AI Quotas, Single-Cycle Usage, and Engineering Lessons from Ultracold Atom Experiments

For developers and researchers, two recent developments may seem unrelated, yet they address the same question: how can long-running tasks remain stable when resources are limited and operating conditions are demanding? One concerns quota policies for AI APIs; the other concerns an ultracold atom experiment aboard a space station. This article explains the technical principles behind both and presents engineering practices that can be reused directly.

1. Dual-Cycle Quotas vs. Single-Cycle Quotas

Some AI platforms previously used a dual-cycle quota model: a short-term limit, such as five hours, combined with a weekly quota. The short-term limit is intended to prevent a single account from sending a large number of requests in a short period and to reduce pressure on the service.

This design is reasonable for ordinary interactive use, but it can create significant problems for batch workloads. Large-scale document summarization, RAG index construction, and long-context fine-tuning data generation may need to run continuously for several hours. Once the short-term quota is exhausted, requests are rejected even when substantial weekly quota remains, forcing the job to wait for the short cycle to reset.

When the short-term restriction is removed, usage is counted only against the weekly quota. Developers can run batch jobs when the system is idle without artificially slowing them down to avoid exhausting a short-term allowance.

Two points require attention. First, platform policies can change, so the current documentation for the relevant account type should be checked before implementation. Second, a single-cycle quota does not mean that concurrency is unlimited. Short-term throttling, including HTTP 429 responses, may still occur.

Dimension	Dual-cycle model	Single-cycle model
Limiting granularity	Short-term limit plus weekly quota	Weekly quota only
Interruption of long-running jobs	The job may stop when the short-term quota is exhausted	No short-term interruption before the weekly quota is exhausted
Sources of throttling	Short-term quota and concurrency limits	Mainly concurrency throttling
Retry strategy	Must distinguish between two quota mechanisms	Must distinguish throttling from quota exhaustion

The applicable account types and reset schedules should be confirmed in the platform documentation. The table above describes only the difference in mechanisms.

2. Handle the Two Types of HTTP 429 Separately

An HTTP 429 response can represent two different conditions.

The first is temporary rate limiting. Waiting and retrying will often succeed. The second is an exhausted weekly quota. Retrying in that case has no practical value and only wastes time and request attempts. If the two cases are not separated, a retry framework may repeatedly call an endpoint that is guaranteed to fail.

The following example separates the conditions. When the quota is exhausted, it raises a non-retryable exception. Only temporary rate limits are passed to the exponential backoff logic.

import os
import requests
from tenacity import retry, retry_if_exception_type, stop_after_attempt, wait_exponential
class RateLimitError(Exception):
    """Temporary throttling; the request can be retried with backoff."""
class QuotaExhaustedError(Exception):
    """The weekly quota is exhausted; do not retry until the quota resets."""
class AIModelClient:
    def __init__(self, base_url: str):
        # Read the API key from an environment variable. Do not write it to code or logs.
        self.api_key = os.environ["AI_API_KEY"]
        self.base_url = base_url
    @retry(
        retry=retry_if_exception_type(RateLimitError),
        stop=stop_after_attempt(5),
        wait=wait_exponential(multiplier=1, min=4, max=60),
        reraise=True,
    )
    def send_request(self, payload: dict) -> dict:
        headers = {"Authorization": f"Bearer {self.api_key}"}
        resp = requests.post(self.base_url, json=payload, headers=headers, timeout=60)
        if resp.status_code == 429:
            # The header name is illustrative only. Follow the documentation for the platform in use.
            remaining = resp.headers.get("X-Week-Quota-Remaining")
            if remaining == "0":
                raise QuotaExhaustedError("Weekly quota exhausted; stop sending requests")
            raise RateLimitError("Request throttled; retry with backoff")
        resp.raise_for_status()
        return resp.json()

Several details are important:

• QuotaExhaustedError is not included in the retry condition, so it is immediately propagated to the caller. The caller can record a checkpoint and terminate the current batch.
• The timeout parameter prevents a request from hanging indefinitely.
• The response header and endpoint in this example are illustrative. They have not been validated against a specific platform and must be adjusted according to the actual API documentation.

Batch processing also requires resumability. After each batch completes, write the IDs of successfully processed samples to a local file or database. When the job restarts, skip those IDs and process only the remaining items. This ensures that completed work is not lost even when the quota limit is reached.

For monitoring, collect at least these four metrics:

• Remaining weekly quota: to determine whether the planned workload can finish within the current cycle.
• 429 response ratio: a sustained increase may indicate that concurrency is too high.
• Average retry count: to expose hidden throttling problems.
• Checkpoint position and completion progress: to estimate the remaining runtime.

Alerting when these metrics exceed defined thresholds is more effective than investigating only after the job has failed.

3. Ultracold Atom Experiments: Starting at 9.7 Picokelvin

A Chinese research team conducted an ultracold atom physics experiment using an experimental facility in the Mengtian module of a space station. According to the team, the experiment achieved an equivalent temperature of 9.7 picokelvin (pK) and maintained observations for approximately seven seconds.

One picokelvin is one trillionth of a kelvin. Room temperature is approximately 300 K, making the ratio between the two roughly 30 trillion to one. Based on this result, the team described it as the lowest reported temperature for artificially produced matter at the time.

The measurement method needs to be clarified. An ultracold atom cloud cannot be measured directly with a conventional thermometer. Its equivalent temperature is typically calculated from the width of the atomic velocity distribution, or momentum distribution. The reported value therefore depends on the measurement method and model assumptions.

The exact measurement conditions, observation window, and uncertainty range should be verified against the team’s published paper. Until the original paper has been checked, readers should understand 9.7 pK as the result disclosed by the team rather than as an independently verified universal record.

Comparing Cooling Methods

Different trapping and cooling techniques involve different trade-offs:

Method	Trapping principle	Main characteristics	Main limitations
Magneto-optical trap	Magnetic-field gradients combined with laser cooling	High capture efficiency; a common starting point for cold-atom experiments	The atom cloud falls under gravity, and the magnetic field can affect some energy levels
Magnetic trap	Confinement through magnetic-field gradients	A relatively deep potential well, suitable for further cooling	Strong magnetic fields can interfere with atomic energy levels
Optical trap, or dipole trap	Optical dipole forces generated by a laser	Does not rely on a strong magnetic field; the potential can be adjusted precisely	The potential well is relatively shallow, making it difficult to maintain the atom cloud for long periods on Earth

An all-optical trap is a type of optical trap. Its main advantage is that it avoids strong magnetic fields, which can help achieve lower temperatures.

Why Microgravity Matters

On Earth, gravity causes the atom cloud to fall. The optical trap must therefore be deep enough to support it, but a deeper trap can introduce greater optical disturbance. In a microgravity environment, the atom cloud is not forced downward in the same way, making it possible to use a shallower optical trap and maintain the cloud under less demanding conditions.

This illustrates a broader engineering principle: removing a dominant environmental constraint can reduce the control effort required by the system. In the AI quota case, removing a short-term quota reduces unnecessary scheduling constraints for long-running batches. In the ultracold atom case, reducing the effect of gravity makes it easier to maintain an extremely cold atomic cloud.