GM. This is Milk Road AI, the only newsletter that reads the fine print so you don't have to.

​Here’s what we’ve got for you today:

Nexo is back in the U.S. - and new clients get 30 days of Wealth Club Premier perks! Higher yields, lower borrowing rates, and crypto cashback - start here.

META IS PLAYING A DIFFERENT GAME

Warren Buffett has one test for any business he considers buying: "How big is the moat?

Margins can be engineered, and management can be replaced. 

But the moat is the only thing that answers: what happens when a well-funded competitor shows up tomorrow? 

For three years, the AI industry has been obsessing over benchmarks.

GPT-5 beats Gemini on reasoning, Claude wins the coding eval, Llama 4 is a disaster.

This is the sports equivalent of arguing about jersey colors while the other team buys the stadium.

Here is what the actual scoreboard looks like:

In total, 3.5B people open a Meta app every single day.

OpenAI has 900M weekly ChatGPT users, which is genuinely impressive, while Anthropic doesn't publish their user numbers, which tells you everything about their user numbers.

But here is the key difference.

OpenAI has to convince you to download a new app, create an account, learn a new interface, and pay $20 a month before you ever interact with their AI.

Meta's AI already lives inside the app you opened seventeen times before lunch.

That is not just a competitive advantage but a completely different ball game.

The money printer funding the moonshot

Before we delve into the details, who is actually paying for all of this?

OpenAI just raised $122B at an $852B valuation and is projected to burn through $665B in cash before it ever turns a real profit between now and 2030.

Meanwhile, Anthropic is running the same playbook at a smaller scale, with training costs expected to exceed $100B before 2030.

OpenAI has Microsoft and a revolving door of investors keeping the lights on, while Anthropic has Amazon and its own crew doing the same.

Meta on the other hand is printing its own money.

In Q4 2025, Meta generated $59.9B in revenue, up 24% year over year, and $22.8B in net income.

That is roughly $171,860 per minute of pure profit, and nobody else in AI can say that. 

It is like showing up to a poker table where everyone else is borrowing chips, and your salary gets auto-deposited into your stack every single hand.

Zuck’s three-part master plan

None of this was an accident; Zuckerberg executed a very specific, very expensive three-part plan in 2025. 

Most people missed it because they were too busy dunking on the Llama 4 benchmarks.

Part 1: Buy the brain

In mid-2025, Meta dropped $14.3B for a 49% stake in Scale AI and hired the supergenius 28-year-old CEO Alexandr Wang to run a brand new Meta Superintelligence lab.

Scale AI isn't just a data company; they employ a literal army of PhDs and mathematicians who act as human tutors for AI models, grading the AI's homework until it stops pattern-matching and actually learns to reason.

But here's the real reason the price tag had nine zeroes.

Wang knows his competition inside and out better than anyone.

Because every major lab relied on Scale AI, Wang knew exactly what each competitor was building and where they were failing.

That is an enormous strategic advantage.

Part 2: Buy the hands

In early 2026, Meta spent roughly $2.5B to acquire Manus AI.

Here is the difference between normal AI and what Manus does. Most AI assistants are a chat window, you type, and it responds. 

The whole interaction is you doing the work with a smart autocomplete watching.

Manus is autonomous, and you give it a goal: "research my competitors, draft the report, book the room," and you walk away.

Think of it as the difference between a GPS that reads you directions and one that drives the car.

Meta has 200M businesses on its platforms. 

If even a fraction pays a monthly fee for an AI agent that runs their customer service, books their appointments, and chases their leads, the revenue math gets offensive very quickly.

(I ran the numbers three times. It didn't get less offensive.)

Meta is already starting to wire Manus directly into its ecosystem. 

Inside Ads Manager, Instagram, and WhatsApp Business, these agents can analyze campaigns, recommend improvements, match brands with creators, and even handle customer conversations automatically.

Instead of jumping between tools, the AI sits in the background doing the work, turning Meta’s entire platform into one continuous operating system for businesses.

If this works at scale, this is a new layer on top of the entire Meta economy.

Part 3: Build the body

This is the part everyone is still sleeping on.

While the internet spent two years dunking on Zuckerberg for torching billions on VR headsets nobody asked for, he was quietly getting the hardware right.

The pivot was simple: stop making people strap computers to their faces and build glasses that a billion people already wear every day.

Ray-Ban Meta glasses sold roughly 7M units in 2025, more than triple the combined total from 2023 and 2024.

The newest version has in-lens AR, live translation, and a neural wristband that reads electrical impulses in your forearm so you can control it with silent hand gestures.

You could walk through a Tokyo market, look at a menu, and have it translated live in your lens before you've finished reading the first line, and it looks like a normal pair of Ray-Bans.

That is a new operating system for reality.

And that’s exactly where the story starts to break.

