Subscribe
18:00Tao calls OpenAI’s Navier–Stokes push “resource extraction”17:10LAPTOP memecoin hits $190.81, then loses 99% inside an hour17:05Hubinger puts the odds of AI killing everyone above 10%; a colleague resigns16:39CancerBench launches; five frontier models tied at zero cancer types cured16:30ElevenLabs preparing 2028 IPO after $11bn round, The Information reports16:30Anthropic retracts its July explanation: Mythos 5 attacked systems knowingly
Hardware2 min read

A 1.6Tbps switch now costs $1,100 and runs on 123 watts, because AI clusters moved onto desks

ServeTheHome measured 1.567Tbps through MikroTik's four-port 400GbE box. That is 98% of line rate from a half-width desktop unit with an Amazon chip doing the management.

In briefThe MikroTik CRS804-4DDQ-hRM has four 400GbE QSFP56-DD ports for 1.6Tbps, lists at $1,295 and street-prices near $1,100, under $0.70 per Gbps1ServeTheHome measured 1.567Tbps on Keysight CyPerf and 1.589Tbps on IxNetwork2The switch idles at 29.3W with a 123W maximum rating, and a cheap used 32-port 100GbE switch can draw 900W or more3

ServeTheHome measured 1.567Tbps through a MikroTik CRS804-4DDQ-hRM on Keysight CyPerf and 1.589Tbps on IxNetwork, from a half-width desktop box with four 400GbE ports that lists at $1,295 and street-prices near $1,100. That works out to $0.69 per Gbps, and to about 98% of nominal line rate.

MikroTik CRS804-4DDQ-hRM throughput (Gbps)
Nominal1,600Keysight IxNetwork1,589Keysight CyPerf1,567
STH@ServeTheHome

This cheap 1.6Tbps switch has 4x 400GbE ports, lowering the price ($0.69/1Gbps) for local AI clusters and those who just want fast networking t.co/W9yGwF9kho

on X · 11.6K views · captured Sep 10, 2026

But the number that should stop you is the power draw. The switch idles at 29.3W and is rated at 123W maximum. ServeTheHome's own point of comparison is a cheap used 32-port 100GbE switch, which "can easily draw 900W or more." Do the division and this box costs 0.077 watts per gigabit against roughly 0.28 for the used bargain — better than three and a half times the efficiency, in a unit you can leave on a desk at 40-odd decibels.

Why does a product like this exist now? Because somebody has to connect eight NVIDIA GB10s in a room without a data centre. ServeTheHome has been running exactly that: the RDMA backend for an 8-way GB10 cluster on one of these for months, and a GB300 Station cluster on another with DR4 optics. Four ports is not a topology. It is a pod. Desktop AI boxes made a market for a four-port top-of-rack, and MikroTik got there first at this price.

Open it up and the bill of materials explains the number. Switching is a Marvell 98DX7335A1, a 5nm Prestera part with 32 lanes of 50Gbps (which is 1.6Tbps and not a gigabit more). Management runs on an Annapurna Labs AL52400, a quad-core CPU from Amazon's chip division, with what looks like 4GB of ECC memory beside it.

An Amazon-designed processor, in a Latvian budget switch, on a desk in a lab. That is where custom silicon ends up.

So here is the hard limit worth knowing before you buy, and it is not the port count. The link between that management CPU and the switch ASIC runs at 10Gbps. Anything you route through the CPU path — firewalling, deep inspection, anything RouterOS does in software — is capped there, whatever the front panel says. ServeTheHome is blunt about it: not enough performance to be a 1.6Tbps firewall.

And the counter-case for spending more is real. Four ports splits into eight 200GbE or sixteen 100GbE, and then you are done. Larger topologies want a proper spine, and the ports are QSFP56-DD rather than the QSFP112 on a GB300 Station (or the OSFP on Spectrum-X gear), so optics and DACs become a separate shopping problem. Getting NCCL and RDMA running took MTU changes, PFC and switch-feature work — though ServeTheHome notes that since roughly February an AI agent can read MikroTik's documentation and produce a working configuration, which is its own small sign of the times.

Our read is that this is the first switch priced as though small AI clusters are an ordinary product category rather than a lab curiosity, and that the price will keep falling, because a 32×50G Marvell part sells into a market far larger than desktop pods. We would expect a four-port 400GbE desktop switch under $1,000 street before the middle of 2027.

