What Google Meridian Costs to Run In House in the First Year

The total cost of standing up and running Meridian internally in the first year is a payroll number with a small cloud bill attached. Google gives the library away under an open-source license, so no line on the budget carries a vendor quote, and that makes it harder to defend than a SaaS contract: a finance partner has no price list to check it against.

We built this budget from the outside in. Every rate comes from a public page you can open, and every quantity of hours is a labeled assumption you should overwrite with your own. On those inputs, a hypothetical first year totals $42,829 for a lean setup, $111,588 for a typical one and $227,787 for a heavy one. Labor is more than 97 percent of all three.

Is there a license fee for Meridian?

The Google Meridian marketing mix model costs nothing to license. The code is public in Google's GitHub repository under the Apache-2.0 license, and the README states the intent plainly: "Meridian is an MMM framework that enables advertisers to set up and run their own in-house models." We found no paid tier, seat price or usage fee on any Google page we read for this piece. Google also describes Meridian Studio as its new Cloud-powered enterprise platform for building and managing Meridian models; that page lists no price, so this budget covers the open-source library only. What you pay for is everything around the code: people who can assemble the history, fit a Bayesian model and explain it, a GPU to run the sampler, and the experiments that keep the model honest.

The four public rates behind every line

A wage understates what an employee costs, so we divide each one by 0.70 to get a loaded rate: $82.57 an hour from the Bureau of Labor Statistics median for data scientists, and $54.10 from the BLS median for market research analysts. The 70/30 split is the average for all private industry workers in the June 2026 compensation release from the BLS, so swap in your own benefits rate if finance has one.

How much of the in-house cost is data science?

Data science is the largest in-house cost of running Google's Meridian. In our typical scenario, hours billed at the data scientist rate come to $92,478 of the $111,588 total, about 83 percent, and three kinds of work sit inside that number.

The first is the pipeline. Google's data collection guide lists media data, media spend, control variables and a KPI as required inputs, names weekly data at the geo level as best practice, and gives its guidance on history: "As a general rule of thumb, historical data should be a minimum of two years' worth of weekly data for geo-level models and three years' of data for national-level models." Someone has to pull that history from every ad platform, reconcile it with finance's revenue figures, map it to regions and fill the gaps. The guide points to Google's MMM Data Platform for exposure data on Google channels, and every other channel is yours to collect. To keep the model to two wage lines, we price pipeline hours at the data scientist rate.

The second is modeling: choosing priors and controls, fitting, checking diagnostics and refreshing through the year. The third is experiment design and readout, covered below. Analysts carry the rest, gathering media plans, checking inputs and turning model output into a budget recommendation. If platform attribution is your current source of truth, expect some of those hours to go to explaining why the model and the dashboards disagree, which is the gap incrementality measures.

What does the internal platform cost in compute?

Meridian's internal platform cost, meaning the compute under the model, is the smallest line on the budget. Google's installation guide puts the hardware advice in two sentences: "We also recommend using a minimum of 1 GPU." and "This project has been tested on V100 and T4 GPU using 16 GB of RAM." A third note covers laptops: "There is no official GPU support for macOS." A team working on MacBooks should plan for a cloud instance.

Google Cloud's GPU price sheet lists the T4 at $0.35 per GPU hour and the V100 at $2.48, on demand in Iowa (us-central1). The same page says: "Each GPU adds to the cost of your instance in addition to the cost of the machine type." The VM, disk and storage therefore bill separately, and we have left them out because they depend on the machine you pick. Leaving them out changes little. Our high scenario runs 2,000 GPU hours on the V100, the pricier of the two cards Google tested, and spends $4,960, about 2 percent of that scenario's total. If the machine type tripled the hourly figure, compute would reach about 6 percent.

The GPU hours are our assumption. We found no published run-time figure we could turn into a yearly quantity, so the scenarios use 100, 500 and 2,000 hours. Replace them after your first few fits.

The lift tests that calibration needs

Meridian will run without experiments. Google's calibration guide describes the default priors as "moderately informative" and then makes the case for testing: "Incrementality experiments are perhaps the strongest basis for formulating your intuition, but translating experiment results into priors isn't a precise formula." Budget for tests anyway. A model nobody has checked against a causal result is hard to defend on the day it tells you to move money.

The test software is open source as well. Google describes its companion library this way: "Meridian GeoX is Google's global, open-source incrementality solution that lets you run transparent, cost-effective, and publisher-agnostic Geo experiments (GeoX)." The calibration guide adds that Meridian's CalibrationBuilder turns experiment results into calibrated ROI priors, with direct integration for Meridian GeoX.

That leaves two costs. One is labor: designing each test, monitoring it and reading the result, which we assume takes 60 data scientist hours per test, with no new tests in the low scenario, two in the typical and four in the high. The other is the media cost of the test, the one line we could not price from a public source. A holdback or go-dark design gives up the sales the paused spend would have driven in the test regions, less the spend you saved, while a heavy-up design adds spend. Both depend on your own revenue and on the very number the test exists to find, so the table leaves that row for you to fill. Our guide to running a quarterly geo-holdout test covers the design choices that set its size, and analyzing an incrementality result covers the readout.

One caution from Google's guide belongs in the plan: "There is no single formula to translate an experiment result into a prior." A test result is evidence to weigh, and the same judgment applies when you use tests to correct platform numbers with a calibration multiplier.

What does Google Meridian cost to run in the first year?

Google Meridian's cost to run in the first year, on the rates above and our assumed hours, lands between $42,829 and $227,787. All three scenarios are hypothetical. Low is a national model on data already in a warehouse, calibrated with past experiments. Typical is a geo-level model with quarterly refreshes and two new geo tests. High is a geo-level model across many channels with monthly refreshes and four tests.

The typical scenario is 1,470 labor hours, about 71 percent of one 2,080-hour work year. That is less than one hire, so in practice it is borrowed from people who already have a job to do, and part of the cost shows up as work that didn't happen elsewhere.

Wage level is the input most likely to move the total. BLS reports that the highest-paid 10 percent of data scientists earned more than $199,130 in May 2025. At that salary the loaded rate is about $137 an hour, which lifts every data science line by roughly 66 percent.

What the first year carries that later years shed

Run internally, Meridian's first-year cost includes work that later years shed: the pipeline is built once, the first model tends to take the most iterations, and the team learns the library on the clock. We have not modeled year two, because any figure for it would be another stack of assumptions. The table still shows where to look: pipeline hours should fall to maintenance, modeling hours should settle at your refresh cadence, and test costs continue for as long as you keep calibrating.

Finance will ask next for the alternative, an agency or vendor quote for the same model. Compare the two line by line, including who owns the data pipeline when a contract ends. If you want a second read of the hours first, our data and analytics team can review them with you, and the Attribution Doctor is a further resource on the attribution side of the same question. Whichever route you take, keep the shape of the table: the rates are public and dated, the hours are yours, and anyone who opens the same pages can recheck the budget.

Sources

• Google, Meridian repository on GitHub (license and README)
• Google, Install Meridian
• Google, Collect and organize your data
• Google, Calibrate treatment priors
• Google, Meridian GeoX
• Google, Meridian Studio
• Google Cloud, GPU pricing
• Bureau of Labor Statistics, Occupational Outlook Handbook, Data Scientists
• Bureau of Labor Statistics, Occupational Outlook Handbook, Market Research Analysts
• Bureau of Labor Statistics, Employer Costs for Employee Compensation, June 2026