu ($2.14 billion) joins for its test-bottleneck exposure.

---

Why the Small-Cap AI Infrastructure Setup Is Different This Cycle

1. Hyperscaler capex nearly doubled instead of plateauing. The roughly $700–730 billion of combined 2026 guidance is not a forecast — it is stated spending plans from four buyers, revised upward mid-year. When the buyers of compute grow their budgets that fast, the sub-suppliers of interconnect, memory, power and test capacity get pulled along regardless of which GPU wins.

2. Memory moved from cyclical recovery into outright allocation. 2026 HBM output is described as effectively sold out by all three merchant DRAM suppliers, TrendForce's projected supply-demand gaps (4.9% DRAM, 4.2% NAND, 5.1% HBM) are the widest since 2011, and server DRAM pricing has risen sharply through the year. Shortages ripple down the toolchain: Cohu reports test-cell utilization back at 80% and orders up 57% year over year — an early-cycle signal for the equipment chain.

3. Small-cap suppliers crossed from design-win announcements into volume revenue. That is the structural change versus our April list. 1.6T transceivers shipped commercially in 2026, NVIDIA cites 1.6T link power falling from ~30W with pluggable transceivers to ~9W with co-packaged optics, and MaxLinear's infrastructure segment grew 145% year over year in Q2 on its PAM4 optical DSP ramp. Announcement-stage stories are becoming income statements.

The honest bear case: four of the seven names below carry SPECULATIVE grades, and the scores themselves need reading with care — Aeluma's +407.6% revenue growth is a fiscal-2025 base-effect artifact, not a run-rate, and MaxLinear's "net cash" runway label sits on a balance sheet that actually carries net debt. Customer concentration is endemic at this size (a single $26 million order moved Cohu's bookings; MaxLinear's DSP sockets sit with a handful of module makers), and every name here is downstream of the same four capex budgets. One quarter of hyperscaler discipline would hit optical, memory, power and test simultaneously.

The Names

ALMU — Aeluma, Inc.

Score: 81.3 (EXCELLENT) | Market cap: $0.24B | Revenue YoY: +407.6% | Rule of 40: 361.7 | Cash runway: 38 months

Aeluma, a UC Santa Barbara spinout, grows compound-semiconductor (III-V) materials on large-diameter silicon wafers — a platform aimed at quantum-dot lasers for silicon photonics, datacom, sensing and defense: the photonics layer of AI datacenter interconnect. Read the list-topping 81.3 score correctly: the +407.6% growth and 361.7 Rule of 40 reflect the fiscal-2025 jump off a tiny fiscal-2024 base, while the three most recently reported quarters were flat at $1.2–1.4 million each — almost entirely government R&D contracts — and management guides full fiscal 2026 to just $4–6 million. The pipeline is real and non-dilutive: six development contracts totaling roughly $5 million, more than $4 million of that from U.S. government agencies including NASA and DoD. But at roughly 45x sales this is a pre-commercial platform story, and dilution ran 15.4% over the trailing 12 months. Fiscal 2026 results are scheduled for September 16. → See full ALMU score card

AIP — Arteris, Inc.

Score: 66.1 (SOLID) | Market cap: $1.02B | Revenue YoY: +22.3% | Rule of 40: -24.7 | Cash runway: net cash

Arteris licenses network-on-chip (NoC) interconnect IP — the on-die traffic network connecting compute, memory and I/O blocks inside AI SoCs and chiplets — which is why gross margin sits at 90.2%. Q2 2026 was a record: revenue of $24.1 million (+46% year over year), annual contract value plus royalties of $99.5 million (+44%), remaining performance obligations of $135 million (+36%). Management raised full-year guidance and pointed to possible non-GAAP operating profitability as early as Q4 2026; the Cycuity acquisition added hardware-security IP since our April list. The -24.7 Rule of 40 records TTM GAAP losses, not the trajectory, and here the net-cash label is genuine: $122 million of cash against $9 million of debt. The caveat: license deal flow is lumpy, and at ~12x sales the market already prices continued 40%+ bookings growth. Q3 results are expected November 3. → See full AIP score card

MXL — MaxLinear, Inc

Score: 65.2 (SOLID) | Market cap: $5.25B | Revenue YoY: +29.7% | Rule of 40: 7.8 | Cash runway: net cash

MaxLinear is the interconnect-silicon name on this list: its Keystone 5nm PAM4 optical DSP (100G per lane) ships into 800G/1.6T optical modules and active electrical cables, with management attributing the ramp to roughly 40% lower power than competing DSPs. Q2 2026 revenue was $168.8 million (+55% year over year), infrastructure revenue was about $85 million (+145%), and the 2026 optical datacenter outlook was raised to $210–230 million. One thing the metrics line gets wrong in spirit: the "net cash" runway label reflects positive free cash flow, but the balance sheet actually carries about $83 million of net debt ($64.8 million of cash against $148.2 million of total debt) — do not read it as balance-sheet strength. Keystone's sockets are concentrated among a handful of module and hyperscaler customers in a PAM4 market dominated by Marvell and Broadcom. Q3 results are expected October 22. → See full MXL score card

MRAM — Everspin Technologies, Inc.

Score: 59.0 (SPECULATIVE) | Market cap: $0.38B | Revenue YoY: +9.5% | Rule of 40: -2.3 | Cash runway: net cash

Everspin is the only public pure-play in MRAM — persistent magnetic memory used as fail-safe write buffers and log memory in datacenter RAID and storage accelerators, plus industrial and aerospace systems. Q2 2026 was a record: revenue of $18.7 million (+42% year over year), MRAM product sales of $15.3 million, gross margin of 53.9% and non-GAAP EPS of $0.11, with Q3 guided to $19.5–20.5 million; the TTM-based +9.5% growth and -2.3 Rule of 40 lag that acceleration by construction. The AI angle got more concrete since spring: an MoU with MaxLinear — also on this list — pairs CXL-attached MRAM with storage accelerators for metadata and write-buffer caching, though that is a memorandum, not a product ramp. Net cash is verified ($43.9 million against $2.8 million of debt), but at an ~$18 million quarterly run-rate, one delayed program shifts a whole quarter. Q3 results are expected November 11. → See full MRAM score card

GSIT — GSI Technology, Inc.

