Hello, and happy Easter to those who celebrate. In case you missed it: We’ve launched a new site focused on AI security, along with a dedicated newsletter – Control Plane. If you enjoy ID Tech’s weekly AI Update, please go ahead and sign up for the newsletter, where I’ll be providing even more in-depth coverage of the AI security space.
Now, on to this week’s Update:
Raspberry Pi is reporting a 25 percent jump in annual earnings, reflecting in part some very strong demand for agentic AI tech. Raspberry Pi is known for making compact computer boards and modules enjoyed by hobbyists, but thanks to advances in AI efficiency, these small computers can increasingly run sophisticated AI systems. That is prompting demand from people who want to run AI agents. At the same time, the broader AI boom is driving demand for Dynamic Random Access Memory, which is physically built into the Raspberry Pi boards.
On that note, Google DeepMind launched Gemma 4, the latest in its open source line of AI models, and it’s available in versions that are small enough to run on phones or Raspberry Pi boards. In launching Gemma, Google explicitly framed it as offering AI for “complex logic and agentic workflows”. We covered the news on Control Plane, where we also noted the startup PrismML’s emergence from stealth and its launch of Bonsai 8B, a 1-bit large language model that fits into 1.15 gigabytes of memory. The upshot is this: standard LLMs use 16-bit precision for storing numbers, but PrismML has found a way to shrink that down to 1-bit. That points to a future where even more sophisticated AI models will run on small computers, phones, and other devices.
Sticking to the memory front, UK-based Fractile is looking to raise about $200 million at a valuation of about $1 billion. The startup has been working on specialized AI chip architecture that uses in-memory compute, which it says is better for running inference than the standard AI chips made by the likes of Nvidia, which are better for AI training. Fractile’s bet is that compute is increasingly shifting from training new models to actually running them, and is pitching its chips as a better way of doing the latter.
Nvidia, meanwhile, is looking to counter the trend toward custom AI chips, and to that end has invested $2 billion in California-based Marvell as part of a broader strategic partnership. Marvell makes custom XPUs, offering things like optical interconnects, silicon photonics, and scale-up networking. AI-focused companies might have sought out these solutions in order to avoid being over-reliant on Nvidia GPUs; now, thanks to the partnership, they’ll still be going through Nvidia to get to them, with Nvidia offering the surrounding compute and interconnect stack through its NVLink Fusion solution, which is designed to help various AI chips work together as if they were one giant chip.
Paris-based Mistral has raised $830 million in debt financing to build an Nvidia-powered data center in Bruyères-le-Châtel, south of Paris – a small town that happens to host a major French nuclear research facility run by the French Alternative Energies and Atomic Energy Commission. Compared to other AI heavyweights like OpenAI and Anthropic, an $830 million raise isn’t much. But Mistral isn’t playing the same game as its rivals. It’s trying to build an AI platform that European businesses and governments can use without routing sensitive data through the US.
SpaceX has confidentially filed for an IPO. That’s big news because the Elon Musk-owned space travel company was merged with xAI – Elon’s AI company – earlier this year, in a deal that valued SpaceX at about $1 trillion and xAI at $250 billion. Bloomberg reports that the IPO effort is now targeting a valuation of more than $2 trillion. Elon has repeatedly spoken of a need to bring AI compute into orbit, and he seems to be serious about it.
The chatbot’s take: I tried to get a bit of grounding on this whole compute-in-space idea:
–
By Alex Perala
