Commodity Tokens, Scale Wins
sam lessin 🏴☠️@lessin
−6.5%
since Sep 30
The idea
I think AI inference is turning into commodity matrix multiplication, so premium token margins get competed away and the winners become the lowest-cost producers at scale plus everyone they depend on. This basket owns the hyperscalers building their own silicon, the custom-chip design and foundry chain behind them, the design and fab tools, the volume inputs like memory and power, and networking and hardware, with the heaviest weight on the scale producers themselves. I deliberately leave out the maker of the premium chips, since that margin is exactly what I think erodes.
What it owns
- Hyperscalers building their own chips become the low-cost producers when AI inference becomes a commodity, competing on volume and efficiency rather than token prices. · Alphabet Inc.Its custom TPUs let it serve AI inference without paying Nvidia's premium, making it the low-cost producer when tokens become commodity matrix multiplication., Amazon.com Inc.AWS pushes Trainium and Inferentia chips to serve compute at commodity prices while maintaining margins on volume, shifting customer workloads away from resold Nvidia GPUs., Meta Platforms Inc.Its MTIA chips let it avoid Nvidia markups on inference for its own products, dropping that premium directly to its cost line as inference scales., Microsoft CorporationIt has distribution to sell AI at scale and is building Maia chips to serve it without the premium-token margins that disappear in commodity markets.29%
- Companies designing and manufacturing custom hyperscaler chips get paid on volume regardless of which company wins the race to build the most efficient silicon. · Broadcom Inc.It co-designs custom ASICs for hyperscalers building their own chips instead of buying Nvidia, a shift that directly drives its revenue., Taiwan Semiconductor ManufacturingIt manufactures every advanced AI chip whether it's a Nvidia GPU or a hyperscaler ASIC, so it gets paid on volume regardless of whose design wins., Marvell Technology Inc.It designs custom AI chips for hyperscalers, directly benefiting from the shift toward in-house silicon as inference commoditizes., Astera Labs Inc.It makes the connectivity stitching custom-silicon clusters together and gets designed into hyperscaler chips regardless of which company builds them., Arm Holdings plcMost custom hyperscaler silicon is built on its architecture, so it collects licensing fees on the chips replacing Nvidia's GPUs.27%
- More companies designing their own chips means sustained demand for the design software and manufacturing equipment needed to build them, independent of token prices. · Synopsys Inc.It supplies the design software that every new custom AI chip is built in, earning more as hyperscalers design their own silicon., Cadence Design SystemsEvery custom hyperscaler chip design uses its software, so it collects licensing revenue as more companies build their own inference silicon., Applied Materials Inc.It sells the equipment that manufactures every AI chip, earning on fab volume no matter whose logo is on the die., KLA CorporationIn commodity markets, yield becomes everything, and it supplies the inspection tools fabs use to maintain the high yields that make them the low-cost producer., Amkor Technology Inc.Custom AI chips need its advanced packaging technology regardless of token prices, and it gets paid per package as hyperscaler volumes grow.16%
- Memory, power, and cooling are consumed based on how many tokens are produced, not what they sell for, so suppliers are paid on volume in a commodity market. · Micron Technology Inc.Every inference token touches memory, and it sells that memory by the gigabyte regardless of what tokens cost, earning on volume alone., Constellation Energy Corp.Commodity tokens at scale require enormous amounts of always-on electricity, driving demand for the nuclear power it supplies under long-term data center contracts., Vertiv Holdings Co.Commodity tokens still require power and cooling infrastructure in every data center, and it sells both to hyperscalers building out at volume., Vistra Corp.Its power generation fleet earns higher returns as data center electricity demand pushes up prices and contract rates, an input cost scale producers pay regardless of token value.12%
- Larger clusters and more data center buildouts in a commodity-token world drive demand for the switches, cables, and assembled systems that wire them together. · Arista Networks Inc.It sells the switches that wire up hyperscaler clusters, earning on data center buildout volume rather than on token margins., Credo Technology GroupIts cables and interconnect ship with hyperscaler clusters regardless of whose chips are inside, earning on deployment volume in a commodity market., Celestica Inc.It assembles the servers and networking systems hyperscalers deploy with their own chips, earning on unit volume as they scale infrastructure.11%
- Cash5%