Theme
AI Inference
Custom silicon and connectivity for inference workloads.
0.421avg corr (252d)
9stocks
Weak co-movementStocks · key Fib levels
- ALAB$319.74last$186.00fib+71.9%vs fibanchored by Amazon AWS · Microsoft Azure · Google Cloud · Meta · NVIDIA · Dell · Supermicro
- ▸2026 Q1: Scorpio P-Series PCIe Gen6 fabric switches sampling to hyperscalers; production ramp guided H2 2026.
- ▸2025 Q4: Disclosed design wins on NVDA GB300 NVL72 platform - Aries + Scorpio attach.
- ▸2025: Sole-source PCIe retimer on multiple AWS Trainium2 / Inferentia server SKUs.
- ▸2025: COSMOS software suite extended to fleet-wide telemetry across all four product lines.
- ▸2024 Q1: IPO at $36 (Mar 2024), priced above range; first pure-play AI interconnect public listing.
- AMD$500.94last$253.00fib+98.0%vs fibanchored by Microsoft · Meta · Oracle · OpenAI · Google · Amazon
- ▸2024-12: $MSFT Azure ND MI300X VM series GA - first hyperscaler MI300 production deployment
- ▸2025-02: $META disclosed MI300X internal inference workload at scale (Llama serving)
- ▸2025-06: $ORCL OCI MI300X bare-metal instances live; multi-billion commitment disclosed
- ▸2025-10: MI325X launch - 256GB HBM3E, shipping to hyperscaler anchors Q4 2025
- ▸2026-Q1: MI350 series (CDNA4, 3nm) sampling; production ramp H2 2026
- ▸2026-Q1: ZT Systems acquisition closed - rack-scale AI systems integration in-house
- ▸2026-04: EPYC Turin (5th gen, Zen 5) shipping; cited in $ORCL, $GOOG, $AMZN earnings as standard fleet CPU
- ARM$262.01last$185.00fib+41.6%vs fibanchored by NVIDIA · Apple · AWS · Google · Microsoft · Meta · Qualcomm · MediaTek · Samsung
- ▸2024-02 - Microsoft Cobalt 100 announced - first MSFT-custom Arm server CPU; Azure deployment broadens
- ▸2024-04-09 - Google Axion (Arm v9) announced for GCP general-purpose compute
- ▸2024-05 - Apple M4 launched (3nm, all-Arm v9 across M-series, A18 Pro across iPhone, S-chip across watch)
- ▸2024-11 - AWS Graviton4 GA; Arm Neoverse V2 powers 50%+ of EC2 compute hours within targeted SKUs
- ▸2025 - NVIDIA Grace + Blackwell shipping at scale (Arm v9 Neoverse V2 paired with every Blackwell GPU); Rubin generation continues the Arm pairing
- ▸2025-09 - Google Axion GA on GCP
- ▸2026-Q1 - Apple M5 (3nm successor, Arm v9.x) + A19 Pro launch maintain Apple's 100% Arm-silicon stack
- ▸2026 - Compute Subsystems (CSS) chiplet IP in production-ready hyperscaler programs; royalty per AI accelerator unit rises as CSS-licensed designs ship
- AVGO$374.45last$336.00fib+11.4%vs fibanchored by Alphabet · Meta Platforms · Apple · undisclosed third hyperscaler
- ▸**Nov 2023** - Closed $69B VMware acquisition; began subscription conversion + portfolio trim
- ▸**Sep 2024** - Disclosed third large hyperscaler XPU customer in development (unnamed)
- ▸**Dec 2024** - Guided AI semi SAM to $60-90B by FY27 across 3 hyperscaler customers
- ▸**2024-2026** - $GOOGL TPU v5/v6/v7 silicon co-design (multi-generation)
- ▸**2024-2026** - $META MTIA v1/v2 inference + training silicon partner
- ▸**2026** - Multi-year $AAPL wireless component extension signed; RF baseline through FY28+
- ▸**Ongoing** - Tomahawk 5 (51.2T) shipping; Tomahawk 6 (102.4T) sampling for AI scale-out fabrics
- CBRS$180.46last- fib$24.6B backlog750 MWanchored by G42 · MBZUAI (Mohamed bin Zayed University of AI) · OpenAI · Amazon Web Services
- ▸2024-09 - First S-1 filed; later withdrawn after CFIUS opened a review of G42's minority stake
- ▸2025-10 - CFIUS review concluded after G42 holding restructured to non-voting shares; path cleared for listing
- ▸2025-12-24 - NVIDIA's $20B Groq asset/license acquisition (read-through: frontier inference market partitioning into a specialist tier; NVDA buying in rather than ceding the lane)
- ▸2026-01-14 - OpenAI deal signed: 750 MW of low-latency compute through 2028. Initial press value $10B; subsequent MRA disclosures place binding commitments at >$20B with expansion provisions up to 2 GW by 2030
- ▸2026-02 - Hiive private secondary implied ~$23B valuation (IPO 3 months later doubled this)
