TokenWatt is operated by Teampulse Solutions LLC, a Delaware limited liability company ("TokenWatt", "we", "us"). This policy explains what we collect and what we don't. Use of our services is also governed by our terms of service.
Summary
| What | Policy |
|---|---|
| Prompts and completions sent to our inference API | Not stored. Not logged. Not used for training. Processed in memory only while the request is served. Details: API data policy. |
| Request metadata (no content) | Time, model, token counts, cost, status and latency, kept for billing, capacity accounting and troubleshooting |
| Data sold or shared for advertising | Never |
| TokenWatt Bench (the measurement tool) | Runs on your machines and sends nothing to us |
| tokenwatt-check (open-source node checker) | Sends nothing unless you add --share; then an anonymized report without hostname, IDs or IP address |
| Node-evaluation requests | The details you enter in the form, kept to reply and to run the evaluation; deleted on request; IP address not stored |
| TokenWatt Monitor | Your email, and the GPU telemetry your nodes send (power, clocks, temperature, utilization, PCIe, ECC, detected events); kept while your account is active |
| This website | Aggregate, cookieless visit statistics; no tracking cookies, no advertising; standard hosting logs |
| Where inference runs | The United States (RunPod Secure Cloud, data center US-NE-1) |
API data policy
This section covers requests to TokenWatt's inference API (https://api.tokenwatt.io/v1), whether you call it directly or through a marketplace or router.
| Topic | Policy |
|---|---|
| Provider | TokenWatt, operated by Teampulse Solutions LLC, a Delaware limited liability company |
| Contact | boshen@tokenwatt.io or hello@tokenwatt.io. Postal address available on request. |
| Prompts and completions | Not stored and not logged (zero data retention). Processed in memory only while the request is served. |
| Training | None. We don't use prompts or completions to train, fine-tune or evaluate models. |
| Where inference runs | United States: GPU servers in RunPod's Secure Cloud, data center US-NE-1 |
| Where metadata is stored | On the serving node (United States) and in Cloudflare R2 |
| Metadata retention | Per-request records on the serving node: deleted after 30 days. Daily usage totals per account (no request content): kept for billing. |
| Sub-processors | RunPod, Cloudflare, Vercel, Stripe, Resend (below) |
Content. The text you send (prompts, messages, system instructions, tool definitions) and the text we generate (completions and tool calls) are processed in memory to serve your request. We don't write them to disk, to logs or to any database, and we don't use them to train, fine-tune or evaluate models. The inference engine (vLLM) runs with request and output logging off, and our serving gateway logs metadata only. To speed up prompts that share a beginning, the engine's prefix cache keeps derived attention state (not text) from recent requests in memory. It is never written to disk, is overwritten as memory is needed for new requests, and is lost on restart.
Metadata. For each request we record:
- a request ID and timestamps;
- the model requested;
- the HTTP status and, for failures on our side, the error type;
- token counts (input, cached input and output) and the resulting cost;
- latency (time to first token and total time) and the response size;
- the channel the request came through (an account key, a marketplace, or our own health checks);
- the user-agent string of the calling software (and the referer header, if one is sent).
Requests reach our gateway through a Cloudflare tunnel, so your IP address is not in our logs. Usage on account keys is also added to that account's daily totals (requests, token counts and cost per model).
Where metadata is stored and for how long. Per-request records stay on the serving node and are deleted after 30 days; the node's logs are rotated and also deleted after 30 days. Account records and daily usage totals (requests, token counts and cost per model, no request content) are kept in Cloudflare R2 for billing and accounting. Aggregated statistics that identify no customer (for example "requests per hour" or "median time to first token") may be published in our reports.
Where inference runs. Only in the United States, on GPU servers we rent in RunPod's Secure Cloud (currently data center US-NE-1), running our own serving stack (vLLM behind our gateway). Requests reach the server through Cloudflare (TLS and Cloudflare Tunnel), entering Cloudflare's network at the location nearest the caller. Content passes through Cloudflare in transit only; we don't store it there.
| Sub-processor | What it does for the API | Data it handles |
|---|---|---|
| RunPod | GPU servers (Secure Cloud, United States) where inference runs | Request content in memory while it is served; per-request metadata on the server's disk |
| Cloudflare | Network, TLS and tunnel for api.tokenwatt.io; R2 object storage | Request content in transit; account records and usage totals at rest in R2 |
| Vercel | Hosts tokenwatt.io and the account service (sign-up, dashboard, key and usage sync) | Account data and usage totals; no request content |
| Stripe | Card payments for prepaid credit | Payment details; we receive only the amount, its status and an identifier |
| Resend | Account email: your API key and dashboard link at sign-up, and our own notices of sign-ups and top-ups | Your email address and the details you give at sign-up; no request content |
Marketplaces. If you reach us through a marketplace or router, its own privacy policy also applies to the data it handles.
