-4.jpg?w=1024)
It takes serious money to challenge Nvidia before your first chip ships. Britain's Fractile has just raised $720 million to try.
On Nov. 11, the European Commission proposed the Digital Fairness Act, which would tighten subscription and cancellation rules while giving users more control over potentially addictive app features.
Fractile is building processors for AI inference, the work of generating responses after a model has been trained. The investment gives Britain's chip-design ambitions a boost, but the company does not expect its first commercial chips and systems until 2027.
UK backing lifts Fractile to a reported $7 billion valuation
According to Reuters, the round was led by Thrive Capital, Founders Fund, Lightspeed, and Redpoint.
Reuters attributed the $7 billion valuation to a person involved in the deal; a UK government announcement put the company's value above £5 billion.
The public investment supports the UK's push to retain AI hardware expertise and skilled jobs, including around Fractile's London and Bristol operations. The government did not disclose the size of its investment in this particular round.
Fractile stacks memory directly above its AI processors
Fractile's design places DRAM directly above processor logic rather than following the conventional approach used in high-bandwidth-memory GPU systems. An investor's account says the arrangement aims to move model data faster while offering more memory capacity than on-chip SRAM.
The company is also developing its systems and software alongside the silicon. Faster memory access could help with long AI responses and growing workloads, but overall performance will still depend on model support, power consumption, and the surrounding software.
It would seem that Fractile isn't alone in the race to make AI inference faster and cheaper. AMD and Cerebras are pursuing different chip designs, while Groq has raised funding to expand its inference cloud. Even Nvidia is reportedly considering a deal with South Korean AI chipmaker Rebellions.
What eWEEK found: Fractile's funding is ahead of its commercial proof
The Fractile investor materials and Nvidia's published software announcement show two companies at different stages. Fractile is preparing a purpose-built inference platform, while Nvidia already has shipping GPU systems and established inference software.
The comparison below examines their hardware designs and commercial readiness rather than independently measured performance.
Comparison | Fractile | Nvidia |
|---|---|---|
| Memory approach | DRAM stacked directly above processor logic | Data-center GPUs commonly use HBM |
| Primary focus | Purpose-built AI inference | AI training and inference |
| Commercial availability | First chips and systems planned for 2027 | GPU systems already deployed |
| Software | Custom systems and software under development | Established CUDA, TensorRT-LLM, and Dynamo ecosystem |
| Published proof | No independent production-chip comparison in the materials reviewed | Shipping platforms and vendor-reported performance results, but no direct Fractile comparison |
Faster memory is only part of the enterprise equation
Fractile's design targets data movement, an important constraint when large models generate long answers or process many requests. However, improved memory bandwidth does not automatically translate into the same gain in tokens generated per second, response time, or operating cost.
The company's claims will need to be tested on working systems using comparable models and configurations.
For enterprise buyers, relevant measurements include time to the first token, output speed under concurrent demand, energy used per token, and total cost per million tokens.
Tests should also account for model size, context length, reliability, and whether existing applications can run without extensive changes.
Software may prove just as important as silicon. Nvidia's Dynamo 1.0 release describes production inference orchestration and integration with widely used frameworks.
Fractile will need to demonstrate comparable operational practicality for the workloads it targets, although the reviewed materials do not yet provide enough detail to judge that question.
UK chip design is not the same as a UK chip supply chain
The government's investment announcement emphasizes high-skilled employment, domestic expertise, and reduced reliance on overseas technology. It does not identify a UK fabrication plant or specify where Fractile's processors will be manufactured, packaged, or supplied from.
For UK and European businesses, having another AI chip supplier could provide more options. However, they would still need to consider manufacturing capacity, delivery timelines, technical support, and where the systems would be hosted.
A chip designed in Britain does not necessarily mean it will be manufactured there or that customer data will stay in the UK.
Fractile's 2027 launch will test the $720 million bet
The new funding gives Fractile capital to pursue a different way of running demanding AI models. It also raises expectations: a promising memory design must become a reliable commercial platform with software, customer support, and pricing that make sense outside a demonstration.
For European organizations planning AI infrastructure, Fractile could become another option to consider, but it has yet to prove itself as an alternative to Nvidia.
Businesses will need to see how its chips perform in real-world applications, how much energy they consume, how easily they integrate with existing systems, and whether they can be delivered reliably.
For now, the $720 million funding shows strong investor interest, but Fractile still needs to prove what its technology can deliver.
As companies like Fractile push to make AI faster and more efficient, knowing how to get the most from AI tools remains just as important. Sharpen your prompting skills with The Neuron Academy's free six-minute How to Talk to AI lesson. You can also explore more AI courses with seven days of free access. Check out all the lessons here.