Welcome back! The industry is shifting this week from open exploration to strategic control. One superpower is putting up a hardware firewall to force a domestic pivot, while a creative software giant is dissolving its own borders to live inside a chatbot. At the infrastructure layer, two research labs dropped competing agents that promise to fix the “hallucination loops” that have held AI back, and e-commerce stores are now being tested by AI customers before humans ever log on.
In today’s Generative AI Newsletter:
China blocks Nvidia’s H200s to prioritize local rivals.
Adobe turns ChatGPT into a free creative starter pack.
Google fights GPT 5.2 with deep research agents.
Shopify lets AI shoppers stress-test your storefront.
Latest Developments
Washington allowed approved Chinese buyers to purchase NVIDIA’s H200, framing it as a controlled release and taking a reported 25% share of NVIDIA’s revenue. Recently, the White House AI czar David Sacks said China is rejecting the chips. The paperwork points out that Beijing can delay, limit or regulate purchases even if Chinese AI teams still need the compute. If China keeps the gate tight, NVIDIA’s estimated $10 billion a year stays theoretical.
Here's what the trail actually shows:
Gatekeeping: China reportedly wants case-by-case permits to regulate the purchase of NVIDIA's H200 chips.
Demand: Firms like ByteDance and Alibaba still want H200s because more GPUs let teams train larger models faster and run more user queries per dollar
Security: Officials consider trust, ‘a hardware requirement’ because hidden access could mean hidden access that monitors workloads, leaks sensitive data
Leverage: Washington trades access for revenue share and control to prioritize financial gains and influence
This narrative aligns with the emerging trend in the AI industry where chips move like controlled goods rather than conventional products. The upside is having more GPUs can lead to better models and cheaper inference, but compliance issues slow down builders and create barriers for approved customers. Companies already diversify suppliers and design fallbacks across multiple chip stacks. As AI evolves, maintaining a balance between security and efficiency will be essential for companies looking to remain competitive.
Special highlight from our network
Poor data costs companies millions every year.
Not just in tools. in bad decisions.
Databricks and mindit.io are hosting a 30 minute live session on how to build real trust in the data behind your analytics and AI.
You will see how to trace data across its lifecycle, monitor quality, and protect sensitive information so leaders can rely on the numbers they see.
If you work in data governance, BI, analytics, platform or AI, this will give you practical steps you can put to work right away.
Special highlight from our network
Join our fast-paced workshops where we don’t just talk about tools; we show you how to use them.
Next up:
Slop-proof AI writing
Tue, Dec 16 · 9–10 am PT
Learn how to verify facts, check sources, and improve tone and clarity with AI.
We’ll demo: Perplexity, Gemini, and YAi Digital.
Creating impactful slide decks with AI
Thu, Dec 18 · 9–10 am PT
Turn rough ideas into polished, high-impact presentations.
Tools: Google Slides, Nano Banana, and Chronicle.
Every session includes access to 100 free AI tools and exclusive discounts.
Adobe has integrated its flagship tools, such as Photoshop, Express and Acrobat, into ChatGPT. Users can interact with these tools as built-in apps in ChatGPT, allowing commands like 'blur the background,' 'fix this logo,' or 'clean up this contract' to be handed over to the appropriate app. Adobe aims to make creative work easier for users and expand its reach to 800 million weekly ChatGPT users. The chat window represents Adobe's most significant agentic AI integration to date.
Here’s how the new setup behaves:
Editing: To edit images or PDFs, type instructions while Adobe apps display tools, sliders and effects.
Pricing: The apps are free within ChatGPT, but there are reports of missing features, delays on Android, and a 20MB image limit.
Data: OpenAI can train its models using customer files unless you enable the 'do not train on my content' setting.
Agents: Adobe considers this its most significant agentic AI advancement to date, competing with Gemini-style smart editing.
Overall, Adobe gets distribution inside the leading generative AI app while users get easier editing with fewer tools to learn. That future looks smooth for users but it also concentrates power to a single assistant between them and their files. The advantage lies in speed and access, however, the downside is that creative workflows may gradually rely on a single AI gatekeeper determining the tools, formats, and business models users encounter.
Google has released an upgraded Gemini Deep Research agent, opening it to developers for the first time and expanding its role beyond static reports. Built on Gemini 3 Pro, the system is designed for long running research tasks that require repeated search, synthesis, and correction across large context windows. The launch arrived the same day OpenAI introduced GPT 5.2, placing two competing visions of agentic research into production at once.
Deep Research capabilities:
Agent Design: Gemini Deep Research plans queries, reads results, identifies gaps, and loops until it reaches a stable answer.
Developer Access: The new Interactions API lets teams embed Google’s research agent directly into their own apps.
Product Path: Google plans to integrate the agent into Search, NotebookLM, the Gemini app, and Google Finance.
Benchmarks: Google open sourced DeepSearchQA and reported strong results on Humanity’s Last Exam.
For researchers, the signal sits below the press release. Google is optimizing for error accumulation, one of the hardest problems in agent design. Long horizon tasks fail when a single hallucinated decision slips through, and Gemini 3 Pro is being positioned as a model that can survive those chains intact. OpenAI’s GPT 5.2 counters with its own claims across standard evaluations, and the timing shows how tight this cycle has become. Research agents are no longer demos. They are becoming infrastructure, and whoever stabilizes them first defines how machines learn to ask better questions at scale.
SimGym is Shopify’s new experiment that lets you watch AI shoppers browse your store before real customers ever arrive. Think of it as a dress rehearsal for your storefront. Instead of guessing whether a redesign, campaign, or navigation tweak will work, you can see how simulated shoppers behave first.
These AI shoppers move through your site like people do. They click around, scroll collections, add items to cart, and sometimes leave without buying. The value is not prediction. It is visibility into friction.
What you can do with it:
Compare themes safely: Test your live theme against another installed theme and see which one drives more add to cart actions
Test big changes early: Try seasonal layouts, promotions, or navigation updates before launch
Watch real behavior patterns: See where shoppers hesitate, loop, or abandon
Read plain feedback: Get written notes on navigation clarity, layout flow, and checkout experience
Try this yourself:
Install SimGym from the Shopify App Store and join the waitlist. Once approved, choose your active theme and one alternative. Start a simulation and let AI shoppers explore both versions. When it finishes, review the winning theme, replay individual sessions, and look for repeated issues.
When it helps most:
SimGym works best if your store already has products and content. Use it to catch obvious friction before traffic spikes. It will not replace real customers, but it can save you from shipping changes blindly.
