It was like watching an endless before and after tennis rallies between hype and disillusionment recently following the news of artificial intelligence.
It is said that the AI investment bubble is shrinking. The revenue is too low to justify your investment. New AI products like the GPT-5 are overwhelming.
But wait, AI is the next great industrial revolution. Our current decisions will shape our future prosperity. Copyright laws need to be abolished to allow AI companies to suck up creative content and train their models for free. Your work will be automated. Data centers drink our water, consume our power, and pollute the atmosphere.
No, it needs to be braceed for an imminent AI and data center crash.
Or maybe not? Australia must compete with ChatGpt and build their own AI to properly refund artists' work.
However, these models can cost billions of dollars to train, which cannot be afforded.
And I'll go to that.
So, what is really going on?
One way to understand mixed signals around AI is to recognize that each one is looking at two very different visions of the future that are struggling to become reality.
One is dominated primarily in the US by big technology. Meanwhile, the ability to build powerful AI models is much more widely shared, with Australia not only being a consumer of the technology, but also a producer.
A big technological vision for the future of AI
A few years ago, the popular stories told about AI were simple, if a bit dark.
It turned out to be as follows: Several large companies emerged as masters of new technology and owned the means of production for most of the world.
AI is everywhere, embedded in every car, phone, refrigerator, all search queries and online ads, but the cost of building an AI model is extremely high.
They use their models to train better models and more, increasing the gap with potential rivals.
This is extremely beneficial. AI companies will do their best. They invent and even own “super intelligence.”
Openai's 7 Gigawatt, $400B ($60.6 billion) Stargate AI data center is under construction in Texas.
(Getty Images: Kyle Grillott/Bloomberg))
In the lofty words of Openai CEO Sam Altman (Openai Makes ChatGpt), AI “may capture a light cone of all future values in the universe.”
Openai is pursuing this vision alongside other leading tech companies.
This week, US chip maker Nvidia announced that it would invest USD 100 billion ($150 billion) in Openai to supply millions of AI chips to the company, allowing it to build vast supercomputers to develop a new generation of AI.
The idea is to crush the competition and maintain Openai's spot as an AI Frontrunner.
Other companies do the same. In June, Meta (the Facebook owner) announced that it would spend hundreds of billions of dollars on building a huge US data center, including one that covers an area that is almost the size of Manhattan.
The numbers are “eye-wonderful,” says Ann Whicker, head of the global technology team at Bain & Company, a management consulting firm, from San Francisco.
A row of natural gas turbines to power Openai's Stargate data center under construction in Texas.
(Getty Images: Kyle Grillott/Bloomberg))
This week, Bain released a report summing the unprecedented size of AI investments.
By 2030, global electricity supply needs to increase by 20% to deliver the extra energy required by data centers.
In the US, new data centers require approximately US$500 billion ($700 billion) of capital investments each year.
AI companies will need to add $2 trillion ($3 trillion) in annual revenue to fund capital expenditures in their new data centers.
“AI is moving forward at a faster pace than we actually saw transitions with other technologies,” says Hoecker.
New data centers are primarily used for the use of “inference” or trained AI models. This is not training new models, but rather asking ChatGPT.
“As AI usage increases, meeting growth will be much faster than training growth,” Hoecker said.
However, technological advances in AI have slowed down.
Given these investments, the idea of nation-states building their own AI models or supporting local businesses may seem like a hopeless ambition.
But that's exactly what's happening. And the model is pretty good.
Earlier this month, Switzerland released a large-scale language model (LLM) similar to ChatGPT. It was trained on its own infrastructure (Alpine supercomputers) and was estimated at around US$50 million ($75 million) in a fluent of 1,000 languages and dialects.
This is possible for inconvenient truths. This undermines the future of AI, dominated by the large tech outlined above.
The technological advances in AI have slowed down.
Established LLM has not improved at expected rates.
The GPT-4 (released in March 2023) was far better than the GPT-3 (June 2020). However, the GPT-5, released two years later in 2025 and estimated to take 10 times more training, did not give the same impression.
Currently, there is little talk of inventing super intelligence, and there is a lot of interest in practical applications that are not exciting (coding software, organizational workflows).
