On October 28, FuriosaAI announced a $620 million Series D and a $2.9 billion post-money valuation. Participants include Cisco Investments. These figures come from the company.
It says RNGD has been in mass production since January. The 20,000-card target this year is a plan, not verified deliveries. Its next generation remains in development.
Architecture and software belong in the same assessment
The architecture description presents a Tensor Contraction Processor: a programmable approach to tensor operations with an emphasis on data locality. The company explains that its compiler plans memory allocation and execution ahead of time to reduce data movement. This describes an engineering approach, rather than an independent efficiency measurement against every alternative.
The October 16 SDK 2026.4 publication supplies specific software context. It describes experimental hierarchical KV caching across NPU, host and shared storage, initially using Mooncake. The feature is disabled by default. The release also describes vision-language model support using standard continuous batching. This earlier SDK material is background to the financing announcement.
What would matter to an actual operator
Our editorial assessment is that accelerator selection should begin with a specific workload. An illustrative example would be an internal document-summarisation service. Its team would measure more than one fast response: different input lengths, several concurrent users and periods of low traffic would all belong in the evaluation. This example is not a test of RNGD conducted by SIExpo.
Caching would be a separate part of that assessment. We would compare repeated and unfamiliar requests, record memory consumption and check behaviour after a restart. An experimental capability should not automatically become part of a production promise. Reproducible settings and a clear account of what comes from the model, compiler and data reuse would make the result easier to assess.
In our view, the investment matters because manufacturing and software support can be addressed as a connected problem. Capital raised does not itself determine a customer's cost per useful answer. That requires a comparable model, output quality, response time and a full accounting of operating costs.
SIExpo has not performed hardware comparisons and does not verify claimed savings for a particular data centre. The established news is the announced funding round, assessed alongside documented architectural and software qualifications.
*AI illustration: an imagined accelerator and server environment, not a product photograph of RNGD.*
*Editorial verification: 30.10.2026, 09:13:25 +0200 (Europe/Sofia).*
