Behind a static IP request sit network policy problems. They show up in a pilot as timeouts, blocked calls and certificate errors.

Security teams usually put it as one question. Can we get a static IP or DNS entry so the service is always reachable at the same address?

Why do enterprises ask for a static IP on an inference endpoint?

Enterprises ask because their firewall works from a list. A default-deny egress policy blocks every outbound connection from the company network unless a rule allows it. Egress means traffic leaving the network. The simplest rule names one address.

An IP address is a machine's numeric address on the network, and a static IP never changes. DNS, the Domain Name System, turns a hostname such as api.example.com into an address.

Inbound or outbound: which address are you actually allowlisting?

Allowlisting means a rule that lets named addresses through and blocks the rest. First ask which way the traffic flows.

DirectionWho starts the connectionWhat your firewall needsExample
Your apps to the providerYour applicationThe provider's hostname or address, on your egress policyCalling a chat completions endpoint
The provider to youThe providerThe provider's source address, on your ingress policyA webhook that reports a finished job

Ingress is traffic entering your network. A provider may offer a fixed address for one direction and not the other.

Our serverless runs show the second direction. A run is a job you start without managing a machine. When a run finishes, we send an HTTPS POST to the callback URLs you supplied. Our callback docs say to gate that endpoint behind a shared secret in the path or a header, because we don't sign callbacks. That check works whatever address the call comes from.

Two rows. Your apps calling the provider pass an egress rule, highlighted, that names the provider's hostname and reaches the inference API. The provider calling you back passes an ingress check that verifies a shared secret in the path or a header and reaches your webhook. A note says a hostname rule survives an address change.
Your egress rule names our hostname, and your webhook checks a shared secret, so neither depends on a fixed address.

Why do managed API endpoints rarely have a fixed IP?

A managed endpoint is an API the provider runs for you on shared machines. Its address is rarely fixed, because the provider moves load, adds capacity and replaces failed servers.

Providers give you a hostname instead. DNS records carry a TTL, the time to live. RFC 1035 defines it as how long a record "may be cached before it should be discarded". When the provider changes the address behind a name, clients pick up the new one once the TTL runs out.

Many hosts also share one address. TLS, the encryption behind HTTPS, has an extension called SNI (Server Name Indication) for that case. RFC 6066 describes it as letting a client name the server it wants on hosts with several "'virtual' servers at a single underlying network address". An address alone often doesn't identify one service.

Is IP allowlisting a security control or an availability risk?

IP allowlisting is a weak security control and a real availability risk. An address rule checks where a connection comes from, not who sent it. Everyone behind the same company gateway carries the same address.

NIST's zero trust architecture, SP 800-207, starts from the same point. It says network location "is no longer seen as the prime component to the security posture of the resource". Authentication, proving who the caller is, comes first.

If a provider changes an address and your rule names the old one, production calls fail until someone edits the rule. Keep address rules only as a second layer on top of API keys or mutual TLS, where both sides present certificates.

What does a default-deny egress policy actually need from a provider?

A default-deny egress policy needs five facts from a provider.

  1. A stable hostname to allow, written as an FQDN (fully qualified domain name), such as api.example.com.
  2. Whether a documented address range exists, and how changes to it are announced.
  3. The ports and TLS versions in use, usually port 443 for HTTPS.
  4. Whether an inspecting proxy, one that decrypts traffic to scan it, ends the TLS connection and opens a new one.
  5. What happens during a failover, when the provider moves traffic to other machines.

Ask your network team whether the firewall can match on the hostname. It reads the name from the DNS lookup, or from the SNI field the client sends at the start of TLS. A hostname rule keeps working when the address behind it changes.

Does token streaming survive a corporate inspecting proxy?

Token streaming survives a proxy that passes each piece of data on at once. It fails when the proxy buffers or closes quiet connections. Streaming means the model sends its answer token by token as it writes it. A token is a short piece of text, often part of a word.

Streaming over HTTP usually uses server-sent events, a format the HTML standard defines as text/event-stream.

nginx, a widely used proxy, shows both defaults. Response buffering is on, and proxy_read_timeout is 60 seconds between two reads. If nothing arrives in that time, "the connection is closed".

