This week we talk about OpenAI, the Stargate Project, and Meta.

We also discuss o1, AGI, and efficiency.

Recommended Book: The Shortest History of Economics by Andrew Leigh

Transcript

One of the bigger news items these past few weeks, in terms of the numbers involved, at least, was an announcement by US tech company OpenAI that it will be starting a new company called the Stargate Project, which will boast a total $500 billion-worth of investment, the first $100 billion of which will be deployed immediately.

All that money will be plowed into artificial intelligence infrastructure, especially large-scale computing clusters of the kind required to operate AI systems like ChatGPT, and the funds are coming from OpenAI itself, alongside SoftBank, Oracle, and MGX, with Arm, Microsoft, and NVIDIA also involved as technology partners.

It’s a big, beefy enterprise, in other words, and the fact that this has been in the works since 2022, it’s official announcement seemingly held back so that newly returned US President Trump could announce it as part of his administration’s focus on American infrastructure and AI dominance, didn’t dim the glow of the now-formal announcement of what looks to be a truly audacious bet on this collection of technologies, doubts about the players involved having the money they’ve promised ready, notwithstanding.

That said, this is far from the only big, billions and tens of billions-scale wager in this space right now.

Last year, Microsoft announced a $30 billion infrastructure fund, in collaboration with BlackRock, and earlier in January of 2025, Google’s CEO said that his company would spend about $80 billion on the same, separate from their commitment to Stargate.

Meta’s CEO Mark Zuckerberg recently divulged that the company would spend somewhere between $60-65 billion on capital expenditures, mostly on AI, in 2025—that’s up about 70% from 2024 spending.

And last December, xAI CEO Elon Musk announced that his company had just raised a fresh $6 billion to build-out more compute infrastructure; and his role at the head of that company is assumed to be part of why he trash-talked the aforementioned Stargate effort, though there’s also a long-simmering animus between him and OpenAI CEO Sam Altman, and the fact that everyone seems to be trying to get in good with Trump—which is probably part of why many of these announcements are happening right now: Trump is in the position to king-make or cripple their respective efforts, so whomever can get in good with him, or best with him, might have an advantage in what’s become a very expensive knife-fight in this most rapidly burgeoning of tech investment loci.

There’s a reason there’s so much money flowing to this space, announcements aside, right now, too: the chatbots that’ve emerged from the GPT, LLM era of AI systems are impressive and useful for many things, and AI powered bots could even replace other sorts of user interfaces, like search engines and apps, with time.

But there are also some more out-there efforts that are beginning to bear fruit.

AI is helping Google’s DeepMind team discover new materials at an astonishing rate—including both the discovery and the testing of their properties, stage.

AI systems are also being used to accelerate drug discovery and trial design, and a company (backed by OpenAI’s Altman) is trying to extend human life by a decade using exactly this process.

Meta has a new tool that enables real-time speech and text translation between up to (depending on the type of translation being done) 101 different languages, and we’re even seeing AI systems meant to detect and track small, otherwise overlooked infrastructure issues, like potholes, at a local level.

And to be clear, this is far from a US government and US-based tech company effort: government agencies, globally, are scrambling to figure out how to regulate AI in such a way that harms are limited but research, investment, and innovation isn’t hampered, and entities all over the place are plowing vast wealth into these projects and their related infrastructure; India’s Reliance Group recently announced it will build what could become the world’s biggest data center, planned to go into operation within two years—a project with an estimated price tag of somewhere between $20-30 billion. And that, all by itself, would more than triple the country’s data center footprint.

So this scramble is big but also global, and it’s partly motivated by the gold rush-like desire to be first to something like artificial general intelligence, or AGI, which would theoretically be capable of doing basically anything a human can do, and possibly better.

That could, depending on the cost of developing and running such a system, put a lot of humans out of work, scrambling the world and its economy it all sorts of ways, and causing untold disruptions and maybe even havoc. That chaos could be very good for business, however, for whomever is able to sell this new commodity of labor to everyone else, replacing most or all of their employees with digital versions of the same—each one cheaper than a comparable human would be to perform the same work.