Score: 58.7 (SPECULATIVE) | Market cap: $0.20B | Revenue YoY: +22.4% | Rule of 40: -47.1 | Cash runway: 51 months

GSI Technology is a legacy SRAM maker — including radiation-hardened SRAM for aerospace and defense — attempting a transition to Gemini-II, its associative processing unit: compute-in-memory silicon that runs AI inference inside the memory array to minimize data movement. The +22.4% TTM growth is largely the SRAM business recovering, including AI-related SRAM orders, not APU commercialization. 2026 brought two external validations: a $2.0 million U.S. Army xTech Phase-II SBIR contract for a ruggedized edge-AI platform on Gemini-II, and Phase I of a Hsinchu County, Taiwan smart-city project with a multi-million-dollar follow-on structure. Note how the roughly 51-month runway is funded — dilution ran 32.1% over the trailing 12 months — and the -47.1 Rule of 40 keeps this deeply lossmaking name SPECULATIVE. If the pilots do not convert into volume orders, GSI remains a subscale SRAM company. Fiscal Q2 2027 results are expected October 29. → See full GSIT score card

AOSL — Alpha and Omega Semiconductor L

Score: 58.3 (SPECULATIVE) | Market cap: $0.74B | Revenue YoY: +5.9% | Rule of 40: 1.8 | Cash runway: net cash

Alpha & Omega supplies power MOSFETs, power ICs and modules — the power-delivery picks and shovels inside AI servers, racks and their cooling fans. Fiscal Q4 2026 (June quarter) revenue was $170.4 million, down 3.5% year over year, but the mix is shifting fast: advanced-computing revenue rose 35% sequentially to a record 31% of the computing segment, and management guided September-quarter AI and server revenue up more than 60% sequentially with total revenue around $176 million. The score reflects what this still is — a modest-growth (+5.9% TTM), thin-margin (23.1% gross) cyclical near 1.1x sales, and the June quarter still produced a non-GAAP loss of $0.13 per share. The balance sheet is genuinely net cash ($181 million against $28 million of debt), and the share count actually shrank 18.5% over the trailing 12 months. Fiscal Q1 2027 results are expected November 9. → See full AOSL score card

COHU — Cohu, Inc.

Score: 56.4 (SPECULATIVE) | Market cap: $2.14B | Revenue YoY: +12.7% | Rule of 40: -0.4 | Cash runway: net cash

Cohu sells semiconductor test handlers, inspection systems and interface products; its AI exposure runs through thermal test handlers for AI processors and HBM inspection — test capacity being one of the under-covered bottlenecks of the memory supercycle. Q2 2026 revenue was $149 million (+38% year over year), the 2026 high-performance-computing outlook was raised from $80–100 million to $100–110 million, total orders rose 57% with a $26 million single-customer Eclipse order, and test-cell utilization returned to 80% — a classic early-upcycle marker. About 53% of revenue is recurring consumables and kits, which cushions the cyclicality the -0.4 Rule of 40 records from the TTM trough. The net-cash label checks out on a net basis (about $167 million), though the $331 million gross debt load is the largest on this list, and equipment orders reverse fastest when capex pauses. Q3 results are expected October 29. → See full COHU score card

---

How to Think About Position Sizing

Seven names, three distinct trades. The volume-revenue cyclicals — MXL, AOSL, COHU — have real AI revenue today: an optical DSP ramp, an AI-server power mix-shift, a test-equipment upcycle. Each also carries a cyclical core (legacy broadband, PC and consumer power, equipment orders) that can swamp the AI line in a downturn, and only MaxLinear's AI exposure is already the growth engine rather than the kicker. The high-margin compounder — AIP — is the structurally cleanest business here, 90.2% gross margin license-and-royalty economics with genuine net cash, priced accordingly and lumpy by nature. The pre-commercial options — ALMU and GSIT, plus MRAM's CXL angle — are where the score grades matter most: Aeluma and GSI are funded by dilution (15.4% and 32.1% over the trailing 12 months), and their theses hinge on contracts converting into commercial volume. Hold those at option-sized weights or not at all.

What the mix does not diversify: hyperscaler capex itself. All seven sit downstream of the same handful of budgets, and the earnings calendar clusters tightly — Aeluma prints September 16, then all six remaining names report between October 22 and November 11. For the industrial side of the same buildout — cooling, piping, power — see our September 2026 data center infrastructure list.

How We Scored

Our model rates every US small-cap stock across eight fundamentals: revenue growth, gross margin, cash runway, debt/equity, P/S ratio, Rule of 40, insider ownership, and 12-month dilution. Sector-adjusted where appropriate, refreshed regularly — this list uses the September 6 snapshot across our 2,200+ stock universe. See methodology and track record or browse all small-cap Technology stocks.

Related Coverage

• Best Small-Cap AI Infrastructure Stocks — May 2026 Update
• 8 Best Small-Cap Data Center Infrastructure Stocks — September 2026
• 7 Best Small-Cap Edge AI Semiconductor Stocks — June 2026
• 7 Best Small-Cap Nuclear & SMR Stocks — September 2026

Data as of September 11, 2026. Updated monthly. Past performance does not guarantee future results. Not investment advice.