Invidious Frenemies

The financial stakes of computing have never been higher. Incestuous deals between hardware makers, software companies, their AI customers, and major Wall Street investors sure look like circular trade. To some degree it is, but that’s not (necessarily!) a three-card monte-style scam.

It happened in automobiles a hundred metaphorical years ago, in postwar defense technologies seventy years ago, and in the fiber internet buildout of the Nineties: Wall Street/Washington spending so gargantuan that buyers and sellers subsidize each other. Rivals, enemies, and frenemies are trying, often clumsily, to cooperate and compete at the same time—anything to keep the good times rolling! Interestingly, they still need Main Street’s political help. This happens with any seemingly limitless new opportunity. It doesn’t require a conspiracy, but, of course, that never precludes one.

My sources for this post weren’t AI, and they weren’t exclusively tech. The Economist and business TV like Schwab, Bloomberg and CNBC—all of them fond of human interest that tells a “hard news” story but makes it relatable. As always, don’t take anything at face value.

One of the TV subjects was the very successful CEO of American Micro Devices, AMD, Dr. Lisa Su, from Queens, New York. (More specifically, her family moved from Taiwan at age 6.) She went to the Bronx High School of Science. (Full disclosure: which I wish I could have attended! But I was across the Whitestone Bridge, five miles away in the faraway Catholic school system.)

When it comes to top-level computer chip manufacturing, I’ll make a dated reference here: in AI-worthy graphics cards, AMD is Avis to Nvidia’s Hertz; a competitive number 2 chipmaker that’s grabbing market share.

The boomer CEO earnestly explained how she’s been able to raise the value of the company fifteenfold in ten years. She proudly showed off what her company is shipping to customers this fall: Helios, a huge open cabinet of shiny racks with processors about the size of two refrigerators. (I thought miniaturization was the hallmark of the modern era. Then reality calls back, “This is insanely miniaturized compared to the acres of space this used to take.”)

Each of these data behemoths costs $5 million and weighs 7000 pounds. They are almost unimaginably intricate, like aluminum sculpture. Each data center, the size of a dozen Costco stores, uses many thousands of them, like the emergingly sentient computer in a hollowed-out mountain that takes over the world in “The Forbin Project.”

Now multiply that by the hundreds of new data centers planned and under construction that have backordered them years in advance. (The make-or-break business prospect whose uncertainty was the subject of my last R> post.) Until now, high-flying Nvidia has had most of those extraordinarily lucrative sales. They are booked up years in advance at about $600 billion a year. That’s just for one company. Now, gutsy AMD, jumping from 10% to 25% of the market segment, has been moving rapidly into the Wall Street winner’s zone.

The investment is not a no-brainer. With this much dough at stake, it’s a definite brainer.

Lisa Su did a brief on-camera walkthrough of a southwestern factory (in today’s jargon, a “foundry” or a “fab”) that manufactures the chips that make all this possible. Americans invented this process, but decades ago conceded the actual making of the chips to overseas producers. There are bottlenecks to conquer to get them back. As of now, even if the deep memory and the rapid processor are made in the United States, they are shipped overseas for final assembly, called “packaging” in this context.

Some basics of chip building are starkly simple: think of a microchip as a factory, manufacturing units of computing instead of cars. It’s not all virtual, but also practical and physical. You need the internal pathways of that chip to be as short and fast as possible. You need the memory to be stacked in 3D like the floors of a skyscraper. Those HBW (high bandwidth) memory stacks are mounted right around the processor on the old chip, so there’s much less delay time.

(In 2026’s frothy market, even those post-market altered microprocessors are now so valuable that black-market Chinese labs are breaking apart old, non-sanctioned gamer chips and gluing in new 2026 bootlegged HBW stacks—in effect repackaging them into unauthorized, but eminently saleable high-speed “Frankenchips”.)

That rapid processor is also under scrutiny. In 1993, I bought an Apple Newton, the first consumer device with an ARM-system chip, a different kind of microprocessor that used very little power and didn’t generate much heat, an ideal combo for a portable device. Mobile phones switched over to them, and later tablets.

