I made some jokes this week about the RAM shortage. I made a joke about hitting Sam Altman with my car. These are the kinds of jokes you make when you’re watching your country fail a test it had five years to study for, and the class curve just got obliterated by China. I also think Sam Altman is kind of a dick, but that wouldn’t justify hitting him with your car. You should not hit Sam Altman with your car. You’ll go to prison… probably. (It really depends how powerful you are, and how much money you have. We are, after all, living in an era where it seems dozens of billionaires engaged in a child sex trafficking ring with Jeffrey Epstein and got away with it, so you may be able to get away with hitting Sam Altman with your car.)
I want to be precise about what happened over the past five years, because precision matters when you’re performing an autopsy on national competitiveness. We failed to meet the industrial standard of the moment. Our largest economic competitor had a better plan. I’m not spinning this story, nor am I offering partisan talking-head punditry. I am giving the kind of blunt, unadorned fact that should drive change. I want policymakers to lose sleep over what I’m writing. Why? Because I have lost sleep over this shit. (This assumes we still believe policymakers sleep with any regularity or conscience.)
Five years ago, anyone with even a cursory understanding of high-performance computing, memory architecture, and GPU-bound workloads could have told you that a super-boom in data-center demand for machine learning was coming. This was not arcane knowledge locked in some subterranean DARPA briefing room. It was visible. It was obvious. I remember the moment it happened for me. I was living in Houston, Texas, and I was at a party, showing off some of the first image-generation models at a BBQ. (That should tell you how nerdy I am. The rest of the dads were talking about college football, and I got on my laptop and said, “Imagine there was a pool full of cats that were Olympics swimmers!” They were bad images, don’t get me wrong. But I could tell we were on the verge of something insane, and it was only a matter of time before things improved.) This was the technological equivalent of the modern NBA adopting the high volume 3-point game, and going to an analytics based approach. The teams that understood scoring efficiency would be the teams that won championships, and the stubborn old-heads would be left shooting contested 14 foot-jumpers.
I wish we could claim some reasonable excuse for this collective blind spot, but we can’t. All our excuses evaporated the moment COVID-19 arrived, and the entire world discovered, simultaneously and uncomfortably, that the bottleneck in global semiconductor manufacturing existed on an island of 23 million people… off the coast of an adversarial nation that considers that island a breakaway province.
That’s Taiwan, for those keeping score. That is where our supply chain went to be fragile. Also, shoutout to Taiwan for figuring it out first. I suppose having an aggressive dickhead as a neighbor can inspire national pride for what is possible, especially when it comes to making sure your allies definitely come to your defense in the event China decides to make good on their promises.
Every OECD nation figured out the supply chain bottleneck in real time. The United States, Canada, the United Kingdom, Germany, France, Japan, South Korea, Australia, New Zealand. All of us looked at our computing infrastructure and realized it had a single point of failure with the geopolitical resilience of a fishnet condom. And then, with the confidence of a civilization that had just identified a civilizational vulnerability, we proceeded to do… absolutely fucking nothing about it.
I do not pretend to fully understand why governments behave the way they do, though I suspect it begins with China having what one might charitably describe as an interesting relationship with Taiwan. If you need me to unpack the sarcasm there, I cannot help you, and you should probably not be making policy.
The Velocity of Everything
Here is my core thesis, stated without embellishment: The speed of adoption is directly tied to cost.
This is true whether you are working in medicine, robotics, defense and national security, automotive manufacturing, industrial production, or construction. Machine learning-driven applications are going to alter the pace of innovation and the velocity at which we deliver work in every single one of these sectors. The operative word(s) in that sentence is/are “going to,” which means we are not yet there, which means it’s still possible to scrape together some modicum of national pride, and shift gears. It’s possible for all of us to put a plan in place, and think of this as a moment to come together for the good of literally… fucking everyone.
I am generally reluctant to invoke the United States’ industrial mobilization during World War II as an analogue for anything. That comparison has been cheapened by overuse, and deployed to sell everything from electric vehicles to meal-kit subscriptions sold by YouTubers. But in this particular case, the comparison is not merely an apt 1:1 comparison. It’s where we need to be. We are competing for the future of who can be the most dominant economy on Earth, and that competition will be won or lost on the basis of computational infrastructure. It’s finally here, y’all. The nerds were right. This is our moment to be the heroes.
