Companies Run Their Own AI
Surya@singhsurya1012
The idea
I believe companies will stop renting AI from the big labs and start running it themselves, training open models on their own data once those models are good enough. This basket owns the layers that bet depends on: chips, memory and networking, the neoclouds renting out GPU access, the power and cooling gear every GPU rack needs, and a specialist helping companies train and fine-tune their own models. It skips application software, since that layer is easy to swap out.
What will be bought
- Whether companies buy or rent, open models run on these chips, memory and network gear. · NVIDIAMost open models, whether rented or run in-house, run on its chips, and this holds as long as it stays the default hardware for running and fine-tuning open models., Advanced Micro DevicesA cheaper chip that runs open models well gets more attractive to companies hosting models in-house to cut costs, and this works as long as it keeps winning real orders from clouds and companies., Arista NetworksIt makes the switches that wire GPU clusters together for companies training their own models, competing with NVIDIA's own networking products., Micron TechnologyEvery GPU running or training an open model needs high-speed memory stacked next to it, and this requires AI demand to keep its memory sold out rather than slipping into a price glut.30%
- Every GPU, owned or rented, needs electricity and cooling, and that equipment is hard to get. · VertivEvery GPU rack needs power gear and cooling, and it supplies both liquid cooling and power equipment as long as data center builders keep ordering faster than they can install., GE VernovaMore companies running their own AI means more data centers, and it makes the turbines and grid equipment those facilities need as long as data center orders keep growing., EatonIt makes the transformers, switchgear and power distribution between the grid and GPUs, and this holds as long as data center demand keeps its electrical gear in short supply., Bloom EnergyWhen the grid can't deliver power fast enough, its fuel cells let a data center make its own electricity on site, and this works as long as data center owners keep choosing on-site power to launch sooner.26%
- Companies that don't buy machines rent GPUs or open-model access from these clouds instead of paying the big labs. · Nebius GroupIt rents GPUs and sells access to open models by the token, following the rent-instead-of-buy path as long as companies keep signing up for that capacity., CoreWeaveIt rents out GPUs instead of selling its own AI model, gaining if companies move off big labs and run AI themselves, though it depends on customer demand growing and its debt staying fundable., IRENIt owns cheap power and data center sites turned into GPU rental capacity, and this works as long as it keeps landing GPU cloud customers rather than staying mostly a bitcoin miner.26%
- Companies need outside help to train open models on their own data with reinforcement learning. · InnodataIt helps businesses train and fine-tune models including reinforcement learning work, and this holds as long as it keeps winning model training contracts from companies building on open models.9%
- Cash10%