Microsoft is building a $100B supercomputer for AGI - Weekly News Roundup - Issue #461
Plus: Israel uses AI to target Hamas; US announces partnerships on AI safety with the UK and EU; Apple explores personal robotic projects; how biomanufacturing can move beyond steel vats; and more!
Hello and welcome to Weekly News Roundup Issue #461. This week’s main story is Stargate, a $100B supercomputer that Microsoft and OpenAI are building to train models that could be considered AGI.
In other news, it was revealed that Israel is using AI-powered databases listing 37,000 Palestinians as potential targets, raising questions about the use of AI in warfare. Elsewhere in AI, Google is planning to charge for AI-enhanced search results and the US announced partnerships on AI safety with the UK and the EU. Over in robotics, reports emerged about Apple’s secretive personal robotics projects and the US Army mulls introducing robot platoons into armoured brigades.
I hope you’ll enjoy this week’s selection!
It is undeniable that AI is the hottest technology right now. Almost every week, there's either a new breakthrough model released or another AI company announces a new investment worth hundreds of millions, if not billions, of dollars. New AI startups, some less than a year old, achieve valuations in the billions of dollars while other companies are propelled into being worth trillions of dollars.
According to a report published by The Information, Microsoft is adding more fuel to the AI bubble by building Stargate - an AI datacentre worth $100 billion designed to train next-generation AI models, possibly even artificial general intelligence (AGI). This US-based supercomputer will be 100 times more expensive than some of the largest data centres today, and Microsoft executives want to launch it as soon as 2028, according to the report. OpenAI and Microsoft are building Stargate in five phases. A phase four system would cost less and may launch as soon as 2026 with the full system being ready by 2030. Currently, Microsoft and OpenAI are in the middle of the third phase.
When fully completed, Stargate could be the most powerful supercomputer ever built. At the time I am writing this, the most powerful supercomputer for AI is Nvidia’s Eos. This supercomputer, which ranks 9th on the Top 500 list of the most powerful supercomputers in the world, has 4,608 H100 GPUs delivering 121.4 petaflops of general-purpose FP64 performance or 18.4 exaflops of FP8 AI performance. Currently, the supercomputer holding the overall title for computing power is Frontier, which has almost 9 million CPUs and GPUs offering 1,194 exaflops of FP64 computing power. Stargate is expected to exceed Frontier's performance.
Why would OpenAI need so much computing power, then? The answer is simple - to train models that could be considered true AGI. As it stands now, the path to AGI will require not only new architectures adding reinforcement learning to large language models to explore multiple different reasoning paths, but also a breakthrough in computing power.
Another reason could be to catch up with Google in terms of available computing power. Despite its slow start to the generative AI game, Google is still a major contender to be the first company to create AGI or at least offer a better product than OpenAI. That’s because Google is the most compute rich company in the world, as SemiAnalysis reports, massively outgunning OpenAI in this area. By building Stargate, Microsoft and OpenAI plan to catch up, if not jump over, Google’s computing resources.
However, making that vision a reality comes with a massive list of technical challenges to solve before full Stargate is operational. I suspect the majority of the $100 billion budget would go into research and development. Despite offering the best GPUs for AI right now, I don’t think Nvidia’s chips will be powering Stargate. I expect Microsoft will eventually design its own high-performance chips designed specifically for crunching numbers inside Stargate. The reasons for that are the independence from Nvidia as the chip supplier (which has proven to be an annoying partner) and the ability to tailor the chips to perform AI operations as efficiently as possible. Google already does that with their Tensor Processing Unit (TPU) chips since 2015. Last year, Microsoft made the first step towards that vision by introducing Maia 100 and Cobalt 100 to Azure’s cloud infrastructure.
Designing custom chips to run Stargate is just one of many challenges facing Microsoft and OpenAI. A supercomputer the size of Stargate would have hundreds of thousands, if not millions, of dedicated AI chips. All these chips will require an enormous amount of power, possibly coming straight from a nuclear power plant, bypassing the national electric grid. At this scale, problems like cooling become major engineering challenges and might require designing completely new solutions, like connecting the chips not with copper wires but with fibre-optic links. Meanwhile, software engineers will have to create new tools not only to run AI models on such massive supercomputer but to also manage the entire datacentre.
The idea of a $100 billion supercomputer just to run AI models perfectly captures the current state of the AI industry. On one hand, there is hope that after multiple false starts and AI Winters, we are now close to solving intelligence. From the perspective of companies like Google, Microsoft or Amazon, spending unimaginable amounts of money on new AI supercomputers makes sense as the return from the investment, the creation of AGI, is worth it. On the other hand, $100 billion just so happens to be the same amount of money the same companies and many more invested into self-driving cars. After years of promising that self-driving cars will arrive “the next year”, we still don’t have fully autonomous cars widely available on our roads. Stargate could be Microsoft's biggest failure, becoming a testament to the perils of technological overambition.
Only time will tell the path Stargate will ultimately take.
If you enjoy this post, please click the ❤️ button or share it.
Do you like my work? Consider becoming a paying subscriber to support it
For those who prefer to make a one-off donation, you can 'buy me a coffee' via Ko-fi. Every coffee bought is a generous support towards the work put into this newsletter.
Your support, in any form, is deeply appreciated and goes a long way in keeping this newsletter alive and thriving.
🦾 More than a human
Chinese mourners turn to AI to remember and ‘revive’ loved ones
A new trend is emerging in China where AI is used to create digital avatars of deceased loved ones. Starting at just 20 yuan (£2.20 or $2.80), companies are offering these digital replicas, allowing mourners to interact with lifelike representations of the departed. For those intrigued by the concept of digital immortality, I've explored this phenomenon in greater depth in a separate article.
