Our Mission
We’re bringing the artificial intelligence of the future to life, today.
Since the dawn of the PC, we’ve designed our computers around the assumption that long distance communication was difficult and costly, with diminishing performance at scale. We spend huge amounts of time, money, and effort molding software around discontiguous hardware to compensate for bottlenecks in communication.
Today’s Leading AI Supercomputers…
Simply can’t keep up.
The amount of compute required to train models has grown far faster than Moore’s law, and the number of parameters required to run the largest model has more than quadrupled in the last year alone.
Don’t scale well.
Thanks to slow communication links, the largest recommendation systems spend more time communicating than computing, while large language models take months to tune for maximum performance. Communication bandwidth is becoming the most important bottleneck in large-scale AI.
Are too complex.
Increasingly complicated software techniques are being used to expand the size of models, becoming a larger fraction of the cost and difficulty of training and deploying new AI capabilities.
We are breaking past these bottlenecks.
Eliminate the bottleneck.
We can build data links that are orders of magnitude more bandwidth-dense and power-efficient than have ever been possible. We’re inserting these links into every part of the computer. Communication will simply no longer be the bottleneck for scaling and performance.
Revisit computer architecture at its core.
Breaking these bandwidth bottlenecks is allowing us to revisit fundamental assumptions about computer architecture—at the hardware, software, and model levels.
Nobody has gotten this right yet. Our approach will vastly simplify both the hardware and software stack, and it will allow us to build and deploy AI algorithms that cannot run on today’s machines.