An August MIT report found that despite investing USD 3 billion ($4.5-6 billion) in generated AI, 95% of organizations are not making profits from AI investments.
Another survey from July found that developers completed tasks 20% slower than working with AI.
Meta's 5-gigawatt, $500 billion “Hyperion” data center is under construction in Louisiana, with a focus on Manhattan here. (Supply: Meta))
Despite the hype, it is still not clear how AI should be used.
“It's not necessarily because everyone assumes that the future is a bigger and more complex model,” said Nicholas Davis, co-director of the Human Technology Institute at Sydney Institute of Technology.
He compares the current moment of AI to the emergence of Noughties' digital cameras.
At the time, the number of “megapixels” that cameras had was an important advertising feature. But when even the most basic cameras had enough megapixels to take decent photos, the focus moved to other features.
At least for basic tasks, the public, commonly available and commonly used open source AI models fill the gap with their global proprietary rivals.
Access to AI, which allows you to do most of the basics that use AI, such as transcription, drafting, and summarizing, is becoming more widely available.
“We're in AI when everyone is focused on megapixels, but at some point we're going to go through a critical pixel limit where an open source model can do everything well,” Davis said.
“Large tech companies are putting billions of dollars in, but the open source model isn't far behind.”
Recession could benefit Australian startups
The slower technological advancement in AI sounds like bad news, but it includes hopeful possibilities.
For example, the rapid rise in open source LLMs could be utilized in countries like Australia.
Some of the best open source models approach GPT-4 performance with specific benchmarks, including language understanding, general knowledge, and problem solving.
These models fine-tune Australian data, run from Australian data centers, create LLMs that reflect national values, trained on national data for better general knowledge, and cannot be suddenly retracted by other countries.
“By adopting major open source models and tweaking, you can do huge amounts,” Davis said.
Another option is to build from scratch, according to Simon Chris, CEO of startup sovereign Australia AI.
“What we need is a simple, large-scale language model that handles all the text-based tasks we have. All chatbots and transcription services.”
Earlier this month, Sovereign Australia AI announced plans to train LLMs with 700 billion to 1 trillion connections (also known as parameters) with networks raised through private capital (also known as parameters).
The number of parameters is one measure of the size of the LLM (slightly similar to the megapixels of a digital camera). The GPT-3 has 1.75 billion parameters. The GPT-4 could be 1.5 trillion yen.
“A trillion parameters may be enough because many AI-using businesses want it,” says Kriss.
“The AI model doesn't need to be too big.“
Another Australian company, MainCode, is set to launch Matilda, Australia's first sovereign LLM later this year.
Dave Lempers, co-founder and CEO of MainCode, says the 30 billion parameters are “more than enough” to provide a model tailored to Australia's needs.
Like Chris, he believes Big Tech pursues an outdated vision.
“I think these are common wisdom. [big AI] Companies were about making their models bigger. I need to throw more [computing power] And.
“And I think what you're looking at is that these companies are starting a particular strategy and feeding that strategy.”
Will the AI bubble burst?
How is the AI bubble trying to pop while the data center shoots like a mushroom?
Bane and company Hoecker said some AI companies were overvalued and could have “road clashes” similar to the dot-com crashes of the late 90s.
“And the question is, when is that bump, and maybe how big is it?”
However, data centers are still in use.
“Even if the economy goes into a recession… data centers will not be fully utilized in a year or so,” she said.
In other words, even a recession does not stop the AI-enabled future from passing.
Gas turbines power US data centers run by Elon Musk's AI company. (Getty Images: Brandon Dill of the Washington Post))
AI is deeply embedded in everyday tasks, from Google to word processing, and running a digital economy requires a huge number of new data centers. Opting out of AI is not an option.
AI Future Hoecker's sights are a mixture of large companies that build unique models of complex tasks, such as operating humanoid robots and operating autonomous software systems (AI agents).
“There will be a continuous push in the industry for super intelligence. And then, “Now, what is the actual use case?”
It's an amazing moment for AI. Thousands of billions of dollars are invested in technology with uncertain use cases. Huge data centers can potentially derail the progress of reducing emissions and derail the line. And there are concerns that the entire US economy is being supported by the AI investment bubble.
At what point do you ask yourself if Big Technology is rushing straight to the future you imagined: Do you want this?
