Google I/O Previews Gemini 3.5 and Nvidia Vera CPU

🔑 Enhanced Key Takeaways

•Google's Gemini 1.5 Pro, a precursor to the anticipated 3.5, introduced a 1 million token context window, capable of processing extensive data like an hour of video or 30,000 lines of code, and utilizes a Mixture-of-Experts (MoE) architecture for efficient complex query processing.
•Nvidia's Vera CPU, featuring 88 custom 'Olympus' cores based on Armv9.2-A, is purpose-built for reinforcement learning and agentic AI, delivering up to 1.2 TB/s of LPDDR5X memory bandwidth and being the first CPU to natively support FP8 precision.
•SandboxAQ has integrated its Large Quantitative Models (LQMs), which are physics-based AI models trained on real-world lab data and scientific equations, with Anthropic's Claude, enabling researchers to access advanced drug discovery and materials science tools through natural language prompts without needing specialized coding skills.
•The AI accelerator market, encompassing GPUs, TPUs, CPUs, ASICs, and FPGAs, is experiencing exponential growth, projected to reach $26.41 billion in 2026, driven by increasing AI workloads and the demand for energy-efficient processing.

📊 Competitor Analysis▸ Show

LLM Market Comparison (as of Feb 2026)

Feature/Metric	Google Gemini	OpenAI ChatGPT	Anthropic Claude
Market Share (Web Visits)	23.9% (growing)	60.5% (declining from 72.5% in Oct 2025)	13.9% (declining from 48.4% in Nov 2025)
Context Window	Up to 2 million tokens (Gemini 1.5 Pro)	Varies by model (e.g., GPT-4)	Varies by model
Multimodality	Native (text, images, audio, video)	Yes (e.g., image generation with Imagen 2)	Yes (integrated with SandboxAQ LQMs)
Architecture	Mixture-of-Experts (MoE) for 1.5 Pro	Transformer-based	Transformer-based
Key Strengths	Long context, multimodal understanding, Google ecosystem integration	Broad user base, plugin ecosystem	Strong reasoning, enterprise focus

AI Accelerator Market Overview (as of 2026)

Type	NVIDIA Vera CPU	Traditional x86 CPUs (Intel/AMD)	TPUs (Google) / ASICs
Primary Use Case	Reinforcement Learning, Agentic AI, HPC	General-purpose computing, some AI workloads	AI training and inference (hyperscale)
Core Architecture	88 custom 'Olympus' Armv9.2-A cores	x86-64 cores	Custom designs for AI workloads
Memory Bandwidth	Up to 1.2 TB/s LPDDR5X	Lower than specialized AI accelerators	High, optimized for AI workloads
Precision Support	Native FP8 support	FP32/FP64 primarily	Optimized for lower precision (e.g., FP8, bfloat16)
Energy Efficiency	High (2x performance per watt vs x86)	Lower for AI workloads	High (especially TPUs)
Market Share	Emerging (part of CPU segment, ~7% of AI accelerators)	~7% of specialized AI hardware	TPUs & ASICs collectively ~26% of large-scale AI infrastructure

Note: The article mentions 'Nvidia Vera CPU' which is a specific product. Nvidia also produces GPUs (e.g., Blackwell platform) which are dominant in the overall AI accelerator market (~58% share). The comparison above focuses on the CPU aspect of Vera against other CPU types and dedicated AI accelerators.

