Goodbye to the Cloud?
In the Fall of 2022, the American software company 37signals announced a bold new objective for all of its applications: to bring their infrastructure back on-premise. This plan primarily targeted their two most prominent products, Basecamp—a popular project management tool—and HEY, their email client, which they still actively marketed. The remaining legacy applications—including Basecamp Classic, Basecamp 2, Highrise, Backpack, Campfire, Writeboard, and Ta-da List—were already in maintenance mode, no longer receiving active development but kept accessible for existing users.
From Cloud Reliance to On-Site Infrastructure
Historically, Basecamp operated mainly using on-premise servers, with hardware maintained by Deft, hosted at their data center. The only components in the cloud were three essential services: the search functionality (via OpenSearch), file storage (using Amazon S3), and a Content Delivery Network (via CloudFront).
HEY, meanwhile, primarily relied on Amazon Web Services (AWS) — specifically Elastic Kubernetes Service (EKS), Aurora/RDS for databases, and ElastiCache — with some exceptions for email processing services and image hosting. The other legacy apps also ran on EKS and utilized RDS for database management, forming a largely EKS-based infrastructure.
Prioritizing Calculation Power Before Storage
By early 2023, 37signals had conducted a detailed review of potential cost savings from migrating their infrastructure back to their own servers. Over the previous year, their total cloud expenses across all applications had amounted to approximately $3.2 million. While they had already optimized costs through internal efforts and committed usage contracts with AWS, the company aimed to cut this expenditure by about $2.3 million in 2023, targeting the eliminations of most cloud costs early in the year. The remaining costs—primarily associated with S3 storage (which managed Petabytes of data, around 8 petabytes with replication across two regions)—were earmarked for reduction in the following year.
The planned infrastructural shift involved deploying a new data center setup, consisting roughly of:
– Two Deft data centers located in Ashburn, Virginia, and Chicago
– Each housing four 48U server racks
– A combination of older and newer servers, totaling around 90 servers per site
The new servers chosen were Dell PowerEdge R7625 models—initially a fleet of twenty—equipped with two AMD EPYC 9454 processors (with 48 cores at 2.75 GHz and 96 threads). The servers offered an impressive total of 7.68 terabytes of RAM, 384 terabytes of NVMe storage, and nearly 4,000 vCPUs.
The cost for these servers was estimated at approximately $600,000, which, amortized over five years—and considering some legacy servers with longer operational lifespans—equated to about $120,000 annually. Additional expenses included rack space rental, networking, and power, bringing the total monthly data center operational costs to an estimated $60,000. This figure accounted for capacity headroom to add more servers as needed.
Putting these numbers into perspective, the annual cost for this on-premise setup was projected around $840,000, representing a significant saving of roughly $1.5 million compared to previous cloud expenses. By setting aside approximately $500,000 as a contingency fund for unexpected costs, 37signals believed they could save around $7 million over five years.
Dedicated Operations Team and Streamlined Management
In September 2023, the company updated its forecasts, now anticipating even greater savings—up to $10 million over five years—based on the observed downward trend in cloud expenditure over the previous six months. The expenses excluding S3 had decreased from $180,000 per month to about $80,000, a trend expected to continue with the gradual phasing out of reserved instances and other commitments.
A few weeks later, 37signals announced the shutdown of their last large OpenSearch logging cluster, a managed service costing approximately $43,000 per month for all applications. Their replacement infrastructure involved hardware investments of around $150,000 for new servers—split equally between the two data centers to ensure redundancy. If operational costs roughly doubled—and assuming the existing operations team remained unchanged—monthly expenses would come to around $5,000 over a five-year lifespan for the new setup. This budget includes a CDN service and contracts for support, mainly relying on open-source software platforms such as MySQL (replacing RDS) and OpenSearch, along with Elasticsearch.
Transition to Pure Storage for Storage Solutions
By early 2024, the company reported that all migrated applications maintained an uptime exceeding 99.99%. Overall, they reaffirmed their goal of saving $10 million over five years, although their S3 costs had increased slightly to about $1.3 million. This uptick was attributed to the growth in storage capacity—currently around 10 petabytes, stored across multiple regions—and network transfer costs associated with data migration back to their own infrastructure. The S3 contract, signed in 2021 with a four-year commitment, remains active until summer 2025.
The migration plan involves switching to 18-petabyte-scale Pure Storage arrays, distributed across their two data centers. Although the upfront cost of these flash arrays is roughly equivalent to a year of S3 expenses, their high density and energy efficiency mean they can be integrated directly into server racks, helping control ongoing costs. The expected recurrent expenses would be limited, enabling additional savings of about $4 million over five years—excluding future infrastructure evolution if required.
Currently, the installation of the storage arrays is underway, with nearly 6 petabytes remaining to be transferred from AWS S3. A dedicated 40-gigabit network link has been established for this purpose. The company estimates the data transfer process will take around three weeks, leveraging the 60-day AWS cost-free data egress window. The hardware cost for the arrays is approximately $1.5 million, with a further $1 million allocated for support, maintenance, and warranty over five years.
Looking Ahead: The Future of 37signals’ Infrastructure
In 2015, during the launch of Basecamp 3, 37signals had purchased 61 Dell servers—model R430 and R630—with an investment of about $500,000. Nearly a decade later, they began a phased decommissioning of these assets, turning some servers into legacy application hosts and repurposing others as the transition to on-premise infrastructure continues. These efforts highlight a strategic shift away from reliance on public cloud providers towards customized, dense, and energy-efficient data centers tailored to their specific needs.
The move to on-premise infrastructure signifies a significant change in their operational philosophy, aiming for cost reduction, greater control, and potentially improved security and compliance by maintaining critical systems within their own facilities. While cloud solutions offer flexibility and scalability, 37signals’s approach demonstrates a deliberate shift towards owning and managing their hardware to meet their long-term strategic goals.
