AI chips ahoy!

This is the second edition of The Angle where AI chip-building news led the round-up in as many weeks – which alone says volumes

Plenty of big news this week – and lots in AI, including Nvidia’s bonkers earnings which I won’t go into in detail since plenty of markets-aimed publications do that much better already. The big theme, appropriately, was a number of different approaches to challenging Nvidia’s dominance when it comes to being the hardware layer powering the AI revolution.

SoftBank’s founder is looking to fund and build a $100 billion AI semiconductor company

• SoftBank’s Masayoshi Son, who is probably most famous recently for pouring absolutely wild amounts of capital into the startup ecosystem via the formerly very spendy SoftBank Vision Fund, is now tracking down $100 billion to build and Nvidia competitor

• The project is codenamed ‘Izanagi,’ which is the name of a Japanese deity who was responsible for the creation of the country’s island archipelago alongside his sister ‘Izanami’

• The creator god moniker is an appropriate one for an ambitious project intended to disrupt Nvidia’s current stronghold on the market

• Masayoshi Son’s hundred billion might not mean much, however, if Sam Altman manages to round up the trillions he’s reportedly seeking for a similar endeavor

Groq’s ‘LPU’ impresses with ChatGPT and Gemini turbocharging

• Groq, a company that shouldn’t be confused with Elon Musk’s Reddit bro-themed chatbot Grok, aims to challenge Nvidia’s throne without trillions or even billions of dollars

• The company is creating special purpose AI chips it calls ‘Language Processing Units’ or LPUs, which it claims are better at running large language models (LLMs) like Open AI’s ChatGPT and Google’s Gemini than Nvidia’s GPUs

• Groq’s founder and CEO Jonathon Ross previously helped set up Google’s dedicated AI chip division, which created the company’s Tensor Processing Unit or TPU, a chip designed specifically to boost training and inference performance for large AI models

Microsoft taps Intel to build it custom chips, likely at least in part with AI in mind

• Intel and Microsoft announced a $15 billion+ deal this week that will see Intel produce custom chips for the Windows-maker

• The details revealed by the company didn’t specify that these were being made specifically with AI in mind, but Microsoft has been pushing the integrated AI features of Windows more and more with recent partner hardware, and the company is believed to be working on in-house designs for CPUs and AI-specific coprocessors for some time now

• Apple’s playbook since taking control of its own chip design first in iPhone, then later in Mac devices, has included building in specific sub-processors that handle local machine learning and AI tasks

• This could definitely help future Windows machines without discreet GPUs do more locally with AI processing

Techstars is shutting down Seattle accelerator; Toronto location’s future uncertain

• Techstars announced a significant change to its overall approach this week, with a refocus that will aim back on cities “with the most robust venture capital communities,” specifically SF, Boston, LA and NY

• Seattle is one of the cities affected by the change, and the accelerator there (the home of Microsoft) will shut down immediately, though its staff says alumni will still get support from the Techstars team

• This follows the closure of Techstars Austin late last year, and earlier today BetaKit’s Josh Scott reported that Techstars Toronto has “paused applications” as its managing director Sunil Sharma departs

• The future is uncertain for the Toronto program at the moment – as it is for many of the other markets where Techstars operates, since the accelerator says that it’ll run programs in 30 other locations beyond the core four mentioned above, but has not yet specified exactly where

This was a huge week in news, which also included Google debuting open source AI models named ‘Gemma,’ – but which come with 67 different licenses it would take a vertical-specific LLM trained just on those to parse. The search giant also suffered backlash for some imagery generated by Gemini that seemed to prioritize diversity in representation of subjects to the exclusion of all other considerations – even when producing images of “Nazis.”

Thanks for reading, and please share if you enjoyed. See you Tuesday for The Angle’s deep dive.