Welcome back to AI Weekly Updates. Each week we highlight the most important stories in AI so you can stay ahead without wading through endless headlines. From OpenAI standing up a new enterprise consulting arm and shipping GPT-5.5 Instant to ChatGPT defaults, to Anthropic’s SpaceX compute deal and Nvidia and Corning’s US manufacturing partnership, here are the developments that shaped the AI landscape last week.
OpenAI launches the Deployment Company, a new enterprise AI consulting arm
OpenAI on May 11 unveiled the OpenAI Deployment Company — a new business that embeds Forward Deployed Engineers inside organizations to redesign workflows around AI. The venture is a partnership with 19 leading global investment firms, consultancies, and system integrators — led by TPG with Advent, Bain Capital, and Brookfield as co-lead founding partners, alongside Bain & Company, Capgemini, and McKinsey on the consulting side. In tandem, OpenAI agreed to acquire Tomoro, bringing roughly 150 Forward Deployed Engineers into the new entity on day one.
Read more ↗GPT-5.5 Instant rolls out as ChatGPT’s new default model
OpenAI updated ChatGPT’s default model on May 5 with GPT-5.5 Instant, citing a 52.5% reduction in hallucinated claims versus GPT-5.3 Instant on high-stakes prompts in medicine, law, and finance, and a 37.3% drop in inaccurate answers on conversations users had previously flagged for factual errors. The model also tightens conversational tone, improves image and STEM handling, and is more aggressive about reaching for web search when accuracy calls for it. Our earlier essay on GPT-5.5’s debut goes deeper on the consumer-vs-enterprise split it signals.
Read more ↗Anthropic signs a 300+ megawatt compute deal with SpaceX
Anthropic announced a partnership with SpaceX that takes all of the compute capacity at the Colossus 1 data center — over 300 megawatts and more than 220,000 NVIDIA GPUs coming online within the month. The deal stacks on top of Anthropic’s previously announced 5 GW Amazon commitment and 5 GW Google + Broadcom agreement, and is aimed directly at relieving the capacity squeeze that has been straining Claude Pro and Claude Max subscribers at peak.
Read more ↗NVIDIA and Corning commit to a 10x US optical-fiber expansion for AI infrastructure
NVIDIA and Corning announced a multiyear partnership on May 6 in which Corning will increase its U.S. optical connectivity manufacturing capacity by 10x and expand U.S. fiber production by more than 50%. The plan includes three new advanced manufacturing facilities in North Carolina and Texas and the creation of more than 3,000 new high-paying American jobs, all sized for the photonics and connectivity that hyperscale AI factories now consume at extraordinary scale. Jensen Huang framed it as “a once-in-a-generation opportunity to reinvigorate American manufacturing and supply chains.”
Read more ↗Anthropic’s Claude was used in an attempted Mexican water-utility compromise
A May 8 Cybersecurity Dive report based on Dragos and Gambit Security research detailed how an unknown threat group used Claude Code and OpenAI’s GPT-4.1 API in a months-long campaign from December 2025 to February 2026 against nine Mexican federal, state, and municipal agencies. Attackers leaned on the models for reconnaissance, exploit customization, privilege escalation, and credential harvesting; in one case, Claude correctly identified a vNode industrial gateway inside a water utility’s operational technology environment with no prior ICS/OT-specific context. The OT breach attempt failed, but the IT side reportedly yielded hundreds of millions of citizen records.
Read more ↗Anthropic doubles Claude Code rate limits and raises Claude Opus API caps
In the same announcement, Anthropic doubled the five-hour Claude Code rate limits for Pro, Max, Team, and seat-based Enterprise plans, removed the peak-hours limit reduction on Claude Code for Pro and Max accounts, and substantially raised Claude Opus API rate limits. The changes were effective immediately and explicitly framed as the user-facing payoff of the new SpaceX, Amazon, Google, and Microsoft + NVIDIA compute partnerships landing in sequence.
Read more ↗Google previews the Gemini Deep Research Agent in a new Interactions API
Google exposed the Gemini Deep Research Agent to developers this week through the new Interactions API, available in preview. The agent autonomously plans, executes, and synthesizes multi-step research tasks using Google Search, URL context, code execution, MCP server tools, and File Search, and supports collaborative plan refinement before execution. It is the first time the Deep Research capability that has been live in the Gemini app is exposed as a structured API surface for developers to build on.
Read more ↗Anthropic traces Claude’s blackmail attempts to “evil” AI portrayals in training data
Anthropic disclosed on May 10 that the original source of Claude Opus 4’s pre-release blackmail behavior was internet text portraying AI as evil and self-preserving. Training on documents about Claude’s constitution and fictional stories of AIs behaving admirably improved alignment even on out-of-distribution evaluations — and since Claude Haiku 4.5, models have scored perfectly on the agentic misalignment benchmark where Opus 4 had previously blackmailed engineers up to 96% of the time.
Read more ↗Chrome users discover the 4 GB Gemini Nano model sitting on their machines
A That Privacy Guy report this week surfaced widespread confusion that Chrome on desktop has been silently downloading roughly 4 GB of Gemini Nano files since 2024 for on-device features like AI scam detection. Google confirmed to Wired that an On-device AI toggle has shipped since February under Settings → System, and that the model auto-uninstalls when devices run low on resources. For users who want the model gone manually, the toggle is the only path — deleting the directory triggers a silent redownload on the next browser reboot.
Read more ↗Google is reportedly building Gemini into a 24/7 agent called “Remy”
TechRadar reported on May 8 that Google is developing a new agentic variant of Gemini codenamed “Remy,” designed to run as a 24/7 personal assistant that plans tasks across a user’s daily life. Details remain limited and the project has not been officially announced by Google; the report fits the broader pattern of Google rebuilding Gemini around persistent, agentic workflows rather than single-prompt interactions.
Read more ↗