Optimization makes waste cheaper. Sobriety removes it. Four layers to question before you start tuning.
In 2013, Yale decommissioned its Becton Data Center, home to 300 servers and storage devices. Its aging heating and cooling systems would have needed a $2.5 million investment. Instead of renovating, Yale shut the site down because it had already moved its information to virtual platforms.
During an architecture audit last year, we found three database instances running 24/7 for a reporting dashboard deprecated in 2022. Instead of optimizing their query patterns or right-sizing the clusters, we dropped the environments and archived the cold storage. Zero operational impact, zero cost, and 100% resource elimination.
That is the difference between making a system run better and asking whether it should run at all.
Optimization fixes inefficiency. Sobriety removes the unnecessary.
Efficiency alone doesn’t cap consumption.
In 1865, William Stanley Jevons observed that more efficient steam engines increased coal use because coal became cheaper to burn. Digital systems follow the same rebound effect.
The International Energy Agency projects that data center electricity demand will roughly double by 2030, driven largely by AI. At the company level, Flexera’s State of the Cloud survey shows estimated cloud waste falling from 28% in 2023 to 27% in 2024 and 2025, then rising to 29% in 2026, the first increase in five years, which Flexera attributes to the added complexity of AI workloads.
Efficiency is necessary. It does not set a limit. Sobriety does.
Four layers of excess
I call this approach Digital Sobriety Engineering: proportionality applied as a design constraint, before anything is built. It builds on existing Green IT and eco-design frameworks (GR491, INR, ADEME/Arcep).
- Storage: why are we keeping this? Deleting requires a decision; keeping doesn’t, so dark data, orphan backups, and copied test databases pile up.
- Compute: what is the minimum that meets the SLA? Elasticity hides overprovisioning. Flexera respondents estimate 27 to 29% of cloud spend is wasted (self-reported, but stable year over year). Yale also closed its Medical Campus Data Center after 15 years and joined the MGHPCC research data center consortium in 2014.
- Transactions: what breaks if this call disappears? One redundant call is negligible; a chatty architecture at scale is not.
- Models: is a smaller approach enough? A large language model suits some problems. For classification or extraction, rules, regular expressions, or a small specialized model often do the job with a fraction of the compute and latency.
Three questions at design time
- Is it necessary? What breaks if we don’t build or keep it?
- Is it proportionate? What is the smallest footprint that meets the requirement?
- Does it expire? When will we review or remove it?
Where sobriety stops
- Resilience: Redundancy can look like excess on paper and still be essential. Sobriety must not introduce fragility.
- Compliance: Keep what the law requires, for as long as it requires, and no longer.
- Prototypes: An oversized setup can be rational for a 48-hour proof of concept. Revisit it before production.
- User experience: Removing resources should never degrade a service people depend on.
Conclusion
Sustainability added after deployment tends to yield marginal gains. Built in during architecture and discovery, it has a far better chance of lasting, because the cheapest resource to run is the one never provisioned.
Before tuning your next pipeline or scaling your next cluster, ask your team:
What can we design out before we optimize what remains?
Sources: Yale Sustainability (sustainability.yale.edu/node/396); Yale Daily News (Becton & Medical Campus Data Centers); Flexera, State of the Cloud 2023–2026; IEA, Energy and AI (2025); Jevons, The Coal Question (1865).