AI Engineering Signal #151
DeepSeek DeepInfer open-source inference runs 155T tokens/day at 80% margin
Signals
DeepSeek DeepInfer open-source inference runs 155T tokens/day at 80% margin
Self-host economics undercut API serving for high-volume workloads; rework capacity and cost models.
Latent Space
eMudhra CA private keys cracked in milliseconds
Revoke and audit eMudhra-issued certificates; trust chain failure is active, not theoretical.
Web
Google open-sources TorchTPU native PyTorch TPU backend
PyTorch code can target TPU silicon without XLA translation; revisit TPU procurement assumptions.
GitHub
Samsung sells out 2027 HBM capacity
HBM prices may jump 200%; lock GPU memory contracts and rework cluster budgets.
Web
Figure F.03 walks nine days straight on treadmill
Watch humanoid procurement specs; 1.2 million treadmill steps raises endurance verification requirements.
Web
Leg-compression therapy improves early Alzheimer's memory in trial
84% improved cognition with gains lasting six months; track this non-drug intervention for dementia care.
Web
The Take
Inference cost compression is now coming from self-hosted open-source serving, TPU tooling, and memory supply risk, so lock contracts and rework budgets. Embodied and clinical claims now need measured endurance and trial data, not demos.
Subscribe
Related Signals