What I’d like to talk about today, though, is a challenge to the currently dominant theory of operation in this industry, and why a new family of AI models is sending many of the tech world’s biggest players into a panic.

—

A lot of the news coming out of the AI world, at the moment, is focused on what are called agents, or agentic AI.

An AI agent is a system that can operate with agency: it can do things on its own. So you could have one of these systems, something you might engage with like a chatbot, but one capable of taking complex instructions, and you could tell it to find the best e-bike for your use case, and it would then take your info, your context, your needs into consideration, do a bunch of research, and maybe even buy and set up the delivery of the bike for you, with limited check-ins required on your part.

A truly agentic AI would operate as sort of a personal assistant, capable of doing anything a human personal assistant would be able to do—sans the physical body, of course—though that could come later.

This is generally seen as a step on the path toward AGI, and perhaps even AI superintelligence, which would be AGI that’s massively smarter and better at everything than any human, all of which also moves these things from the realm of “tool to be wielded by humans”, toward something more like a robot that can do all the things it’s supposed to do, without a human present; a different category of product and service.

This type of AI, with this level of capability, is generally considered to be really expensive to make—to train, in the industry parlance—and to use, because of how much computing power is required to run the code required to leverage these sorts of smarts.

In 2020, ChatGPT-3 cost somewhere between $2-4 million to train. Its successor, ChatGPT-4, which was deployed in 2023 cost more like $75-100 million.

That’s a lot more money. The model is a lot more powerful, granted, but the scaling laws that have seemed to be at play in this space, the increase in cost between generations of AI, have suggested that getting another capability leap comparable to what we saw between ChatGPT-3 and 4 would cost something like a billion dollars, and even that might give us a jump, but not the same staggering growth in performance that we saw between those generations.

The are arguments to be made that the size and type of dataset matter, here, and that the culling of said datasets, and how the models are tuned to use the data and respond to things are also vital, perhaps as much or more so than the initial training.

Companies like OpenAI have also figured out all sorts of ways to wring more performance out of less training and compute, including things like allowing the AI to reference other sources—basically doing a web search or checking wikipedia and similar references, in addition to knowledge that already exists in its training dataset—or allowing them to “think” longer, giving them more time to work through a problem or task, which tends to lead to better results, even with weaker—in terms of training and compute power—systems.

Ultimately, though, most of these companies seem to be assuming that more money churned into more infrastructure and compute capability will be necessary, to make these things better at doing science and solving global problems, at maybe running military campaigns-scale issues, but also at replacing humans as employees—creating more agentic, ultimately, they hope, AGI-level systems.

So that’s a big part of why there’s so much money sloshing around in the AI world right now: all these companies want to build the biggest, baddest model, they would love to develop AGI and put everyone out of work, and they assume that more money will equal more potency, so if they don’t start building now, they risk being left behind in a couple of years when all their competitor’s snazzy new assets are available and powering their AI systems—which could allow their competitors to get there first, and there’s a general assumption that it’s important to be first or close to first on this, as truly AGI-level, or beyond AI could theoretically allow them to refine their own systems faster, which could secure them a permanent lead over their opposition, moving forward.

Though the US is generally considered to be in the prime position in that particular race, so far, China has been investing a lot in this space, as well, and many of their investments have been similar to those of their Western competitors; dropping lots of money on the issue, building big infrastructure, and so on.

They’ve been hindered quite a lot by Western, especially US, sanctions, though, and that’s made it more difficult, not impossible, but more time-consuming and expensive for them, to get the highest-end chips optimized for AI systems, like those made by NVIDIA.

This has forced them to take some different approaches to their international peers, and while many of those approaches still involve huge price tags and build-outs, some of them have instead focused on a less-celebrated aspect of the industry: that of smaller models that are a lot more efficient, achieving gains that are out of proportion to their training and operating costs.

