Sunk Cost
How long until a local LLM machine pays for itself against API bills
Emerged finds new products worth a look, daily — what’s new today
Before renting compute, a founder can see which accelerator gives the most memory and FLOPS per hour across 24 providers in one place.
Flopper gathers datasheet numbers for about 221 GPUs and accelerators, from memory and bandwidth to throughput at each precision from FP64 to FP4, plus 82 multi-GPU systems. It tracks around 385 hourly rental prices across 24 cloud providers, maps roughly 786 datacenters, and adds calculators for FLOPS, LLM memory needs and power. Comparisons are side by side, and users can submit specs. Free to browse, with sponsorship slots. The maker notes real throughput runs near 85% of advertised in their own testing.
More like this
ai-infra-book
An open textbook that sizes LLM inference and training from first principles
Phobos
A Rust kernel language that compiles its own GPU code and serves LLMs with it
Warp
Run trillion-parameter models on a laptop by streaming weights from NVMe
Is AI Profitable Yet?
A sourced ledger of what each AI company has spent versus what it has earned
Lokutor
Voice agents that run entirely on CPUs, in the cloud, on-prem or on-device