Self-serve API accounts. If you sign up on our website we keep your email, the name, company and intended use you give us, hashes of your API key and dashboard link (we can't recover either from storage), your credit balance, purchases and daily usage totals (requests, token counts and cost per model), in Cloudflare R2. We use these to run your account, bill you and contact you about the service. Payments are handled by Stripe, which receives your card details under its own privacy policy; we receive only the payment amount, its status and an identifier. Account records are kept while your account is active and for up to 7 years after the last payment where tax and accounting rules require it; ask us to delete an account at hello@tokenwatt.io.
TokenWatt Bench
Bench is software you run on your own machines. It sends requests to your inference endpoint and reads hardware telemetry (GPU and node power, clocks, temperatures). It's read-only, doesn't load model weights and doesn't store prompt or completion text. It sends nothing to TokenWatt. Results stay on your machines unless you choose to share or publish them.
tokenwatt-check
tokenwatt-check is our open-source node checker. It runs on your machine and sends nothing unless you add --share. With --share, it sends one report to tokenwatt.io/api/check-report, and we keep only:
- the GPU model, driver version, memory size, power and clock limits, PCIe link and memory-error counters;
- the measurements: compute, memory bandwidth, clocks and power during the test, disk and download speed;
- CPU count, RAM size and CPU model;
- the provider you name with
--provider, if any, and the PASS/WARN/FAIL results.
We don't keep hostnames, GPU UUIDs, file paths or your IP address, and we can't link a report to you. We publish shared reports only as aggregate statistics per GPU model, for example the distribution of clocks under load.
TokenWatt Monitor
When you sign up for Monitor, we store your email (and name, company and GPU description if you give them) and create an upload key and a private dashboard link. We keep only hashes of the key and the link, so we can't read or resend them.
- What your nodes send: every five minutes, per GPU: model name, power limit and draw, SM clock, temperature, utilization, PCIe link, uncorrected ECC error count, and the events the checker detected. Also the checker's version, a node name only if you set one, and an anonymous node id (a hash of the hostname, not the hostname). Nothing about processes, files, users or environment variables.
- Why: to show your dashboard and email you alerts. We may use telemetry, aggregated and anonymized so that no account, node or provider can be identified, in our published research on how rented GPUs behave.
- Where: in private storage at Cloudflare (R2). Alert emails go out through Resend.
- How long: while your account is active, and deleted within 30 days of a request to hello@tokenwatt.io.
Node-evaluation requests
When you request an evaluation on the evaluation page, we store what you enter (name, email, company, role, GPUs, where they run, goal, the option you're interested in, models and your message), when you sent it, your browser's user-agent string and, if present, the campaign tag of the link you came from. We don't store your IP address.
- Why: to reply, scope and run the evaluation you asked for. We don't add you to mailing lists or share your details with anyone outside TokenWatt.
- Where: in private storage at our hosting provider (Vercel). A copy is emailed to our team through our email provider (Resend) so we can reply quickly.
- How long: while we're in contact and for up to 24 months after, unless you ask us to delete it sooner. Write to hello@tokenwatt.io to see or delete your request.
- Self-run evaluations: if you run our evaluation script yourself, it shows you every file before asking to upload. The archive (GPU readings including GPU UUIDs and serial numbers, benchmark results and logs, the hostname, CPU/RAM/disk sizes and software versions; never your files, credentials or environment variables) goes to private storage at Cloudflare (R2) through a one-time upload link that expires after seven days. We keep it, and the report made from it, on the same terms as your request. The report is emailed to you through Resend.
- Evaluation results belong to you. We never publish them without your written permission.
This website
- Visit statistics without cookies. We use Vercel Web Analytics to count page views, referring sites, countries and device types in aggregate. It sets no cookies, doesn't build profiles and doesn't follow you across sites; visitors are distinguished by a hash that changes daily and isn't stored. We run no advertising pixels and no other trackers.
- Local storage: your light/dark theme choice is saved in your browser and never sent to us.
- Hosting logs: our hosting provider (Vercel) processes standard request logs (such as IP address, user agent and pages requested) to serve and protect the site, under its own privacy policy.
- Fonts are loaded from Google Fonts, which receives standard request information from your browser.
- Email: if you write to us, we keep the conversation to reply and follow up. We don't add you to mailing lists without asking.
Published reports
Public reports and lab notes contain measurements from hardware we rent or from partners who agreed to publication. We never publish a customer's results, configuration or identity without written permission.
Security
Traffic to our endpoints is encrypted in transit (TLS). Access to infrastructure and to metadata is limited to the people who operate the service, with credentials kept out of source code and logs.
Changes and contact
If we change this policy we'll update this page and the effective date above. For material changes affecting the inference API, we'll also notify marketplaces and direct customers before the change takes effect.
Questions or requests: Teampulse Solutions LLC (TokenWatt) · Delaware, USA · hello@tokenwatt.io · boshen@tokenwatt.io.