e=edge. It would be amazing to have a true edge device to run router OS software and bridge L3 and L2 with 4 physical ports (2 visible bonded ports). Small single CPU, 4-8G RAM, small Storage, Lots of NIC ports. Think pFSense, Vyos, Mikrotik, etc..
Intel Ice Lake-based c3.small.x86 (iGPU)
With Ice Lake's iGPU having VP9 hardware encoding support it would be a very efficient option for some of our workloads instead of burning lots of CPU cores (and wall clock time would be lower as well).
Support Arista Routers as Custom Appliances on Metal
Similar to how Pure Storage is deployed today, allow customers to order their own dedicated Arista to use for high-performance interconnection environments.
100Gb Networking Option for Server Configurations
Building block servers with minimum 100Gb ports. Includes fast storage and NVME interconnection. Allows 100Gb to the internet. Allows high-perfomance VMware, NoSQL DB, low-latency key/value stores, object stores etc.
GPU Config for Graphics-Based Workloads
GPU enabled server configuration for workloads that are graphics intensive, eg broadcast servers seeking to produce streams at 60FPS at 4K.
SW RAID Selection
Allow any on-demand instances to be able to select SW RAID when provisioning, either via API or portal selection.
mini-servers to give the sweet, sweet, taste of equinix metal
Something that gives people just enough metal to get a feel, without a massive up-front expense. The old intel atom c2550 | c2750 (rangeley) series were ideal for getting people familiar with cloud metal, bgp etc, and all the awesome things packet/equinix brings to the table, but without the high price point of the larger servers. I don't know if this is possible, but we still hang onto our reserved ones, they're perfect little puppies.
I'd love to have 256-way servers available, for maximum performance. These would be incredibly competitive, surpassing the best available offerings from AWS and GCP at the moment (whose normal offerings top out at 96 and 224 CPUs respectively). While I'm sure some applications will want a huge amount of memory to go with that, in my case I'd be happy with around 2GB per CPU. It'd likely make sense to have a "high compute" version in that range, as well as an offering with 4+GB per CPU for more memory-intensive workloads.