Damnang
· Damnang’s Substack
· June 01, 2026 at 22:34
· ⏱ 8 min read
| Read on Substack ↗
Summary
IBM possesses a genuine technical moat in AI inference through its Spyre chip (75W, 128GB, 4.2 TOPS/W), integrated into mainframes for secure, low-latency transactions, plus software lock-in via Red Hat and watsonx ($24.6B ARR). However, the recent 30% stock surge in two weeks was driven by a resurfaced Trump video and a quantum announcement, not fundamentals, and the author questions whether the rally justifies an AI-era re-rating given unchanged business metrics.
•IBM's Spyre inference chip is a 75W PCIe card with 128GB LPDDR5, 204 GB/s bandwidth, and 4.2 TOPS/W, leading the NVIDIA L4 in efficiency within the single-slot inference category.
•Spyre took eight years and five generations of test chips, reflecting a positional moat in regulated finance/government where data must never leave the mainframe.
•The z17 mainframe (Telum II processor) includes an on-chip AI accelerator and a DPU to handle light inference in milliseconds and offload heavy models to Spyre cards (up to 48 cards, 6.1 TB accelerator memory).
•IBM's software ARR reached $24.6B, up 10% YoY, anchored by Red Hat OpenShift (~$2B ARR) and watsonx, creating high switching costs.
•IBM stock dropped 13% in a single day in February 2026 (worst in 25 years), then surged 30% in two weeks after a quantum foundry announcement and a viral Trump video from December 2025, despite no change in fundamentals.
•The inference chip market is a price war among d-Matrix, Groq, Furiosa, Rebellions, and Qualcomm, but IBM avoids that fight by positioning Spyre inside mainframes only.
Article directly compares IBM's Spyre to NVIDIA's L4 inference card, showing Spyre leads in TOPS/W (4.2 vs lower) and has 128GB capacity vs 24GB, highlighting a competitive gap in the low-power single
Article directly compares IBM's Spyre to NVIDIA's L4 inference card, showing Spyre leads in TOPS/W (4.2 vs lower) and has 128GB capacity vs 24GB, highlighting a competitive gap in the low-power single-slot inference niche, though IBM is not playing in NVIDIA's core GPU market.
Risk: Niche comparison may not translate to revenue impact; NVIDIA's dominance in high-power training/inference remains unchallenged.
The article validates IBM's three-layer technical moat (Spyre chip, mainframe integration, software lock-in) as real and difficult to replicate, particularly for regulated industries needing on-premis
The article validates IBM's three-layer technical moat (Spyre chip, mainframe integration, software lock-in) as real and difficult to replicate, particularly for regulated industries needing on-premise secure AI inference, supporting a long-term thesis despite near-term valuation caution.
Risk: The author notes the stock may be expensive after the 30% surge and that 'being solid and being a good buy right now are two different things' — implying near-term downside risk.