
Unless you live under a rock, you have likely heard of Project Stargate. Project Stargate is a $500 billion AI infrastructure joint venture that was announced by Donald Trump, despite the venture being in the works long before he took office. Indeed, Sam Altman, CEO of OpenAI, has publicly stated that the project wouldn’t be possible without Trump, which is odd, to say the least, considering it looks like he has nothing to do with the project. But here’s the thing: if you have been paying attention to the actual progress of AI, this huge push makes absolutely zero sense!
Let’s start with what Stargate actually is. At least for now, Trump’s administration isn’t putting a penny into the project. Instead, it is a joint venture between OpenAI, US cloud computing specialist Oracle, investment group SoftBank, and the tech investment arm of the UAE government, MGX. Essentially, over the next four years, SoftBank and MGX will supply $500 billion into this joint venture, and Oracle and OpenAI will use the money to build US-based AI infrastructure, like data centres. Initially, it seemed these data centres would serve a multitude of US-based AI companies, until it was recently announced that they would only serve OpenAI. I wonder why.
So, why are they doing this?
During the announcement, Trump said it was to keep the future of technology in the US, alluding to them trying to win the AI race. Unfortunately, this is a total waste of money and effort.
Let’s start off with the scaling issues regarding AI. The only reason Stargate costs so much is that OpenAI wants to build AI models substantially larger than the ones they already have, which requires gigantic data centres that are far larger and more expensive than anything that currently exists. However, all AI companies are already running into significant diminishing returns when scaling up AI models.
I have covered this topic extensively before (read more here). Simply put, an AI trained on 1 TB of data might perform 10% better than when trained on 500 GB of data. But increase that training data to 2TB, and you might only see a 5% increase in performance over the 1 TB version. This trend now seems to be a universal law of AI programming, and it has become known as the “Efficient Compute Frontier.” Companies like OpenAI are now at the point where the models they are developing are orders of magnitude larger than the current generation but only display imperceptibly small gains in performance.
This means that, even with $500 billion, OpenAI doesn’t stand a reasonable chance of producing an AI any better than the ones they already have.
And there are still a litany of other problems. For one, growing evidence shows that making AIs larger worsens them. Studies have proven that as these AIs get larger, they get worse at broad, simple tasks, which are critical for using them in a commercial setting, and only marginally better at more specific complex tasks. There is plenty of real-world evidence supporting this notion, such as OpenAI’s latest o1 model being worse at basic tasks than their old ChatGPT-4o model and Tesla’s latest FSD software, which has started cutting corners and hitting curbs.
Then there is the data issue. AI companies are already struggling to find enough high-quality data to train their AIs. They can’t train their AIs on AI-generated data, as it destabilises the model, and they used to get their data by scraping the internet. However, recently, the internet has been flooded with AI-generated content, rendering their primary data sources useless. To use this gargantuan AI infrastructure, OpenAI will need much more high-quality data than it already has. Where on Earth are they going to get that from without introducing AI-generated content and ruining their model?
And the huge question is profit. OpenAI is currently not profitable at all. They are already burning through cash at a staggering rate and had to be bailed out by Microsoft last year to avoid bankruptcy. Even if OpenAI stopped all development expenditure, their current models cost so much to run that they would still operate at a net loss. So how will spending nearly 40 times more than Microsoft invested in them solve this? As I have already stated, the AI Stargate’s vast infrastructure will not be a significant improvement over what we already have; if anything, it might be worse. So how can OpenAI produce the profits to service and repay this immense investment? I’m not sure if they can.
Then there is the power issue. Where will they get the power to keep this giant data centre going? AI companies are already stressing the US power grid to a breaking point.
This is where Trump comes in.
Trump plans to impose 100% import tariffs on Taiwanese computer chips. The idea is to force them to make them in the US. For reference, Taiwan makes around 90% of the world’s advanced computer chips. Your laptop, phone, car, and tablet likely all run on Taiwanese silicon.
When Trump was last president, he threatened similar tariffs, spooking TSMC (Taiwan’s primary chip manufacturer) to build a US plant in Arizona, which is set to open this year. So enacting such a tariff would be a massive boost for this site. The US is the world’s third-largest importer of computer chips, so losing these customers would be a huge blow for non-US-based chipmakers! However, the chipmakers in the US, like TSMC’s Arizona plant, can charge extortionate prices, still undercut the competition, and grow their profit margins dramatically!
This is where I need to put my tin-foil hat on. It looks like Oracle (via AMD) will be a major customer of this Arizona site. Not only that, but SoftBank also owns 90% of ARM, which owns the architecture those AMD chips use. SoftBank has also recently walked away from a deal with Intel to make an AI chip and is now penning a deal with TSMC, and it looks like that will involve this Arizona plant.
Here is my theory. These companies don’t care if OpenAI makes a next-gen AI or even hope to profit from the Stargate project. They can see that Trump’s tariff on Taiwan chips could make them a tonne of money if they can get their foot in the door of US chip making. Funding Stargate can further squeeze the US market by skyrocketing local demand, as this $500 billion project could easily use up most of the US’s chip-making capacity. This could raise prices so much that locally made chips could be as expensive, if not more expensive, as paying the tariff on Taiwanese chips. This will dramatically drive up the profits of all the Stargate investors (except OpenAI), as they all have significant shares in chip manufacturing and design companies (except OpenAI). OpenAI, on the other hand, has its value bumped even more than it already is because of its association with such a large project.
This would have dire consequences for the US public and the US economy, as all electronic devices would become extortionately expensive overnight, limiting people’s ability to access computing power. However, Trump has repeatedly shown that he doesn’t care about this. He is happy as long as he and his billionaire pals get richer.
If I take my tin-foil hat off, this theory runs on many assumptions. Not the least that these companies would need to have absolute faith in Trump’s word to make this scheme pay off, which has been proven to be worthless over the years. Mind you, the power demand from Stargate might justify his push to keep drilling for oil; after all, renewables have now been proven to be cheaper, more profitable, better for the economy, more land efficient, and cleaner. He can only justify a fossil fuel push if power demand suddenly spikes… So, maybe his word is more solid here?
Either way, if you look at how AI actually works and the science behind how it will develop going forward, Stargate makes zero sense. I can’t help but feel there is something seriously fishy afoot. My little pet theory might be entirely wrong. But nonetheless, we should maintain a healthy amount of scepticism over this project. After all, it is doomed to fail as an AI project. The people pushing it will undoubtedly know that. So what are its actual aims?
Thanks for reading! Content like this doesn’t happen without your support. So, if you want to see more like this, don’t forget to Subscribe and help get the word out by hitting the share button below.
Sources: Reuters, Forbes, BBC, The Guardian, Will Lockett, Will Lockett, Will Lockett, Will Lockett, TH, AMD, TH, CIO, CNBC, Ampere, OEC, CFO Brew, Techzine