🛠️ Technical Deep Dive

Nvidia Vera CPU: Features 88 custom-designed 'Olympus' cores, which are NVIDIA's first fully bespoke implementation of the Armv9.2-A instruction set.
Nvidia Vera CPU Memory: Utilizes server-class LPDDR5X memory, delivering up to 1.2 terabytes per second (TB/s) of memory bandwidth and supporting up to 1.5 TB of memory capacity.
Nvidia Vera CPU Interconnect: Employs NVIDIA NVLink™ Chip-to-Chip (C2C) connectivity and the NVIDIA Scalable Coherency Fabric (SCF) for high-bandwidth data flow between cores and other components.
Nvidia Vera CPU Precision: It is the first CPU to natively support FP8 (8-bit floating-point) precision, which harmonizes with the precision used by modern GPUs for AI workloads, reducing latency during inference.
Nvidia Vera CPU Architecture: Designed as a chiplet-based CPU, with a single compute chiplet containing all 88 CPU cores, surrounded by chiplets for I/O and memory, ensuring a single NUMA domain for efficient access.
Google Gemini (General): A family of multimodal large language models (LLMs) capable of processing text, images, audio, video, and computer code simultaneously.
Gemini 1.5 Pro Architecture: Incorporates a sparse Mixture-of-Experts (MoE) Transformer-based architecture, which allows the model's total parameters to grow while keeping the number of activated parameters constant, enhancing efficiency and performance for complex queries.
Gemini 1.5 Pro Context Window: Offers an expanded context window starting at 1 million tokens, with experimental capabilities up to 2 million tokens, enabling the model to analyze and reason over very long documents, videos, or codebases.
SandboxAQ Large Quantitative Models (LQMs): These are physics-based AI models trained on proprietary, physics-grounded data generated through high-fidelity simulations, including quantum chemistry calculations, molecular dynamics, and microkinetics.

🔮 Future ImplicationsAI analysis grounded in cited sources

The integration of quantitative AI models with large language models will significantly accelerate scientific discovery.

By allowing natural language interfaces like Claude to access physics-based LQMs from SandboxAQ, the barrier to entry for complex scientific simulations in drug discovery and materials science is lowered, enabling faster hypothesis-to-discovery cycles.

Specialized CPUs like Nvidia's Vera will become critical for scaling agentic AI and reinforcement learning workloads in data centers.

Vera's custom Olympus cores, high memory bandwidth, and native FP8 support are specifically engineered to address the control-heavy, latency-sensitive demands of agentic AI, leading to more efficient AI factories.

Google's continued advancements in Gemini's context window and multimodal capabilities will drive more sophisticated and integrated AI assistants.

The ability of models like Gemini 1.5 Pro to process and reason over vast amounts of diverse data (text, video, audio) in a single prompt paves the way for AI assistants that can understand and act on highly complex, real-world scenarios.

⏳ Timeline

2023-05

Google I/O announces Gemini development and unveils PaLM 2.

2023-12

Google officially launches Gemini 1.0 (Ultra, Pro, Nano versions), integrating Gemini Pro into Bard and Gemini Nano into Pixel 8 Pro.

2024-02

Google releases Gemini 1.5 Pro in limited preview, featuring a 1 million token context window.

2024-05

Google I/O 2024 announces Gemini 1.5 Flash and the 6th generation Trillium TPU, along with updates to Gemini Advanced.

2025-01

Google releases Gemini 2.0 Flash as the new default model.

2026-01

Nvidia unveils the Vera CPU (also known as Vera Rubin Platform) at CES 2026, designed for agentic AI.

2026-03

Google releases stable versions of Gemini 3.1 Pro, Deep Think, Flash, and Flash Lite.

2026-05

SandboxAQ integrates its Large Quantitative Models (LQMs) with Anthropic's Claude for drug discovery and materials science.

Google I/O Previews Gemini 3.5 and Nvidia Vera CPU

⚡ 30-Second TL;DR

Key Points

🧠 Deep Insight

🔑 Enhanced Key Takeaways

LLM Market Comparison (as of Feb 2026)

AI Accelerator Market Overview (as of 2026)

🛠️ Technical Deep Dive

🔮 Future ImplicationsAI analysis grounded in cited sources

⏳ Timeline

📎 Sources (32)

👉Related Updates

Korea to Inject $14B into AI and Data Infrastructure

Nvidia may guarantee $250bn for OpenAI data center

Autonomous Intelligence and the End of Productivity Software

AI Video Industry Shifts from Hype to Maturity