The nations that can supply their own data centers with sufficient CPUs, GPUs, and RAM, and that can generate enough energy to power those data centers, will be the nations that set the terms of the next century. Everyone else will be purchasing access to someone else’s future at someone else’s price.
The Bipartisan Art of Fumbling the Bag
Ready for the shitty part? I’m going to delivery some weapons grade fuckery here. At the very moment we need a whole-of-nation approach to what is functionally the next industrial revolution, we find ourselves in a stagnated political environment where gridlock is the default setting. A whole-of-nation approach to anything sounds like the meth-pipe of destiny. You still believe in functional governance? Here’s the meth-pipe of destiny for you. Take a rip from your cheap 7-11 lighter, and hope your teeth stay in your head.
We need new fabrication facilities. We need the semiconductor supply chain to exist on this continent in a meaningful way. We need energy infrastructure that can support the thermal appetite of a machine learning economy.
Underneath all of that, we need another generational push of electrical engineers, semiconductor designers, specialists in advanced computing architectures, experts in x-ray crystallography and materials science. Every one of these models is propelled by advanced mathematics, which means we need a coordinated, sustained investment in mathematical education starting right now, today, in middle schools and high schools and universities, so that the next generation of kids who love math have somewhere productive to take that love.
I am really not trying to be sanctimonious about this, because the failures that brought us here are genuinely (tragically) bipartisan. This is not a situation where one party dropped the ball and the other was heroically trying to catch it. Both parties fumbled. The ball rolled into the parking lot. Nobody noticed because they were too busy posting on social media about it, trying to score cheap political points.
All of these accumulated, incremental, bipartisan failures are why we could not meet the moment. And those failures cascade all the way down to the very electrons powering our current high-performance computing applications.
Maybe I am acutely aware of this because I am someone who builds gaming computers and therefore understands how much power these workloads demand. So consider this: even a mid-grade, non-frontier model consumes roughly as much energy as your refrigerator. And let me be clear, it is genuinely thrilling that I can run certain models on my MacBook. That is a technological achievement worth celebrating. But those are not the bleeding-edge models that will determine competitive advantage between nations. Those are the equivalent of a middle school classroom spitwad, when the contest requires a Bren gun.
If you are running multiple GPUs, and many serious workloads require exactly this, you are talking about the power consumption of your clothes dryer. Or multiple clothes dryers. Now perform the following thought experiment with me: Imagine that tomorrow, every American suddenly needed the equivalent power draw of four additional clothes dryers in their home. I can tell you, without consulting a single utility executive, that this would place enormous and probably catastrophic strain on our electrical grid. That is the scale of what we are discussing, except we are not discussing it with nearly the urgency it warrants.
The Energy Problem. The Goddamn Grid.
I said this years ago, and I will say it again in a hoarse tone, after shouting this exact shit for years, because I have always known I was correct. Investments in small modular reactors, manufactured at scale in factories, would have been able to meet the data center energy demands we currently face. The key phrase in that sentence is “would have been.” Past tense. Conditional. The tense of regret.
I will also make one explicitly political observation, because this is a policy decision being made by the current administration. For reasons that I cannot fathom, and I have genuinely tried, we are now attempting to revive coal power plants to meet these needs. Coal, which scales terribly. Coal, which will not meet the demands of our data centers. Coal, which is to the energy needs of a machine learning economy what a garden hose is to a five-alarm fire.
We need to be building power supplies adjacent to our data centers. Transmission losses are a real and measurable phenomenon. Every mile of copper between generation and consumption is a mile of lost efficiency. This is not a matter of ideological preference. It is physics. Physics does not negotiate.
The Historical Lookback
I say all of this out of a deep and sincere sense of patriotism. I want the nation I live in to be competitive, to be at the frontier, to be at the bleeding edge of what is scientifically possible through high-performance computing and next-generation machine learning. I look at the moment we are living in and I see something analogous to the Industrial Revolution, or more precisely, to the birth of the internet. The speed and pace of innovation is changing in a way that most of us cannot fully appreciate because we are submerged in it.
This is historically borne out. The people living through the Industrial Revolution understood that things were changing. They could see the smokestacks. They could feel the rhythm of their days shifting. But because they were inside the transformation, living it hour by hour, they consistently underestimated the speed and totality of the change. They paced themselves for a marathon when they were running a sprint.