This Bag of Cells Could Grow New Livers Inside of People
A biotech company LyGenesis announced that an initial volunteer has received an injection of donor cells to turn one of their lymph nodes into a second liver. This method leverages the regenerative capabilities of lymph nodes, repurposing them as bioreactors to grow new liver tissue. Early animal experiments have shown promising results. If the experiments in humans prove to be successful, this new method could save the lives of thousands of people waiting for liver transplants.
Salt-Size Sensors Mimic the Brain
Researchers at Brown University have developed tiny silicon sensors, each the size of a grain of sand, that mimic the brain's functionality. These microsensors act like neurons, transmitting data upon detecting changes, potentially revolutionizing the monitoring of brain activity. This neuromorphic approach signifies a step towards scalable, low-power monitoring of brain activities and could lead to not only better brain-computer interfaces but also low-power devices for monitoring a wide range of physiological signals.
🧠 Artificial Intelligence
‘The machine did it coldly’: Israel used AI to identify 37,000 Hamas targets
Israeli intelligence sources revealed the use of an AI-powered database, named Lavender, during its bombing campaign in Gaza. Lavender, developed by the elite intelligence Unit 8200, could identify potential targets linked to Hamas, listing up to 37,000 Palestinian men at one point. This tool, along with another AI system called the Gospel, was used to accelerate the target selection process. Intelligence sources revealed that the system's efficiency led to a reduction in human input to merely formal approvals, sparking a debate on the moral and legal challenges of AI in warfare.
Google set to charge for internet searches with AI, reports say
Google is reportedly considering charging for AI-enhanced search features, marking what could be the most significant change to the company's revenue model in its history. Experts suggest this radical shift is a natural consequence of the substantial costs associated with providing such a service and could encourage other players in the sector to adopt some form of subscription model to cover these expenses.
US, Britain announce partnership on AI safety, testing
The US and the UK have announced a new partnership on the science of AI safety. This collaboration involves joint AI model testing and international cooperation on AI safety standards. It aims to tackle both national security and societal concerns posed by AI, including the development of more capable generative AI models and preventing its misuse in scenarios like bioterrorism.
EU and US set to announce joint working on AI safety, standards & R&D
The EU and the US are set to announce cooperation on AI safety, marking a significant step in aligning transatlantic strategies on AI's risks and potential. This collaboration will cover AI safety oversight, standardisation efforts, and promoting AI for the public good in developing regions, aiming to support regulatory implementation and foster beneficial AI applications globally. This follows a similar US-UK safety agreement, mentioned previously.
Navigating the Challenges and Opportunities of Synthetic Voices
In this blog post, OpenAI shares insights from testing Voice Engine, an AI model capable of replicating voices from a 15-second sample which OpenAI is already using to power the preset voices available in the text-to-speech API as well as ChatGPT Voice and Read Aloud. The post includes examples of Voice Engine's ability to recreate voices in various languages. OpenAI does not plan to release Voice Engine in the near future and uses this post to start a discussion on addressing potential issues arising from voice-cloning models.
▶️ Making AI More Accurate: Microscaling on NVIDIA Blackwell (7:59)
In this video, Dr Ian Cutress from TechTechPotato explains in more detail the concept of microscalling and reduced precision formats that Nvidia is introducing in Blackwell chips. These techniques aim to increase the number of operations the chip can perform without compromising the accuracy of the AI models running on these chips.
If you're enjoying the insights and perspectives shared in the Humanity Redefined newsletter, why not spread the word?
🤖 Robotics
Apple Exploring 'Mobile Robot' That 'Follows Users Around Their Homes’
Mark Gurman revealed in his column at Bloomberg that Apple is working on a number of “personal robotics” projects. One of these projects is described as a "mobile robot" that would "follow users around their homes," while another is said to be an "advanced table-top home device that uses robotics to move a display around". Apple is exploring home robotics as the next big thing after their effort to build an electric car failed. These are long-term, moonshot projects that are unlikely to be publicly revealed any time soon.
Army mulls introducing robot platoon into armored brigades
The US Army may introduce a drone and robotics platoon into its armoured brigade combat teams. If implemented Army-wide, the new platoons would lead to a dramatic increase in the use of robotic systems, especially ground robots. These robotic platoons won't be autonomous, as the technology is not yet sufficiently advanced. The integration of robotic platoons seeks to improve operational efficiency and protect infantry lives.
Bipedal robot takes a beating, keeps on hiking
Chinese robotics company LimX Dynamics presents P1, a bipedal robot that can easily navigate difficult terrain such as dense forests, ditches or uneven, rocky slopes, all while being kicked or hit with a stick.
🧬 Biotechnology
Beyond Steel Tanks
The big dream of synthetic biology is the ability to grow everything (and I mean everything) solely through biological processes. However, current methods of using bacteria in large steel vats may not be good enough to scale production to meet market demands. In this article published at
Thanks for reading. If you enjoyed this post, please click the ❤️ button or share it.
Humanity Redefined sheds light on the bleeding edge of technology and how advancements in AI, robotics, and biotech can usher in abundance, expand humanity's horizons, and redefine what it means to be human.
A big thank you to my paid subscribers, to my Patrons: whmr, Florian, dux, Eric, Preppikoma and Andrew, and to everyone who supports my work on Ko-Fi. Thank you for the support!
My DMs are open to all subscribers. Feel free to drop me a message, share feedback, or just say "hi!"
Cool.
I have to admit that the part I like to browse through the newsletter most is the final links, especially on robotics. I was especially struck by Apple's move, which is now a widespread topic of discussion. Thanks for the intriguing links!