---
title: "How Float Runs an AI Energy Company on a 3-Person Team with Tiger Data"
published: 2026-06-30T07:00:05.000-04:00
updated: 2026-06-30T07:00:04.000-04:00
excerpt: "How Danish startup Float hit 99.3% compression on 1Hz smart meter data with Tiger Data, powering appliance-level energy analytics for hundreds of homes."
tags: Dev Q&A
authors: Team Tiger Data
---

> **TimescaleDB is now Tiger Data.**

_Danish startup achieves 99.3% compression on 1Hz smart meter data, powering real-time appliance-level energy analytics for hundreds of homes._

* * *

[_Float_](https://float.energy) _is a Danish AI and energy startup that collects 1Hz smart meter data from hundreds of homes and disaggregates it into per-appliance consumption using a proprietary ML model. The entire system depends on one architectural bet: that compression on their time-series database would be high enough to make the storage economics work at scale. Co-founders Jens Brandt Nellegaard (CEO) and Victor Grabow (CTO) share how they evaluated every major time-series database, why 90% compression was the hard floor for viability, and what happened when they hit 99.3% on Tiger Data._

## About Float

Most people have no idea what their individual appliances cost to run. In Denmark, most energy providers send a monthly PDF bill, and that is the entire customer experience. The market has seen no meaningful innovation in 10 to 15 years. Fraud and a lack of transparency have been so widespread that the Danish government introduced new consumer protection regulation effective January 2026.

The underlying data to solve this problem actually exists. Most European smart meters have a standardized consumer interface that can output total household electrical load down to one-second resolution. But total load is not very interesting on its own. What consumers need is a breakdown by individual appliance - to understand which one is wasting electricity, which one is running at peak price, which one is approaching the capacity limit of a fuse. Academic research on energy disaggregation goes back to 1992. Nobody had solved it in a commercially scalable way.

Float built three things to close this gap: a proprietary hardware module that plugs into the consumer port on a European smart meter, a signal processing and neural net pipeline that classifies appliance-level consumption from the raw waveform, and a consumer-facing app with a proactive AI energy agent. The system collects roughly 15 measurements per second per household, each at 1Hz resolution. One second is the hard floor - Victor explains: _"At one-minute resolution, the model would break, because what we track are the changes. If there are too many appliances turning on within the same minute, it would be very hard to differentiate them."_

Jens Brandt Nellegaard and Victor Grabow co-founded Float in 2022. After nearly three years of R&D, they achieved a proof of concept in December 2024, secured their energy provider license in December 2025, and are now rolling out a private beta to 350 pre-vetted customers. The company has just three people.

## The Challenge

Float started on Azure managed Postgres with the TimescaleDB extension. The team had Azure credits early, so it made sense at the time. But the Apache version available on Azure did not include compression, and that turned out to be a dealbreaker.

Every customer generates roughly 15 measurements per second. Float samples voltage, frequency, and total load across each phase entering the home. At 1,000 homes, that is 15,000 data points per second, continuously. Without compression, the storage cost alone would break their business model. “_We are an energy company with a flat-rate subscription fee,”_ says Jens. _“We pass through the spot market price one-to-one with no markup. If storage cost per user exceeds what the subscription supports, the economics collapse._” 

> _We also tried InfluxDB before settling on TimescaleDB. We ran into ingestion issues, and we needed SQL,_” says Victor. “_When you are a three-person team building an asset-centric microservice platform, you cannot afford a database that requires a proprietary query language and limits how you join and query data across domains._

On top of the storage problem, Float needed continuous aggregates. The Danish DSO delivers settlement data at 15-minute resolution. Float collects data every second. To generate a live energy bill and compare it against the grid operator's numbers, the system needs to aggregate raw 1Hz data down to 15-minute windows constantly. On managed Postgres without TimescaleDB's full feature set, that meant writing and maintaining batch jobs - more infrastructure overhead for a team that was already stretched thin across hardware, ML, and a licensed energy company.

## Why Tiger Data

> _We tested pretty much every time-series database on the market. We think Tiger Data is the best solution for our use case. - Victor Grabow, CTO, Float_

The team researched time-series solutions extensively and discovered TimescaleDB through the compression and continuous aggregate features. The compression was compelling enough that it was just a matter of time before they needed the full capability. When they found Tiger Data - the company behind TimescaleDB - the managed cloud service made the path clear.

Two features drove the decision. First, **_Compression_**. The team had modeled the unit economics and needed at least 90% compression on the time-series data for the subscription model to work. Anything below that and storage costs per user would exceed what the flat-rate fee could support. Second, **_Continuous Aggregates_** - materialized views that update incrementally as new data arrives. Float runs aggregations constantly, converting 1Hz readings to 15-minute settlement windows, calculating threshold-based alerts on voltage and frequency, detecting outages, and triggering duration-based notifications like flagging an oven that has been running for four hours. Continuous aggregates handle all of this without batch jobs or scheduled pipelines.

