{"id":15,"date":"2026-09-13T15:33:17","date_gmt":"2026-09-13T15:33:17","guid":{"rendered":"https:\/\/tensorcurve.com\/?p=15"},"modified":"2026-09-13T16:16:06","modified_gmt":"2026-09-13T16:16:06","slug":"hourly-rate-versus-total-cost","status":"publish","type":"post","link":"https:\/\/tensorcurve.com\/?p=15","title":{"rendered":"A lower hourly GPU price can still cost more"},"content":{"rendered":"

Start with published inputs<\/h2>\n

Use the H100 base and one-month discount documented in our term-curve guide<\/a>. Consider a single-GPU project over a normalized 720-hour window. Compare flexible usage at $3.25 per hour with a calculated commitment rate of $3.1850.<\/p>\n

For this budgeting exercise only, assume that every hour in the window is payable under the commitment, while flexible spending stops when usage stops. Actual calendar hours, eligibility, payment timing and cancellation terms need supplier confirmation. This is a sensitivity calculation, not a prediction of an invoice.<\/p>\n

Calculate the full obligation<\/h2>\n
Modeled commitment = $3.1850 \u00d7 1 GPU \u00d7 720 hours = $2,293.20<\/strong><\/div>\n

Flexible spending equals $3.25 multiplied by used hours. Under these assumptions, the bills meet at 705.6 hours, or 98% utilization. A small published discount leaves little room for idle capacity.<\/p>\n

\n\n\n\n\n\n\n\n\n
Utilization<\/th>\nUsed hours<\/th>\nFlexible cost<\/th>\nModeled commitment<\/th>\n<\/tr>\n<\/thead>\n
50%<\/td>\n360<\/td>\n$1,170.00<\/td>\n$2,293.20<\/td>\n<\/tr>\n
90%<\/td>\n648<\/td>\n$2,106.00<\/td>\n$2,293.20<\/td>\n<\/tr>\n
98%<\/td>\n705.6<\/td>\n$2,293.20<\/td>\n$2,293.20<\/td>\n<\/tr>\n
100%<\/td>\n720<\/td>\n$2,340.00<\/td>\n$2,293.20<\/td>\n<\/tr>\n<\/tbody>\n<\/table>\n<\/div>\n

The fully utilized saving is $46.80 in this normalized month. That is a useful starting point for deciding whether operational flexibility is worth more than the discount. It is not evidence that a commitment is unsuitable for every workload.<\/p>\n

Measure productive hours, not just allocated hours<\/h2>\n

A GPU can be allocated without doing useful work. Data preparation, debugging, checkpoint recovery and idle time between runs can all consume a paid window. A budget should distinguish allocated hours from productive hours and explain which measure its utilization assumption uses.<\/p>\n

At 50% productive utilization, the committed bill divided by productive GPU-hours is $6.37, even though the contracted rate is $3.1850. This is not a change to the supplier\u2019s tariff. It is the effective cost of useful capacity in this example.<\/p>\n

The simple break-even rule is committed rate divided by flexible rate. It applies only when billing periods, hardware, availability and extras are comparable. If the flexible option has a minimum charge or the committed option has a separate upfront fee, include those terms before solving for the crossover.<\/p>\n

Extend the budget beyond compute<\/h2>\n

Add storage, data transfer, support and software charges as separate line items where applicable. Include the operational cost of interruptions or migration if they matter to the workload. Do not assume that an hourly price includes these services merely because it appears on the same product page.<\/p>\n

For calendar contracts, use the hours specified by the supplier\u2019s billing terms. Our Pricing Lab uses a normalized 720-hour month to make its planning arithmetic reproducible. That convention is not a statement that every real monthly bill covers 30 days.<\/p>\n

Make uncertainty visible<\/h2>\n

Instead of reporting one utilization estimate, calculate low, central and high cases. Keep the committed payment fixed and vary productive hours. This shows whether the apparent saving depends on a very optimistic workload schedule.<\/p>\n

A useful procurement note reports the listed rate, the total obligation, the break-even utilization and the consequences of finishing early. It should also identify whether access is assured. Cost and availability answer different questions, and both can determine whether a rental works for the project.<\/p>\n","protected":false},"excerpt":{"rendered":"

A worked example of utilization, commitments and the true cost of useful compute.<\/p>\n","protected":false},"author":1,"featured_media":0,"comment_status":"closed","ping_status":"closed","sticky":false,"template":"","format":"standard","meta":{"footnotes":""},"categories":[3],"tags":[],"class_list":["post-15","post","type-post","status-publish","format-standard","hentry","category-cost-analysis"],"_links":{"self":[{"href":"https:\/\/tensorcurve.com\/index.php?rest_route=\/wp\/v2\/posts\/15","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/tensorcurve.com\/index.php?rest_route=\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/tensorcurve.com\/index.php?rest_route=\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/tensorcurve.com\/index.php?rest_route=\/wp\/v2\/users\/1"}],"replies":[{"embeddable":true,"href":"https:\/\/tensorcurve.com\/index.php?rest_route=%2Fwp%2Fv2%2Fcomments&post=15"}],"version-history":[{"count":1,"href":"https:\/\/tensorcurve.com\/index.php?rest_route=\/wp\/v2\/posts\/15\/revisions"}],"predecessor-version":[{"id":19,"href":"https:\/\/tensorcurve.com\/index.php?rest_route=\/wp\/v2\/posts\/15\/revisions\/19"}],"wp:attachment":[{"href":"https:\/\/tensorcurve.com\/index.php?rest_route=%2Fwp%2Fv2%2Fmedia&parent=15"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/tensorcurve.com\/index.php?rest_route=%2Fwp%2Fv2%2Fcategories&post=15"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/tensorcurve.com\/index.php?rest_route=%2Fwp%2Fv2%2Ftags&post=15"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}