The GPU gold rush is over; the efficiency audit has begun.
The Summary
- SambaNova CEO Rodrigo Liang says AI infrastructure is shifting from training to inference, efficiency, and ROI, marking the end of the "build at any cost" era
- Enterprises like JPMorgan are running AI behind their own firewalls, not renting cloud compute
- The company just closed Series F funding with strategic investors betting on private AI deployment
The Signal
The AI infrastructure market is bifurcating in real time. SambaNova's Series F close signals what CFOs have been whispering for months: training giant models was the spectacle, but inference is where the actual work happens. And inference at scale requires completely different economics.
Liang's pointing at something specific here. When JPMorgan runs AI infrastructure behind its own firewall, that's not paranoia. That's a calculated bet that the data moat matters more than the model moat. Banks don't need the frontier model. They need a model that knows their customers, their risk profiles, their regulatory constraints. That model doesn't need to live in OpenAI's data center.
"Inference, efficiency and return on investment" are the new table stakes for enterprise AI deployment.
This tracks with what we're seeing across the enterprise stack:
- Inference costs are 10-100x training costs at production scale
- On-premise deployment lets you skip the API tax on every query
- Strategic investors mean customers are now backers, which means real revenue traction
The coexistence argument Liang makes about frontier models and private systems isn't kumbaya. It's market segmentation. Frontier models are research platforms and consumer product engines. Private systems are where regulated industries, governments, and anyone handling sensitive data will actually deploy AI. Those are different businesses with different unit economics.
SambaNova's pitch is hardware optimized for inference workloads behind the firewall. That's a narrow wedge, but it's the wedge that matters if you believe enterprises will prioritize control and cost over access to the latest GPT variant. The strategic investor piece is the tell. When your customers invest in your Series F, they're not betting on your technology. They're betting on their own AI roadmap, and you're on it.
The Implication
If you're building in AI, the question isn't "can I access the best model?" anymore. It's "can I run inference at a cost that makes the unit economics work?" Watch for more enterprise AI companies to offer on-premise options and custom silicon optimized for specific workloads. The cloud API era for enterprise AI might be shorter than anyone thought.