Route inference to the cheapest, coldest power.
ColdShift is an API gateway that sends each request to the datacenter with the cheapest off-peak energy and lowest cooling load.
client = OpenAI(base_url="https://api.coldshift.net/v1")A kilowatt hour has no fixed price.
Energy price and cooling cost move by region, hour, and weather. Inference is one of the few workloads that can follow them.
Integration is one changed line.
Point your client at the ColdShift endpoint and change nothing else.
Code diff with one changed line: the provider base URL is replaced with the ColdShift endpoint. Tabs show the same change in Python, TypeScript, and cURL.
from openai import OpenAI client = OpenAI( base_url="https://api.openai.com/v1", base_url="https://api.coldshift.net/v1", api_key=os.environ["API_KEY"],)Routing never blocks the first token.
The route resolves before inference begins, off the token path. Time to first token is unaffected.
Simulated three stage router pipeline. Telemetry streams feed a scoring stage where candidate regions re-rank as conditions change, and the winning region receives the dispatched request.
Live feeds of regional weather, grid stress, carbon intensity, energy prices, and spot compute prices.
Candidate regions are ranked on delivered cost per token, cooling load, and capacity headroom.
The request ships to the winning region with your provider settings unchanged.
An earth-scale load balancer for AI inference.
Where the margin comes from.
ColdShift buys compute where energy is off-peak and cooling is cheap, then resells it just under standard rates. You keep the discount and ColdShift keeps the spread.
Modeled from observed regional energy price and cooling load spreads. Not measured from ColdShift production traffic.
Datacenter electricity is growing four times faster than everything else.
Every point of growth widens the spread between the cheapest hour on the grid and the most expensive one.
Global datacenter electricity in 2024, about 1.5 percent of world electricity.
Source: IEA, Energy and AIBase case for 2030, about 3 percent of world electricity. About 1,200 TWh by 2035.
Source: IEA, Energy and AIGrowth to 2030, more than four times faster than all other sectors combined. AI-optimized datacenter demand more than quadruples by 2030.
Source: IEA, Energy and AIUS datacenters used about 4.4 percent of US electricity in 2023 (176 TWh). Projected 6.7 to 12 percent by 2028 (325 to 580 TWh).
Source: LBNL, 2024 United States Data Center Energy Usage ReportObjections
coldshift ask "Does this add latency?"
No. The route resolves before inference begins, off the token path. The decision budget is 8 ms.TARGET
coldshift ask "What about data residency?"
You pin an allowed region list per API key and requests never leave it. The enforcement tooling ships in the batch.TODO(founder)
coldshift ask "What if a region goes dark?"
The scorer prices a dark region as unavailable and traffic moves to the next candidate. Automatic failover is in-batch work.TODO(founder)
coldshift ask "Which models are supported?"
Anything served behind an OpenAI-compatible endpoint. A named provider list goes up once contracts are signed, not before.TODO(founder)
coldshift ask "How is this different from a proxy or load balancer?"
A load balancer spreads traffic for server health inside one region. ColdShift prices energy, cooling, and spot compute across regions, then moves the workload to the cheapest one.
coldshift ask "What happens to my provider contracts?"
Bring your own keys and ColdShift routes inside the regions your contracts already cover. Pooled capacity bought through ColdShift is in-batch work.TODO(founder)
coldshift ask "When can I use it?"
The backend gets built in the batch. Leave an email and you get the first endpoint.
Join the waitlist