The 1–2 year commitment question, answered honestly

“Why would I commit for two years?” is the single best question a buyer can ask us, and we take it as a compliment every time. A company that asks it is already thinking about compute the right way. Here’s our honest answer — including the cases where we’d tell you not to sign.

The core claim is simple: if you’d bet your job that you’ll still need these GPUs in month 18, committing is just prepaying a cost you were going to pay — at a materially lower rate. Everything else is risk management.

01 / WHEN IT MAKES SENSE

• 01Sustained utilization. Your fleet runs hot, month after month — average utilization above 70–80%, with no meaningful idle weeks.
• 02The bill is real. You’re already spending $50k+/month on GPU compute. At this scale the commitment math usually wins by enough to matter in the board deck.
• 03Production workloads. The committed capacity serves production inference, recurring training, or steady generation pipelines — not experiments you’re still discovering.
• 04Growth is the shape. Demand is flat-to-growing over 12–24 months, so committed capacity gets absorbed rather than stranded.

02 / WHEN IT DOESN’T

• 01Still finding fit. If your workload mix changes every quarter, you need flexibility, not a rate.
• 02Spiky or seasonal. Bursts that come and go belong on cloud elasticity — locking in idle capacity is the one way to make this model lose money.
• 03Short runway. If the next 12 months are uncertain for the business itself, infrastructure commitments are the wrong kind of bet.
• 04Tiny fleets. A handful of GPUs won’t carry commitment economics. The model earns its keep at production scale.

We mean this sincerely: we’d rather lose a deal than sell a commitment that strands you. A stranded commitment isn’t revenue — it’s a story we don’t want told about us.

03 / THE RISKS, STATED PLAINLY

GPU capacity generations. A two-year commitment could, in theory, strand you on last year’s silicon. The practical antidote: commit the baseload you know runs on today’s architecture, and keep growth on flexible capacity that can absorb new generations as they ship.

Forecast error. You might over-commit. The antidote is baseload math: commit the floor you can prove from trailing data — the 10th-percentile week, not the average — so the worst case is that extra demand stays on the cloud.

Counterparty risk. You’re trusting a provider to be there for 24 months. This is why agreements specify node configuration, networking, SLA, uptime target and support in writing before anything is signed — and why we publish the spec we’d want to read.

04 / THE DECISION RULE

Forget the spreadsheet for a minute. The rule: if month-18 demand feels like a forecast you would defend to your board, a commitment turns variable spend into a fixed line item at a lower rate. If it feels like a guess, keep the flexibility and revisit when the fog clears.

Our job is to model the difference with your real numbers — and when the honest answer is “not yet,” to say so. The companies that commit when it fits tend to do it again. The ones that commit when it doesn’t tend to tell everyone. We know which outcome gets chosen.

Long-term leasing isn’t for everyone. For the fleet that’s hot, growing, and production-proven, it’s simply the arithmetic winning over habit. And when it doesn’t make sense, the job of an infrastructure partner is to say that first.