The Tax on Clinical Enthusiasm

In the second week of November last year, the chief financial officer of a 280‑bed hospital in regional New South Wales opened a cloud invoice and discovered that the line marked “data transfer out” had grown from $1,420 in February to $6,180 in October. Patient volume had grown 4%. The radiology workload had grown 9%: a new teleradiology partner had come online, and a quality‑assurance programme had begun pulling priors from the archive. Nothing on the invoice explained the difference. She described the moment, later, as feeling like someone had moved the goalposts during the match. Someone at a different hospital had described the same dynamic to me, on a different invoice the year before, more sharply: as a tax on clinical enthusiasm. The phrase travelled. We borrowed it.

The bill was not anomalous. It was a faithful record of what happens when you take a workload designed around the unpredictable behaviour of human beings — 11 radiologists opening whatever they need, whenever they need it — and price it on a variable‑cost ladder built for SaaS backends whose revenue scales with their usage. Hyperscaler healthcare cloud, as currently priced, punishes clinical success.

This piece is about what the picture looks like once you load the same hyperscaler estate the way a radiology group actually buys it, and what changes when you do.

The pricing model beneath every hyperscaler invoice has an elegant logic. You pay for compute by the hour. You pay for storage in small units. You pay for snapshots, for cross‑zone replication, for the gateway through which your traffic leaves the cloud. You pay, separately, for the support tier that gets you a human on the phone at 3am. Each charge maps to a unit of infrastructure cost the provider actually incurs. In a world of elastic, self‑service web applications, it is the cleanest model anyone has yet invented.

The problem is that a radiology PACS is not an elastic, self‑service web application. It is a clinical substrate. When a mammographer opens a three‑year‑old priors study at 7:12am because today’s screening needs comparing against a 2023 baseline, she is not behaving like a well‑disciplined consumer of cloud resources. She is behaving like a mammographer. When a teleradiology partner brings on 40 more nightly reads, the egress meter rises, and the multi‑AZ replication tax rises with it. None of this is misuse. It is the workload working correctly.

For most of the last decade, comparisons between healthcare cloud options have been published with one structural sleight of hand: the hyperscaler line is compute only, and the comparison line is fully loaded. This is convenient if you are selling private cloud. It is also dishonest.

We rebuilt the comparison the other way. The hyperscaler is fully loaded: Linux PACS on a 3‑year reservation, multi‑AZ storage replica because clinical resilience is non‑negotiable, NAT Gateway in a high‑availability pair, an Application Load Balancer with a real web application firewall, backup retention at 35 days, basic KMS, monitoring, audit logging, threat detection, and the production support tier a clinical workload actually buys. Nothing missing. Nothing inflated.

The numbers move.

What changes, when you load the hyperscaler honestly, is not that 3verest becomes cheaper than AWS on every line. We don’t. The per‑vCPU rate, the per‑GB on cold archive, the breadth of the service catalog: these are places where the hyperscalers are remarkable engineering at scale, and we say so. What changes is the shape of the comparison. Three lines, in particular, that are absent from the headline analysis but present on every actual invoice.

First, multi‑AZ resilience is not free anywhere except 3verest. AWS charges for the second‑AZ replica volume in full. Azure adds a 50% uplift on Premium SSD for zone‑redundant storage. Google Cloud bakes it into the regional disk SKU at a higher per‑GB rate than the zonal equivalent. None of this is on the front‑page VM price.

Second, NAT Gateway data processing is the silent killer. At 25 TB of monthly egress — a modest Medium PACS estate — the per‑GB data‑processing fee at the NAT layer adds roughly $1,200 a month on top of the internet egress itself. Most comparisons miss the line entirely. The invoice does not.

Third, the support tier a production clinical workload requires costs money on every hyperscaler. Azure Professional Direct is a flat $1,000 a month, which is fine on a Large estate and brutal on a Small one. AWS Business Support blends to between 7% and 10% of spend. Google Cloud Enhanced sits at 4% of spend with a $500 floor. None of this is optional, and none of it disappears.

These are not edge cases. They are the lines a CFO finds in October, when the invoice does not match the procurement assumption from March.

Fully loaded — same operational scope on every side, Linux PACS, 3‑year committed compute, multi‑AZ for clinical resilience, production support, the works — a Medium PACS workload in Sydney runs $10,512 on AWS, $12,619 on Azure, $9,682 on Google Cloud, and $7,449 on 3verest. The gap holds across regions. In Belgium and Frankfurt, Medium lands at $9,190 on AWS, $8,581 on Google Cloud, and $7,449 on 3verest. In US East: $8,715, $8,246, $7,449.

The egress curve is the lever. At 10 TB of monthly egress in Sydney, 3verest sits approximately level with the cheapest hyperscaler stack. At 25 TB, the gap is 23%. At 80 TB — a multi‑site PACS with an active teleradiology footprint and AI inference workflows — 3verest is between 56% and 70% cheaper than the loaded hyperscaler equivalent. The egress meter is a step function, not a curve, and Australian outbound rates compound it. AWS Sydney charges $0.114 per GB after the first 100 GB free. Azure Sydney $0.12. Google Cloud Sydney $0.19. 3verest does not have an egress meter.

There is one place we do not win on raw unit price, and it is honest to say so. Cold archive — AWS Glacier Instant Retrieval at four‑tenths of a cent per GB‑month, with retrieval at 30 cents per GB — is structurally cheaper than anything a sovereign cloud at sub‑petabyte scale can build. 3verest’s DICOM Object Archive sits at 4 cents per GB landed, with retrieval included and no minimum retention. The trade is predictability for unit price. For active radiology archives where priors get pulled routinely, predictability tends to win on 3‑year TCO once retrieval volume is honestly forecast. For deep cold tier with sub‑2%‑per‑month retrieval, the hyperscaler archive wins. We say so plainly. We let customers layer their archive where it makes sense.

There is also the question of which cloud a healthcare buyer is actually allowed to use. For workloads under HIPAA, IRAP PROTECTED, HDS, NHS‑DSPT or EU‑residency obligations, the relevant peer set is not commercial AWS US East. It is the sovereign and restricted‑jurisdiction tiers: AWS GovCloud, Azure Government, Azure Local, Google Cloud’s sovereign overlays delivered through T‑Systems in Germany and S3NS in France.

The premium runs between 15% and 40% above commercial. The service catalog runs between 12 and 24 months behind. Against that peer set — which is the comparison most large healthcare buyers will actually make once procurement starts — 3verest is between 30% and 50% cheaper at Medium scale, ships on a full healthcare‑tuned service catalog, and runs with a named engineering team rather than a sovereign‑partner ticket queue layered over a hyperscaler operations centre. Sovereign hyperscaler tiers, in our experience, carry the worst of both architectures: hyperscaler operational variability plus a sovereign premium minus most of the service catalog. 3verest was built sovereign‑first. It is a different cost structure, not an uplift on a commercial one.

Bill shock at this scale is not a billing error. It is a product decision. And its absence — a flat, predictable, behaviour‑independent number — is also a product decision. The question for any health system thinking about cloud is not which vendor has the better unit price on an isolated component. It is whose product has the shape that aligns with how radiologists actually work.

The CFO from the regional New South Wales hospital signed a 36‑month order form with us in March of this year. In the eighth month of that contract, her cloud line was a dollar lower than the month before, and a dollar lower than the month after. The goalposts had stopped moving.