Hardware Podcasts

Inside Google’s massive AI capex (live)

Catalyst · 23 Apr 2026 · 35m

What this episode covers

Amin Vahdat, Google’s chief technologist for AI infrastructure, explains how Google is shifting from training to inference, enabling smaller, distributed data centers and lowering reliability requirements to double compute capacity. He details Google's strategy of using behind-the-meter power as a bridge to grid interconnection and their milestone agreement with utilities for one gigawatt of demand response. Vahdat emphasizes that co-optimizing chip design, building cooling, and power generation creates superefficient 'AI factories,' while identifying power, chips, and construction as simultaneous rate limiters for growth.

Most relevant to

large enterprises

Topics

Mentioned

Companies
Google, DeepMind, EnergyHub
Products
Gemini
Chips
TPU, GPU

Related episodes from other shows

Other shows on the same ground, closest first.

  • Moore's Lobby · 58m

    Ep. 27 | Two Google Senior VPs of Engineering: From Shipping Containers to Today's Data Centers

    Google Fellows Luiz Barroso and Amin Vahdat discuss the evolution of Google's data center infrastructure, from early 'broom closet' colocation facilities and shipping container prototypes to modern, energy-efficient warehouse-scale systems. They explain how Google shifted from relying on commodity hardware to designing custom domain-specific accelerators like TPUs for machine learning and video processing to overcome the limits of Moore's Law. The conversation also covers the critical importance of energy proportionality, Power Usage Effectiveness (PUE), and the cultural philosophy of 'healthy hubris' that drives Google's engineering decisions.

    Energy & climate techChips, FPGAs & semiconductors
    8 chapters

  • The Circuit · 38m

    Episode 8: AI Infrastructure and Silicon Demands

    Ben Bajarin and Jay Goldberg discuss the immense compute demands of training large foundational models like GPT-3 and GPT-4, noting that Nvidia’s A100 and H100 GPUs dominate this $300 billion infrastructure market due to CUDA’s software moat. They analyze the bottleneck in inference scalability, arguing that while training is a small fraction of the market, inference costs will drop as incumbents like Google, Amazon, and Apple leverage their own silicon, leaving little room for new AI accelerator startups. The hosts conclude that generative AI represents a powerful new category of productivity software rather than a fundamental shift in human-computer interaction, driving data center Capex through natural upgrade cycles rather than complete rebuilds.

    Chips, FPGAs & semiconductorsConnected products & IoT
    8 chapters

  • The Amp Hour · 1h 05m

    All Heat, No Useful Work

    Dave Jones and Chris Gammell discuss the massive energy demands and infrastructure challenges of AI data centers, noting that 100% of input power becomes heat and requiring extensive cooling and diesel backup. They cover supply chain constraints, including a three-year backlog for transformers and a 40% surge in bareboard PCB prices driven by AI server demand. The conversation shifts to practical electronics work, where Chris details building a power-injecting Tag-Connect debug board and using Claude to implement SMP firmware updates over web serial, while Dave shares his frustration with a solar monitoring service's price hike and his son's struggles with Linux at school.

    PCB design & layoutEmbedded firmware
    7 chapters

  • The Circuit · 43m

    Ep 160: More good News for Intel, AI Model Trends, Future of Datacenter

    Ben Bajarin and Jay Goldberg analyze Intel's recent strategic moves, including a partnership with Tesla for its 'Terafab' project and co-design work with Google on its IPU, interpreting these as validation for Intel's advanced packaging and foundry capabilities. They discuss the broader semiconductor market's valuation disconnect, noting how investors are shifting focus from established compute giants to obscure optical and mineral supply chain bottlenecks. The hosts also examine the future of data centers, arguing that extreme co-optimization of power (800 VDC) and cooling is becoming essential to unlock compute capacity, and cover Anthropic's rapid product rollout, including its Word connector and 'Mythos' model, while highlighting rising token costs and capacity constraints.

    Chips, FPGAs & semiconductorsEnergy & climate tech
    6 chapters

  • EE Times Current · 18m

    Designing the AI Factory from the Rack Up

    Neeraj Paliwal, Senior Vice President of Silicon IP Product Management at Synopsys, explains why the AI rack has replaced the data center as the fundamental unit of design and economics. He details the shift from 8kW to 130kW+ power densities in systems like NVIDIA's GB200 NVL72, driven by the need to optimize cost per token across distributed GPUs. Paliwal highlights that the industry is moving past single-die limits toward 3D stacking, custom HBM memory integration, and high-speed die-to-die interfaces like UCIe to overcome the memory and I/O bandwidth walls.

    Chips, FPGAs & semiconductorsBatteries & low-power design
    5 chapters

The summary, chapters and topics are generated from the episode’s audio by this directory, not written by Catalyst. The episode itself belongs to its publisher.