NVIDIA Fireside Chat at Google Public Sector Summit 2025

Short summary

A Google Cloud / NVIDIA fireside chat described a deep, multi-year technical partnership to deliver AI infrastructure, software optimizations, and production-ready models (notably Gemini) to commercial and public‑sector customers. Deployments include public cloud, air‑gapped Google Distributed Cloud, and fully on‑prem environments for sensitive workloads.

Key technologies, products and features

• Gemini for Government

  • The same Gemini model used in Google products (Search, Maps, YouTube) made available to agencies.
  • Integrated with enterprise systems and optimized for secure, high‑performance deployments.

• NVIDIA hardware (names corrected where necessary)

  • Blackwell family (GB200 / GB300 racks referenced).
  • Grace family (including Grace Blackwell references).
  • RPU600‑class accelerator family.
  • These are offered through Google Cloud and for on‑prem deployments.

• Distributed inference

  • NVIDIA software (referenced as “Dynamo” / distributed inference) to shard models across many GPUs (example: using all 72 GPUs in a GB rack).
  • Enables larger models, improved performance, and lower cost per inference.

• Stack and software optimizations

  • Low‑level engineering collaboration (NVLink tuning, protocol buffers, JAX optimizations) between Google and NVIDIA to maximize performance and cost efficiency.

• Guardrails / safety tooling

  • NVIDIA‑ and Google‑adopted guardrail approaches to keep agents on task and reduce hallucinations.

• Integration and connectors

  • Built‑in connectors to common enterprise sources: Office 365, SharePoint, ServiceNow, Slack.
  • Includes data agents, data science tooling, and Google Dataflow/Dataproc integrations for querying organizational data.

• Multimodal agent capabilities

  • Form and image extraction, automated transaction flows (examples: permit workflows, mortgage processing, credit‑card issuance).
  • Tool chaining (e.g., check credit score → execute backend steps).

• Deployment options

  • Public cloud.
  • Google Distributed Cloud (air‑gapped/offline).
  • Fully on‑prem with Blackwell GPUs for sensitive or classified environments.

Operational and strategic concepts

• AI factory metaphor
  • The partnership framed AI as a factory where inputs = data and outputs = tokens/answers. Inference (token generation) is described as the primary driver of value and revenue.
  • Emphasis on scaling tokens cheaply: cost per token down, compute capability up, and model context/insight improving.

Inputs = data, outputs = tokens/answers; inference (token generation) is the primary driver of value and revenue.

• Practical adoption advice

  • Start now and pick quick‑win projects (e.g., form processing, transaction automation).
  • Leaders should be hands‑on users of the tools.
  • Implementations are often measured in weeks or months, not years.

• Productivity impact examples

  • Internal data search and summarization via data agents.
  • Automated flows handling a high percentage of mortgage transactions and same‑day credit operations in commercial examples.

Guides, tutorials and operational recommendations

1. Get started

   • Pick a small, high‑impact use case (form processing, data access, permit workflows).
   • Connect the model to existing data sources using built‑in connectors and agents.
   • Deploy on the appropriate mix of cloud, on‑prem, or air‑gapped infrastructure.

2. Security and compliance

   • Use air‑gapped or on‑prem Blackwell deployments for sensitive/classified workloads.
   • Prefer integrated perimeter and audit features rather than bolting security on afterward.

3. Cost and performance optimization

   • Leverage joint Google–NVIDIA optimizations (JAX, NVLink tuning, distributed inference).
   • Choose the right hardware configuration (GB200/GB300 racks, RPU family) through Google Cloud.

Noted caveats from the transcript

• The subtitle transcript contained some mis‑transcriptions (names and product identifiers). Likely corrections applied in this summary include Thomas Kurian (Google Cloud CEO) and corrected Blackwell/Grace hardware references.

Main speakers / sources

• Ian Buck — Vice President, Hyperscale and High Performance Computing, NVIDIA
• Thomas Kurian — CEO, Google Cloud
• Also referenced: Jensen Huang (NVIDIA CEO) and Sundar Pichai (Google CEO) as originators or initiators of partnership ideas.