The AI Bottleneck Nobody Talks About Starts With a Single Crystal

Everyone is talking about GPUs and power. Far fewer people are talking about the substrate, the physical material every chip is built on. On this episode of What Breaks Next, I sat down with Cameron Khorsandi to go one layer deeper than most AI supply chain conversations ever go.

I’ll admit it: before this conversation, I didn’t know that every chip starts as a crystal. I knew foundries were hard to build. I didn’t realize how much of the AI buildout rests on a slow, unforgiving process run by a small number of people who know how to grow a perfect one.

The expensive problems live upstream

Cameron spent almost a decade at SAP in the Office of the CTO, working with large manufacturers on how they actually use their own data. In practice, that meant a lot of time inside operations.

The pattern he kept seeing: manufacturers are very good at making things, and much less good at knowing why a particular batch came out the way it did. The most interesting problems were never in the software layer. They were upstream, in the physical process, where data was never collected, thrown away, siloed, or living only in someone’s head.

Crystal growth, he says, is the most extreme version of that problem. It’s the first step in making every chip, it’s enormously expensive to get wrong, and it still depends on the judgment of a very small number of very experienced people. Bucket Labs exists to close that gap.

Every chip starts as a single crystal

Whether it ends up in your phone, your laptop, or a server in a data center, every chip begins the same way. Someone melts a material, grows it into a single-crystal cylinder roughly the size of a forearm or larger, and slices it into wafers.

If that crystal has a defect, the wafers cut from that section are scrap. And you don’t find out until the run is over.

That makes standing up a new material a game of trial and error, where each trial costs weeks or months of furnace time. As Cameron put it, you find out you were wrong long after you could have changed anything.

Bucket Labs builds software that lets crystal growers test a process change on a computer in a few minutes. The model is built from the physics of the customer’s actual furnace, and it gets more accurate with every completed run. The goal is simple: give crystal growers more attempts to learn from.

You can buy a furnace. You can’t buy a qualified process.

This was the insight I most wanted our audience to hear. With enough capital, anyone can buy another furnace. What you cannot buy is a qualified process for a new material at a new diameter, because the runs that get you there take weeks or months.

So the shortage doesn’t show up as a furnace count. It shows up as qualified substrate. A supplier can double its furnaces to meet AI demand and still be unable to ship crystals good enough to become chips, because the process isn’t yielding.

The real constraint, then, isn’t money. It’s the number of attempts a team gets to learn from. If you have a pipeline of run data and a model of your furnace, you’re no longer bound to the furnace’s calendar to test an idea.

What breaks first: people, then yield

I asked Cameron what breaks first when a material supplier tries to scale for AI: the furnaces, the process knowledge, the people, or the yield. His answer was people, then yield. The furnaces are the easy part.

You can’t compress the time it takes to make an experienced crystal grower. It takes six months to a year to get someone to confidence. That isn’t because the training is bad. It’s because you learn at the rate of your runs, and you only get so many chances to watch something grow, go wrong, and understand why.

So a supplier that doubles its furnaces often discovers it has the same number of people who can tell when a run is going sideways. Equipment scales. Judgment doesn’t. And yield on a new process is almost entirely a function of how fast you learn.

That raises a harder question: what happens when those senior process engineers retire? Cameron’s answer is that if every past run has been captured in a clean, well-documented dataset, the knowledge no longer walks out the door with one person. New engineers can test a recipe change against a model built from all previous runs, instead of waiting a full cycle to see what happens.

Building the rails before running the train

Today Bucket Labs works engagement by engagement, structured as proofs of concept. The long-term goal is a foundation model for crystal growth simulation that is agnostic to material and growth method.

Customers start in very different places. One is standing up a brand-new material with zero historical runs. Another has 15 years of data and no in situ monitoring tying any of it together. Some keep run data in Excel; others in a database managed by IT. No off-the-shelf product could serve all of them on day one.

So Bucket Labs meets each customer where they are. The team sets up the data pipelines, builds in situ monitoring, and adds predictive maintenance so a team knows when a furnace is likely to go down and can schedule around it. Each engagement produces a fine-tuned physics model for that customer.

Materials differ, too. Indium phosphide is seeing a surge in demand; silicon remains the workhorse. They use different furnace types and have different cycle times. Every engagement teaches the team more about how processes vary, and that makes the underlying physics model stronger. In Cameron’s words, they’re building the rails before running the train.

The artifact is not the asset

Crystal growth is a secretive, conservative industry. So how does an AI startup earn trust? Cameron’s answer surprised me: Bucket Labs doesn’t care about recipes.

Customers own their data and their model. Bucket Labs licenses it back through its software, and customers are free to tweak or fine-tune it. The company builds it so that if Bucket Labs disappeared tomorrow, the customer could keep using it.

Why be so relaxed about handing the model over? Because the model alone degrades. Furnaces age, feedstock changes, recipes move, and yields drift. What matters is whoever is watching the model, correcting it, and rebuilding it as the process changes. As Cameron put it, “the artifact is not the asset. The asset is accumulation.”

He separates two things that often get blurred together. Protecting a customer’s process means their data and model stay with them. Building a repeatable product means getting better at the method: capturing runs, simulating the physics, and updating models. Bucket Labs never needs to share one customer’s recipe to serve the next one better.

For anyone building AI for heavy industry, that’s a useful frame. A contract clause alone won’t win trust. A business model that doesn’t depend on taking the customer’s secrets will.

Where the real leverage lives

The lesson for founders, operators, and investors is that scaling physical systems is not a procurement problem. Hardware is the easy part. The binding constraints are learning cycles, scarce expertise, and data that was never captured.

Bucket Labs is currently raising, but Cameron was clear that the ask is narrower than capital. The company is looking for strategic partners who can open doors in industry, academia, or ideally both.

In industry, the hard part isn’t the pitch. It’s getting in front of the handful of people at each company who actually own the growth process, and earning their trust takes time. In academia, university growth labs are where new methods and materials get worked out first, and they’re far more open with their data. For a company building a model that generalizes across materials, that access is a must-have.

In Cameron’s framing, a smaller check paired with the right phone call is worth more right now than money alone. If that’s you, reach out.