• The Californian startup is partnering with Google Cloud to build two new super-powerful supercomputers 
• The initiative is based on Nvidia’s most powerful GPU chips
• The goal is to realise artificial general intelligence (AGI) within a short timeframe 
• The duo are constructing an automated AI software engineer and researcher that can see and understand an organisation’s entire code repository

San Francisco, California-based generative AI (GenAI) startup Magic has forged a partnership with Google Cloud to build two extremely powerful cloud-based supercomputers, using chips from Nvidia, on the hyperscaler’s platform. 

It is claimed the new machines will have 60 exaflops of performance. One exaflop is a quintillion or 1,000,000,000,000,000,000 (that’s a 1 followed by 18 zeros) floating point operations per second (FLOPS). That is effectively the equivalent of 60 billion individuals each having command of 1 billion state-of-the-art scientific calculators and running a complex computation at precisely the same microsecond. 

The project was announced as Magic concluded a $320m round of funding from a variety of investors. The company noted in this blog (buried some way down the text) that it has now raised a total of $465m, including the $320m from “new investors Eric Schmidt, Jane Street, Sequoia, Atlassian, among others, and existing investors Nat Friedman & Daniel Gross, Elad Gil, and CapitalG.”

Magic, which is barely two years old, develops AI models both to write software and to automate software development tasks. According to this blog published recently by the company, Magic’s partnership with Google Cloud will result in the production of the Magic-G4 supercomputer, which will be based on Nvidia H100 Tensor Core graphics processor unit (GPU) chips. Magic says it has 8,000 of the hugely powerful chips, each of which cost between $25,000 and $40,000, in stock. 

The company’s next planned supercomputer, the Magic-G5, will run on Nvidia’s even more powerful Grace Blackwell GPUs, which should initially become available later this year and  scale up to “tens of thousands of Blackwell GPUs over time”. (It’s worth recalling that earlier this year, and despite the fact that Nvidia has a cloud-based service of its own called DGX Cloud, the company’s co-founder and CEO Jensen Huang declared “we are not a cloud company – we are a platform company”.)  

Eric Steinberger, the founder and CEO of Magic, said his company’s goal is to build artificial general intelligence (AGI), a machine representation of the generalised cognitive abilities of a human being expressed in software and able to perform any task that a human being is capable of performing. Steinberger stated that achieving AGI “will take a lot of compute” – a gigantic understatement. Nonetheless, AGI is regarded as being ultimately achievable and the CEO added, “We are excited to partner with Google and Nvidia to build our next-gen AI supercomputer on Google Cloud. Nvidia’s system will greatly improve inference and training efficiency for our models, and Google Cloud offers us the fastest timeline to scale, and a rich ecosystem of cloud services.” 

According to Google Cloud, more than 50% of the world’s GenAI startups are utilising its cloud platform because it is trusted, powerful and robust.

Currently AI models “learn” in two different ways: First via initial training and then via in-context inference. With inference, an already-trained machine learning model draws conclusions from new data. An AI model able to infer can do so without seeing or knowing examples of the desired result and thus inference is often referred to as an “AI model in action”.

In AI, a ‘token’ is a basic unit of data or “building block” processed by algorithms. It is a component of a larger dataset, which may represent words, characters, phrases, images or video. Tokens are particularly pertinent to natural language processing (NLP) and machine learning (ML). Meanwhile, a context window measures the number of tokens that the AI model can process at one time. The purpose of a context window is to aid AI models in finding or recalling information whilst they are working.

In Magic’s partnership with Google’s Cloud’s AI platform, the two new cloud-based supercomputers will support Magic’s intent to develop code assistants (which greatly reduce the time it takes to write code by offering real-time suggestions, automating routine tasks and predicting what the next lines of code will be, and thus accelerate development cycles) with a context window reaching 100 million tokens. That equates to the amount of data generated by a decade of human speech. Complicated, isn’t it?

Magic’s ultimate goal is to build an “automated AI software engineer and researcher that can see and understand an organisation’s entire code repository and complete large tasks over long time horizons.” To do this, it is training “frontier-scale” large language models (LLMs) with ultra-long context windows and other advanced capabilities. 

Ian Buck, vice president of hyperscale and high-performance computing (HPC) at Nvidia, commented: “The current and future impact of AI is fuelled to a great extent by the development of increasingly capable large language models. Powered by one of the largest installations of the Nvidia GB200 NVL72 rack-scale design to date, the Magic-G5 supercomputer on Google Cloud will provide Magic with the compute resources needed to train, deploy and scale large language models, and push the boundaries of what AI can achieve.” 

Whether it will also be able to pull a holographic rabbit from a digital hat is another matter entirely…

– Martyn Warwick, Editor in Chief, TelecomTV