Fractile just raised $720M for chips

What happened: The London and Bristol chip startup, founded in 2022, raised a $720 million Series C round.

• The round was led by Thrive Capital, Founders Fund, Lightspeed Venture Partners and Redpoint, with the UK's Sovereign AI Fund and British Business Bank also investing.
• The UK government put the valuation above £5 billion, about $6.6 billion by its conversion, while Reuters and the Financial Times reported about $7 billion.
• That is up from a $220 million Series B in May at around $1 billion. The chips stack DRAM directly on the processor, and the company projects up to 100x performance and a 90% energy cut.

Why it matters: Chip startups are drawing very large rounds, which shows where investor money and talent are headed. The performance and energy numbers are company projections, so treat them as a reason to watch, not a reason to plan.

What happened: Reuters reported that Northwestern Polytechnical University, a Xi'an school with ties to the People's Liberation Army, bought two Nvidia H200 chips through a reseller.

• The purchase was listed in an October 21 procurement notice at 874,700 yuan, about $123,000, from Shaanxi Bosuan Network Technology rather than Nvidia itself.
• The university has been on the US Entity List since 2001, and it also gained access to Nvidia processors through five contracts this year for rented computing capacity.
• Reuters reviewed 26 Chinese procurement and corporate records from 2026 and found other universities buying restricted chips or renting Nvidia-powered services from domestic suppliers.

Why it matters: Export rules are being tested through resellers and rented capacity, not just direct sales. If you buy GPU compute, ask your provider where its chips came from, because customers and partners will likely ask you the same questions.

What happened: ElevenLabs released Scribe v3 for recorded audio and Scribe v3 Live for real-time use such as voice agents and live captions.

• Scribe v3 cuts word errors 27% against Scribe v2 on the FLEURS benchmark, with average word error rate falling from 16.94% to 12.38% across 102 languages.
• Scribe v3 Live has a median latency of 150 ms, realtime speaker labels, and turn detection that reads conversation context instead of silence alone.
• Scribe v3 labels up to 32 speakers, accepts keyterm prompts of up to 1,000 terms, and has a separate library that holds up to 10,000 terms.

Why it matters: Voice agents depend on speed and accuracy, so the 150 ms median is worth testing. These are ElevenLabs' own benchmark numbers, so run Scribe v3 on your own audio before you switch.