Summary

  • Volantis announced an $88m Series A and says its first integrated A-1 inference engines are planned for customer delivery in 2027.
  • The company reports wafer-scale optical-link measurements, but its public system-level cost, power and latency comparisons lack enough test detail for buyers to reproduce them.

The new money changes Volantis’s capacity to build and commercialize A-1. It does not, by itself, settle whether the product performs as advertised. That distinction matters because the company’s public material describes evidence at different levels: link measurements, product specifications, performance comparisons and a future customer-delivery target.

Volantis said on 1 October that it had raised $88m in a Series A co-led by Lachy Groom and Abstract Ventures. The release names additional venture and angel investors and says the funds will support A-1 and its photonic memory architecture. The company’s own post says cumulative funding has reached $97m. Neither disclosure states a valuation. The release says first integrated inference engines are planned for customer delivery in 2027; it does not report a shipment, customer acceptance or operating deployment.

The product page lists 10 terabytes of memory, 240 terabytes per second of memory bandwidth, 10 terabytes per second of off-wafer input/output, a 20-kilowatt power envelope and a 15U form factor. It also advertises 15 times the tokens per dollar of Nvidia Rubin, six times the tokens per watt for low-latency models with more than one trillion parameters, and over 30 times lower latency for large-model inference. Those are company comparisons. The public page does not identify a benchmark model, run configuration, software version, price boundary, latency percentile or independent test report that would let a buyer reproduce the ratios.

There is a narrower measurement claim. Volantis’s technology page reports clean optical links, a bit-error rate below 1e-12 at wafer scale, more than ten times electrical reach and link energy below one picojoule per bit. Those results, as stated by the company, concern links at wafer scale. They are not a whole-system tokens-per-watt result. A finished inference engine also draws power in compute, memory, optical/electrical conversion, host systems and supporting equipment; the disclosed link figure alone does not set that total.

The scale claims also need a workload definition. A simple illustration: twenty trillion weights stored at exactly four bits each occupy about ten terabytes before metadata, runtime state or key-value cache. That arithmetic neither disproves Volantis’s design nor establishes that one A-1 can run the stated model. The company has not specified the model, precision, whether “20 trillion” means total or active mixture-of-experts parameters, or how weights and runtime state are distributed.

The next useful evidence is therefore not another funding total. It is an A-1 result that names the model and precision, prompt and output lengths, context, batch and concurrency, latency measure, software stack, comparison system, and complete power and cost boundaries. Until then, the raise is evidence of financing and the wafer figures are company-reported link evidence; neither is a substitute for a reproducible customer-workload result.

Sources: Volantis Series A announcement; Volantis company funding post; A-1 product specifications and comparisons; photonic technology and wafer-scale link claims.