• Everyday AI
  • Posts
  • Ep 823: The U.S. vs China AI Cold War Is Starting: What It Means and How It Impacts You

Ep 823: The U.S. vs China AI Cold War Is Starting: What It Means and How It Impacts You

U.S. and China to hold AI talks, Gemini launches 3 new models, Claude Cowork brings recordable skills, Gemini Notebook's big updates and more.

Author

Everyday AI
July 21, 2026

 

Outsmart The Future

Today in Everyday AI
8 minute read

🎙 Daily Podcast Episode: The U.S. and China are both moving to control AI access. Here's how that could affect your AI strategy. Give today’s show a watch/read/listen.

🕵️‍♂️ Fresh Finds: Google announces Gemini 4 pre-training, Anthropic settles for AI copyright, OpenAI model breaks out of its sandbox and more. Read on for Fresh Finds.

🗞 Byte Sized Daily AI News: U.S. and China to hold AI talks, Judge grants massive settlement against Anthropic, OpenAI model broke out of its sandbox and more Read on for Byte Sized News.

💪 Leverage AI: China and the U.S. are both moving to limit AI access. Here's what businesses should do before new rules reshape the AI market. Keep reading for that!

↩️ Don’t miss out: Miss our last newsletter? We covered: Meta is considering a $10 billion AI deal with Anthropic, Moonshot paused new Kimi K3 subscriptions, and Microsoft is bringing AMD's AI racks to Azure. And more. Check it here!

Ep 823: The U.S. vs China AI Cold War Is Starting: What It Means and How It Impacts You

A U.S. vs China AI cold war is starting, and most business leaders have no idea they're already in it.

China's open models just closed the gap with America's best, oftentimes at a fraction of the price.

Now both governments are moving to wall off their AI within days of each other.

Why? Because this was never about benchmarks. It's about power y'all.

We break it all down on today's show and help you figure out the 101 of the AI war between U.S. and China.

Also on the pod today:

• China limiting AI model exports 🇨🇳 
• US may ban Chinese AI 🇺🇸🚫 
• GLM 5.2 shakes up pricing 💸 

Listen on our site:

Click to listen

Subscribe and listen on your favorite podcast platform

Listen on:

Here’s our favorite AI finds from across the web:

New AI Tool Spotlight –  Lev8 Turns the live web into people and business intelligence, Cart AI is The AI agent that handles checkout, Routine AI is all your work in one place, powered by AI

Grok in Excel — Grok now works right inside Excel, letting you ask questions, write formulas, and run scenarios without leaving your spreadsheet.

TSMC Chip Prices — TSMC is hiking chip prices by up to 10% starting in 2027 to cover higher costs.

Hugging Face Hackers — Hackers broke into Hugging Face by exploiting a security flaw, grabbing internal credentials and possibly more.

Sony AI Lawsuit — Sony just hit AI music startup Udio with a second lawsuit, claiming the company illegally used over 30,000 songs for training.

Gemini 4 — Google announced it is already pre-training Gemini 4.

X on Android — The Android X app just got a full rebuild, boasting smoother performance and more reliability.

Sakana Fugu-Cyber — Sakana just launched Fugu-Cyber, a new orchestration model that matches the best in real-world cybersecurity benchmarks.

AI and Mathematics — Anthropic’s Fable 5 may have found a counterexample to the famous Jacobian conjecture, shaking up algebraic geometry.

AI and Legal — Anthropic just settled for $1.5 billion over using pirated books to train its AI. 

AI Research — OpenAI paused a research model after it broke out of its sandbox, even sneaking onto GitHub. Curious how it pulled it off?

1. U.S. and China plan September AI talks before Xi’s Washington visit 🇨🇳

According to Reuters, Washington and Beijing are preparing their first official AI dialogue under President Trump, with talks expected in September ahead of President Xi Jinping’s planned September 24 U.S. trip.

Treasury Secretary Scott Bessent is expected to lead the U.S. side as both countries confront concerns that advanced AI could sharpen military, cyber, and economic competition.

2. Google Launches Gemini 3.5 Flash Cyber, 3.6 Flash, and 3.5 Flash-Lite ⚡️

Google released three Gemini models Tuesday, led by a limited pilot of Gemini 3.5 Flash Cyber, built to find and patch software flaws for governments and trusted partners.

The move is Google’s clearest answer yet to Anthropic’s early lead in automated cyber defense, while Gemini 3.6 Flash and lower-cost Flash-Lite target coding and high-volume business work. Arriving just before Alphabet’s earnings, the rollout shows Google is betting that lower costs, custom chips, and tighter hardware-software design can help it catch up in an increasingly crowded AI race.

3. Claude Cowork Adds Screen-Recorded Skills for Pro, Max, and Team Users 🎥

Anthropic has introduced “Record a skill” in Claude Cowork, letting subscribers teach Claude a repeatable desktop workflow by recording and narrating it rather than writing instructions or code. The feature, similar to the ‘Record and Replay’ feature from OpenAI’s Codex, captures actions such as clicks, typing, and spoken guidance, then saves the process as a reusable skill that can be run later.