- ▸2026-03-13 - AWS announced Cerebras as first cloud provider for disaggregated inference on Amazon Bedrock; Trainium + WSE pairing claims "5× more high-speed token capacity in the same hardware footprint"
- ▸2026-05-13 - IPO priced $185 (above $150-160 range, walked up from original $115-125); book reportedly 20× oversubscribed
- ▸2026-05-14 - Day 1 open $350, close $311.07 (+68.2%); $5.55B raised on 30M shares; $48.8B fully-diluted valuation; 2026's largest US tech IPO
- ▸2026-Q3 (est) - Accelerated lockup: 60M+ shares unlock by Q2 2026 earnings release, ~2× IPO float hitting tape inside 90 days (vs standard 180-day expiry mid-November 2026)
- ▸Snapshot 2026-05-18: price ~$311, mcap ~$48.8B fully-diluted (~157M sh × $311), 2025 revenue $510M (+76% YoY from $290M), 2025 GAAP net income $237.8M but operating loss $145.9M
- ▸Customer concentration: G42 (24% of 2025 rev, was 85% in 2024) + MBZUAI (62%) = ~86% UAE-linked, both flagged as related parties in the S-1; apparent 2025 diversification away from G42 was reallocation between connected Abu Dhabi entities, not new customer acquisition
- ▸Backlog concentration: ~$24.6B disclosed, of which ~80% is OpenAI; revenue recognition gated by OpenAI's own datacenter buildout pace (power + zoning, not chips)
- ▸Setup tension: first specialist inference accelerator on the public tape, but the entire forward business is explained by three customers (G42, MBZUAI, OpenAI) through 2028; GAAP profitability is non-operating (the $237.8M net income line is not a sustainable run-rate while operating loss continues); fast lockup creates structural supply pressure ~Q3 2026; inference TAM is contestable and workload-specific (21× B200 advantage on Llama 3 70B reasoning doesn't generalize to training, embeddings, or image gen)
- ▸Editorial deep-dive: /articles/cerebras-cbrs-ipo
- CRDO$207.97last$148.00fib+40.5%vs fibanchored by Microsoft · Amazon Web Services · Intel Foundry · Samsung Foundry · TE Connectivity
- ▸**Microsoft** - anchor AEC customer; multi-year ramp into Azure AI back-end fabric; disclosed >40% of revenue concentration in FY25 10-K
- ▸**Amazon (AWS)** - second hyperscaler AEC design win; Trainium/Inferentia cluster fabric ramp through CY26
- ▸**Third US hyperscaler** - disclosed in FY26 guide as ramping; widely believed to be $META or $GOOGL
- ▸**Intel Foundry** - 5nm/3nm SerDes IP licensing engagement
- ▸**Samsung Foundry** - SerDes IP partnership on advanced nodes
- ▸**TE Connectivity** - joint AEC + connector reference designs for OCP/OAI form factors
- INTC$96.98last$60.00fib+61.6%vs fibanchored by Amazon Web Services · Microsoft · Dell · HP · Lenovo · Department of Defense · IBM Cloud
- ▸2024-02: CHIPS Act award - $8.5B grant + $11B loan for Arizona/Ohio/New Mexico fabs
- ▸2024-09: Amazon Web Services ($AMZN) - multi-year 18A custom AI fabric chip + Xeon 6 supply deal
- ▸2024-09: Microsoft ($MSFT) - 18A custom chip design win (disclosed at Direct Connect 2024)
- ▸2025-01: Department of Defense - Secure Enclave / RAMP-C trusted-fab Phase 3 award
- ▸2025-Q2: Altera carve-out - Silver Lake acquires 51% at $8.75B EV
- ▸2026-Q1: Gaudi 3 - IBM Cloud ($IBM) and Dell ($DELL) as launch infrastructure partners
- MRVL$188.30last$126.00fib+49.4%vs fibanchored by AWS · Microsoft · Google · Meta · Hyperscalers (undisclosed third + fourth)
- ▸2021 - Inphi acquisition closed ($10B) - coherent-optical DSP leadership
- ▸2024 - AWS Trainium2 + Trainium3 custom-silicon programs ramping
- ▸2024 - Microsoft Maia 100 / successor program co-designed with Marvell
- ▸2025 - Coherent-optical DSPs inside 800G + 1.6T pluggables shipping in volume
- ▸2025-Q4 - Hyperscaler AI silicon revenue mix climbing; legacy networking mix shrinking as share of revenue
- ▸2026-Q1 - Quarterly print; AI silicon revenue commentary the watch
- ▸2026 - Multi-generation custom AI ASIC programs visible 2-3 years forward
- RMBS$101.42last- fibanchored by Micron · SK Hynix · Samsung · AMD · Marvell
- ▸**2026-Q1**: DDR5 server chipset revenue at multi-year high; MRDIMM design wins ramping into 2H26.
