GM. This is Milk Road AI, where we follow the chips, the chaos, and the chess moves shaping the future of AI.
Here’s what we got for you today:
✍️ Did the U.S. just hand china a cheat code?
🎙️ The Milk Road AI Show: The Macro Force of the Decade: How AI, Energy & Geopolitics Are Merging w/ Kim Isenberg
🍪 Meta’s new “avocado” model might shake up the AI race
Okara is a privacy-first AI workspace for teams to collaborate with 20+ open-source AI models using common context, memory and knowledgebase. Create a Workspace to get access to a private collaborative AI chat.
Prices as of 10:00 AM ET.
AMERICA’S AI BACKFIRE: WHAT DID WE JUST DO?
Have you ever played the game of Whac-A-Mole and no matter how hard you smack the thing, it pops up somewhere else smiling at you?
Yeah, that’s the U.S. chip ban on China right now.
For three years, the U.S. strategy was simple: "Starve the Dragon".
We banned the sale of Nvidia's most powerful chips to kill China's AI progress before it could even start.
Spoiler Alert: the plan face-planted.
So guess what the White House just did, it flipped its entire narrative.
President Trump has authorized Nvidia to sell its powerful H200 AI chips to "approved customers" in China.
But there is a massive catch.
Uncle Sam is taking a 25% cut of every sale. A literal AI tax.
And just as the approval dropped, Beijing pulled its own Uno Reverse Card.
The Financial Times reports China is drafting rules to restrict access to the H200, even though they desperately need it.
Buyers will need to justify why Huawei (China’s top domestic chipmaker) can’t meet their needs before touching an American chip.
The U.S. tried to play the "Drug Dealer" (keep them addicted to Nvidia).
China just checked itself into "Rehab" (forced domestic adoption).
And it only gets crazier:
Both governments are slamming the brakes for completely different reasons but they’re arguing over a chip most people don’t even understand.
To see why this matters, you need to know what the H200 actually is.
Nerd corner: Why the H200 actually matters
When you train a giant AI model, the GPU doesn’t slow down because it's bad at math, it slows down because it runs out of data to process.
Picture a factory:
The GPU is a super-fast worker.
The data is the material it needs.
The memory system is the supply truck delivering that material.
If the truck is slow, the worker keeps stopping.
Not because he’s lazy but because he has nothing to work on.
That’s what slows down AI training on older chips.
The H200 fixes exactly that.
It uses a newer memory technology called HBM3e, which is basically a hypercharged supply system for the GPU.
It brings:
Much bigger loads (almost double the memory).
Much faster delivery (40% more bandwidth).
Almost no downtime (up to 45% faster training).
This is why every AI lab wants the H200 on their racks.
And here’s the part nobody seems to get: The H200 is not outdated even though it's a year old.
It belongs to Nvidia’s Hopper generation and 18 of the 20 most powerful public GPU clusters on Earth still run primarily on Hopper.
Blackwell (Nvidia’s next-generation GPU architecture after Hopper) is coming eventually, but it won’t take over until 2026 or later because it only began shipping in volume in mid-to-late 2025.
The H200 isn’t a small upgrade, it’s a different species.
On Nvidia’s own performance metrics, the H200 delivers roughly six times the compute performance of the crippled H20 China used to be allowed to buy.
This chart says it all: the H20 barely clears export thresholds, while the H200 blows past them like a drag racer passing a scooter.
This is the strongest chip the U.S. has EVER approved for export to China.
When teammates use AI tools separately, the important conversations and decisions get lost. No one can see what others have learned, so work gets repeated and knowledge disappears.
Okara Workspace solves that, it is a private, shared AI chat workspace with common knowledge, context and memory; built for teams that care about privacy.
Access 20+ open-source AI models for text and image generation
Brainstorm together in one shared chat
Debug and solve technical problems with unified memory
If you're building in Web3, security shouldn't be optional. Choose AI that aligns with your values.
AMERICA’S AI BACKFIRE: WHAT DID WE JUST DO? (P2)
The great debate: Genius or suicide?
This move has split Washington into two camps and they both think the other side is delusional.
Team Genius (Nvidia, economists, pragmatists)
Their argument:
If you cut China off, they’ll build their own chips faster.
