Oxide Computer's $445 million Series D goes to making more on-prem racks

A software company running an Oxide rack can split it into separate virtual networks, one environment per engineering team, with a built-in firewall controlling the traffic between them, SiliconANGLE reported ^[12]. Under that layer, compute, storage and networking come as one pre-integrated package ^[7].

The pitch is an on-prem cloud. By Oxide's account, the new money goes to parts and factory output. "The accelerating pace of on-prem cloud computing projects has required us to scale manufacturing capacity by 20x over the past 12 months, and yet demand still exceeds supply," said Steve Tuck, Oxide's co-founder and chief executive ^[4]. "This funding will allow us to aggressively invest in the materials and manufacturing scale our customers need for years to come," he said ^[5]. Counting the February round, Oxide has raised $645 million across its last two rounds ^[14].

For the person sizing a deployment, the rack totals matter more than the per-sled figures. Fully loaded, one rack carries up to 3,456 cores and 7,200 terabytes of raw flash ^[15][16]. The storage writes three copies of every record to different flash modules ^[11]. That puts usable capacity at about 2,400 terabytes at most, before any other overhead ^[17].

Much of the rest of the design is integration work a team would otherwise do itself. A custom Oxide chip on each sled manages temperature and power and installs firmware updates ^[9]. The two switches keep a continuously updated map of the rack's network and send traffic around a malfunctioning component ^[10]. A background checksum hunts for data errors in storage ^[11].

None of that helps a project whose rack has not shipped. SiliconANGLE's report does not include a price or a delivery time. Tuck himself describes a company that has more orders than it can build ^[4].

Two questions sort the decision for a given team. The first is whether the workload runs on CPUs and flash today. The second is whether the project can absorb a wait from a vendor whose CEO says demand exceeds supply ^[4]. The buyer this round is built for has a CPU-only workload and a flexible start date, because the money goes to the capacity that buyer is waiting on ^[5]. With a CPU-only workload and a fixed go-live date, the decision turns on whether Oxide will commit to a delivery date in the contract. A workload that needs GPUs now falls outside the rack as SiliconANGLE describes it, because the graphics-card sled is still a plan ^[8][13].