Case in point are the new DeepSeek R1 models, which are a collection of AI models that were cheap to make, released free for public use and editing, and which seem to beat OpenAI’s o1 reasoning models—which are very much not free, and which were a lot more expensive to develop—on some of the most widely used performance benchmarks.

These models apparently cost something like 3-5% what OpenAI spent on its o1 model, a mere $5.6 million, and again, they’re free to use, but also open source, so anyone who wants to build their own business or new AI atop them can do so; and their API costs are more than 90% lower than o1’s, so it’s also a lot cheaper to use these models for development purposes than OpenAI’s options.

This isn’t the first time a Chinese company has taken a look at what folks are doing in the west and then massively undercut their efforts by amplifying the efficiency many fold. Also, again, there’s a constraint on Chinese companies’ ability to get the latest and greatest AI hardware, which incentivizes this path of development, and they also have a super competitive tech industry in China, which tends to force a lot of their sub-industries, like batteries and solar panels, to iterate rapidly and push costs as low as functionally possible.

This family of models was made as kind of a side project by someone who’s been competing within that somewhat brutal evolutionary context, and the rest of the world, by comparison, just hasn’t had the same forcing functions influencing its development path—so this level of efficiency with this level of performance has been, up to this point, unheard of. And as a result, these DeepSeek models have sent the US and other western tech industries into a tizzy.

And it makes sense that these people would be panicking: they have spent, and are intending to continuing spending heavily on next-gen AI infrastructure, and this type of model, trained for basically nothing, demonstrating this level of performance? It calls all those investments into question, even to the point that some commentators—without evidence, so there’s no reason to believe this is the case—have wondered out loud if this might be some kind of psyop by China to kill the US’s AI industry, basically making it look like a bad investment, if these kinds of results can be achieved so inexpensively elsewhere.

Again, that’s almost certainly not what’s happening here, but these models have reportedly landed like a live hand grenade in the offices of the US AI industry, with folks in big tech companies frantically trying to figure out how DeepSeek does what it does, and then surreptitiously copying whatever they can to try to get ahead of this, build their own version of the same and maybe work those findings into their planned investments.

Meta in particular has apparently been on edge about this, as they’ve tried to own the free, open AI model space with their Llama family of AI models; which have been generally well received, but apparently DeepSeek’s earlier model, v3, was already messing with their heads, surpassing what they were able to do with llama, and this new family, the R1 family, has them worried they won’t be able to hold onto that position, and might not even be able to compete, despite their tens of billions of dollars worth of investment.

What’s more, something this effective and efficient can be run by a lot of companies that would otherwise have had to rely on entities like OpenAI and Meta, which have the computing infrastructure—all those big buildings they’re constructing at a frantic rate and high cost—to handle the larger models.

Non-AI companies that want to use these systems, though, could theoretically just buy their own, smaller setup and run their own AI, in-house, which would alleviate some security concerns related having all that stuff processed off-site, but it would also almost certainly be cheaper over the long-term, compared to just paying someone like Google or OpenAI for their services, forever.

All of this has resulted in a fair bit of volatility in the US stock market, which has been heavily reliant on AI-oriented tech stocks for growth over the past year, with NVIDIA in particular taking a hit, due to the possibility that heavyweight chips might not be vital to creating high-end AI systems.

There are downsides to DeepSeek, of course, perhaps most obviously that this model, having come from China, is laden with censorship about exactly the sorts of things you would expect: Tianammen Square, China’s government and it’s many well-documented abuses, and so on. There could be more issues, too, that the folks who look into such things will discover after spending more time with this family of AI, though thus far, the response has generally been very positive, even with those caveats.

Either way, this challenges the assumption that the US or any other country can stifle another nation’s, or group’s, AI ambitions with hardware sanctions.

It also suggests that, if this general approach could be replicated, we may see a lot more models that are cheap and easy to run, but which are also effective enough for a lot of those next-step, higher-end utilities. And that would allow AI to spread a lot more quickly, more people being able to wield more powerful tools, while also potentially doing away with many of the moats—the defendable, unique value propositions—these larger tech companies assumed they would have by building and controlling the pricy infrastructure they assumed would be necessary to spin-up AI systems of that calibre.

