Every article in this series has started the same way. We type a question into Earnings Desk, our fictional research app, and a few seconds later, an answer appears. Over five articles (see below), we’ve followed that answer from the app to the power plant.

One stop remains, and it’s the one folks outside of trading and technology circles rarely think about. Nvidia (NVDA), Advanced Micro Devices (AMD), and the cloud giants design the chips we covered in our second article, but none of them makes its most advanced ones. Somebody else does, and right now that somebody can’t keep up.

Bob Byrne’s AI Series:

• Who Gets Paid When You Ask AI a Question?
• Meet the Chips Behind Every AI Answer
• Why Fast AI Chips Still Have to Wait
• AI Needs Somewhere to Plug In: 7 Stocks That Have the Power
• The AI Boom Has an Electric Bill. Here’s Who Gets Paid.

Designers and Builders

Think of a chip design as a recipe. Nvidia is the celebrity chef everybody’s heard of, but it doesn’t own a stove. Most of the world’s leading AI chips are cooked at Taiwan Semiconductor Manufacturing (TSM), better known as TSMC.

That kitchen is running at redline, judging by TSMC’s October 15 report. Third-quarter revenue came in at $46.7 billion, up 41% from a year earlier, and CEO C.C. Wei said AI demand is still running well ahead of what the company can supply.

TSMC isn’t the only kitchen in town, but it’s the one with the line out the door. Intel (INTC) wants other chefs cooking in its kitchens, and it has heavyweight backers in its corner, with the U.S. government holding a stake and Nvidia investing $5 billion. Samsung, better known for its memory chips, is cooking for others too, including a $16.5 billion deal to make Tesla’s (TSLA) next-generation AI chips at its Texas plant.

And before anyone starts cooking, the recipe has to be written down precisely, and that’s where Synopsys (SNPS) and Cadence (CDNS) come in, with the software engineers use to lay out and test a design before the first chip is made.

The Machines That Make the Machines

If TSMC is the kitchen, the equipment inside is where things get insanely expensive, with a single top-of-the-line machine costing close to $400 million. And just a handful of companies supply nearly everything on the factory floor.

ASML Holding (ASML) is the only company that makes the extreme ultraviolet lithography machines used to print the most advanced chips. Picture a printing press that draws circuits with light, thousands of times thinner than a human hair. ASML said on October 14 that third-quarter sales jumped 55% to 11.7 billion euros, and some customers already want to place orders for 2029 and 2030.

China would love a printing press of its own, and Reuters reported last December that a team including former ASML engineers had built a prototype EUV machine in a guarded Shenzhen lab. ASML has never sold one of these machines to China, so Beijing is pushing for working chips by 2028, though insiders told Reuters that 2030 is more realistic. ASML’s own first prototype took close to two decades to become a commercial product, so its lead looks safe for a while.

Printing is only one step, though, and Applied Materials (AMAT) and Lam Research (LRCX) make the equipment that lays down ultra-thin layers of material and etches away what isn’t needed until the circuits take shape. KLA (KLAC) is the quality-control crew, selling tools that hunt for defects along the way. One speck of dust can wreck a chip headed for some of the priciest computers on the planet.

A Round Trip to Taiwan

You’d think an AI chip is just one chip, but it’s more like a construction crew working shoulder to shoulder. Nvidia pairs its processors with high-bandwidth memory from SK Hynix (SKHY), Micron (MU), and Samsung, all squeezed into one tight package so information can move quickly between them.

TSMC handles much of that work with a process called CoWoS, short for chip-on-wafer-on-substrate. The engineering goes over my head, but an engineer explained it to me like this. A construction crew stands on a thin silicon platform, and that platform gets bolted onto a sturdier base that plugs into the rest of the computer. Almost all factories that can do this are in Taiwan, so when TSMC’s Arizona plant made its first U.S.-built Nvidia Blackwell wafer a year ago, it still had to ship the wafer back across the Pacific for packaging.

Amkor (AMKR) is working to change that with a roughly $7 billion packaging and testing campus in Peoria, Arizona, expected to start production in early 2028. Then every chip gets tested before it ships, since nobody wants to find a bad one after it’s bolted into a server. Teradyne (TER) makes the machines that run those tests, and it said about 70% of its first-quarter revenue came from AI demand.

Following One Answer All the Way Down

So let’s ask Earnings Desk one more question and follow the money.

Our $20 subscription pays the app, which pays a model provider like OpenAI or Anthropic. That provider rents computers from Amazon (AMZN), Microsoft (MSFT), or Nebius (NBIS), packed with Nvidia chips that TSMC built on ASML equipment. Seagate (STX) storage and Arista (ANET) networking keep the information moving, Applied Digital (APLD) supplies the building, and Constellation (CEG) keeps the power coming. Each one gets paid for a different job, and each one depends on the others.

If I had to boil this series down to one lesson, it’s that the bottleneck keeps moving. First everybody scrambled for GPUs, then memory got tight, then networking, and chip factories are still booked years ahead. But right now the biggest squeeze is finding right-sized data centers that can be built and powered in months, not years.

That’s also where lenders are getting nervous, because funding a 500-megawatt to multi-gigawatt campus is a tough sell when its power won’t arrive for years. After Oracle’s force majeure notice on Project Jupiter, debt tied to its $18 billion construction loan was reportedly trading below 90 cents on the dollar. It’s no wonder OpenAI and Anthropic have been shopping for 20- to 30-megawatt sites, as we saw in our fourth article. Whoever owns today’s bottleneck usually has pricing power, and that’s worth remembering any time you size up an AI stock.

Call It a Bear Trap

There’s plenty to be nervous about in the AI trade.

Most of these stocks are priced well ahead of where the businesses are today, and heck, plenty are priced ahead of where they’ll be in a couple of years. But I don’t see the recent volatility as the businesses breaking down. Call it a bear trap, call it whatever you like. We’ll still have a 10-year yield above 5% and the war in Iran to deal with, but once the midterms are behind us, I expect buyers to step on the gas into year-end.

What gets me excited is that everything we’ve covered in this series has been about AI answering questions. Tesla (TSLA) just unveiled a smoke-spewing Roadster that corkscrews through the air on thrusters, and Elon Musk called it “the best of the last of the human-driven cars,” which tells you where he thinks everything else on the road is headed. Add in the robots, self-driving trucks, and factory machines that will put AI’s answers to work, plus the AI apps businesses and consumers haven’t even seen yet, and we’re still in the early innings.

When this series started, we just wanted a quick answer from Earnings Desk before the opening bell. Before long, we’ll be asking AI to run the warehouse, staff the help desk, and keep the factory line moving, and every one of those jobs will need more chips, more power, and more of the companies we’ve spent this series getting to know.

It reminds me of the internet, where the companies laying fiber and building routers made plenty of money, but the names that changed how we live were Google (GOOGL), Expedia (EXPE), Uber (UBER), and Airbnb (ABNB), businesses that couldn’t have existed before the web. Nobody can name AI’s Uber yet. When we can, that’s when all of this will start to feel real, and I suspect it’ll build on everything we’ve covered in this series.

At the time of publication, Byrne was long NBIS & APLD.