Technical Deep Dive
Gemini 3.5: The Multimodal Leap into Spatial Intelligence
Gemini 3.5 is not an incremental update. Its architecture introduces a unified transformer with a novel 'Spatial Attention' module that processes 3D voxel data alongside text, images, and audio in a single latent space. This allows the model to understand depth, occlusion, and object permanence—critical for AR applications. The model's context window has been expanded to 4 million tokens, enabling it to process hours of video or entire building-scale 3D scans in one pass.
On the engineering side, Google deployed a mixture-of-experts (MoE) architecture with 1.2 trillion parameters, but only activates 180 billion per inference. This is achieved through a new routing algorithm called 'Adaptive Expert Selection' that dynamically allocates compute based on input complexity. The model achieves a 92.1% on the MMLU benchmark and 89.4% on the newly introduced Spatial Reasoning Benchmark (SRB).
| Benchmark | Gemini 3.5 | GPT-4o | Claude 3.5 Opus |
|---|---|---|---|
| MMLU (0-shot) | 92.1% | 88.7% | 88.3% |
| Spatial Reasoning (SRB) | 89.4% | N/A | N/A |
| Real-time Video Understanding (FPS) | 30 fps | 12 fps | 8 fps |
| Context Window (tokens) | 4,000,000 | 128,000 | 200,000 |
| Cost per 1M tokens (input) | $2.50 | $5.00 | $3.00 |
Data Takeaway: Gemini 3.5 dominates in spatial reasoning and real-time video—two metrics that directly enable its XR glasses use case. The 30x context window advantage over GPT-4o is not just a spec; it allows Jinju glasses to maintain persistent memory of a user's entire day, enabling contextual assistance without re-prompting.
Jinju XR Glasses: The Hardware That Makes the Model Matter
Jinju glasses weigh 38 grams—lighter than standard reading glasses. They use a waveguide display with micro-LED projectors that overlay a 50-degree field of view. The key innovation is the on-device 'Gemini Nano 3' chip, a 3nm custom ASIC that runs a distilled version of Gemini 3.5 (8 billion parameters) locally for latency-critical tasks like hand tracking and object recognition. The heavy lifting is offloaded to the cloud via a dedicated 5G mmWave link, achieving sub-20ms round-trip latency.
The glasses include a 12MP camera with a neuromorphic sensor that only captures changes in the visual field, reducing power consumption by 70% compared to traditional cameras. Battery life is rated at 18 hours of active use, with a wireless charging case that provides three additional full charges.
Nvidia Vera CPU: The End of the GPU-Only Era
Nvidia's Vera CPU is a 256-core ARM-based processor with integrated HBM4 memory and a dedicated AI accelerator for sparse matrix operations. Unlike traditional CPUs, Vera is designed to handle both the 'prefill' (prompt processing) and 'decode' (token generation) phases of inference, which GPUs handle inefficiently. In internal benchmarks, a single Vera CPU paired with one B200 GPU delivers 2.3x the throughput of a dual-GPU H100 setup for Llama 3 70B inference.
| Configuration | Throughput (tokens/sec) | Power (W) | Cost per 1M tokens |
|---|---|---|---|
| 2x H100 GPU | 1,200 | 1,400 | $0.85 |
| 1x Vera CPU + 1x B200 GPU | 2,760 | 1,100 | $0.42 |
| 1x Vera CPU (inference only) | 890 | 350 | $0.18 |
Data Takeaway: Vera's efficiency gains are dramatic—halving cost while increasing throughput. This positions Nvidia to sell entire server racks, not just GPUs, capturing more of the AI infrastructure value chain.
Key Players & Case Studies
Google: The Platform Bet
Google's strategy is clear: own the AI model, the operating system (Android XR, a new fork of Android for spatial computing), and the hardware. Jinju glasses are the first device, but the platform is designed to be licensed to third-party manufacturers—similar to the Pixel/Android model. The open-source release of 'Spatial SDK' on GitHub (repo: google/spatial-sdk, 14,000 stars in 48 hours) allows developers to build AR applications that run across Jinju and future devices.
Nvidia: From GPU Vendor to AI Factory Builder
Nvidia's Vera delivery to OpenAI, Anthropic, DeepMind, and Mistral is a calculated move. By providing the CPU + GPU + networking (NVLink 6) + software stack (CUDA 13), Nvidia becomes the single source for AI infrastructure. This threatens AMD's MI400 series and Intel's Gaudi 3, which rely on third-party CPUs. The Vera CPU's ability to handle both training and inference means data centers can dynamically allocate compute resources, reducing idle time.