Because the same strategy that could make Meta unstoppable could also be its biggest weakness

META IS PLAYING A DIFFERENT GAME (P2)

Alright, I'll stop cheerleading.

Because there are real cracks in this plan that deserve more than a footnote.

Crack #1: The AI godfather walked out

Yann LeCun, a Turing Award winner, Meta's Chief AI Scientist for twelve years, resigned in November 2025.

When Zuckerberg hired Wang and handed him the keys to Meta's entire AI operation, LeCun, one of the most celebrated computer scientists alive, was suddenly required to report to a 28-year-old whose previous company labeled data for other people's AI models. 

He didn't take it well, and Bloomberg reported he "clashed with others internally." 

So what did he do? He announced he was starting AMI Labs, Advanced Machine Intelligence and he raised $1.03B from global investors before most people had even heard the company's name.

That is the market telling you it thinks LeCun might be right.

His argument, which he has been making loudly for years, large language models are a dead end. 

They are useful but fundamentally incapable of actual reasoning, planning, or understanding the physical world.

He said it bluntly at an MIT symposium before he left: "Within three to five years, world models, not language models, will be the dominant architecture in AI, and nobody in their right mind would use LLMs of the type we have today."

World models are his alternative. 

AI that learns like a child by watching, interacting, and building an internal model of reality, not predicting the next word but what happens next in the world.

AMI has four goals: understand the real world, have persistent memory, reason and plan, and be controllable and safe, which is a polite way of saying everything that today's AI cannot reliably do.

Meta will be a partner in the new company, which is either a sign of goodwill or Zuckerberg quietly hedging his own bet. Probably both.

When the founding scientist of your AI research lab quits, raises a billion dollars to build the thing he thinks replaces your entire strategy, and sets up shop on three continents, that is not a speed bump.

That is the smartest person in the building tapping out before the bell.

Crack #2: The models are late, and they know it

This one dropped this week, and it changes things.

Axios published a scoop today: Meta is preparing to release its first new models built under Wang, with plans to eventually open-source versions of them. 

Wang's vision is to democratize access, go wide, and own the consumer while OpenAI and Anthropic chase Fortune 500 contracts and government deals.

On the surface, good news. Meta is back, and the open-source flag is still flying, but read the fine print.

The flagship model codenamed Avocado was supposed to ship in March, but it didn't. 

The New York Times reported that it underperformed on coding, reasoning, and writing compared to Google and others. Meta quietly shelved the launch.

And when the new models do finally arrive, the open-source version is the stripped-down one. 

The real models stay locked.

Axios put it diplomatically: "Meta knows its new models may not be competitive across the board, but believes it will have areas of strength that appeal to consumers."

That sentence is doing an enormous amount of heavy lifting.

"May not be competitive across the board" is the corporate equivalent of telling someone their soufflé is great for a soufflé.

There is one genuinely interesting angle buried in the Axios story, though. 

Wang is deliberately ceding the enterprise market to OpenAI and Anthropic, both of whom are deep in government contracts and corporate boardrooms and going the other direction entirely. 

Building AI for the other 8B people who don't have a procurement budget.

It's the Android playbook, open the OS, close the crown jewels, win through volume, not margin.

Whether the models are good enough to pull that off is the $115B question.

The verdict

So is Meta building the most dangerous AI moat in the world or the most expensive catch-up plan in tech history?

Both, and that is exactly what makes this interesting.

The distribution is unmatched, the cash machine is printing, and the glasses momentum is accelerating.

The 200M businesses waiting for a Manus agent they can actually afford are real.

But the model quality gap is also real. 

LeCun's exit is real, Avocado's delay is real and the fact that they're eyeing Google and Anthropic as a bridge while Llama 5 bakes is real.

Buffett's moat question isn't "who has the best product?"

It's "what happens when a well-funded competitor shows up tomorrow?"

OpenAI showed up, Google showed up, Anthropic showed up.

3.5B people still opened a Meta app before breakfast.

That's your answer.

Alright, that’s it for this edition of Milk Road AI. We want to hear from you.

GLOBAL CHAOS IS RISING… SO WHY ARE WE BUYING? 📈

In today's episode, we sat down to break down how to position through geopolitical risk, an oil spike, and the accelerating AI compute arms race.

Here's what you'll hear:

Don't sleep on this one 👇️

YouTube | Spotify | Apple Podcasts

BITE-SIZED COOKIES FOR THE ROAD 🍪

Intel is joining Musk’s Terafab chip project with Tesla and SpaceX. It will help manufacture high-performance chips, though details remain unclear.

Anthropic unveiled Mythos, a new AI model for cybersecurity via Project Glasswing. It’s designed to help partners find and fix software vulnerabilities.

Google Maps can now use Gemini to write captions for photos users post about places. The feature is designed to make contributing reviews and local content easier.

MILKY MEMES 🤣

ROADIE REVIEW OF THE DAY 🥛