WhiteFiber launches a two-site GPU supercluster as performance testing continues

Why it matters

WhiteFiber is selling cross-site connectivity as a way around the power and space limits of a single AI campus. Its materials distinguish the 136 Tbps design figure from the earlier R&D result and say final fiber testing to confirm guaranteed commercial specifications is underway, leaving customer-scale performance to be demonstrated.

WhiteFiber CEO Sam Tabar (@samirtabar) is putting a new infrastructure bet on the idea that AI clusters should span data centers rather than stop at their walls. On September 23rd, WhiteFiber made Continuum, a system linking two facilities into one logical GPU supercluster, commercially available. Its launch announcement says the architecture delivers 136 terabits per second across an 83-kilometer link, with 0.9 milliseconds of guaranteed round-trip latency.

The launch-day materials qualify the status of those figures. The announcement states that the architecture delivers 136 Tbps and 0.9 milliseconds, while a company blog post updated on launch day says full-fiber spectrum testing is underway to confirm guaranteed commercial specifications. The blog describes 136 Tbps as the design's capacity after additional wavelengths were brought online and identifies the earlier 111.2 Tbps result as research and development testing. The release does not provide an independent benchmark or test methodology. WhiteFiber has announced commercial availability; confirmation of the guaranteed commercial specifications is still pending.

Tabar's role also calls for precision. WhiteFiber's investor biography lists his path through law, Bank of America Merrill Lynch, FullCycle and Fluidity, where he was co-founder and chief strategy officer. He later served as Bit Digital's chief strategy officer and now leads both Bit Digital and WhiteFiber. WhiteFiber's SEC filings describe a business built from Bit Digital's high-performance computing operations and contributed to WhiteFiber before its August 2025 IPO; they do not establish Tabar as WhiteFiber's founder. His quote in the announcement captures the premise he is now selling: "geographic distance as a ceiling on what a cluster can do."

The cluster crosses 83 kilometers

Continuum's initial installation links two HITRUST-certified QTS facilities with 12 strands of Zayo dark fiber. DriveNets supplies the AI network fabric, while WEKA provides storage and memory infrastructure through NeuralMesh. WhiteFiber says the arrangement makes the separated sites operate as one logical cluster, rather than as two installations connected by a conventional network link. WhiteFiber has also submitted patent applications covering its implementation.

That architecture targets a recognizable constraint in AI infrastructure: one campus can run out of power or space before an operator can justify building another. WhiteFiber says Continuum can pool capacity across sites for training and inference, let telecom and metro facilities contribute otherwise isolated resources, and provide failover if a location goes offline. WhiteFiber also says sensitive workloads can remain within their originating jurisdiction while compute is pooled across sites. Those are commercially useful aims, but the announcement supplies no workload-level results showing how a distributed training job performs during ordinary operation or a site failure.

WhiteFiber's own explanation makes the performance question central. Its blog says the initial research test reached 111.2 Tbps and 0.9 milliseconds over the 83-kilometer link, while the launch announcement promotes 136 Tbps. Moving data at high aggregate bandwidth is one piece of a distributed cluster; the launch materials do not quantify the training workloads, synchronization overhead or GPU utilization that would show whether the two sites behave like one in practice. Without those results, the figures describe the network link more clearly than the compute gains available to a customer.

A product launch on top of a concentrated business

Continuum enters a business that already has reported revenue, but it has not yet been shown as a separate source of sales. WhiteFiber reported $50.8 million in consolidated revenue for the six months ended June 30th, 2026, including cloud and colocation services. The filing does not identify Continuum as a separate source of sales or report customer demand for the product. Those figures cover WhiteFiber's broader operations; they do not establish Continuum revenue or customer demand for the new product.

That context shapes the launch's commercial test. Tabar is extending WhiteFiber's pitch from providing AI compute and data-center capacity to connecting infrastructure across locations. WhiteFiber is asking customers to trust a new layer of its offer with workloads that may be constrained by reliability, compliance or campus capacity. WhiteFiber's existing operations give it a platform from which to sell that idea, while the launch materials leave the customer economics and independently validated performance to be proven in deployment.

WhiteFiber says Continuum is available to enterprise customers now. The next evidence that will matter is practical: published test conditions for the 136 Tbps claim, workload-level performance across the link, and customer deployments that show whether multi-site compute can be sold as a useful alternative to expanding a single campus. For Tabar, the bet is that a distributed cluster can become a product customers buy, rather than an impressive network demonstration.