Sonnet Xu

bullish on reciprocating engines for BTM compute.

large turbines are optimized for centralized power generation, but reciprocating engines excel at modular BTM deployments, w/ shorter lead times and capacity that scales w/ GPUs.

BTM bypasses parts of the grid queue, not air permitting.

gas access and a viable permitting path can be worth more than proximity to the grid.

NVIDIA increasingly extends beyond the GPU.

the newest reference architectures cover substations, power distribution, cooling plants + facility layouts.

NVIDIA doesn't need to manufacture this infrastructure to influence how it's built; a bigger NVIDIA ecosystem means a bigger NVIDIA moat.

NVIDIA is doing the architectural work of a hyperscaler without its capital intensity - look at its reference architectures.

packaging hyperscaler expertise for everyone else has compelling economics - NVIDIA can pull through more networking and systems revenue, make its hardware harder to substitute, and influence billions in adjacent infrastructure spending.

all while others finance, build, and operate the data centers.

hyperscalers vertically integrate to lower costs. NVIDIA vertically integrates the architecture to expand what it can sell.

why does it matter that GPUs aren't fungible?

@shadeformai has been very vocal about how identical GPUs can deliver vastly different performance.

Jeff, COO of Shadeform, shared that understanding the quality of compute is central to marketplaces aggregating GPU supply.

@shadeformai looks at everything from geopolitics to teams supporting clusters.

as supply fragments, GPU SKUs tell you less and less. aggregation creates liquidity, but quality assurance makes that liquidity is useful.

the companies that know the most about GPU reliability have the least incentive to share what they know.

GPU validation is surprisingly closed. basic tools are open source, but deeper testing methodologies often aren't.

why? because transparency can be expensive.

without independent validation, the fungibility narrative persists. identical GPU specs are sold equivalently despite very different reliability and performance.

without quality verification, we overindex on the spec.