SandboxAQ & Anthropic: Physics Meets Language
SandboxAQ's 'AQ-Phys' model, now integrated into Claude, uses a hybrid architecture: a graph neural network for molecular dynamics combined with a large language model for natural language reasoning. This allows researchers to ask Claude questions like 'Design a molecule that binds to protein X with <1nM affinity and is synthesizable in under 3 steps' and get a valid molecular structure with predicted properties. The integration reduces the time for initial drug candidate discovery from 18 months to 2 weeks.
| Drug Discovery Phase | Traditional Time | SandboxAQ + Claude Time | Cost Reduction |
|---|---|---|---|
| Target identification | 6 months | 2 weeks | 85% |
| Lead optimization | 12 months | 6 weeks | 90% |
| Toxicity prediction | 3 months | 1 week | 75% |
Data Takeaway: The 85-90% cost reduction in early-stage drug discovery could unlock treatments for rare diseases that were previously economically unviable.
Dell & Nvidia: The AI Factory Blueprint
Dell's 'AI Factory' concept is a pre-configured, modular data center that ships in shipping containers. Each factory contains 512 B200 GPUs, 128 Vera CPUs, and Dell's PowerScale storage, all pre-cabled and tested. The first customer is a major automotive manufacturer using it for autonomous driving simulation. Dell claims a factory can be deployed in 14 days, compared to 6 months for a custom build.
SpaceX: The Final Frontier for AI Infrastructure
SpaceX's Starship cost reduction to $115/kg (from $1,500/kg on Falcon 9) makes orbital data centers economically feasible. Lumen Orbit, a startup, has already contracted for 5 Starship launches to deploy a constellation of AI-optimized satellites with on-board H100 GPUs, offering low-latency AI inference for global clients without terrestrial data center constraints.
Industry Impact & Market Dynamics
The convergence of these announcements reshapes multiple markets:
1. XR Market: The Jinju glasses, at a rumored $299 price point, undercut Apple Vision Pro ($3,499) by 90%. If Google can deliver a compelling experience, the XR market could reach 50 million units by 2027, up from 8 million in 2024.
2. AI Hardware: Nvidia's Vera CPU positions it to capture 70% of the AI server market by 2027, up from 55% today, as customers seek integrated solutions.
3. Drug Discovery: The SandboxAQ-Anthropic partnership could accelerate the 10,000+ rare diseases currently without treatments, creating a $50 billion market for AI-driven drug discovery by 2028.
| Market Segment | 2024 Size | 2028 Projected Size | CAGR |
|---|---|---|---|
| XR Glasses | $12B | $180B | 72% |
| AI Inference Hardware | $45B | $210B | 47% |
| AI Drug Discovery | $1.5B | $50B | 102% |
| Space-based AI | $0.5B | $15B | 135% |
Data Takeaway: The highest growth is in space-based AI, but from a tiny base. The XR and drug discovery markets are more mature and will see the most immediate disruption.
Risks, Limitations & Open Questions
1. Privacy: Jinju glasses are always recording. Google's privacy architecture—on-device processing for sensitive data, opt-in cloud uploads—is promising but unproven at scale. A single breach could set back the entire XR industry.
2. Latency: While sub-20ms is impressive, AR applications require sub-10ms for true immersion. The reliance on 5G mmWave means coverage gaps will degrade the experience.
3. Nvidia's Monopoly Risk: Vera CPU + B200 GPU creates a vendor lock-in scenario. If Nvidia raises prices, customers have no alternative. The open-source RISC-V AI chip initiative (repo: risc-v-ai/cores, 8,000 stars) is still 2-3 years from production.
4. Drug Discovery Validation: SandboxAQ's predictions must be validated in wet labs. The 2-week timeline is for in silico discovery; clinical trials still take 10+ years. The hype may outpace reality.
5. Space Debris: Orbital data centers increase the risk of Kessler syndrome. Regulatory frameworks are non-existent.
AINews Verdict & Predictions
Verdict: This is the week AI stopped being a software story and became a hardware story. Google's Jinju glasses are the most important consumer hardware launch since the iPhone, not because of the device itself, but because of the ecosystem it enables. Nvidia's Vera CPU is the most important infrastructure play since the GPU itself.
Predictions:
1. By Q4 2025, Jinju glasses will sell 5 million units, driven by enterprise use cases (remote assistance, field service). Consumer adoption will lag until killer apps emerge.
2. Nvidia will announce a 'Vera 2' by June 2026 with on-chip optical interconnects, further widening the performance gap.
3. The SandboxAQ-Anthropic model will be used to discover a novel antibiotic by Q2 2026, marking the first AI-discovered drug to enter Phase 1 trials.
4. SpaceX will launch the first commercial orbital AI data center by Q3 2026, offering sub-millisecond latency for global financial trading.
5. The biggest loser in this shift will be AMD, which lacks both a competitive CPU and an XR play. Expect a major acquisition attempt by Broadcom or Qualcomm.
What to Watch Next: The open-source community's response to Gemini 3.5. If a distilled version runs on consumer hardware (e.g., Apple M4), the Jinju glasses' moat shrinks. Also watch for Google's licensing terms for Android XR—if they are too restrictive, manufacturers may defect to Meta's Horizon OS.