If you sell to China, you keep them dependent on Nvidia and CUDA (Nvidia’s GPU software).
CUDA is so deeply integrated into Chinese AI startups that switching to Huawei would be painful and slow.
So instead of China sprinting toward self-reliance, they stay in Nvidia’s world, paying the U.S. a 25% “Trump Tax” for the privilege.
In this view, the U.S. isn’t helping China, it’s taxing China’s AI future and starving Huawei of demand.
Team Suicide (Pentagon, security hawks, Congress)
This group says:
Huawei can’t match the H200 until 2027 or later
China’s chip production capacity will be 1–2% of U.S. output in 2026
Letting China access frontier compute creates direct national security risks
In their eyes: Selling H200s doesn’t slow China, it rescues them.
It lets China skip years of painful R&D losses for a 25% fee.
And the scariest part?
AI progress = compute capacity.
More GPUs → bigger models → faster breakthroughs.
And opening this H200 export channel directly shifts the global balance of AI power.
One analysis shows:
If the U.S. keeps exports nearly closed → the U.S. and close partners control ~75% of global AI compute by 2026, while China stays stuck in the low teens.
With conservative H200 exports, China’s share rises meaningfully.
With aggressive exports → China’s share jumps to around one-third of global compute, while the U.S. share drops sharply.
In other words: Every H200 that crosses the Pacific changes the future AI power map.
The leak: Operation Gatekeeper 🕵️♂️
The final nail in the “total ban” coffin?
The ban was leaking like a cheap roof.
The DOJ uncovered a massive smuggling ring called Operation Gatekeeper, and the details read like a crypto heist Netflix would greenlight overnight.
A Florida company called Janford Realtor, yes, Realtor, wasn’t selling condos. They were buying millions in GPUs.
The path looked like this:
Buy restricted chips in the U.S.
Ship them to fake companies in Singapore and Malaysia
Repackage them as “furniture”
Fly them into China
In just a few months, over $160 million in banned chips moved through this pipeline.
Chinese firms wanted these GPUs so badly they were out here paying 3x–4x markup.
You don’t pay triple retail for something you kind of need.
You pay it when your entire AI strategy is gasping for air.
China can pretend to play hard to get, but the second this deal hits the table, they’re signing papers with both hands.
Pride doesn’t train LLMs. Compute does.
And when they cave? Nvidia’s revenue goes vertical, their earnings sheet catches fire, and Jensen Huang shows up to the next keynote flexing so hard that it blinds half the front row.
Alright, that’s it for this edition of Milk Road AI, but we’d love to know what you think.
Hit reply and tell us:
Are you Team Drug Dealer (Smart move, keep them addicted and tax 'em!) or Team Bailout (We just armed the enemy and saved them from failure)?
Stay thirsty, my friends. 🥛
THE AI POWER SHIFT IS HERE ⚡
In today’s episode, we sat down with Kim Eisenberg, co-founder and chief editor of Superintelligence, to talk about how AI, energy, and geopolitics are colliding to reshape the global economy.
Here’s what you’ll hear:
Why AI + robotics are creating real job displacement risks, not just worker “augmentation”.
How compute and energy bottlenecks are becoming the biggest limiting factors in global AI rollout.
Why the U.S. and China are pulling ahead while Europe lags behind.
Which sectors and technologies investors should watch as AI-driven capex booms.
Hit play and get the full scoop. 👇
Incogni helps to remove people's private information online. Sign up with code MILKROAD for 55% off.
Bridge helps businesses send, store, accept, and launch stablecoins instantly. Learn how Bridge is powering stablecoin-backed money movement.
Wanna get in on the next iteration of AI domains, before everyone else? The Singulant just launched .AI4 domain names – check them out here.
Google’s Gemini is gaining fast on ChatGPT after its v3 release boosted downloads and traffic. Monthly active users jumped toward 350M while ChatGPT grew only slightly.
Meta is gearing up to launch its new “Avocado” AI model in Q1. It will upgrade Meta AI across Instagram, Facebook, and Messenger with stronger multimodal features.
Google is baking Gemini directly into its 2026 Ads system. It will generate ad creative, automate targeting, and shift Google’s entire ad platform into an AI-native engine.
