> For the complete documentation index, see [llms.txt](https://docs.thetpa.ai/deai-tokenized-processing-architecture/llms.txt). Markdown versions of documentation pages are available by appending `.md` to page URLs; this page is available as [Markdown](https://docs.thetpa.ai/deai-tokenized-processing-architecture/addressing-networking-isolation.md).

# Addressing, networking and isolation

Every TPA has a short, readable address, and every address maps by formula to one private network IP. That one design choice gives TPAI three things at once: names people can remember, a network that needs no name server to find a TPA, and isolation that validators can apply TPA by TPA.

## TPA addresses

A TPA address is **three letters and three digits**, like `ABC-042` **\[Founder statement]**. It is the address of an AI endpoint, not a wallet. Wallets on TPAI Chain use `TPAI21-` addresses controlled by two or more private keys (see [TPAI Chain](https://docs.thetpa.ai/deai-tokenized-processing-architecture/tpai-chain.md)).

$$
26^3 \text{ letters} \times 10^3 \text{ digits} = 17{,}576{,}000 \text{ TPA addresses}
$$

Why this format:

* **Readable.** Seven characters, easy to say, type and remember, like a license plate.
* **Compact.** 17,576,000 addresses fit in 25 bits.
* **A fixed pool.** 17,576,000 is the total, ever. When a TPA is deactivated, its address is freed and reused by a new owner only after a cooling-off period, so nobody can pose as the old TPA **\[Planned]**. The length of the cooling-off period is **\[Open]**.
* **Endpoints, not agents.** The limit applies to TPA endpoints. Many AI agents, models and (from Phase 2) dApps can live inside one TPA.

## From address to IP, with no DNS

A fixed formula turns every TPA address into an index, and every index into a private IP address, so finding a TPA needs no name server **\[Planned]**. The private network routes traffic to wherever that TPA runs, and every connection is checked against the TPA's verified identity, not just its IP **\[Planned]**.

| Step                | Rule                            | Example                                |
| ------------------- | ------------------------------- | -------------------------------------- |
| Letters to a number | A = 0 … Z = 25, read as base 26 | ABC → 0 × 676 + 1 × 26 + 2 = 28        |
| Address to index    | letters × 1,000 + digits        | ABC-042 → 28 × 1,000 + 42 = **28,042** |
| Index to IP         | index added to 10.0.0.0         | 28,042 → **10.0.109.138**              |

The example is illustrative; the exact mapping is part of the published TPAI Chain specification **\[Open]**. Measured latency figures are **\[Open]**.

## The address space

User TPAs live in one private **/8 address space** (`10.0.0.0/8`, as defined in [RFC 1918](https://www.rfc-editor.org/rfc/rfc1918)), the range Kubernetes commonly uses for pods. A /8 holds 2^24 = 16,777,216 addresses. Earlier materials called this "NAT Type 8"; the standard term is a /8 block.

| Range              | Addresses      | Use                                                                                            | Status                      |
| ------------------ | -------------- | ---------------------------------------------------------------------------------------------- | --------------------------- |
| AAA-000 to YVH-215 | 16,777,216     | User TPAs, one IP and one network segment per TPA                                              | \[Founder statement]        |
| YVH-216 to ZZZ-999 | 798,784        | Reserved system TPAs: Stage's hosted base models, TPAI Chain validator nodes and network needs | \[Planned], with TPAI Chain |
| **Total**          | **17,576,000** |                                                                                                |                             |

**Why the numbers fit.** Of the 17,576,000 TPA addresses, the first 16,777,216 map one-to-one to the 16,777,216 IPs of the /8 network, and also to the 16,777,216 isolated network segments that VXLAN provides (24 bits). Every user TPA can therefore have its own IP and its own segment, and the remaining 798,784 addresses are reserved **\[Analysis]**.

**System TPAs.** System TPAs hold Stage's hosted base models (each base model sits in its own TPA) and TPAI Chain validator nodes **\[Planned]**. The reserved range is a fixed block at the end of the address space; earlier versions of this paper said any TPA could be designated a system TPA. The overlay network is private, inside TPAI infrastructure, not the public internet **\[Planned]**.

## Each TPA is an isolated container

When a TPA is activated, the platform starts an isolated container for it, in the preferred design a Kubernetes pod, and gives it its own IP in the /8 overlay **\[Founder statement]**. The pod is a lightweight gateway: it holds the TPA's address, keys, settings and isolation rules, and forwards requests to shared inference servers where the models run. With S-LoRA, many fine-tuned adapters share one GPU **\[Founder statement]**.

## The chain decides, the network enforces

Isolation is an AI safety control, and on TPAI Chain it is decided on chain **\[Planned]**.

* **Its own segment by default.** In the preferred design, VXLAN provides up to 16,777,216 isolated network segments (classic VLANs allow only 4,094). By default every TPA runs in its own isolated segment.
* **Shared segments by choice.** An owner can place several of their TPAs in a shared segment when they need to work together, so several models or agents can cooperate quickly on a complex task.
* **Separation by policy.** TPAs that must never interact sit in segments that cannot see each other. This stops AI models or agents from combining in unsafe ways through orchestration.
* **Validator controls.** With network policies, the 1,000 validators can quarantine any TPA into isolation at any time, block traffic between TPAs, or isolate a whole group, for example every TPA running a frontier Stage model.
* **Datacenters enforce.** Bonded datacenters are contractually required to apply the chain's network decisions.

{% hint style="info" %}
**What isolation cannot do.** For a TPA that connects to an outside AI provider's API, a quarantine cuts TPAI's access to that model. It cannot stop the provider's model itself **\[Analysis]**.
{% endhint %}

## TPA subscription

On TPAI Chain every TPA pays a flat subscription of **$10 a month**, priced in US dollars and paid in TPC, whatever the TPA cost **\[Founder statement]**. Like a domain name, a TPA is kept by paying for it.

**When it starts.** No subscription is charged on Ethereum. It starts when TPAs move to TPAI Chain at mainnet on 11 October 2027. TPAs bought before 29 March 2027 get **12 months free** after mainnet, so their first payment is due in October 2028; after that every TPA pays the same $10 a month **\[Founder statement]**.

| Stage          | What happens                                                                                                                                                                        |
| -------------- | ----------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- |
| Missed payment | A grace period, during which the owner can pay what is owed                                                                                                                         |
| Hold           | The TPA is put on hold; the owner can recover it by paying what is owed plus a recovery fee                                                                                         |
| Lost           | If it is never recovered, the owner loses the TPA and everything bound to it (fine-tuned models, agents and data), and its address returns to the pool after the cooling-off period |

The owner's wallet and TPC are not affected, since a TPA is not a wallet. The exact grace and hold periods and the recovery fee are **\[Open]**.

The subscription splits **50/50**: half to validators and half to Community & Ecosystem, which refills the pool that pays staking and Meta Energy rewards (see [TPCoin](https://docs.thetpa.ai/deai-tokenized-processing-architecture/tpcoin.md)) **\[Founder statement]**.

## Engineering notes

* An AWS VPC IPv4 block can only be /16 to /28, so a full /8 of pod addresses needs an overlay network (for example Calico or Cilium) or IPv6 **\[Analysis]**.
* Upstream Kubernetes guidance is about 150,000 pods per cluster, so 16.7 million TPA endpoints need many clusters joined into one overlay, routed across every datacenter. Designing and proving that routing, and planning IPv6 for growth, is **\[Open]**.
* Part of the GPU pool will run in Stage's own AI compute datacenters in Canada; see [TPAI Chain](https://docs.thetpa.ai/deai-tokenized-processing-architecture/tpai-chain.md) **\[Planned]**.
