Thursday Essay: So Who's Winning the AI Arms Race?

“The future is already here — it’s just not very evenly distributed.” —William Gibson

Artificial intelligence turned the software business model upside-down so badly that everyone from the guy who just bought his first share of SpaceX to the guys writing billion-dollar checks has to wonder when those infrastructure investments start generating little things like profits. 

Microsoft CEO and co-founder Bill Gates might have been the first in the industry to truly understand (or at least exploit) the near-zero marginal cost of software sales. Writing an operating system or even just an app isn't cheap, but distribution is. Write once, sell millions. Each additional software sale, Gates realized, cost little more than the price of writing the same code to floppy disk and a colorful box to put it in.

That's how Gates became the world's first tech-world billionaire in 1987, and many more would follow.

OpenAI launched the AI arms race with the November 2022 release of ChatGPT to the public. In the following four years, the sums spent just on infrastructure are beyond mind-boggling. The original Big Four data-center builders (about to be joined by SpaceX's xAI) are Microsoft, Amazon, Alphabet (Google), and Meta (Facebook). Between them, they've spent roughly ONE POINT SIX TRILLION DOLLARS just since the start of 2023.

Adjusted for inflation, that's roughly 50 Manhattan Projects or nearly half a dozen Apollo programs.

But if that weren't absurd enough — although it does help that we're much richer now than we were during World War II or Space Race I — of that ONE POINT SIX TRILLION DOLLARS, the Big Four spent roughly half of it in 2026 alone.

SpaceX is still a bit of an also-ran on AI spending, but that will change quickly as Starship comes online with hundreds of launches each year to build the company's Starmind orbital complex.

Hang on though because I'm not quite finished with the Big Scary Numbers before we get to the really weird stuff.

And Another Thing: Today, software companies don't even need to bother with discs or boxes, and the buyer pays for the download bandwidth. Nifty, eh?

AI capital expenditures (CapEx) accelerated from $140 billion in 2023 to around $800 billion this year, and are expected to continue accelerating. Goldman Sachs projects AI CapEx of about $1.2 trillion next year, and an additional $1.4 trillion in 2028. Over the next five years, Goldman projects roughly $7.6 trillion in AI infrastructure investment (compute, data centers, power).

Assuming President Donald Trump gets his big defense spending increases — perhaps unlikely, if the Dems take Congress next month — we'll spend somewhat less, about $7 trillion, on national defense during those same five years.

That's just in the U.S. The other big spender is China, of course, and I'll get to them shortly. 

But unlike Bill Gates printing free money with every box of Windows 3.1 discs, that's hardly where the big AI spending ends. A large language model (LLM) isn't some program you install from a floppy disc or download for the App Store — although that's changing, which I'll also get to further down — it's a jillion-dollar token-generating machine consuming huge amounts of electricity. Which my local utility company assures me each month isn't free.

A token is just a bite-size chunk of text that LLMs break your prompts into for processing. There's no actual intelligence, just the data center examining your tokens and making very sophisticated predictions about which tokens to generate in response and assemble into a coherent reply, whether you want to know how much xAI spent on CapEx this year, edit a 2,000-word essay, or generate that ANNA KOURNIKOVA NUDE photo spammers teased you with back in the '90s. Or so I'm told.

Tokens are generated with computing power — these days just called "compute" — which my MasterCard charges on Grok and GPT assure me isn't free, either. And I still have to pay for the bandwidth.

Then there's training each new model, which almost seems like a pittance at "merely" a billion or three dollars per. New models, usually involving trillions of training tokens, are constantly in development because, make no mistake, this is absolutely an arms race.

So who's winning?

I hate to hedge, but it depends entirely on what you want to measure.

Let's start with the superpower AI arms race between China and the U.S.

Beijing's path to AI dominance is a bit like King Gillette's old "give away the razors, sell razor blades" business model. At the Party's insistence, Chinese firms remain focused on smaller, "open-weight" LLMs like Qwen, DeepSeek, and Kimi that can be almost given away, undercutting American firms. But like China's Belt and Road Initiative (BRI), those models are meant to create permanent dependencies on Chinese firms for the services and infrastructure that make those models most useful and on the bleeding edge. 

The lock-in effects, Beijing hopes, enhance China's global power vis-a-vis the U.S., while giving Chinese tech firms an edge over ours.

China's biggest advantage might be power generation. Beijing builds coal plants, nuclear plants, solar arrays, you name it, like there's no tomorrow. Or more accurately, China invests in power generation like AI is the future, and if there's one thing Commies understand, it's power.

China's biggest weakness is that AI is inherently disruptive, and the Chinese Communist Party has, shall we say, a low tolerance for disruptions. As previously discussed, it remains to be seen whether the CCP can thread the needle between frontier-class AI and maintaining its political monopoly.

Who wins? Dunno. Absent some AI apocalypse that brings down the CCP, I suspect there's room enough in the world for both countries to succeed to at least some degree.

Ready for the juicy stuff? Because there's no way I did all that research into America's trillion-dollar AI spending spree without delivering the goods.

And Another Thing: I won't get into why the world's appetite for AI is, at least for the foreseeable future, insatiable. That would take an entire book — which Glenn Reynolds already wrote — so for the purposes of this essay, just take it as a given. 

In no particular order, thumbnail sketches of how the Big Four operate.

Alphabet: Google's Gemini chatbot kinda sucks, despite serious spending and access to a Mongol horde's worth of user data. That aside, perhaps Gemini's real job isn't as a chatbot, but to enhance the company's core search and AdSense businesses. The company also designs its own chips and models, reducing dependencies on rivals. Results are mixed so far though. 