Our worked example uses Qwen3.8 27B on our serverless API, where you pay per token and run no machines. Time to first token is the wait before the first word. Both speeds are 7-day medians from our status checks, read on its model page on 30 October 2026. The 3,000-token answer is hypothetical.

time to first token  = 0.291 s        (Qwen3.8 27B, 7-day median)
output speed         = 45 tokens/s
3,000-token answer   = 0.291 s + 3,000 / 45 = 67.0 s
streaming            = one token every 1 / 45 s = 22 ms
longest silent answer inside 60 s = (60 − 0.291) × 45 = 2,687 tokens

Without streaming, the server sends nothing until the whole answer is ready. That silence lasts 67 seconds, and a 60-second proxy closes the connection. A silent answer fits inside 60 seconds only up to 2,687 tokens.

With streaming, data arrives every 22 milliseconds, so the timer never fires. A buffering proxy undoes that. It holds every token, so your user sees the first word after 67 seconds instead of 0.29. While it waits, its read timeout keeps counting, so nginx's 60-second limit can still close the connection. So streaming must be on and buffering off, which in nginx means proxy_buffering off.

Calculated example for a hypothetical 3,000-token answer on Qwen3.8 27B at 291 ms to first token and 45 tokens per second: the user sees the first data after 0.291 seconds when streaming passes through, after 67.0 seconds behind a buffering proxy, highlighted, and after 67.0 seconds without streaming. A dashed line marks a 60-second proxy read timeout, which both 67-second cases cross.
In our calculated example with a hypothetical 3,000-token answer, a buffering proxy delays the first data from 0.291 to 67.0 seconds.

The example is optimistic: its speeds are medians for short requests, and real long requests are slower, so the silence lasts even longer.

Our serverless docs recommend streaming for long outputs because it "keeps data flowing over the connection". They allow up to 300 seconds of generation per request. Test it from inside your company network before the pilot.

What does 'custom DNS' mean, and which version do you need?

Custom DNS means one of two things. Meaning (a) is a vanity CNAME, an alias record in your own domain that points to the provider's hostname. It changes the name your developers type, not the path the traffic takes.

A vanity name also runs into certificate checks. The client checks that the server's certificate names the host it asked for. RFC 9525 calls the name the client expects the "reference identifier". If the provider's certificate doesn't list your alias, the client reports a certificate error instead of a friendly name.

Meaning (b) is a private DNS name that resolves only inside your network. It points to a private address reached over a private link, such as a VPN. That is a networking product, not a DNS record, and our private endpoint guide covers the deployment side. Procurement often wants (b) and writes down (a), so ask which is meant.

If you need one address of your own today, run it yourself. Each of our GPU virtual machines gets its own IP address, shown by the VM status command, with root access. A reverse proxy on that machine, a server that forwards requests, gives your clients one address. The security guide of vLLM, an open-source serving engine, recommends this setup, with the proxy adding authentication, rate limiting and logging. Our gateway guide shows how.

What should you ask a provider before you design around a fixed address?

Ask these before anyone writes a firewall rule:

  • Do you offer a fixed inbound address, or a documented address range with change notices?
  • Which address do your callbacks and webhooks come from, and how do I verify them?
  • Is there a private connection option, and which product does it belong to?
  • What TTL do your DNS records use, and how should my client reconnect after a change?
  • What happens to my allowlist during a failover?

A precise "no, here is what to do instead" beats a vague yes, because it tells your network team what to build.

Here are our answers from our docs. We are Lyceum, a sovereign European AI inference provider for open-source models. Our serverless models run in Paris and Finland, and prompts are not stored.

QuestionWhat our docs and website say
Hostname to allowServerless: https://api.lyceum.technology/openai/v1. Dedicated endpoints: https://api.lyceum.technology/api/v2/external, with your deployment ID as the model name
Callbacks to youAn HTTPS POST to the URLs you supply, checked with a shared secret you set
One address of your ownYour own proxy on a GPU virtual machine, which has its own IP address
Live behaviourOur status page shows live status, per-model latency and past incidents

Dedicated endpoints run one model on GPUs reserved for you. For anything the table doesn't cover, such as private connections, ask our engineers. Business customers get a direct line with contractual response times.

Bring your egress, allowlisting and DNS requirements to us before you design around them.

Sources

  1. rfc-editor.org
  2. rfc-editor.org
  3. csrc.nist.gov
  4. html.spec.whatwg.org
  5. nginx.org
  6. rfc-editor.org
  7. docs.vllm.ai