DGX Spark/Guides/Headless setup
Guide · November 2026

DGX Spark headless setup: first boot without a monitor, finding it, and provisioning several units

By Samuel Seidel · Published November 7, 2026 · Sources checked November 7, 2026

A DGX Spark you own can be set up with no monitor or keyboard: power it on, join the Wi-Fi hotspot whose SSID and password are printed on the sticker with the Quick Start Guide, finish setup in a browser, and the Spark then joins your network and answers at a name like spark-abcd.local, per NVIDIA's first-boot guide (updated 4 November 2026). Expect up to about 10 minutes of updates after the screen says it is rebooting. The October 2026 release, which began rolling out on 21 October, made headless use easier: a DGX Dashboard Wi-Fi Assistant joins new networks in appliance mode, the display memory reserve drops to 512 MB with no display attached, and native cloud-init can provision a unit from a USB stick with no setup screens at all. None of this applies to a rented GPUwerk Spark, which arrives racked and set up.

ReleaseWhat changed for headless or multi-unit setupSourceDate
LaunchNetwork appliance setup: hotspot, captive portal, browser setup from another computerFirst boot guideOct 2025
April 2026Option to skip the setup screens for IT provisioning; installs and updates from USB or a local repository; cloud-init in custom ISOs; fix for "unable to detect network during system setup"Release notesApr 2026
June 2026Over-the-air updates no longer installed by default during setup, so first boot finishes soonerRelease notesJun 2026
July 2026cloud-init over DHCP, HTTP and TFTP as well as USB; display reserve switchable between 2 GB and 4 GBRelease notesJul 2026
October 2026Wi-Fi Assistant in the DGX Dashboard; Auto display reserve (0.5 GB headless); native cloud-init via the nvidia-spark-cloud-init package; fix for fans not responding in headless idleOctober announcement2026-10-21

Read 7 November 2026. NVIDIA says these releases apply to the DGX Spark Founders Edition; partner GB10 systems "will offer the same updates in their next software release", per the October announcement. The package name is from NVIDIA's cloud-init guide.

Can you set up a DGX Spark without a monitor?

Yes; NVIDIA calls it setting up "as a network appliance", and it needs only a second computer with Wi-Fi and a browser. The first-boot guide gives the order:

  1. Plug in Ethernet now if you will use it; the Spark powers on as soon as the supply is connected, so cables go in first.
  2. Power on. The Spark opens a Wi-Fi hotspot. Join it from your laptop with the SSID and password on the Quick Start Guide sticker.
  3. A captive portal opens; if it does not, browse to the setup address printed on the same guide.
  4. Pick language and time zone, accept the terms, create your user, choose Wi-Fi (skipped if Ethernet already has internet).
  5. The Spark joins your network and tears the hotspot down; your laptop rejoins your own Wi-Fi and the setup page continues there.
  6. The Spark downloads and installs its software and may reboot more than once. Do not power it off.

NVIDIA's warnings are specific: no captive-portal Wi-Fi (hotel, airport) and no phone hotspots during setup, and if your laptop cannot find the Spark again after it joins your network, you "must connect a display/keyboard/mouse to continue". The usual reasons it cannot are client isolation on the Wi-Fi, the laptop sitting on a different network, or mDNS blocked on a corporate LAN. On a guest or office network, use Ethernet or the local setup path.

What goes wrong with the hotspot, and how did owners fix it?

Three failures come up again and again on NVIDIA's forum, and only one has a software fix.

No hotspot. It exists only until first setup has finished. An NVIDIA moderator said so on 21 October 2025 (thread 348457), and an owner repeated the advice in October 2026 for a unit "set up or tested before, for example by the seller" (thread 385762). One owner got the hotspot back by running a system recovery with a keyboard and display attached, then cutting power at the final "press return" screen and booting again with nothing connected. That still needs the peripherals once.

"Unable to Join" or "Reason 500". In March 2026 an owner, building on another user's patch, traced setup refusing every network to the setup wizard's HTTPS check against connectivity-check.ubuntu.com: if that request fails, setup treats a working Wi-Fi or Ethernet link as offline (thread 364255). NVIDIA's April 2026 release notes list a fix for network detection during setup. If you still hit it, check that your network does not block that hostname.

