37signals leaves the cloud

In October 2022 the company behind Basecamp and HEY announced it was leaving Amazon Web Services. It had spent $3.2 million on cloud in 2022. It bought its own servers, put them in racks it already rented, moved seven applications out during 2023 without growing its operations team, and by 2024 had cut the bill to $1.3 million, all of it file storage still under contract. It is the clearest public example of the rent-or-buy decision from Module 33a, and of the conditions that make buying win.

1. The situation

37signals runs Basecamp, HEY (an email service) and several older applications: a mid-sized software business with a long history and a steady, predictable load. By 2022 it was spending $3.2 million a year on cloud. Just under a million of that was storing about 8 petabytes of files in Amazon S3, replicated across regions; about $2.3 million was everything else: application, cache, database and search servers, much of it as managed database (RDS) and search (OpenSearch) services. For HEY alone, database and search cost over half a million dollars a year.

The company already rented eight racks across two data centres from a hosting provider (Deft), at about $60,000 a month, with space to spare.

2. What mattered

• Cost, with a load that doesn't swing: the cloud's elasticity was insurance they didn't need. DHH's own conditions for the cloud are "very early days" and "enormous fluctuations in load"; he says neither applied to Basecamp.
• Performance per dollar: new AMD Zen 4 servers with fast NVMe drives against rented instances.
• No bigger operations team. The cloud hadn't let them shrink it; the exit mustn't make them grow it.
• Independence: deploying the same way on any hardware. They built their own deployment tool, Kamal, after "a brief detour down a blind alley with an enterprise Kubernetes provider".

3. The options

4. The decision

Decision: Leave the cloud: compute and databases in 2023, file storage when
  its contract ends (2025)

Context: $3.2m cloud spend in 2022 ($2.3m compute/databases/search, ~$0.9m S3).
  Steady load. Eight racks already rented in two data centres, with spare space.
Decision: Buy ~4,000 vCPUs of Dell servers (2,000 per data centre, for
  performance and redundancy), deploy with our own tool (Kamal), run managed
  services (databases, search) ourselves. Move S3 data to our own storage arrays
  when the four-year S3 commitment expires.
Consequences:
  + Projected ~$1.5m a year saved on compute alone; faster hardware
  + Same deployment tool on any machines
  - We buy capacity ahead of need; no instant scale-up
  - We operate databases and search that AWS operated for us
  - Savings arrive only as one-year and four-year commitments run out
Revisit: if load becomes highly irregular, or a new product's growth is
  unknowable (the HEY launch case)

5. How it works: where the money went

The yearly spend, step by step, from the posts. Box text shows the annual run rate of each part.

6. The numbers

DHH's February 2023 estimate: $600,000 of hardware over a conservative five years is $120,000 a year; with the racks' $720,000 a year for power, space and bandwidth, $840,000 a year in total, against $2.3 million in the cloud. Try the same comparison with different facts.

Five-year totals. Defaults are 37signals' figures: $2.3 million a year of cloud compute, $600,000 of hardware, racks at $720,000 a year, no extra staff (their ops team stayed the same). Assumed: an extra engineer costs $200,000 a year; spiky load means buying hardware for a peak three times the average.

7. What happened

The applications were out of the cloud by mid-2023. The bill didn't drop at once, because compute and databases were on one-year commitments: by September 2023 monthly cloud spend excluding S3 was down from about $180,000 to under $80,000. The hardware ended up costing about $700,000 and was paid back by savings during the second half of 2023. In 2024, the first clean year, the cloud bill was $1.3 million against the original $3.2 million run rate, and all of it was S3, still under its four-year contract until summer 2025. For storage they planned two data centres of Pure Storage flash arrays (18 PB in total), for about one year's S3 bill, again fitting in the existing racks. The projected total: well over $10 million saved over five years, with the same operations team.

DHH is explicit about the limits: the comparison is "never fully apples-to-apples", someone without existing racks must rent them, and the cloud still makes sense very early on or with enormous swings in load.

8. Lessons

• Elasticity is a characteristic you pay for. With a steady load it's insurance against a risk you don't have (Module 05, 22a).
• Capex against opex. $700,000 once, written off over five years or more, against $2 million a year of rent (22a, 33a).
• The facts made it work. Existing racks with spare power, a steady load, and a team that already ran the systems. Change those and the answer can flip; the lab shows how far.
• Commitments delay savings. Reserved instances and multi-year storage contracts kept the bill high for months after the move: put contract end dates in the migration plan.
• Portability made the move cheap. Containers and their own deployment tool let the same apps run anywhere (lock-in as an option).

9. Your turn

1. Do the math (15 min). Take your largest steady workload. What does it cost a year in the cloud? What would the hardware and space cost over five years?
2. Find the commitments (5 min). Which reserved instances, savings plans or contracts would keep you paying after a move, and until when?
3. Name what would flip it (5 min). Which fact (load shape, team, racks) would make staying in the cloud right for you?

10. Check yourself

11. Sources and further reading