- ▸**2026**: HBM4 controller IP licensed to multiple AI accelerator customers (undisclosed hyperscaler ASIC programs).
- ▸**2025**: PCIe 6.0 and CXL 3.0 controller IP shipped - qualified into early adopters.
- ▸**2025**: Security IP design wins in automotive ADAS and post-quantum crypto programs.
- ▸**Ongoing**: Patent license renewals with $MU, $SK Hynix, Samsung - royalty backbone.
Co-movement matrix (252d)
ALAB | AMD | ARM | AVGO | CBRS | CRDO | INTC | MRVL | RMBS | |
|---|---|---|---|---|---|---|---|---|---|
| ALAB | · | 0.48 | 0.41 | 0.49 | 0.32 | 0.64 | 0.37 | 0.48 | 0.50 |
| AMD | 0.48 | · | 0.55 | 0.44 | 0.20 | 0.46 | 0.52 | 0.50 | 0.53 |
| ARM | 0.41 | 0.55 | · | 0.50 | 0.03 | 0.40 | 0.40 | 0.52 | 0.54 |
| AVGO | 0.49 | 0.44 | 0.50 | · | 0.20 | 0.55 | 0.35 | 0.47 | 0.44 |
| CBRS | 0.32 | 0.20 | 0.03 | 0.20 | · | 0.15 | 0.33 | 0.28 | 0.28 |
| CRDO | 0.64 | 0.46 | 0.40 | 0.55 | 0.15 | · | 0.38 | 0.43 | 0.48 |
| INTC | 0.37 | 0.52 | 0.40 | 0.35 | 0.33 | 0.38 | · | 0.43 | 0.51 |
| MRVL | 0.48 | 0.50 | 0.52 | 0.47 | 0.28 | 0.43 | 0.43 | · | 0.58 |
| RMBS | 0.50 | 0.53 | 0.54 | 0.44 | 0.28 | 0.48 | 0.51 | 0.58 | · |
About themes
Themes are editorial groupings - narrative buckets used for technical screens. They exist alongside bubbles, which are validated empirically by capital-flow co-movement.
When a theme's avg correlation is high, the editorial story is also tradeable as a bloc. When it's low, the narrative groups stocks that don't actually move together - useful for thinking, less useful for trading.
Related research
·8 min read
After SpaceX: the 2026 IPO pipeline, ranked by what's actually filed
·8 min read
Broadcom (AVGO): what it does, how it makes money, and why it owns custom AI silicon
·8 min read
What is TSMC (TSM)? How it makes money, and why every AI chip runs through it
·8 min read
Nvidia & AMD's China chip loophole just closed - and the demand math barely moves
Frequently asked questions
- What are the top stocks in the AI Inference theme?
- QuantAbundancia's AI Inference theme tracks 9 stocks ranked editorially: ALAB, AMD, ARM, AVGO, CBRS, CRDO, INTC, MRVL, RMBS.
- Do AI Inference stocks move together?
- The AI Inference theme has a 252-day average correlation of 0.42 across its members - weak co-movement. Editorial themes are narrative groupings, not capital-flow validated; tradability follows correlation, not the story.