Lost after joining. In November 2025 an owner power-cycled a Spark that never appeared on his LAN after the hotspot closed, and was left with a unit frozen on its welcome screen with dead USB ports (thread 349832, later marked solved without details). NVIDIA's rule covers it: never cut power during the install. Wait the full 10 minutes, then look for it as below.

How do I find the Spark's IP address, hostname and user?

The hostname is spark- plus four hex characters, also printed on the Quick Start Guide, and the user is the one you created during setup; there is no default login. NVIDIA's cloud-init guide confirms the naming: a setup service rewrites /etc/hostname to spark-<4 hex> on units that boot into the setup screens. To test name resolution, the manual SSH playbook uses:

ping -c 3 spark-abcd.local
ssh <YOUR_USERNAME>@<DEVICE_HOSTNAME>.local

If the ping says "cannot resolve", mDNS is blocked and you need the address from your router's client list. With an Ethernet cable connected you can log in by that address and read the Wi-Fi interface's IP with the command an NVIDIA moderator posted in thread 348457: ip -f inet a show wlP9s9. Give the Spark a DHCP reservation while you are in the router; the remote access guide covers SSH keys, the DGX Dashboard on port 11000 and Tailscale from there.

How do I move a headless Spark to another Wi-Fi network?

On the October 2026 release, use the Wi-Fi Assistant in the DGX Dashboard; on anything older, come in over Ethernet and use NetworkManager. NVIDIA describes the Assistant in one line, as a way to join a new Wi-Fi network "when DGX Spark is used in appliance mode", and the dashboard docs add an administrator switch for it (wifiAssistant.enabled in /opt/nvidia/dgx-dashboard/settings.json). It needs the dashboard to be reachable, so you are still connected some other way while you change networks. The pre-October route, from NVIDIA's moderator in thread 348457:

sudo nmcli d wifi connect <wifi_name> password <password>
ip -f inet a show wlP9s9

Then unplug Ethernet and reconnect by hostname or the new Wi-Fi address.

What should I change after a headless first boot?

Give back the display memory, check the fans, and make sure Wi-Fi and the boot menu behave without a screen.

Display reserve. The Spark reserves 2 GB of system memory for the display by default. The UEFI settings page lists Auto (0.5 GB on a headless system), 0.5, 1, 2 and 4 GB under Advanced, Advanced Menu, Display Reserved Memory. By my calculation, going from 2 GB to 0.5 GB frees 1.5 GB: 1.2% of a 128 GB machine, 2.3% of a 64 GB one. Small on 128 GB; on the 64 GB model it is room for more KV cache, which the 64 GB vs 128 GB guide sizes. Run free -h before and after to confirm it.

Fans. Owners reported fans that never spun in headless mode, one logging a GPU at 52 C drawing 3.72 W with the case too hot to touch in August 2026 (thread 378945). The October release notes list "Fans Stop Responding in Headless Mode" as fixed. Update, then put a hand on the box after an hour idle.

Wi-Fi drops. Several owners lost Wi-Fi on headless units after 16 to 20 idle hours in late 2025 (thread 353777). One owner's log analysis traced it to the Wi-Fi connection belonging to the login screen's session, which NetworkManager removed when that session idled out; the thread starter later stayed stable on a wired connection with an HDMI dummy plug. For a server, use Ethernet.

Boot menu. An owner found GRUB_TIMEOUT=0 on a unit set up in network appliance mode, so a failed kernel update left no menu to pick the old kernel from (thread 368939). The fix for that, and the kernel hold, are in the software updates guide; the wider recovery path is in the recovery and reinstall guide.

How do I provision several Sparks the same way?

Use cloud-init from a USB stick: since the October 2026 release a Founders Edition reads a FAT32 partition labeled OEMDATA at first boot, and if its user-data creates a user, the setup screens never appear. NVIDIA's custom installation guide sets four conditions, and a unit that misses any of them boots normally without logging an error:

Preparing the stick, per the same guide:

sudo wipefs -a /dev/<partition>
sudo mkfs.vfat -F 32 -n OEMDATA /dev/<partition>
sudo mkdir -p /mnt/oemdata
sudo mount /dev/<partition> /mnt/oemdata
sudo install -m 0644 <seed-directory>/meta-data /mnt/oemdata/meta-data
sudo install -m 0644 <seed-directory>/user-data /mnt/oemdata/user-data

The meta-data file can be one line, instance-id: oem-spark-01. Read NVIDIA's skip-setup user-data example before copying it: it creates a sudo user with a password, sets ssh_pwauth: true and writes an sshd drop-in with PasswordAuthentication yes, sets the hostname to the same placeholder as the username, deletes the stock ubuntu user, runs an optional hook.sh to install packages and firmware, and finally creates /etc/cloud/cloud-init.disabled so it never runs again. For a fleet I would change the hostname per unit, add SSH keys and switch password logins off. Results land in /var/log/cloud-init-output.log.