Available in the Claude desktop app’s plus menu for Pro, Max, and Team plans, it pushes Cowork further toward practical task automation for people who would rather show than tell.

4. Gemini Notebook Rolls Out Collections for Better Notebook Organization 📓

Google is gradually adding Collections to Gemini Notebook, formerly NotebookLM, giving users a way to group related notebooks without forcing everything into rigid folders.

According to Android Authority, notebooks can appear in multiple Collections or remain in the standard My Notebooks list, making the feature optional and flexible. The update addresses a growing problem for frequent users: long, cluttered notebook lists as research and projects pile up.

5. Microsoft Commits Billions to Mistral’s European AI Data Center Expansion 🏢

Microsoft on Tuesday announced a multibillion-dollar commitment to support and use Mistral’s expanding European data centers, though it did not disclose an exact figure. Mistral plans to scale the infrastructure with thousands of NVIDIA Vera Rubin GPUs, while its Medium 3.5 and OCR 4 models join Microsoft Foundry and Medium 3.5 enters Copilot Studio.

The tie-up gives regulated businesses more control over where their AI runs, including on their own systems without an internet connection, and expands Microsoft’s European infrastructure push without a new equity investment in Mistral.

A US-China AI cold war just kicked into high gear. There's a good chance your business is ALREADY a participant.

Wait, us?

Reuters says China is weighing limits on who gets its best AI models, and just days later, the Trump administration signaled it might restrict Chinese AI models here at home.

Same fight, two capitals, one week.

Wait, isn't this just a nerdy model benchmark story?

Nope.

This might be the biggest AI story most executives aren't tracking right now, and the ones paying attention are already moving.

So what's actually at stake? Whoever wins this race writes the rules your AI budget and roadmap have to live inside.

Don't worry, you don't need a geopolitics degree, just three moves before this gets messier.

On today's episode of Everyday AI, we broke down why open source needs a data center, why the Chinese discount disappeared, and what to inventory before rules harden.

Don't get stuck in the middle. Let's get after it.

1. China's Open Models Got Too Big To Run 🖥️

GLM 5.2, Kimi K3, and Qwen 3.8 pushed things into trillion parameter territory this year.

So what changed?

These models got so big they need a small data center, or a full rack of NVIDIA GPUs, just to run. That's less a laptop problem now and more a procurement one.

Translation: open source just split into two tiers, consumer and enterprise, and only enterprise is winning the frontier race.

If your plan was saving money by self-hosting a free Chinese model, check the compute first, or you're still paying API prices for a discount that isn't REAL.

Try This

This week, check whether any Chinese open source model your team uses can run on your own hardware, or whether you're just paying cloud pricing for it anyway.

Pull that number and put it side by side with GPT-5.6 Terra or Luna pricing. Nine times out of 10, "free" isn't actually cheaper once you add up the infrastructure.

Do it before renewal.

2. The Cheap Chinese AI Discount Just Died 💰

For years, the pitch for Chinese open models was simple: almost as smart, way cheaper. That math just broke.

On cost per task, Anthropic's models run about $2.75 a task, OpenAI's GPT-5.6 Terra runs about 82 cents, and Kimi K3 and Qwen 3.7 now cost MORE than Terra, not less. Wayyy more than anyone expected.

On the "smart and cheap" quadrant that Artificial Analysis tracks, exactly three companies show up now: OpenAI, Grok, and Meta.

So what's the big deal here?

If you picked up a Chinese model purely for the discount, that discount might be gone, and you're prolly still paying API rates for it. That's a budget line worth checking before next quarter's renewal.

Try This

Pull up your current AI vendor bill this week. If any line item is a Chinese model you're accessing through an API instead of self-hosting, compare it directly against GPT-5.6 Terra or Luna pricing.

Nine times out of 10 now, the "cheap" Chinese option isn't actually the cheap option. Either way, you walk into budget season with real numbers, not a vibe.

3. Inventory Your Chinese AI Models Now ⚠️

Here's the gnarly part.

Even Microsoft, sitting on massive stakes in both OpenAI and Anthropic, is reportedly testing DeepSeek to cut costs inside Copilot Cowork, which is basically Microsoft admitting the multi-model future already got here.

Both governments, meanwhile, are closing in from opposite directions, Beijing outbound, Washington inbound.

Once model weights are out, they can't be pulled back, but new restrictions can still make support and legal cover a lot harder to get.

So what should you actually do?

Inventory where Chinese open models are running in your stack today, route those tasks by cost, security, and policy risk, document your sourcing, and have a backup model ready so a policy change doesn't stall your whole workflow.

Try This

Before Friday, list every Chinese open source or open weight model touching your business, whether it's self-hosted, accessed via API, or baked into a tool you already pay for.

For each one, note why you're using it and the fallback US model. That's your insurance policy if restrictions harden fast.

That gives legal and security something real to review, not a guess.

Reply

or to participate.