From Chips to Copilots: The Week AI Got More Powerful — and More Cautious
Notes by SDG ·
Theme of the week
“Infra arms race + safety blueprints + assistants growing up.”
Cloud providers, frontier labs, and enterprise platforms all pushed on three fronts this week:
- Massive AI compute & infra deals
- Safety & sovereignty (especially for teens and regulated industries)
- Next-gen assistant UX (Copilot, ChatGPT, Sora, Gemini, etc.)
1. Infra & Market Pulse
Google’s new AI chip + Anthropic megadeal
Google unveiled a new AI chip platform with ~4× performance vs its prior generation, aimed squarely at inference efficiency at scale. At the same time, it locked in a “megadeal” with Anthropic, giving Anthropic access to up to 1 million TPUs as it chases a long-term target of $70B revenue and positive cash flow by 2028.
Why it matters:
- Confirms that custom AI silicon + exclusive capacity deals are now a core competitive moat.
- Anthropic gets a guaranteed runway of compute; Google gets a flagship frontier-lab customer on its stack.
AWS × OpenAI: multi-year, multi-billion partnership
OpenAI and AWS announced a multi-year strategic partnership: AWS will provide EC2 UltraServers with hundreds of thousands of chips, scaling to tens of millions of CPUs, in a deal reportedly worth $38B.
Why it matters:
- This meaningfully diversifies OpenAI’s infra beyond Microsoft Azure, even as the MS partnership remains central.
- For AWS, it’s a huge AI marquee win and a validation of its custom infra story in the frontier-model era.
Data sovereignty as a selling point
Microsoft announced that Microsoft 365 Copilot interactions can be processed in-country for customers in 15 markets — starting with Australia, UK, India, and Japan by end of 2025, then expanding to 11 more countries including Canada, Germany, Italy, Spain, Sweden, Switzerland, UAE and the US in 2026.
So what:
- “Where is my AI data processed?” is now a core buying criterion, especially for public sector and regulated industries.
- Expect “AI sovereignty” to show up in more RFPs and marketing decks.
2. Frontier Lab Watch (OpenAI, Anthropic & Co.)
OpenAI’s “AI progress and recommendations”
OpenAI published a long policy/vision essay laying out how they see the next few years:
Key points:
- They expect AI systems to go from doing hour-scale tasks now to day/week-scale tasks in the next few years.
- They suggest small scientific discoveries by ~2026 and more significant ones by 2028, while acknowledging uncertainty.
- They call for:
- Shared safety standards across frontier labs
- An “AI resilience ecosystem” (analogous to cybersecurity)
- Ongoing reporting on real-world impacts and stronger public oversight
Teen Safety Blueprint + parental controls
OpenAI also released a Teen Safety Blueprint, pitched as a framework for making AI safer and more supportive for teens. It emphasizes age-appropriate design, stronger safeguards, and continuous evaluation.
In parallel, they highlighted recent launches like:
- Parental controls in ChatGPT (time limits, content controls, turning off memory/image gen, etc.).
- Work toward age-prediction systems to better tailor teen experiences.
There’s also external scrutiny: some reporting suggests that despite improvements, ChatGPT’s behavior on acute mental-health prompts can still be inconsistent and risky, and experts are calling for more robust safeguards and evaluation.
Anthropic doubles down on Google Cloud
Beyond the TPU mega-deal above, Anthropic and Google reaffirmed their expanded partnership on Google Cloud TPUs and services, explicitly to “define the frontier of AI” and support surging Claude demand from enterprises and AI-native startups.
Takeaway: the frontier-lab race is increasingly an infra + safety + enterprise-go-to-market race, not just a parameter count race.
3. Assistants & Copilots
Copilot: “best of AI search” rollout
Microsoft announced a new AI search experience inside Copilot across copilot.com, iOS/Android apps, and Edge:
- More prominent, clickable citations and the ability to see aggregated sources
- A dedicated search mode inside Copilot for deeper, more precise retrieval
- In Edge, a new Copilot Mode in the address bar lets you tab into Copilot search directly
Signal: Copilot continues to blur lines between search, browser, and assistant, while trying to differentiate on trust and source transparency.
ChatGPT & product updates
- ChatGPT Go promo in India – Eligible customers in India can get 12 months of ChatGPT Go free, redeemable via web or Android (iOS coming shortly). ChatGPT Go is the low-cost plan with more messages, bigger uploads, and advanced data analysis.
- Clarification on medical & legal advice – Despite confusion around updated terms, OpenAI confirmed that ChatGPT will still answer health/legal questions, but reiterates it is not a substitute for professionals and shouldn’t be used to treat others or pose as an expert.
- Mental-health handling updates – GPT-5 Instant has been updated to better detect and respond to distress, with routing of sensitive content to safer models, though external reporting suggests there’s still a gap between policy and real-world behavior.
Sora, Aardvark & other assistant moves
A recap of the week’s curated AI-news roundups highlights several OpenAI & ecosystem updates:
- Sora 2 – Expanded beyond early access; available in the US, Canada, Japan, and South Korea, with a credit-pack model for additional generations.
- Aardvark (OpenAI) – A GPT-5–based AI security research assistant that can scan codebases, build threat models, and propose patches, with free scanning for non-commercial OSS.
- Chrome + Gemini – Chrome Canary adds Gemini-powered Nano Banana image gen & Deep Search, nudging Chrome toward an AI workspace model.
4. Tools, Infra & Developer Stack
Microsoft launches its own image generator: MAI-Image-1
Microsoft’s first in-house image model, MAI-Image-1, is now live in Bing Image Creator and Copilot (but not yet in the EU).
- Strong at food, nature, and complex lighting/reflections
- Focuses on a balance of speed + quality
- Sits alongside DALL·E 3 and GPT-4o in Bing’s image tools
This marks another step away from total reliance on OpenAI models inside the Microsoft ecosystem.
Google developer stack updates
From Google and the broader ecosystem:
- Magika 1.0 – Google’s AI-based file-type detection tool hits a stable 1.0 release, rewritten in Rust and widely used (1M+ monthly downloads).
- Vertex AI Agent Builder updates – New capabilities to take agents from prototype → production with better context management, observability, evaluation, and agent identities.
- Gemini Nano Banana 2 – Next-gen image model planned, promising faster, more diverse visuals with built-in watermarking.
Data foundations for enterprise AI
A Salesforce “State of Data & Analytics” report says 84% of tech leaders feel their current data can’t support their AI ambitions, citing incomplete/poor-quality data and silos; those using zero-copy architectures are significantly more likely to succeed with AI.
Takeaway for enterprises: “AI strategy” without a data modernization strategy is still theater.
5. Governance, Risk & Opinion
Beyond OpenAI’s policy pieces, a few notable “temperature checks” on AI risk and timelines this week:
- Hinton’s warning (again) – Geoffrey Hinton reiterated that AI might eventually outsmart/overpower its creators and argued we may need a “Chernobyl moment” to take the risks seriously.
- AGI timeline debates – Commentators discuss Anthropic’s internal timelines suggesting Nobel-level AI researchers by ~2027 vs OpenAI’s somewhat more conservative forecasts (~2028) for automated AI research.
- Both threads underscore the tension between rapid capability scaling (backed by those giant infra deals) and the governance structures that are still catching up.