Show Notes

https://www.microsoft.com/en-us/research/story/ai-meets-materials-discovery/

https://semianalysis.com/2025/01/23/openai-stargate-joint-venture-demystified/

https://openai.com/index/announcing-the-stargate-project/

https://techcrunch.com/2025/01/24/stargate-will-use-solar-and-batteries-to-power-100b-ai-venture/

https://www.ft.com/content/4541c07b-f5d8-40bd-b83c-12c0fd662bd9

https://www.politico.com/news/2025/01/23/trump-staff-musk-conflict-00200311

https://www.nytimes.com/2024/12/24/technology/elon-musk-xai-funding.html

https://www.cnbc.com/2025/01/22/trump-had-phone-call-with-openais-sam-altman-last-week.html

https://www.whitehouse.gov/presidential-actions/2025/01/removing-barriers-to-american-leadership-in-artificial-intelligence/

https://apnews.com/article/trump-ai-artificial-intelligence-executive-order-eef1e5b9bec861eaf9b36217d547929c

https://restofworld.org/2025/global-ai-regulation-big-tech/

https://www.wsj.com/tech/ai/meta-spending-ai-facebook-data-centers-9452a88f

https://www.reuters.com/technology/meta-invest-up-65-bln-capital-expenditure-this-year-2025-01-24/

https://www.bloomberg.com/news/articles/2025-01-23/billionaire-mukesh-ambani-plans-world-s-biggest-data-center-in-india-s-gujarat?embedded-checkout=true

https://www.bloomberg.com/news/articles/2025-01-24/apple-enlists-company-veteran-kim-vorrath-to-help-fix-ai-and-siri?embedded-checkout=true

https://www.ft.com/content/25a473ea-9f87-474a-8729-bc5287df853a

https://spectrum.ieee.org/machine-translation

https://techcrunch.com/2025/01/24/elevenlabs-has-raised-a-new-round-at-3b-valuation-led-by-iconiq-growth-sources-say/

https://www.bloomberg.com/news/articles/2025-01-24/vc-lightspeed-bets-big-on-ai-megadeals-backing-anthropic-xai-mistral

https://www.ft.com/content/7dcd4095-717e-49f8-8d12-6c8673eb73d7

https://www.nytimes.com/2025/01/23/technology/ai-test-humanitys-last-exam.html

https://every.to/chain-of-thought/we-tried-openai-s-new-agent-here-s-what-we-found

https://www.platformer.news/openai-operator-ai-agent-hands-on/

https://techcrunch.com/2025/01/23/openai-launches-operator-an-ai-agent-that-performs-tasks-autonomously/

https://www.axios.com/2025/01/19/ai-superagent-openai-meta

https://blog.google/technology/google-deepmind/google-gemini-ai-update-december-2024/

https://arstechnica.com/ai/2025/01/china-is-catching-up-with-americas-best-reasoning-ai-models/

https://www.macrumors.com/2025/01/27/deepseek-ai-app-top-app-store-ios/

https://www.statista.com/chart/33114/estimated-cost-of-training-selected-ai-models/

https://sherwood.news/tech/a-free-powerful-chinese-ai-model-just-dropped-but-dont-ask-it-about/

https://www.axios.com/2025/01/17/deepseek-china-ai-model

https://www.nytimes.com/2025/01/23/technology/deepseek-china-ai-chips.html

https://archive.ph/WMUbb

https://x.com/pmarca/status/1882719769851474108

https://venturebeat.com/ai/tech-leaders-respond-to-the-rapid-rise-of-deepseek/

https://archive.ph/vDsQ4

https://archive.ph/CrbGO

https://x.com/nealkhosla/status/1882859736737194183

https://x.com/samfbiddle/status/1882882950368473161

https://x.com/samfbiddle/status/1882884223008493887

This is a public episode. If you'd like to discuss this with other subscribers or get access to bonus episodes, visit letsknowthings.substack.com/subscribe