This Week In AI: September 21-27 2025

The week of September 21–27, 2025 was another reminder of how fast the AI landscape is shifting. OpenAI doubled down on compute with NVIDIA in a $100B partnership while launching Pulse, its first truly proactive ChatGPT feature. Anthropic cleared a major legal hurdle with a $1.5B settlement and announced a global hiring surge. DeepMind advanced embodied intelligence with Gemini Robotics 1.5, xAI landed its first U.S. federal contract, and Meta signaled its next act: bringing AI from glasses to humanoid robots. The pace, scale, and ambition are staggering, here’s what mattered most this week in AI.

OpenAI: Infrastructure Moonshots + More Proactive ChatGPT

OpenAI spent the week turning the dial to “max” on both infrastructure and product.

• NVIDIA partnership (Sept 22): OpenAI and NVIDIA announced a strategic partnership to deploy at least 10 gigawatts of NVIDIA systems, with NVIDIA intending to invest up to $100B as capacity is deployed. First GW lands in H2 2026 on NVIDIA’s Vera Rubin platform. This is framed as infrastructure for training/running the next gen on the path to superintelligence.
• Stargate expansion (Sept 23): OpenAI, Oracle, and SoftBank added five new U.S. data-center sites under the Stargate program. With Abilene, TX and ongoing CoreWeave projects, OpenAI now cites ~7 GW of planned capacity and $400B+ in investment over the next three years.
• ChatGPT Pulse (Sept 25): Product preview that flips ChatGPT from purely reactive to proactive: Pulse performs overnight research against your context (chat history, optional Gmail/Calendar integrations) and delivers a daily set of personalized, scannable briefings. Rolling out first to Pro on mobile.
• Real-world evals (Sept 25): OpenAI also published a research note on measuring model performance on real-world tasks, continuing the industry push beyond synthetic benchmarks.

Why it matters: The NVIDIA and Stargate moves signal that the compute arms race is accelerating with concrete build-outs and financing plans, measured in gigawatts, not just GPU counts. Pulse, meanwhile, nudges assistants toward agentic, context-aware workflows that “work while you sleep,” a theme echoed elsewhere this week.

Anthropic: Legal Milestone + Global Hiring Surge

Two big headlines shaped Anthropic’s week.

• Copyright settlement, preliminary approval (Sept 25): A U.S. federal judge preliminarily approved the $1.5B class-action settlement with authors, one of the first major resolutions in the wave of copyright suits against AI firms. Final approval still pending a claims/notice process.
• International expansion (Sept 26): Anthropic said it will triple its international workforce, 5× its applied-AI team, and add 100+ roles across Dublin, London, and Zurich while opening its first Asian office in Tokyo, citing that ~80% of Claude’s consumer usage is outside the U.S.

Why it matters: The provisional legal green light reduces immediate litigation overhang and may set a template for other settlements. The hiring push underscores enterprise demand for Claude (particularly strong in coding) and a strategic shift toward non-U.S. growth.

Google DeepMind: Robots Get More Agentic

DeepMind introduced Gemini Robotics 1.5 (a VLA model that “thinks before acting”) and Gemini Robotics-ER 1.5 (a VLM that plans multi-step missions and calls tools), with Robotics-ER 1.5 available via the Gemini API now and Robotics 1.5 to select partners. The models are pitched as enabling transparent planning and general-purpose physical tasks across embodiments.

Why it matters: This is a step toward embodied, tool-using agents, bringing reasoning and action closer together. It also keeps DeepMind squarely in the lead pack for robotics + LLM convergence.

xAI: Government Contract + Cheaper “Fast” Model

• U.S. federal contract (Sept 25): Reuters reported xAI secured a contract with the U.S. General Services Administration to make Grok available to federal agencies under the government’s AI adoption program. Details include agency access and pricing specifics relative to competitors.
• Grok 4 Fast (around Sept 21): xAI announced Grok 4 Fast, positioned as a faster, cost-efficient reasoning model with reduced “thinking token” use and large context capacity; multiple outlets covered the debut.

Why it matters: Beyond the daily leaderboard, landing federal adoption is a distribution and credibility win. Cheaper, leaner models suggest a coming split between “flagship max-IQ” and “good-enough + efficient” tiers across the industry.

Meta: From Glasses to Robots

Meta kept pushing a hardware+AI thesis this week.

• Robotics push: Reporting highlighted Meta’s new bet on humanoid robotics (“Metabot”) and a software-first platform strategy (think “Android for robots”), tied to its Superintelligence lab’s work on world models. Early but notable signal.
• Consumer AI devices context: Fresh analysis revisited Meta’s heavy Reality Labs investment, Ray-Ban Display glasses with live captioning/gestures at $799, and the company’s long-term play to challenge Apple in consumer AI hardware.

Why it matters: Meta’s hardware burn continues, but the glasses-as-AI-first device narrative is solidifying, and the robotics angle aligns with the industry’s pivot toward embodied AI.

Microsoft & The Enterprise Stack: Quiet But Steady Updates

Microsoft’s week wasn’t about splashy model drops but incremental product improvements, e.g., AI auto-categorization in the Windows Photos app (for Copilot+ PCs) and ongoing Copilot roadmap refinements across M365.

Why it matters: These are the kinds of “boring but everywhere” enhancements that make AI feel native across enterprise workflows, important even if they don’t move markets overnight.

Cohere: More Fuel For Enterprise AI

Cohere announced a $100M second close to its latest round to scale security-first enterprise AI and global operations (Sept 24).

Why it matters: The enterprise LLM segment remains well-funded, Cohere’s positioning emphasizes security, privacy, and control, a lane that continues to resonate with regulated industries.

Personally

Over the past week, I leaned heavily on AI across both work and play. I used it to write and verify a lot of code, to draft blog posts and YouTube scripts, spin up stylised cartoon thumbnails and cyber-robot images, and even pull together valuations and piece counts for the LEGO collection. It powered everything from SEO-friendly descriptions to LinkedIn and X posts, helping me keep the Redmond’s Forge story consistent across channels. On the professional side, I tapped into AI for research, synthesising the week’s AI news, sharpening strategy notes, and preparing for upcoming advisory conversations with 3 companies. In short, AI has become both my creative partner and my productivity stack, quietly running in the background to help me move faster and stay sharper.

The Bigger Picture: Three Clear Themes

1. Agentic assistants are going mainstream.
   OpenAI’s Pulse formalizes an assistant that acts unprompted (within your constraints). DeepMind’s Robotics models plan before acting in the physical world. Expect near-term product waves where assistants research, schedule, draft, and execute small tasks by default.
2. Compute as destiny.
   The OpenAI-NVIDIA tie-up and Stargate site expansion show that the frontier is now measured in gigawatts and siting strategy (power, cooling, transmission, regulatory throughput), not just model cards. This will shape which labs can train the next frontier models and how quickly.
3. Regulation and real-world risk management are crystallizing.
   Anthropic’s preliminary settlement approval and its rapid non-U.S. growth story capture two forces at once: legal risk is being priced and structured, while demand, especially outside the U.S., is exploding.

What to Watch Next Week

• Agent features across consumer assistants: Will more players mirror Pulse-style daily briefings?
• Robotics demos leveraging Gemini Robotics 1.5, especially third-party partner videos or dev blogs.
• Enterprise procurement momentum: follow-on contracts after xAI’s GSA deal; any new public-sector pilots from Anthropic/OpenAI/Cohere.
• Power & siting: concrete permits, grid interconnect updates, and supplier wins tied to those five Stargate sites.