Buy Azure Dedicated Host reserved instances

This is the rare Advisor recommendation whose entire point is to lower the bill. A Dedicated Host that runs continuously on pay-as-you-go is paying the most expensive possible rate for compute it was always going to consume. A reservation trades a one- or three-year commitment for a discount that Microsoft documents at up to 72 per cent against pay-as-you-go. The saving is not hypothetical: Advisor simulates the host's actual hourly usage over the last 7, 30 and 60 days, models the cost with and without a reservation, and only raises the recommendation when a reservation would have saved money.

Frame each recommendation as a payback calculation, not a purchase request. The relevant numbers are the committed amount, the modelled annual saving, and the break-even point, the month at which the discount has repaid the commitment. A three-year term carries the deepest discount but the longest lock-in; a one-year term saves less but de-risks a workload you are less sure about. The decision is which term the workload's expected life justifies, not whether to save at all.

The cost model has two distinct buckets. Reservations on hosts that genuinely run around the clock are close to risk-free savings: you were paying for those hours anyway. Reservations on hosts that might be retired or resized carry commitment risk, because an unused reservation is sunk cost. Azure softens this: there is currently no early termination fee, exchanges between compute reservation types are allowed, and refunds are self-service, though a USD 50,000 rolling-twelve-month cancellation cap applies. Finance owns sorting the risk-free saves from the ones that need a shorter term or a second look.

Priya is the finance partner supporting the platform team. Advisor has flagged three Dedicated Hosts as reservation candidates, with modelled annual savings of roughly 18,000, 11,600 and 4,100 in account currency. The instinct is not to approve a single blanket reservation budget; it is to turn each recommendation into a payback line. For each host she asks for the committed amount, the modelled annual saving and the break-even month, then sorts the list by certainty rather than by saving size.

The two production hosts have run continuously for over a year and underpin regulated workloads that are not going anywhere, so a three-year term is close to risk-free: the hours were always going to be billed, and the reservation simply buys them at up to 72 per cent off the on-demand rate. The pilot host is different. It saves less and its future is uncertain, so Priya recommends taking the one-year reservation to capture most of the saving while keeping the lock-in short. The finance pack reads: 'Two risk-free three-year commitments with payback inside the first year, plus one short-term saving on the pilot; the only thing standing between us and the discount is a purchase decision, and the commitment risk is bounded because cancellation and exchange are available with no current termination fee.'

The cost model here is the inverse of most security work: doing nothing is the expensive option. Every month a steadily running host stays on pay-as-you-go, the difference between the on-demand and reserved rate, up to 72 per cent on compute, is spend you could have avoided. Unlike a sunk one-off, this is recurring, so the cost of inaction compounds with every billing cycle the recommendation is left open.

The remediation cost is a commitment, and that commitment is where finance earns its keep. For a host that genuinely runs around the clock and will keep doing so, a three-year reservation is close to risk-free: you were paying for those hours anyway, and the discount simply lowers the rate. The risk lives only in the uncertain hosts, and even there it is bounded. There is currently no early termination fee, compute reservations can be exchanged between Virtual Machines, Dedicated Host and related types, and refunds are self-service, subject to a USD 50,000 rolling-twelve-month cancellation cap. The finance role is to attach a payback and a confidence level to each recommendation, commit the certain hosts to the longest justified term, and put the uncertain ones on a shorter term rather than skipping the saving entirely.

Dedicated Hosts are some of the most expensive line items in an Azure bill, and they are usually always-on by design. Paying the on-demand rate for compute that never switches off is leaving money on the table, and Azure Advisor flags exactly where. The recommendation is a savings opportunity, not a risk.

The leadership question is not 'should we cut costs?' It is 'are we committing to the spend we have already decided to make, at the lower rate Microsoft offers for committing?' A one- or three-year reservation converts a predictable, recurring bill into a documented discount of up to 72 per cent, with no change to how the workload runs. The trade is commitment for price.

This is a finance decision dressed as a technical one. The only real judgement is term length against workload life: a three-year commitment for the hosts you are certain will still be running, a one-year commitment where you want the saving but less lock-in. The defensible end state is that every steadily running host is on a reservation matched to its expected life, and the few uncertain ones are reviewed rather than blindly committed.

At a quarterly cost review the CFO asked the obvious question about the cloud bill: are we paying full price for anything we already know we will run for years? The honest answer at the time was that the Dedicated Hosts, some of the heaviest line items in the bill, were all on pay-as-you-go, even though they had been running untouched for over a year. Advisor had been quietly recommending reservations against them the whole time.

The team's response was to stop treating reservations as a one-off finance chore and start treating them as a standing discipline. They listed every host Advisor flagged, confirmed which ones were certain to keep running, and committed those to three-year reservations matched to their expected life, taking the lower one-year term on the workloads they were less sure about. The bill for that compute dropped to the reserved rate immediately, with no change to a single workload. The next time the CFO asked the question, the answer had changed from 'we are paying on-demand for everything' to 'the always-on hosts are on reservations at up to 72 per cent off, and the uncertain ones are on short terms we review each renewal.' That shift, from paying the most expensive rate by default to paying the committed rate by design, is the saving this capability delivers.

Almost every cloud cost programme that delivers real savings does it the same way: it stops paying on-demand rates for compute that runs all the time. Dedicated Host reservations are exactly that lever, and Advisor has already done the analysis of which hosts qualify. Leaving the recommendation unactioned is a standing decision to pay the most expensive rate for the estate's most expensive, most predictable compute.

The cost of acting is a bounded commitment with self-service exit options; the cost of not acting is an indefinite premium of up to 72 per cent on compute you have already chosen to run. This is the rare class of recommendation where the upside is a documented, recurring saving and the downside is a manageable, time-boxed commitment, which is exactly the trade leadership should be quickest to approve.