22V Research · Jordi Visser
The Inference Inflection — Thematic Basket
A 22-company equal-weight basket spanning chips, memory, networking, data centers, cloud, and edge/embodied-AI enablers — Visser's bet that AI workloads are pivoting from training to real-time inference at scale.
Source report: "The Inference Inflection: Where Real-Time AI Meets Real-World Opportunity"
Published: May 15, 2025
Constituents: 22 (equal-weight)
Archived from ai.22vresearch.com
AI Compute & Inference Chips
GPUs, accelerators, and edge inference silicon.
- NVDA
NVIDIA
Market leader in training & inference GPUs; inference now driving data-center growth.
- AMD
Advanced Micro Devices
Momentum with MI300 and distributed inference models.
- INTC
Intel
Gaudi AI chips gaining traction; Ethernet & inference-optimized silicon on roadmap.
- QCOM
Qualcomm
Dominant at the edge — phones, automotive, IoT — with custom low-power inference chips.
- LSCC
Lattice Semiconductor
FPGA-based low-power inference plays at the edge.
Memory & Storage
HBM, DRAM, and edge memory beneficiaries.
- MU
Micron Technology
Leader in HBM3E and low-latency DRAM used for inference workloads.
- WDC
Western Digital
Drives inference-related data center storage needs.
- STX
Seagate Technology
Supports growing edge/cloud storage requirements tied to inference.
Networking & Optics
Ultra-low-latency, high-bandwidth connectivity for inference.
- CIEN
Ciena Corporation
Optical backbone provider; strong AI data-center build exposure.
- ANET
Arista Networks
Top high-speed networking player for hyperscalers deploying inference.
- MRVL
Marvell Technology
Custom silicon and optical interconnects powering inference networks.
- AVGO
Broadcom
Dominant in networking chips, PCIe, and switching — core enabler of inference scale-out.
Data Centers & Infrastructure
Capacity, power, and cooling for inference-heavy facilities.
- EQIX
Equinix
Hyperscaler growth and latency-sensitive inference demand driving record colocation needs.
- DLR
Digital Realty
Real-time AI workloads shifting inference into new facilities.
- VRT
Vertiv Holdings
Power and cooling infrastructure for inference-heavy data centers.
AI-Optimized Cloud Providers
Large-scale inference customers and usage-based cloud beneficiaries.
- MSFT
Microsoft
Azure's 100-trillion token quarter is the proof point.
- GOOG
Alphabet
Google Cloud inference usage expanding fast.
- AMZN
Amazon
AWS seeing a shift toward inference-driven demand.
Edge & Embodied AI Enablers
Robotics, smart devices, FSD, and the next leg of inference.
- AMBA
Ambarella
Edge AI chips used in automotive and vision systems.
- ON
ON Semiconductor
Inference-capable image sensing and power semis for AI-native systems.
- NXPI
NXP Semiconductors
Automotive AI and edge inference control.
- SYNA
Synaptics
AI-powered interfaces and inference at the edge — phones, consumer, IoT.
Why this basket
Visser's thesis: training-era AI narratives have run their course. The next leg of value creation is inference at scale — perpetual, usage-driven workloads that scale with every user interaction, every agent call, every embodied-AI decision. The basket is intentionally semiconductor-heavy: an equal-weight chart of these names tracks closely to the SOX index.
Each name covers a distinct slice of the inference stack so the basket captures the full buildout: silicon → memory → fabric → real estate → cloud → edge.
*Equal-weight performance vs. SPX cited by Visser in the Oracle, Inference, and the Persistence of Disbelief follow-up report.