Podcast
SemiAnalysis
SemiAnalysis video conversations on semiconductors, AI infrastructure, GPUs, datacenters, and compute markets.
Original source
Ep. 024 - SpaceX's 10GW Plan Drives $300B ARR by 2027 (Datacenter, Energy)
SemiAnalysis argues SpaceX’s AI/datacenter plan is fundamentally an economics play: if AI inference can sustain roughly $100M per MW-year in revenue, then rapid gigawatt-scale buildout becomes enormously valuable. The episode then focuses on whether the real constraint is demand or execution—sites, power, turbines, permitting, labor, and supply chain readiness.
Ep. 023 - Everyone Leaves Google, Elon Forecasts 1T ARR, Reflecting On GPT-5 | Jon from Asianometry
This episode centers on three big themes: whether GPT-5’s first release mattered less than later variants, what Google’s leadership churn means for its AI future, and how U.S.-China component bans could backfire in hardware supply chains. It also veers into a long technical digression on AI security, local inference, and a very specific story about using coding agents to build a custom video editor.
Ep. 020 - Anthropic vs OpenAI Usage, Margins, Meta Compute, Future of MSL (Tokenomics)
The episode focuses on how AI usage is actually metered inside companies, arguing that coding dominates token burn while many common tasks are too cheap to justify tight restrictions. The hosts then connect those usage patterns to AI company economics, concluding that API-heavy businesses like Anthropic look far more attractive than consumer-heavy products, while Meta’s compute strategy and cloud distribution shape the next competitive phase.
Did China just beat Intel?
SemiAnalysis argues SMIC’s N+3 chip for Huawei is a real technical achievement: it reaches near-EUV-like density using DUV, but with much worse cost, complexity, yield, and efficiency. The episode’s bottom line is that China is not catching TSMC or Intel at the leading edge, but domestic chips may still be “good enough” for strategic workloads like phones, inference, and networking.
![Artwork for [Emergency Episode] Moonshot’s Kimi K3 has Arrived! China has a Frontier Model](https://i.ytimg.com/vi/hPb3SESIjdU/hqdefault.jpg)
[Emergency Episode] Moonshot’s Kimi K3 has Arrived! China has a Frontier Model
SemiAnalysis argues Kimi K3 is a genuine frontier contender, likely top-three globally and impressive enough to pressure how people think about open vs. closed AI. The episode focuses on serving constraints, 2.8T-parameter infrastructure needs, pricing, and whether Moonshot’s open-weight strategy is narrowing the gap to closed labs.
Ep. 019 - Inside the STEEL Lab: From Package to Transistor (Teardown Lab)
SemiAnalysis’ STEEL lab walkthrough explains how chip teardowns move from package inspection to transistor-level analysis using X-ray, polishing, FIB, SEM, and TEM. The episode uses Huawei/SMIC examples to show how reverse engineering reveals process-node details, architecture choices, and the growing importance of advanced packaging.
Training a 400B Model on 2,048 Blackwell GPUs for $20M | Researcher Conversations at GTC
RCAI’s Lucas Atkins explains why his team moved from post-training into pre-training, arguing that owning the full stack is increasingly necessary for enterprise compliance, product control, and customization. He also details RC’s research organization, Trinity’s multi-company build, and why the team chose B300s to accelerate training despite immature tooling.
The true cost of a GPU cluster
The episode argues that GPU-hour pricing is a misleading proxy for AI infrastructure cost. Real economics depend on useful work per dollar, or goodput, and on hidden costs like storage, networking, support, setup, and failures.
The GPU Power-Performance Curve Most Clusters Ignore | Researcher Conversations at GTC
Pebble argues that GPU power tuning for AI clusters is non-linear, so more power can eventually reduce tokens per watt instead of improving throughput. The company uses telemetry-driven Kubernetes tooling to dynamically cap per-GPU power and clocks, while also exploring grid-responsive data centers that can flex load without breaking SLAs.
Designing Data Centers for 400kW GPU Racks | Researcher Conversations at GTC
Radiant says its neo-cloud platform is built to scale megawatt-class AI infrastructure globally, with Brookfield helping solve land, power, and capital constraints. The interview focuses on 400 kW GPU racks, behind-the-meter power, and a software stack that provisions bare metal into AI services, VMs, Kubernetes, and high-SLA operations.