AI News Flash · Daily Brief

Google pushes Gemini 3.5 Pro to July after missing its own June deadline.

By and · Editorial team, AI News Flash

Platforms

Google pushes Gemini 3.5 Pro to July after missing its own June deadline.

Google has postponed the launch of Gemini 3.5 Pro to July 2026, slipping past the June deadline Sundar Pichai publicly committed to at Google I/O. The company cited a need to refine three specific areas: coding performance, token efficiency, and long-task reliability, drawing on feedback gathered during a limited Vertex AI enterprise preview. A Google spokesperson declined to confirm or comment on the revised schedule. The delay is notable given how recently the June target was set and how directly Pichai named it, raising questions about the competitive pressure Google faces from OpenAI and Anthropic in the enterprise model market.

Why it matters: Enterprise teams evaluating Gemini 3.5 Pro for production workloads must now extend timelines and reassess vendor comparisons.

Technology & Research

NVIDIA Expands Nemotron 3 With Open Speech, RAG, and Safety Models

On June 23, Nvidia extended the Nemotron 3 family well beyond its original Super and Ultra language models, releasing three new open model categories: Nemotron Speech, a leaderboard-topping automatic speech recognition model; Nemotron RAG, designed for retrieval over complex technical documents; and Nemotron Safety, aimed at trustworthy AI applications. The release also ships the dataset and training code for Llama Embed Nemotron 8B. Alongside the models, Nvidia contributed more than 10 trillion language training tokens and 500,000 robotics trajectories to the open-data commons. Together, these additions give developers a complete open stack covering perception, retrieval, reasoning, and safety, all built on a single model family with public training recipes.

Why it matters: Developers can now build production AI pipelines across speech, retrieval, and safety using a single open Nvidia model family with reproducible training.

Nvidia's Alpamayo 2 Super brings a 32B reasoning model to Level-4 autonomous driving.

Nvidia's Alpamayo 2 Super is a 32-billion-parameter chain-of-thought vision-language-action model targeting Level-4 autonomous vehicle development, tripling the parameter count of the original 10-billion-parameter Alpamayo 1. The model adds full 360-degree surround perception and ships alongside two companion tools: AlpaGym for closed-loop reinforcement learning training and OmniDreams for photorealistic scenario generation. Alpamayo 2 Super generates both driving trajectories and causally linked reasoning traces without requiring human annotation, and can act as a teacher model distilled into compact on-vehicle models running on DRIVE AGX Thor hardware. The model family has been downloaded nearly 400,000 times since launch, with JLR, Lucid, Uber, and Berkeley DeepDrive among the early adopters.

Why it matters: Autonomous vehicle teams can now access a fully open, annotation-free reasoning model that distills directly onto production Nvidia DRIVE hardware.

Regulation & Policy

UK FCA signals AI-era regulatory overhaul for financial services

FCA chief executive Nikhil Rathi used a June 25 address at techUK to call for a structural rethinking of how UK financial regulation applies to generative and agentic AI, stating that rules designed around human-led decision-making may need fundamental revision. The speech marks the FCA's most direct acknowledgment to date that its current supervisory framework is poorly matched to AI-driven financial products and services. The signal comes shortly after the UK Competition and Markets Authority's June 12 order requiring Google to offer publishers an AI content opt-out, underscoring a coordinated push by multiple UK regulators to assert authority over AI systems across different sectors.

Why it matters: Financial services firms deploying AI-driven products in the UK must now anticipate regulatory restructuring that could redefine compliance obligations.

AI Stocks

(NVDA) Jensen Huang tells shareholders Nvidia would choose US security over commercial deals

At Nvidia's annual shareholder meeting this week, CEO Jensen Huang told investors that when commercial opportunities conflict with US national security interests, Nvidia would choose American security, a direct response to sustained pressure surrounding chip export controls. The meeting also brought a notable financial disclosure: Nvidia's data-center Ethernet switch business generated $2.1 billion in revenue in Q1 FY2026, highlighting how the company is expanding its AI infrastructure presence beyond GPU sales. Shareholders approved all management proposals. Nvidia separately posted its FY2026 quarterly earnings slides on June 24, 2026. The statements come as export-control policy continues to shape competitive dynamics across the global AI chip market.

Why it matters: Nvidia's public commitment to US security priorities signals that export restrictions will continue to constrain its addressable market in key regions.

Agentforce and AIP post record growth, then Salesforce and Palantir stocks drop sharply.

Salesforce reported Q1 FY2027 revenue of $11.13 billion, up 13% year-over-year, with its Agentforce platform reaching $1.2 billion in ARR, a 205% year-over-year increase. Palantir posted 85% revenue growth and a Rule of 40 score of 145 in Q1 2026. Despite those results, CRM shares fell roughly 14% and PLTR dropped roughly 22% following the reports, part of a broader derating of enterprise software that pushed peers ServiceNow, Adobe, and Salesforce to 52-week lows in the same week. Investors appear to be rotating out of high-multiple AI software names even as underlying adoption metrics accelerate. Palantir CEO Alex Karp added pressure midweek by publicly warning that AI firms are upsetting enterprise clients.

Why it matters: The selloff signals that strong AI adoption metrics alone no longer justify premium valuations, forcing enterprise software firms to demonstrate margin discipline.