One caveat from the review that no spec sheet will tell you: the rear-heavy power supplies will tip a small rack over.

Sources

01
The MikroTik CRS804-4DDQ-hRM has four 400GbE QSFP56-DD ports for 1.6Tbps, lists at $1,295 and street-prices near $1,100, under $0.70 per GbpsThe exciting part of this is that there are four 400GbE QSFP56-DD ports for a total of 1.6Tbps of network bandwidth. The list price is $1295, but these often sell at a street price of $1100 or so, making this under $0.70/Gbps.” — servethehome.com · reported · Sep 10
02
ServeTheHome measured 1.567Tbps on Keysight CyPerf and 1.589Tbps on IxNetworkKeysight CyPerf Throughput: 1.567Tbps (DR4 to DR4 optics) Keysight IxNetwork Throughput: 1.589Tbps (4x DR4 to 16x DR1 optics)” — servethehome.com · reported · Sep 10
03
The switch idles at 29.3W with a 123W maximum rating, and a cheap used 32-port 100GbE switch can draw 900W or moreTaking a look at the power consumption, we get around 29.3W at idle. ... The maximum rated power consumption is 123W, which is far less than one of the 350W power supplies puts out (92W without attachments). Noise-wise, the switch with…” — servethehome.com · reported · Sep 10
Show all 10 sources
04
ServeTheHome runs the RDMA backend for an 8x NVIDIA GB10 cluster on one unit and a GB300 Station cluster on anotherWe have a number of these now and have been running our RDMA network backend for the 8x NVIDIA GB10 cluster for several months now, using one. We also have an NVIDIA GB300 Station cluster running on another (using DR4 optics).” — servethehome.com · reported · Sep 10
05
The switch chip is a Marvell 98DX7335A1, a 5nm Prestera part with 32 lanes of 50Gbps, and management runs on a quad-core Annapurna Labs AL52400 with 4GB of memoryThen we have the quad-core Annapurna Labs AL52400 with 4GB of what appears to be ECC memory onboard, given that there are five DRAM packages. ... On that switch chip, there is a Marvell 98DX7335A1. This is a 32-lane 50Gbps switch (or 4x…” — servethehome.com · reported · Sep 10
06
There is only a 10Gbps link between the management CPU and the switch chip, limiting anything traversing the CPU pathMany may wonder if, given RouterOS has the functionality, if this CPU means that this would be a good 1.6Tbps firewall box. To be clear, this is not enough performance for that much bandwidth, and there is only a 10Gbps link between the…” — servethehome.com · reported · Sep 10
07
The ports are QSFP56-DD, not the QSFP112 of a GB300 Station or the OSFP of Spectrum-X switches, and can be split into 2x 200GbEAs a quick note, these are not the QSFP112 ports you would find in something like the NVIDIA GB300 station. These are not the OSFP ports you would find in higher-end switches like the NVIDIA Spectrum-X series or many ConnectX-8/…” — servethehome.com · reported · Sep 10
08
Getting NCCL and RDMA working required MTU and PFC changes, and since February or March 2026 AI agents can produce a working MikroTik configurationWe managed to get NCCL and RDMA networking working over the switch by changing MTU sizes, enabling switch features, and more. Since Feb-Mar 2026, AI agents have been able to read MikroTik's documentation, configure the ports and…” — servethehome.com · reported · Sep 10
09
The rear-heavy power supplies will tilt a small rackFor example, you may see the support in this half-width rack photo. That is because the rear-heavy power supplies will cause this mini rack to tilt.” — servethehome.com · reported · Sep 10
10
ServeTheHome posted the review on X on September 3, 2026 quoting $0.69 per GbpsThis cheap 1.6Tbps switch has 4x 400GbE ports, lowering the price ($0.69/1Gbps) for local AI clusters and those who just want fast networking” — x.com · primary · Sep 10
Up next · Keep readingHardware · 3 min read

A20 Pro is the first 2nm phone chip. The number that matters is 32 Neural Engine cores after three years at 16

Apple doubled the Neural Engine, widened memory bandwidth 50 percent, moved the DRAM off the thermal path and tripled the vapor chamber. Every one of those decisions is about running models on the phone.

Continue ↓