AI Is Cheap Until You Run It Wrong
The cost of running AI workloads is not the headline price of a model. It is the sum of
choices made without measuring them: a larger model than the task needs, a hosted service
billed per token when the volume justifies dedicated hardware, idle capacity paid for monthly
and used twice.
This is an expense review. It happens to be an expense review of a newer line item than the
ones we normally audit, which is precisely why almost nobody is checking it yet.
What we review
- Model selection — whether the work genuinely requires the largest available model, or whether a smaller one at a fraction of the cost produces equivalent output for that specific task.
- Hosting model — per-token API pricing versus dedicated or on-premise serving, and the volume threshold at which one overtakes the other. The crossover is calculable, and it is usually lower than people assume.
- Throughput and utilisation — paying for capacity that is provisioned but unused is the most common and most avoidable line item.
- Batch versus interactive — work that does not need an answer in the next two seconds can often be run far more cheaply.
- Tool and licence sprawl — overlapping AI subscriptions bought by different departments solving the same problem.
- Data egress and storage — the charges that appear after the decision is made and nobody is tracking.
How the review runs
- Inventory. Every AI-related cost currently being incurred — tools, APIs, hosting, contractors, internal time.
- Utilisation. What is actually being used, by whom, and for what. This step routinely turns up subscriptions nobody can explain.
- Right-sizing. Each workload matched to the cheapest configuration that meets its requirement — tested where it matters, not assumed.
- Written recommendation. What to keep, what to change, what to cancel, and the expected effect on monthly cost.
Infrastructure
Where dedicated capacity is the right answer, we work with
Standard Compute
for high-throughput GPU infrastructure. It is a sensible default for batch processing,
model serving, and workloads that would otherwise be billed per token at a rate that no
longer reflects the volume.
Thryft Advisors receives a referral fee if you sign up with Standard Compute through this
link. It does not change our recommendation.
We are not tied to one provider. If the right answer is a cloud you already
have a committed spend agreement with, that is the recommendation. The review is impartial
because the finding is the finding.
Who this is for
- Businesses whose AI or API spend has grown without a corresponding growth in output.
- Companies running batch work — document processing, data enrichment, content generation — on per-token pricing.
- Anyone who has received a GPU hosting quote and cannot tell whether it is competitive.
- Buyers who suspect they are paying for capability they are not using.
What it costs
The review is priced by scale — how many workloads, how many providers, and how much is
currently being spent. We don’t publish a rate card. Tell us what you’re running and we’ll
give you a figure in writing before anything starts.