Microsoft: Microsoft used to be the Windows company, but in revenue, it's been the Azure cloud services company for longer than most consumers probably know. Redmond now sells AI compute as part of Azure — the bet being that once companies are locked in, they’ll keep paying Redmond for the related applications and the ongoing compute.

Amazon: Amazon Web Services evolved from world-class hosting to the raw compute power behind ChatGPT and other services. Amazon has its own model built behind the Alexa brand, but the company's focus is on renting compute to whomever can pay. The company's most recent quarterly report — AWS revenue up 37% and operating profit up 64% — suggests the model works just fine. 

Meta: The Facebook parent is a bit of an outlier, using its custom LLMs to improve on what it already does so well — keep users endlessly engaged and scrolling through their social media feeds, and hoovering up all the resulting data to serve them up tailored ads. But with its AI enhancements, those ads are tailored less like Nordstrom made-to-measure and more like Savile Row.

Models will come and go, but companies like Palantir with profitable niche uses like warfighting are the names that might still be around in 20 years. The Big Four remind me of Levi Strauss or Wells Fargo during the California Gold Rush. Speculators went broke more often than not, but the companies who sold them supplies and services made, well, bank.

The more I look into all this, despite the trillion-dollar buildouts, the less concerned I am about a dotcom-style bubble. Oh, there will likely be a major shakeout — avoiding one seems impossible with so much money getting spent so quickly — and you'd hate to be the one holding the wrong stocks if and when it happens.

But maybe we get through it without a repeat of the 2000 or 2007-2008 crashes.

Maybe.

If the big paydays begin before investors lose confidence, we're not in an AI bubble. If those profits fail to materialize before investors get overly nervous, I hope you're good on ammo and Top Ramen.

If there's an American firm I don't have much faith in, it's the one that perhaps should have found a way by now to exploit its first-mover advantage: OpenAI and its ChatGPT, the world's first publicly available AI chatbot that was worth a damn.

I might have some personal bias here, because my best friend works in IT, is required by his company to spend OpenAI tokens like a 12-year-old boy playing Skee-Ball for the first time, and can't say CEO Sam Altman's name without adding "snake-oil salesman."

That said, OpenAI's business model seems to be, "We lose money on every token, but make up for it on volume." GPT remains the most engaging LLM I've ever used, but I've spent the last few weeks working to eliminate my dependency on it because of all the big names in AI, to me they're the biggest risk. I was a seriously heavy user up until recently, and still never managed to hit whatever GPT's token limit is — all for $20 a month. That hardly seems sustainable. 

Maybe OpenAI's problem is how little it owns of its own stack. The company that started the AI horserace still rents everything but the horse. In a cutthroat race for profitability, OpenAI's suppliers — chiefly Microsoft Azure for GPT compute — will earn their profits before Altman can earn his. And I probably don't have to remind you that Microsoft's Copilot model competes with OpenAI's.

Which brings us to SpaceX, the company that aims to come closer to anyone else at owning everything from chip production to power production to that same bandwidth I keep paying for every month.

And Another Thing: SpaceX's obvious political risk is a future Democrat-held White House. Bent on payback for DOGE and X, a Dem president could tighten the regulatory screws on SpaceX's launch cadence. But by the time that becomes a live issue, Starship and Starmind may already be too far along to claw back. We'll see.

When, say, Microsoft wants to build a new Azure datacenter, the company has to line up GPU orders from Nvidia (currently sold out through the week after Judgment Day), get permission from the local municipality to build, lock in a long-term power agreement with the utility (or build the generation yourself), improve the local electrical infrastructure (assuming the local officials have their act together), source water for cooling, hire a construction firm to do the work, etc. 

That's a lot of money going out the door, and — I can't emphasize this too strongly — permission to ask.

Musk's vision is this:

• Fab his own AI chips at the massive Terafab facility currently under construction in Grimes County, Tex.
• Source the solar panels in-house from another SpaceX plant in Bastrop, Tex.
• Assemble the assorted parts into Starmind satellites at the company's Gigasat facility, also under construction in Bastrop. 
• Launch them dozens at a time into Low Earth Orbit on SpaceX Starship rockets, at hundreds of low-cost flights per year.
• Enjoy free solar power, no water or real estate expenses, and no meddling local authorities.

As for cost? "An AI satellite is essentially a lot of solar cells, a radiator, and you still need some laser links, but you don’t have all of the super complex antennas that you have on a Starlink satellite," Musk claimed earlier this year. "Given the two, the easier one to design for is the AI satellite."

What people seem to have the biggest problem comprehending is that we're quickly leaving the era when space was the most absurdly expensive way to do anything. While the Space Shuttle gave us capabilities worthy of a superpower, like putting seven crewmembers into orbit to get real work done, like repairing space telescopes and assembling the International Space Station, it was also by far the most expensive way to put anything into orbit.

Meanwhile, the so-called Old Space firms like ULA and its rocket suppliers remained fixated on small numbers of lucrative government launches, with the occasional private launch thrown in for spice.

SpaceX already makes billions selling Starlink satellite services. Between owning the whole stack, fewer upfront costs and regulatory hurdles, and cheaper launch costs — Starship will be much less expensive than Falcon 9 per kilogram to LEO — it's easy to see Starmind turning the Big Four into the Big Five.

Or maybe the Big One.

Last Thursday: The Middle Kingdom's Middle Income Trap