The Frontier

Steve Hou joined Silicon Data almost two months ago as Head of Research, aiming to bring data and derivatives like futures contracts to the physical AI compute market.

Silicon Data seeks to enable hedging against risks in the AI physical compute market, which Steve Hou estimates to be hundreds of billions, if not trillions, in size.

Steve Hou believes futures contracts are a natural hedging tool for data center providers, compute providers, and companies buying compute, offering revenue certainty and enabling bolder acquisition strategies.

The AI compute market, currently dominated by a few players like OpenAI and Anthropic, is expected to fragment as open models and enterprise AI drive broader demand, especially for inference.

Steve Hou's Token Expenditure Index, an expenditure-weighted price index similar to the PCE for AI, tracks token price dynamics aggregated by usage patterns, reflecting consumer behavior and quality-price tradeoffs.

The Token Expenditure Index primarily covers independent developers and small/medium enterprises using public routing platforms, making it a leading indicator for price sensitivity rather than total market demand.

Steve Hou notes the index's recent plateau reflects a shift towards 'token efficiency' and substitution among models, not a decrease in overall token demand.

Drawing on Jevon's paradox, Steve Hou argues that while frontier models' margins may face pressure from cheaper alternatives, the resulting market expansion will still lead to overall growth and profitability.

Steve Hou's GPU Rental Index tracks on-demand rental rates for chips like the H100, A100, and B200, aggregating contracts from various providers to create apple-to-apple comparisons.

The A100 chip's strong and increasing rental rate signals robust inference demand, indicating even older chips maintain value as workhorses for less computationally intensive tasks.

The GPU forward curve, initially backwardated (downward sloping), has shifted upward and flattened into contango, implying cloud providers are less willing to discount long-term contracts due to firm demand and supply shortages.

Memory (DRAM) prices are surging due to models' increasing memory hunger and context length, but Steve Hou anticipates algorithmic innovations like Kimi's memory efficiency improvements will eventually temper linear demand growth.

Steve Hou expects increased demand for storage and memory from multimodal AI (voice, video), noting that conversational AI advancements, like the ability to interrupt, generate significantly more data.

Steve Hou predicts US-China geopolitical decoupling will drive separate but parallel growth in AI capex, with both nations doubling down on investment and potentially restricting cross-border model access.

The next major driver for AI is genuine enterprise adoption and demonstrable return on investment (RORI), enabled by cheaper models that allow companies to experiment without prohibitive token budgets.