Delos Data’s Dan Daly tackles rack-scale networking with 36 OSFP ports

At Computex 2026, Delos Data made a very specific bet: if AI chip startups want to compete with Nvidia and AMD at rack-scale, they should not start by reinventing the entire networking stack. The company is showing off a modular chassis built around 36 OSFP ports, arranged as nine ports per each of four OAM sockets. The point is to make it easier to wire a “rack-scale computer” that behaves like one giant GPU, without forcing new chip companies to also become experts in mechanical packaging, thermal management, power delivery, and the massive interconnect demands that rack-scale architectures require.

CTO Dan Daly told The Register that the platform is designed to be more flexible in how big you make your scale-up domain, “maybe you want a scale up domain of 100 or maybe you want it a scale up domain of one.” That flexibility is not marketing fluff, it is a wiring philosophy. Delos is keeping the system relatively simple compared to the custom integrated backplanes, power delivery, and cooling that Nvidia and AMD have built into their own rack designs. Instead, Delos’ chassis, at least from the front, looks more like a switch than a conventional GPU server, and it uses OSFP front-panel ports so customers can choose the cabling and switching approach that fits their deployment.

To understand why this matters, look at how rack-scale changes the networking math. The Register notes that a typical eight-GPU HGX node only needs one or two ports per GPU, which is relatively manageable by current server networking norms. But at the extreme end of rack-scale, Nvidia’s GB300 NVL72 needs 18 400 Gbps ports per GPU. That is where “building a chip” stops being the whole job. For startup chip teams, the networking problem can become the hidden tax that slows everything down, because the system needs enough interconnect bandwidth and power delivery to keep accelerators fed, while still fitting into a single rack that functions as one cohesive compute entity.

Delos tries to lower that barrier by choosing a front-panel interconnect that is meant to be pluggable. The company assumes 200 Gbps SerDes, which yields 3.6 TB/s per chip of interconnect. The Register adds that this equals Nvidia’s new Rubin GPUs, which is an implicit positioning move: Delos is trying to show that you can hit the same scale of internal bandwidth without locking yourself into a proprietary rack ecosystem. OSFP matters here because it allows customers to use standard DACs or pluggable transceivers, connecting to switches. While OSFP is usually associated with Ethernet, the Register reports you can run just about anything through them, including UALink, Ultra Ethernet, PCIe, or other options.

From a deployment standpoint, the architecture would be wired like any other hyperscale system, but denser. Delos is not the only idea floating around the market for rack-scale reference designs. The Register notes AWS appears to be repurposing Nvidia’s MGX form factor for its Trainium 3 rack systems, and AMD’s Helios rack is now an OCP standard. The difference, in Delos’ framing, is modular flexibility. Daly’s comments emphasize that the design can plug into different types of switches, potentially even optical circuit switches (OCS), depending on how you build your scale-up domain.

The networking topologies are where things get even more consequential. Using existing packet switches from Broadcom or Marvell, the Register says Delos’ design could support 512-1,024 accelerators in a single layer fabric, depending on whether you use 200 Gbps or 100 Gbps SerDes. If you go beyond a single layer, using multi-layer fabrics, OCS, and/or 2D/3D toruses, the compute domain could scale even further, while still using off-the-shelf components. That phrase, “off-the-shelf,” is doing heavy work for executives and boards. It suggests time-to-deploy advantages and potential cost control, because startups can iterate on chip designs while relying on a broadly available systems building block for the rack-scale interconnect.

Still, there is a trade-off. The Register flags that OSFP can create power consumption concerns for larger compute domains, especially because pluggable optics can be power hungry at scale. It notes this is one reason Nvidia has taken so long to embrace optical scale-up, contrasting copper reach limitations with copper’s lower power draw. Delos, according to the report, is already exploring versions that use near package or co-packaged optics out to MPO-style connectors rather than OSFP, implying an evolution path as deployments demand higher optical efficiency.

Delos is also pairing the hardware pitch with software, and that is another place where rack-scale can quietly derail teams. At large scale, physical and logical topologies can vary significantly by workload, which means traffic patterns and resilience strategies differ. Delos says it has developed a software orchestration platform for configuring and monitoring switched fabrics or meshes to enable dynamic rerouting in the event of a link failure. At Computex, the company showed a feature dubbed Nonstop AI networking where attendees could pull links at random and see the network automatically react and correct itself.

Strategically, this is more than a product demo. It is an attempt to reduce the “time tax” that rack-scale imposes on AI chip startups, especially around networking enablement. If Delos can make rack-scale compute domains modular and resilient, it could change how startups allocate engineering effort, and how investors underwrite deployment risk. In a market where custom racks from Nvidia and AMD set the baseline expectations for performance and integration, Delos is trying to create a faster, more flexible on-ramp for competitors. The stakes are simple: whoever makes rack-scale networking less of a bespoke science experiment wins speed, learning cycles, and ultimately customer mindshare.