ARM chips were fast, though only limited attempts were made to use them in full-sized Windows laptops. But starting in 2020, Apple switched to using its own custom ARM chips, called Apple Silicon. This year, Microsoft Windows and Qualcomm teamed up to promote the new Snapdragon processor, bringing greater speed and energy efficiency to Windows. When things are humming in the computer field, everything seems to improve at once.

The backstory: Apple had long felt let down by its longtime chip supplier Intel, a pioneering company that dominated the field for so long that it had become complacent, lazy, and sometimes dismissive of merely small-to-mid-sized customers, as Apple was at the time. Intel’s chips underperformed and generated too much heat, forcing the fastidious, minimalist Apple designers to make laptops thick enough to include bigger cooling fans. Steve Jobs, who considered himself an expert judge of industrial design, hated this and resented being subordinate to the needs of a vendor much larger than Apple, something he wasn’t accustomed to.

Apple had a premium top line that proudly boasted Nvidia’s graphics processing unit, the gamer chip whose 2007 CUDA software would supercharge AI. But considering Nvidia’s hyper-platinum reputation today, it’s surprising to find occasional manufacturing blunders. In 2008, Apple and other companies had to spend big money to rectify an Nvidia screw-up: a chip that got so hot it melted the soldering holding it to the board. Nvidia’s contracts stipulated they pay only limited damages. Apple was only one-twelfth of its size and wealth today, and Jobs had to accept it.

After the abrasive Steve Jobs left the scene in 2011, relations between Apple and Nvidia improved. But in 2014 Apple spitefully pulled Nvidia’s Apple development license just before Nvidia was to release the expensively developed, Apple-approved Mac software drivers for their distinctive CUDA graphics chip architecture. An air of chilly politeness settled between them.

That was finally ruptured in 2020 when Apple began a rapid conversion to self-designed processors. (Like Nvidia and seemingly nearly everyone else, Apple made the design, and TSMC actually built the product in Taiwan.) They’d never done that but learned the craft by manufacturing chips for its iPods, iPhones, and iPads. The phones gradually grew not just in terms of screen size, but computing power.

By 2026 Apple’s phone chips were so good that the hot-selling MacBook Neo hit stores with a phone’s A18 processor, with few complaints from users or critics. Apple Silicon, their design brand, spent the 2020s creating and evolving their next and latest line of M-series Macintosh computer GPUs, and they’re the only truly high-quality, relatively affordable consumer package for today’s AI.

That was until IFA Berlin 2026 in September, one of the world’s major trade shows for consumer computing. Suddenly the world is awash with Windows and Linux computers that are remarkably similar conceptually to the envied market leader, Apple’s Mac Mini. All of a sudden, the Apple formula—RAM shared between a computer’s calculating chip and its graphic chip—was the blueprint for many, many different imitators of Apple’s uncopyable design.

Here’s a human interest story: Despite Nvidia’s almost unimaginable success, becoming the most valuable company in history, they have a PR problem with many young people. After all, the company was basically founded on providing the best gaming hardware in the PC industry. Increasingly, though, the company has turned its attention to maxing out its higher-priced data center GPUs, basically just phoning it in with its original customer base. Gamers correctly sensed that their old buddy Nvidia had a wandering eye and was slipping away from the community.

Worse, the sudden 2024-26 AI rush to buy up everything the company makes meant that even consumer-level chips and cards cost four times what they did only a few years ago. They acquired a gamer reputation for gouging, not all deserved.

So Nvidia quietly started a policy of sending a company rep, a salesperson-evangelist, to visit major gamers’ conventions around the country. The company didn’t have a booth or a stand, just a guy with a show badge. But he broadcast his availability on social media: "I’ll be in the convention hall at [such-and-such a location]. If you come up to me in person, I will sell you whatever graphics card you want at Nvidia’s list price."

That may not sound like much, but it’s the difference between $1500 and $5000. Not exactly trivial. Sure, there are caveats: you have to be there in person—Nvidia still isn’t giving anything away for free. It’s not a food bank. But it’s an effort to help; if you really need them to run AAA game titles and are able to go to the trouble, they’ve done something for their old fans to make it possible.