Dr. Ryan Ries here. Happy almost-Fourth! I’m writing this one with the grill already on my mind, so instead of some of the typical AI headlines you’d see from me, I thought this week we could talk about something more tied to the upcoming holiday.
It’s easy to talk about AI like it fell out of the sky in 2023. It didn’t.
Every model you and I use today is standing on roughly 80 years of American invention — a chain of breakthroughs that mostly looked like toys or niche curiosities when they first showed up. So in honor of the Fourth, here are the five U.S. inventions I’d argue made modern AI possible. (A fun one to start: the term “artificial intelligence” itself was coined right here, at the Dartmouth workshop in 1956.)
1. The Transistor (1947)
Bell Labs, 1947. John Bardeen, Walter Brattain, and William Shockley built the first working transistor: a tiny switch that flips between on and off, replacing bulky, hot, unreliable vacuum tubes. It won them a Nobel Prize, and it remains the single most fundamental building block of everything digital.
In regards to AI, every model is, at the bottom, just an unfathomable number of these switches flipping. A single modern Nvidia data-center GPU packs somewhere north of 200 billion transistors. The entire field is what happens when you make this one invention small enough and cheap enough.
2. The Integrated Circuit (1958)
A decade later, Jack Kilby at Texas Instruments and Robert Noyce at Fairchild (later a co-founder of Intel) worked out how to put many transistors on a single chip. That’s the integrated circuit, the invention that gave us Moore’s Law, Silicon Valley, and the entire economics of cheap compute!
Intelligence, it turns out, scales with how many cheap transistors you can cram together. The IC is the reason “just add more compute” has been a winning strategy for 60 years running.
3. The Internet (1969)
ARPANET sent its first message between UCLA and Stanford in 1969, funded by the U.S. Defense Department. A few years later, two Americans, Vint Cerf and Bob Kahn, gave it TCP/IP, the common language that turned a research project into the internet.
Models need two things the internet uniquely provided: an ocean of training data, and a way to deliver the result to a billion people instantly. No internet, no training corpus, and no ChatGPT-in-your-browser moment.
4. The Nvidia GPU (1999)
On August 31, 1999, Nvidia shipped the GeForce 256 and coined a term that stuck: the GPU. It was built to render video-game graphics quickly. Then in 2006, Nvidia released CUDA, which let people use those chips for general-purpose math. It turned out the math behind pretty explosions is the same math that trains neural networks.
This is the accidental engine of the entire AI boom. A gaming card became the most strategically important hardware on the planet. So important that, as you’ll see in the Quick Hits below, even Google can’t get enough of them.
5. The Transformer (2017)
In 2017, researchers at Google published “Attention Is All You Need”, introducing the transformer architecture. It’s the “T” in GPT, and it’s the design under the hood of Claude, ChatGPT, and Gemini alike.
The transformer is what finally turned decades of accumulated hardware horsepower into something that could read, write, and reason. Seventy years of switches and chips were waiting for the right idea, and this was it!
Quick Hits
Alright, I lied. I can’t leave you without a few real headlines. Here are a few things that happened this week:
-
Google can't get enough chips. Google has started rationing Meta’s access to its Gemini models, citing a compute shortage, and is now paying roughly $920M a month to rent 110,000 GPUs tied to Elon Musk’s SpaceX and xAI. With a $460B cloud backlog and Sundar Pichai openly calling the company “compute-constrained,” it’s the clearest sign yet that demand has badly outrun supply.
-
Open-weight models are closing the gap. Open, self-hostable models from DeepSeek, Zhipu (GLM), and MiniMax are now near the top of the open-model leaderboards, giving cost-conscious enterprises a real alternative to closed APIs for the first time. More control, more privacy, lower long-term cost, but the trade-off is engineering overhead and security work you own yourself.
-
Copilot's billing. GitHub’s switch to usage-based billing has developers reporting 10–50x cost jumps and threatening to walk. Ouch. A useful reminder that “AI per seat” and “AI per token” are very different line items on your budget.
My Thoughts
Two thoughts to close us out this week:
-
All the inventions on that list looked unimportant or had a completely different purpose on day one. The transistor was for hearing aids and radios, the GPU was for gamers, and ARPANET was a niche defense project. If you feel behind on AI right now, remember that the people who built its foundations mostly had no idea what they were building toward. The foundational stuff never looks foundational at the time (which is a pretty good argument for staying curious about the “toys” of today).
-
Looking at the people behind all these breakthroughs, you’ll notice how many of them came to America from somewhere else. American invention has always been a team sport, built by people from all over the world who came here to build. That’s definitely worth celebrating.
Enjoy the fireworks. The transistors in the phone you’ll film them on say you’re welcome.
Let’s Talk
When you’re back from the long weekend and wondering what AI projects would be most impactful for your organization, or how to keep your AI costs sane, our team at Mission Cloud will be here to help. Reach out to our team here.
Happy Fourth,
Ryan
Now, time for this week’s AI-generated image and the prompt I used to create it.
.png?width=464&height=309&name=Designer%20(3).png)
.png)
