Solutions
Your own inference endpoint. Your own limits.
Teams of two to twenty shipping agents on frontier models, hitting a usage ceiling mid-task, with margin set by someone else’s price list.
Why now
Your AI capacity is set by someone else’s dashboard.
The ceiling moved again
On 29 August 2026 Anthropic announced that Claude Code’s weekly limits would reset on 14 September 2026, a 17% reduction from the temporarily boosted level then in force. Teams that build on a metered frontier plan build on a number they do not control. [1]
SourceAnthropic plan-limit change announced August 29, 2026, effective September 14, 2026. The full citation and link are in the footer notes.
What it means for you
If your agents run on a metered frontier plan, your capacity is a line in someone else’s pricing page, and it can move during your launch.
What waiting costs
A limit cut in the middle of a launch is an outage you cannot escalate. Every prompt that leaves your network is also data you no longer control.
Why Triple X
Triple X gives you a private inference endpoint with limits you set, on hosted GPUs or your own, with Sealed for the prompts no operator may read. Standard APIs, so the agent code does not change.
In plain English
If your product runs on someone else’s AI quota, your roadmap does too. Triple X lets you run the models you choose, at the capacity you choose, inside a network you own, and prove no one read the prompts.
What you get
Four things, each a product you can open.
A private inference endpoint
An open model you choose, answering inside your private cloud. Prompts, outputs and training data stay put.
OpenKubernetes for the agent runtime
Standard clusters, up in minutes, with an MCP server so the agent can run its own cloud.
OpenSealed and a receipt per run
For governed agents: data unreadable in use, and a record of what ran.
OpenHourly, not per token
GPU capacity quoted in writing. Measure cost per million tokens against the API bill in a pilot.
Open
Which way in
Start where it fits. Move later without a new vendor.
Triple X Cloud nodes
No hardware of your own. Join the waitlist and start when your seat opens.
Your own devices
Triple X connects your racks or devices to a private cloud of your own.
Both
Sensitive work on your devices, the rest on ours, one private cloud.
Objections you will hear
What someone in the room will say.
Self-hosting is a full-time job.
The platform runs the nodes, the patching and the failover. You pick the model and call an endpoint. The difference from an API is who owns the box and who sets the limit.
Open models are behind the frontier.
For most agent steps, tool calls, classification, extraction and routing, not enough to matter. Keep the frontier for what needs it.
What about our training data and prompts?
They stay in your private cloud. Sealed adds a hardware boundary so even the operator cannot read them in use. Nobody trains on what nobody can see.
Is this cheaper?
At agent volume, owned capacity billed hourly usually is. The pilot measures it on your workload; we do not quote a percentage.
What yes looks like
The list a buyer brings to the table.
- Their own open-weight model, no usage ceiling set by a vendor
- A per-token cost measured on their own workload
- Prompts, outputs and training data never leave their private cloud
- A 60-day pilot with a pass mark they set
- A frontier model kept for the hard cases
- No exit fee if the numbers do not hold
Get started
Measure tokens per dollar on your own workload.
Ten nodes, one open model, 60 days.
- 01
Join the waitlist.
Tell us what you want to run and where. No credit card.
TakesTwo minutes
- 02
Run the pilot.
Ten nodes, 60 days, a pass mark you set in writing.
Takes60 days
- 03
Decide on day 60.
Extend, or leave at the transfer rate and nothing else.
WhoYou
We never ask for your root credentials, API keys or bank details by email.
The first 60 days
A pilot with a pass mark you set.
Ten nodes, one open model, one agent, and a per-token number the team writes down before the first call.
Week 1
Your private cloud, cluster, written GPU quote
Create your private cloud and a cluster for the agent runtime. Get the GPU quote in writing and load one open model on a private endpoint.
Weeks 2 to 4
One agent's traffic, counted
Route one agent's tool calls, classification and extraction to your endpoint and count every token. The pass mark is the team's: cost per million tokens, p95 latency, and no prompt or output outside your private cloud.
Day 30
Beside the API invoice
Set the measured cost per million tokens beside the API invoice for the same month. Keep the frontier model for the calls that failed the quality bar.
Day 60
Decision
Rent more capacity, add Sealed and receipts for governed agents, or stop with nothing owed beyond the transfer rate.
Questions
What buyers ask before they sign.
In the order they come up. Every answer ends with where to check it.
10 answers
All questionsCan we point the agent at an endpoint and keep the rest of our stack?
Yes. You load an open model of your choice and call an HTTPS endpoint inside your private cloud. The agent runtime runs on standard Kubernetes. A CLI, SDKs in TypeScript, Python and Go, and an MCP server let the agent provision its own capacity.
Proofthe Developers page and the AI and GPU page.
What does a GPU-hour cost?
On quote, in writing. The quote names the operator, the facility, the card class, the count, the availability window, the price per GPU-hour and the minimum term. Provisioning time and hardware downtime are not billed, and that is written into the quote.
Proofthe GPU answer on the Pricing page and the written quote for your private cloud.
Is it cheaper than the API per million tokens?
At agent volume, owned capacity billed by the hour usually is. Triple X quotes no percentage. The pilot counts tokens on your workload for a month and sets the number beside your API invoice.
Proofyour pilot's token count against your own API invoice for the same month.
Who sets our limit?
You do, by the capacity you rent. There is no weekly cap that a vendor resets, as happened on September 14, 2026 with a 17% reduction. Account ceilings on machines and clusters exist and are raised on request in 1 to 2 business days.
Proofthe August 29, 2026 limits announcement and the account limits on the Pricing page.[1]
Do our prompts, outputs and training data ever train someone else's model?
They stay in your private cloud. The platform underneath sees that a run happened, how large it was and what it cost. Sealed keeps prompts and data unreadable in use, including to the operator. Nobody trains on what nobody can see.
Proofthe access table on the Security page and the SOC 2 Type II report under NDA.[2]
What is the receipt per run, and who reads it?
A record of what ran, where, and that it was sealed, written after each run. For governed agents it goes to your customer's auditor or your own incident review. It is a record of what the agent did, not a log of what it said.
Proofthe four steps and the honest limits on the Sealed page.
What happens when the GPU node fails mid-job?
Work restarts on another node inside your private cloud under the policy you set. Under the GPU quote, hardware downtime is not billed. Keep state in a managed database or object storage so the restart picks up where it left off.
Proofthe failover rules on the Reliability page and the downtime clause in the GPU quote.
Can we run on GPUs we already own?
Yes. Your devices join your private cloud and the same endpoint, console and API apply. Triple X quotes the platform charge in writing. Mix them with hosted capacity for bursts, in one private cloud.
Proofthe Sovereign page and a written quote from Triple X for your private cloud.
What if the numbers do not hold at day 60?
You stop. No exit fee, no minimum term beyond what a GPU quote states. Leaving costs the outbound transfer rate. The weights, the prompts and the data were yours throughout.
Proofthe no-minimum-term rule in the published pricing and the no-exit-fee clause in the contract.[3]
What is not included?
No frontier model and no model weights: you choose and load the open model. No evaluation suite. GPU availability is on quote, with a window. Sealed is available for inference today; other Sealed workloads are arranged with an engineer.
Proofthe AI and GPU page and the honest limits on the Sealed page.
No questions under this term yet. Ask Triple X.