Repricing the Frontier.
Currently, frontier large language models cost billions to train and serve, an unsustainable amount of money. Arcware’s mission is to make the unsustainable, sustainable and the impossible, possible. We serve models for 50% less than market rate not because of venture money or subsidies, but because our proprietary systems allow us to serve AI models on dramatically cheaper hardware with no performance or speed cost. These systems were developed by Arcware via extensive research, research we expect to use to make efficiency of all hardware radically more efficient.
This puts Arcware in a position to be a unique mix of a Neocloud and Neolab, a position we are grateful to be in. We plan to use our research capabilities to not only improve AI’s infrastructure, but AI itself. We plan to release our own open-weight model to be competitive with the SOTA models of today, and to put American open-weight models back on top for the first time since the days of Meta’s Llama.
Our home is in the San Francisco Bay Area, and all of our models are hosted in the same region. We also hope to expand to make on-prem AI cheaper and easier to install for all teams, something we will work to deliver.
We care deeply about our home in the Bay, and we see first-hand how the price people pay for frontier intelligence is not just monetary but also ecological. Due to the massive machinery, electricity, and cooling systems used to serve these models the environment can sustain heavy damage. With our reduced hardware footprint, we have extraordinary ecological benefits.
Compared to an average data center, Arcware’s cloud hostings is 59.6x more energy efficient, 98.3% less heat, and ~119x lighter. At a proposed scale of 1,000 of our systems, Arcware would save the annual household electricity of well over ten thousand US homes, and could avoid roughly 30 to 266 million liters of onsite cooling water annually (again per 1,000 systems).
Our core mission is to make infrastructure sustainable again, for the Intelligence Age.