AI's Efficiency Dividend
Nidhal Selmi@nselmi
−11.2%
since Oct 6
The idea
I believe the next phase of AI spending moves away from raw training capacity toward making inference cheap, local, routed, and measurable, as usage scales and customers start paying real costs. This take owns public companies positioned for that shift: efficient and custom silicon, edge compute and routing infrastructure, and observability tools that tie AI spend to outcomes, while avoiding leveraged GPU-cloud and data-center buildout plays. It holds 15 positions capped at 12% each with 10% cash, meant to be held for years.
What it owns
- Custom chips and design tools that cut the cost per unit of useful AI work let customers replace general-purpose compute with silicon tailored to their specific inference tasks. · BroadcomIt builds custom AI chips for hyperscalers designed to run specific inference workloads more cheaply than general-purpose GPUs, representing spending shifting toward efficient silicon., Arm HoldingsIt licenses power-efficient chip designs that run most on-device and much data-center inference, earning royalties on the shift to efficient compute across devices and clouds., Taiwan SemiconductorEvery efficient custom chip in this take gets manufactured at its advanced nodes, which are what make each generation of inference silicon cheaper per unit of work., Marvell TechnologyIt pairs custom AI silicon for cloud customers with optical interconnect that moves data between chips efficiently, addressing both silicon design and network efficiency., Cadence Design SystemsIts design software powers the creation of efficient custom chips across the industry, so more custom silicon projects mean more revenue regardless of which specific chips win.29%
- Networks and edge platforms that move and serve AI workloads closer to users cost less to operate than routing everything through centralized cloud data centers. · CloudflareIts Workers AI platform runs inference near users and routes requests to the cheapest capable model, providing the edge routing layer that enables efficient inference spending., Arista NetworksIt makes the Ethernet networking that moves data efficiently between AI clusters, reducing the cost per unit of useful compute rather than requiring more raw capacity., Akamai TechnologiesIts distributed network runs compute and inference closer to users at lower cost than centralized clouds, a direct play on spending shifting toward efficient inference delivery., Credo TechnologyIt sells low-power cables and connectivity chips that cut the energy cost of moving data inside AI systems, reducing the literal power consumption of inference infrastructure.24%
- Software that measures AI spending and performance helps customers prove their AI investments deliver measurable value, which determines what they choose to pay for. · DatadogIts LLM monitoring tools show customers what their AI spending costs and what it returns, making the measurable outcomes that drive decisions about where to allocate AI budgets., DynatraceIts platform ties AI and cloud spending to actual performance outcomes, helping enterprises prove their AI bills buy something useful and measurable., ElasticIts search and vector tools power retrieval-augmented generation, the technique that lets smaller and cheaper models give useful answers by retrieving relevant data instead of brute-forcing larger models.19%
- Chips and platforms that run AI models on devices or at the edge reduce the cost and latency of inference by eliminating the need to send every query to a distant data center. · QualcommIts Snapdragon chips run AI models directly on phones and PCs rather than sending queries to cloud servers, shifting inference spending toward on-device processing., Lattice SemiconductorIt makes small, low-power programmable chips that let factories, cameras and laptops run AI locally on tight power budgets, enabling efficient inference in embedded gear., AmbarellaIt sells chips that run computer vision inference inside cameras or vehicles themselves, which costs less and responds faster than streaming video to a distant data center.18%
- Cash10%