NVIDIA's First In-House CPU 'Vera' Arrives at Top AI Labs, SandboxAQ Democratizes Drug Discovery with Claude, and Big AI's Regulatory Meddling Draws Scrutiny

NVIDIA's First In-House CPU "Vera" Arrives at Top AI Labs

Last Friday, NVIDIA VP Ian Buck personally handed over "**Vera**", NVIDIA's first in-house designed CPU, to three of the world's top AI labs. The units went to **Anthropic** in San Francisco, **OpenAI** in Mission Bay, and **SpaceXAI** in Palo Alto, with a delivery to **Oracle Cloud Infrastructure** in Santa Clara following on Monday.

Vera is a CPU optimized for agentic AI workloads, reportedly running agent sandboxes **50% faster** and enterprise data queries **up to 3x faster** than previous CPUs. In a keynote at Dell Technologies World, Michael Dell predicted that global AI infrastructure investment will reach **$3–4 trillion** by 2030, with token consumption growing 3,400% over the same period. Jensen Huang said demand is "increasing parabolically."

Today, more than **5,000 companies** — including Lilly, Samsung, and Honeywell — run AI workloads on Dell AI Factories with NVIDIA.

SandboxAQ Brings Quantum Chemistry Drug Discovery Models to Claude

SandboxAQ has made its quantum chemistry drug discovery models available through **Claude**. While companies like Chai Discovery and Isomorphic Labs compete to build better models, SandboxAQ concluded that "accessibility is the real bottleneck," making it possible to run drug discovery simulations through Claude's interface — no computational science expertise required.

The move is expected to push the democratization of AI-driven drug discovery a step further. Running quantum chemistry simulations used to require HPC expertise, but the models can now be driven with natural-language instructions.

"Big AI" Is Gutting Regulation with the Same Playbook as Big Tobacco and Big Oil — The Register

According to an investigative report by The Register, major AI companies are undermining the effectiveness of AI regulation by using **the same lobbying tactics once employed by the tobacco and oil industries**.

Specific tactics cited include promoting industry-led self-regulatory frameworks to sidestep statutory regulation, and providing research funding that suits regulators. Critics point out the structural similarity to how the tobacco industry once downplayed health impacts and the oil industry delayed climate action.

As debate over balancing AI safety and innovation intensifies, there are growing calls for deeper scrutiny of how this kind of industry influence is warping the shape of regulation.

MetaAgent-X: Breaking the Automated Orchestration Ceiling with Multi-Agent RL

A new study published on arXiv, "**MetaAgent-X**", proposes a framework that jointly optimizes the automated design and execution of multi-agent systems through **end-to-end reinforcement learning**.

Previous automated multi-agent systems either explored only at test time or optimized only the meta-level designer while keeping executor agents fixed — running into the "frozen-executor ceiling." MetaAgent-X integrates script-based system generation, collection of execution rollouts, and credit assignment over both designer and executor trajectories, achieving stable and scalable training.

It is being watched as a significant step toward raising the performance ceiling of automated multi-agent systems.

AI-Generated Code Now Proves Its Own Correctness

A research team at CodeMetal has announced a method by which AI-generated code can prove its own correctness through **formal verification**. Conventional AI code generation has mostly followed a "works well enough" approach, but this method attaches mathematical proofs to generated code, guaranteeing that it satisfies its specification.

This could lead to a fundamental improvement in trustworthiness for AI-generated code in safety-critical systems.

RAM Prices Could Plummet in H2 2027 on the Back of China's Massive Investment

A former memory-industry executive has suggested that **massive investments** by Chinese companies in DRAM production capacity could drive a sharp fall in RAM prices in the second half of 2027.

CXMT (ChangXin Memory Technologies) recorded a **1,688% profit surge** in Q1 2026 and is accelerating investment in HBM packaging and next-generation DDR5, with plans to raise monthly wafer output from roughly 280,000 to over 300,000. If Chinese companies expand supply capacity as planned, oversupply-driven price drops could happen within a year.

In the local LLM community, expectations are growing that falling RAM prices will help spread local inference environments for large models.