I watched this video the other day. If you have three minutes, watch it first.

The next AI race has already started. More than $60 billion has been committed to it. And most of us - founders, investors, everyone using AI daily - are still arguing about whether we're in an AI bubble while the actual game has moved somewhere entirely different.

There are five distinct layers being funded right now, and they connect in a specific way. Each one solves a problem the previous one couldn't. Understanding them is the difference between spotting the next Anthropic and backing a Cursor wrapper.

Layer 1 - Real-world data collection. Now that AI models have effectively been trained on the entire text of the public internet, the industry has run out of the easy stuff. The next generation of models needs new kinds of data - specialist, embodied, human-corrected, and often physical.

Layer 2 - New AI architectures. Large language models are hitting a ceiling. The next breakthrough won't come from more of the same. It'll come from architectures that model the physical world, not just describe it.

Layer 3 - Physical AI. Once new architectures understand physics, you can build AI that acts in the real world. Robotics, manufacturing, industrial automation, autonomous vehicles.

Layer 4 - The trust layer. Before AI agents write cheques, make trades, or run factories, we need to know who they are, what they're allowed to do, and how to prove what they did.

Layer 5 - Sovereign AI. Whoever controls the compute, the models, and the data that a country's economy runs on will have leverage of a kind we haven't seen since the oil age.

See below to continue…

We built something for you.

You've told us the same thing at every event for two years: you meet great people, and then they slip away. The name goes missing, the follow-up never happens, the intro you promised gets forgotten.

SoPhy fixes that. It builds a profile for everyone you know - filled automatically from your inbox, calendar, events and messages - and reminds you to follow up before a relationship goes cold. Ask it anything about anyone, straight from iMessage.

The first integrated version launches late August. Waitlist members get first access.

Join the waitlist

Continue…

The most important company here that most people have never heard of is Mercor - a San Francisco startup connecting AI labs to human domain experts (cardiologists, lawyers, PhD-level scientists) who correct AI outputs and generate specialist training data.

Mercor raised $350 million in October 2025 at a $10 billion valuation. By June 2026 it was in talks for a new round at a $20 billion valuation, having reached $2 billion in annualised gross revenue - up from around $760 million at the end of 2025. It pays out roughly $4 million per day to more than 30,000 experts on its platform.

The rest of the layer gets stranger. Indian factory workers are being paid to wear cameras while doing manual tasks. Startups are cleaning homes for free in exchange for recording the work. The frontier labs have run out of high-quality internet text to train on, and they'll pay almost any price for the equivalent from human experts.

The point: companies competing purely on model quality will lose to the labs. Companies competing on unique, hard-to-collect training data are building something the labs will eventually need to buy.

Within three weeks of each other in early 2026, two of the most decorated AI researchers alive each raised roughly a billion dollars to build what they both call a "world model."

Fei-Fei Li's World Labs closed $1 billion in February at a $5.4 billion valuation, backed by Autodesk, Nvidia, AMD, and Andreessen Horowitz. Three weeks later, Yann LeCun's AMI Labs raised $1.03 billion at a $3.5 billion pre-money valuation - the largest seed round in European history - backed by Jeff Bezos, Nvidia, and Samsung. More than $3 billion has flowed into world model startups in H1 2026 alone.

Both bets are on the same thesis: large language models are not the path to intelligence. The path runs through models that learn the physics of the real world rather than the statistics of text.

The implication for anyone building on top of an LLM is straightforward. If world models become the new frontier, products built on today's language models don't automatically survive the transition. Some will. Most won't.

Once you have a world model, you can build AI that operates in the real world. This is the layer Jeff Bezos has bet his post-Amazon career on.

Prometheus emerged from stealth in June 2026 with a $12 billion Series B at a $41 billion valuation, backed by JPMorgan, BlackRock, and Goldman Sachs. It's building what it calls an "artificial general engineer" - AI that automates the design and manufacturing of physical products, from jet engines to drug compounds. Bezos is reportedly seeking a further $100 billion for a holding company that would acquire legacy industrial firms for Prometheus to automate.

This layer also includes humanoid robotics, autonomous vehicles, and industrial AI. Global humanoid robot funding alone hit approximately $2.4 billion in 2024. Q1 2026 has already blown past that.

The winners of software AI are not automatically the winners of physical AI. Different architectures, different data, different capital requirements. A whole new set of category leaders is being chosen right now.

Before we let AI agents move real money, sign real contracts, or run real infrastructure, three questions need answers about every agent that shows up. Who is this? What is it allowed to do? How do I prove what it did?

None have clean answers yet, which is why the trust layer is where a lot of infrastructure money is going. Oasis Security raised a $120 million Series B in early 2026 specifically for "non-human identity" governance - managing AI agents as identities inside enterprise systems. Y Combinator's latest batch is full of trust-layer startups: Alter (zero-trust identity for agents), Tolmo (agentic security), and others building the plumbing that will eventually let a bank let an AI agent execute a trade.

Gartner projects the AI cybersecurity segment growing from $10.82 billion in 2024 to $172 billion by 2029 - a 73.9% compound annual growth rate.

Most founders treat trust as something you add at the end of the build. It isn't. It's infrastructure. The companies building it now will be as unavoidable in five years as Stripe is today.

Sovereign AI means running AI inside infrastructure you control, on models you own, with data that never leaves your jurisdiction. The alternative is renting all of that from a handful of US companies. Which, right now, is what almost every business on Earth does.

The scale of the shift is enormous. The US and China alone now capture around 65% of global AI investment. Saudi Arabia's Project Transcendence has committed more than $100 billion. India has outlined more than $200 billion of sovereign AI infrastructure commitments. The EU has 19 operational AI Factories. Deutsche Telekom's Munich Industrial AI Cloud already delivers 80% feature parity with US hyperscalers.

If sovereign infrastructure is the platform layer of the next decade, every AI company built today is effectively choosing which sovereign stack it plugs into. That choice determines which markets it can serve, which regulations it can meet, and which customers will trust it.

Sixty billion dollars has already been deployed across these five layers. That's the leading indicator of where value will accrue over the next decade.

The question worth asking is not "which layer will win." It's "which layer am I actually in, and does it match where value is accruing?"

A wrapper on a foundation model is Layer 0 - and Layer 0 is not on this map. Nothing on this map protects it from being replicated in six weeks by the next model release.

Data collection infrastructure or specialist training data is Layer 1 - one of the most defensible places to be.

New architectures or world models is Layer 2 - where the biggest asymmetric returns are likely to sit.

Physical AI, robotics, or anything that touches atoms is Layer 3 - where category leaders are being chosen now.

Identity, trust, or governance for AI agents is Layer 4 - quiet layer, huge future.

Sovereign infrastructure or national compute is Layer 5 - complicated politics, unprecedented tailwinds.

If none of these fit, that itself is the answer. It means the AI conversation being had is about last year's opportunity while the actual money has moved somewhere else.

The map is still forming. All five layers have room. Every one has unclaimed positions. Knowing they exist is the only precondition for competing for one.

✅ Know a founder or investor still treating AI as one big undifferentiated category? Forward this their way. The sooner more of us can see the five layers, the sooner we all stop competing for the wrong opportunities.

(👉 Vote now - we’ll share the results in next week’s issue. All votes are anonymous.)

(Last week’s poll results)

Access here and please share it with others.