> “_We chose Tiger Cloud, the fully managed service on Azure, because it was a question of speed,_” Victor explains. “_We needed to get up and running fast and offload infrastructure management entirely._ [_Encore_](https://encore.dev)_, our DevOps platform, provides ephemeral environments on Google Cloud, and Tiger Cloud's database branching fits naturally into that workflow._”

## The Float Energy Data Stack

Data starts at the smart meter. Float’s IoT module plugs into the standardized consumer interface port and captures voltage, frequency, and total load across each phase at 1Hz resolution. The module sends readings to Azure IoT Hub, which the team kept from the original Azure setup as a stable ingestion endpoint for all devices.

From there, a bridge connector forwards the stream into Google Cloud, where Encore deploys Float’s microservices. The team moved off Azure Event Hub eventually because it was expensive. Google Cloud streaming services handle the same throughput at lower cost. The streaming layer batches incoming measurements from all households every second and inserts them per batch into Tiger Data.

Tiger Data stores the raw 1Hz time-series readings and runs continuous aggregates for threshold-based monitoring: voltage spikes, frequency changes, mean and max calculations, outage detection, and duration-based appliance alerts. All raw data is retained for ML training purposes through the private beta phase, with tiered storage planned as the fleet scales.

The Float app reads processed data to show customers their real-time energy breakdown per appliance. New customers see total wattage immediately on connection. Appliance-level breakdown takes roughly three to four weeks as the model trains on their home's specific patterns. The agentic orchestration layer on top handles billing, onboarding, customer service, and proactive notifications - flagging forgotten ovens and irons, inefficient appliances, and dangerous load conditions approaching fuse limits.

![Float's data architecture: 1Hz readings flow from the IoT module through Azure IoT Hub and Google Cloud streaming into Tiger Data, which serves the ML pipeline, consumer app, and agentic platform.](https://storage.ghost.io/c/6b/cb/6bcb39cf-9421-4bd1-9c9d-fa7b6755ba0e/content/images/2026/06/float-tiger-data-architecture-diagram.svg)

__Float's data architecture: 1Hz readings flow from the IoT module through Azure IoT Hub and Google Cloud streaming into Tiger Data, which serves the ML pipeline, consumer app, and agentic platform.__

## What Compression Enabled

On Tiger Data, Float is seeing 99.3% compression on its time-series data. Victor puts it directly: 

> _Compression needed to be in the high nineties range to not break our business model. So that was a great outcome. - Victor Grabow, CTO, Float_

That number unlocked three things that would not have been possible at lower compression ratios.

### The Business Model Works

At 15,000 data points per second across 1,000 homes, uncompressed storage would generate terabytes of raw time-series data per year. Float passes through the spot market electricity price to customers at cost with no markup. Revenue comes from a flat-rate subscription fee. If storage cost per user climbs above what that fee can support, the entire model collapses. At 99.3% compression, it does not. The subscription covers infrastructure with margin to spare, and that margin holds as the fleet scales.

### Full Data Retention for ML Training

Float's disaggregation model needs weeks of 1Hz training data per household to learn each home's specific appliance signatures. At lower compression, the team would face a choice: retain all raw data for model training or keep storage costs viable. At 99.3%, they retain everything. All raw 1Hz readings from the entire private beta fleet are available for the ML pipeline, with tiered storage planned only as the fleet scales past the beta phase.

### Real-Time Billing Without Batch Infrastructure

The Danish grid operates on 15-minute settlement windows. Float collects data every second. Continuous aggregates bridge that gap, converting 1Hz readings into the 15-minute intervals the DSO requires for bill reconciliation. Danish energy prices swing up to 80% between peak and off-peak hours, which makes the freshness of the aggregation directly valuable to customers. Because continuous aggregates update incrementally as new data arrives, Float's live energy bill is always current, i.e., no scheduled batch jobs, no pipeline maintenance, no lag.

### A Three-Person Team Running a Licensed Energy Company

Float holds an energy provider license in Denmark. That means billing, customer service, onboarding, regulatory compliance - operational overhead that traditional energy companies staff with dozens of people. Tiger Cloud's managed infrastructure is part of what makes this possible. The team does not manage database operations, storage provisioning, or aggregation pipelines. That overhead is handled. When asked about team size, Jens's answer was simple: _"Three. We have three people... and an army of agents. This is the future."_

## Looking Ahead

Float is targeting 1,000 additional private beta customers within the next 12 months, with a seed round, two additional hardware variants for full Danish grid coverage, and expansion into two more countries.

The next major integration is EV charging - starting with Tesla's telemetry API, enabling smart charging during cheap price windows. The bigger thesis is that a fleet of homes measured at 1Hz resolution can trade power on the spot market more efficiently than any energy company operating at 15-minute resolution with a 24 to 48-hour delay. As Jens puts it: _"Ultimately we are trying to make the home not a burden for the grid, but a partner of the grid."_

The architecture decision that compounds as Float scales is not the compression ratio itself. It is that everything runs on a single Tiger Data instance: the raw 1Hz readings, the continuous aggregates for billing, the training data for the ML pipeline, the anomaly detection queries. No split architecture to maintain, no query paths to reconcile as the fleet grows from 350 homes to 1,000 and beyond. The data model does not change - it just gets bigger.