Together AI

Today’s “state-of-the-art” models were trained on yesterday’s compute, and most teams are running that compute at 35-45% efficiency. The real frontier isn’t model architecture — it’s the explosive growth in usable compute, compute efficiency, and what it unlocks for AI natives. 

The models teams benchmark against today were often trained 12-18 months ago, on clusters running at a fraction of today’s efficiency. Meanwhile, the compute landscape has already shifted: 

10x larger clusters are already being built
2-3x faster chips are replacing prior generations
3x+ efficiency gains are emerging from new kernels and existing optimizations 

As Dan Fu, VP of Kernels at Together AI — the AI Native Cloud — put it on a recent episode of the The MAD Podcast with Matt Turck,

“Models are lagging indicators. You can see up to two orders of magnitude more compute available compared to the models we’re indexing on today.” 

In this session, Dan will expose the hidden tax on your GPU investments and how research-backed optimization is the difference between burning money and unlocking existing capacity.  While most teams obsess over model releases, the real edge comes from squeezing 2-3x more performance out of existing hardware through better kernels, smarter memory management, and infrastructure co-design.

For builders, the message is clear: the constraint is shifting from models to infrastructure strategy. The winners will be those who build systems and teams that can absorb rapid growth without breaking down — and who understand how to turn massive compute gains into real production scale.