Units on an older release cannot use this until they are reflashed with an October 2026 or later recovery image, which erases the drive; NVIDIA's guide then adds OEMDATA as a second partition on the recovery stick. On OEM machines cloud-init is "enabled unless the OEM removes it during the image build", so check the maker's notes. The alternatives are PXE boot, which needs your own DHCP, TFTP and HTTP servers and either Secure Boot off or NVIDIA's signed grubnetaa64.efi.signed enrolled, and the agentless SSH tooling in NVIDIA's enterprise lifecycle guide for everything after first boot.

What applies to a rented Spark?

None of the first-boot steps, because you never see the machine's first boot. A rental is a Docker container on a Spark that is already racked and running DGX OS; you open it in the console's browser terminal or over root SSH with your own key. The nearest thing to cloud-init there is the startup hook: an executable /workspace/.gpuwerk-init that runs as root on every container start, documented on the tenant environment page. Use it to start your model server after a restart.

Which route should you take?

For one Spark at home, the hotspot flow is fine; do it near the router, on your own Wi-Fi, and wait out the updates. For a Spark that will live in a rack or a closet, I would borrow a monitor for 20 minutes, use Ethernet, and set the display reserve to Auto while you are in the UEFI anyway. For three or more units, write one cloud-init seed with keys instead of passwords and test it on a single machine first, since a bad seed fails silently. The setup guide picks up from a booted, networked Spark.

FAQ

Can I set up a DGX Spark without a monitor and keyboard?

Yes. NVIDIA's first-boot guide has a network appliance mode: power the Spark on, join the Wi-Fi hotspot whose SSID and password are on the sticker with the Quick Start Guide, and finish setup in a browser. The Spark then joins your Wi-Fi, installs updates for up to about 10 minutes, and is reachable by its spark-xxxx.local name.

My DGX Spark hotspot does not appear. What now?

The hotspot only exists until initial setup has finished, an NVIDIA moderator said on the developer forum in October 2025, so a unit set up before (by a seller, say) will not show it. Look for the exact SSID on the sticker, wait a few minutes after power-on, or connect a monitor and keyboard and use the local setup path instead.

How do I find my DGX Spark's IP address when it is headless?

Try its mDNS name first, for example ping spark-abcd.local, which NVIDIA's playbook uses. If your network blocks mDNS, look the Spark up in your router's client list. With an Ethernet cable in, you can SSH in by IP and read the Wi-Fi address with ip -f inet a show wlP9s9, as an NVIDIA moderator showed on the forum.

How do I move a headless DGX Spark to a different Wi-Fi network?

Since the October 2026 release, which began rolling out on 21 October, the DGX Dashboard has a Wi-Fi Assistant for joining a new network in appliance mode. On older software, NVIDIA's moderator gave this route: plug in Ethernet, SSH in, and run sudo nmcli d wifi connect with the new network name and password.

Can I set up many DGX Sparks automatically?

Yes, with cloud-init. Founders Edition units on the October 2026 release or later read a FAT32 USB partition labeled OEMDATA at first boot; if its user-data creates a user, the setup screens are skipped. Older units must first be reflashed with an October 2026 or later recovery image. NVIDIA also documents PXE boot for network installs.

Sources

Every figure above is attributed to its source in the sentence or table where it appears. Pages were read on 7 November 2026 unless a date says otherwise.

A GB10 machine that arrives ready.

The GPUwerk store sells GB10 machines checked before they ship, and can host one for you so there is no first boot to do at all.

GB10 machines shipped by DHL Express from Czechia, or hosted in GPUwerk's EU datacenter and rented out for you.

Browse machines Or rent a Spark that is already set up