A similar phenomenon attended the advent of the internet. In retrospect, we were almost certainly too slow to adopt networked computing, too slow to build digital literacy, too slow to understand that the world was being rewired underneath our feet while we were still debating whether email was a trend that would go away.
This would be an excellent time to learn from the proximate past. When you recognize that you are living through a period of rapid technological industrialization, the correct response, as a matter of national strategy, is to understand that the nations which win these transitions are the nations that adopt first, coordinate deliberately, and drive industry with intentionality rather than inertia.
On the subject of “winning” and what “winning” actually looks like.
Winning means vertically integrating supply chains with a coherent national strategy for how those supply chains interact. It means legislating those supply chains in ways that are efficient and competitive rather than merely protective. In our case, I am talking about the United States (because I’m here), but our decisions ripple outward across the globe, and so the stakes extend well beyond our borders.
What we do not need is where we’re headed: A monopolistic concentration of computational power in a handful of enormous technology firms. The people who are winning right now and the companies that are winning right now are simply able to burn enough capital on compute. They are winning out of sheer spending power.
What we need is hyper-competition among small and medium-sized businesses, the kind of competitive ecosystem that allows all of us to drive the pace of innovation through superior outcomes rather than superior spending. The goal is not to see who can burn the most capital. The goal is to see who can build the best thing. The government can level the playing field and democratize computing costs, meaning the ideas that have the highest efficacy should be what drive our national strategy, not the companies with the most spending power.
And I should note, for the sake of historical honesty, that the United States has done this before, albeit imperfectly. I am not trying to single out Alcoa Aluminum, but Alcoa Aluminum is the instructive example. Alcoa was effectively created as a wartime monopoly during World War II because the government recognized that affordable, reliable access to aluminum was a matter of national survival. This was a reasonable decision at the time. (The less reasonable outcome was that Alcoa survived the war and became a de facto peacetime monopoly, which was not the plan and was not, by any serious measure, the optimal result. Once again, it’s possible to learn from our history here and not repeat the same result.)
We are not fighting World War II. I use the analogy because it clarifies a principle: Critical pieces of our supply chain will require a coordinated, whole-of-government approach to remain globally competitive. The question is whether we can execute that approach without accidentally creating the next Alcoa, and whether we have the political will to try at all.
The blunt rotation you have to get right.
This isn’t one of those situations where I’m trying to avoid sounding like a dick, because certain things you say in life just make you sound like a dick. I don’t want China to win the next industrial revolution, because I would like free democratic nations to win the next industrial revolution. Although I’m definitely saying that wrong, because we’re not anticipating the “next industrial revolution,” it’s happening right now before our eyes.
Over the past decade, China has demonstrated massive industrial capability. It’s no longer just predicated upon their ability to copy the designs of American, European, South Korean, and other OEC nations’ technologies.
The era of the cheap Chinese knock-off isn’t quite done, but it’s rapidly fading. We have to stop infantilizing China.
The Chinese Communist Party is acutely aware of the stakes, and they are making investments into what they understand as a massive leverage point and competitive advantage. They’re building their fabs, their infrastructure, their schools… everything is built around winning this moment. Their unique form of governance allows them to execute decade-long plans with brutal efficiency through a whole-of-government approach.
I know that most OEC democratic nations do this poorly through fragmentation and bipartisan bickering, but we don’t have time for that right now. This week, using a frontier model from Anthropic called Claude Code Opus 4.6, I was able to do the work that used to require three to four months and three to four people… with just myself… in five days.
I really can’t put a fine enough point on this: I said it five years ago on my TikTok when the first large language models began circulating in academic circles, and on the edges of computer science forums. Where we are today in artificial intelligence is splitting the atom, the internet, the industrial revolution, the automobile, pasteurization, vaccines, the Haber-Bosch process, sterilization, anesthetics, semiconductors, mapping the Human Genome, cRISPR-Cas, shipping containers, refrigeration, putting a man on the moon, and sliced bread…
… all rolled into one blunt. It’s a super-blunt of technology, so dank and sticky, it would have knocked out Mac Dre in his prime.
We as a nation, are fucking up this blunt rotation. It’s critical to our national security, and all our future economic success, that we get this blunt rotation right.
… and don’t even get me started on quantum computing. That’s gonna have to be in another post.