Analytical briefings for engineers, PMs, and CTOs
Arjun Mehta 
3 min readEvery major hyperscaler is running a significant fraction of AI inference on custom silicon — Google TPU v5, Meta MTIA, Microsoft Maia, Amazon Trainium. NVIDIA retains dominance in training, but the inference market is structurally bifurcating.
4 min readControlled studies show AI coding assistants reduce time-to-completion by 26–55% — but gains concentrate in code generation and do not extend uniformly to debugging, architecture, or code review. The contexts matter enormously.
3 min readIntel’s Hala Point and IBM’s NorthPole deliver 22–100x energy efficiency gains over GPUs on specific sparse workloads — but no standard programming model, a thin software ecosystem, and narrow task compatibility have kept neuromorphic chips out of commercial AI deployments.
4 min readThe AI chip shortage didn’t end — it transformed. The bottleneck shifted from wafer supply to CoWoS advanced packaging, which TSMC has sold out through 2026. NVIDIA locked up over 70% of available capacity, leaving hyperscalers and startups scrambling for alternatives.
4 min readOver 150 AI-designed drug candidates are in clinical trials globally — a 7x increase since 2022. AlphaFold 3 has compressed pre-clinical discovery from 4–6 years to 12–24 months. The first FDA approvals of AI-primary drug candidates are 3–5 years away.
4 min read2025 will be remembered not for the volume of AI announcements but for five shifts that had real structural weight. Models got smarter, cheaper, and started taking action — but the more important story was what each shift revealed about where the field was actually heading.