Under GxP,
$2,401,346 a year
measured.

19 changes went in, 18 of them counted at $1,796,936 a year net, 9.8% of the bill: waste removed under a register the owners signed, an idle p4d.24xlarge released and SageMaker GPU notebooks stopped when idle, non-production schedules, gp2 volumes moved to gp3 in the research accounts, S3 gateway endpoints, and lifecycle rules on the data lake. The tar job for the oldest sequencer runs ran March 9 to 12. On Thursday, March 12, gp2 to gp3 went onto the GMP systems before their freeze began, as a GxP standard change with an executed installation checklist and notice to QA; the log-retention and Config changes, L-2 and L-3, went into qualified accounts the same way.

S3-4 went in too, but objects move tiers only after 30 days without access, so nothing can be measured before week six. It counts nothing toward the gate, and its first measurement window opens May 11. One change came out: L-4 would have limited CloudTrail data events on research buckets to writes, and Rosalind Tran, the Chief Information Security Officer, reversed it on Wednesday, March 11, before it was applied, because the read events feed the insider-risk detection for sequence designs. The statement of work counts a change approved at the readout and reversed later at its estimate, $35,194 a year, so the guarantee cannot be met by approving a change and never making it.

Four changes worth $729,373 a year net, which brought the total in place to 22 changes and $2,526,309 a year at their estimates, 13.8% of the bill, before anything was measured. The genomics Batch pipelines moved to Spot on Monday, March 16, but only after the price table showed that the On-Demand instance type they had been using saved too little on Spot to justify the interruptions: the new environments draw from a cheaper instance type, with the type they run on today as a second pool, and the estimate prices every Spot hour at the cheaper type’s mean and adds an allowance for retried work. The genomics Lustre file system had been provisioned at 64 TiB and 1,000 MB/s per TiB for a peak that never came: its 90-day peak was 14.8 GB/s, and the replacement is 48 TiB at 500 MB/s per TiB (24 GB/s), above the peak. The next step down would have sat below it, so it was rejected.

On Tuesday, March 17, the GxP change advisory board approved a weekday schedule on the GMP validation environments, which hold no GxP records, and it ran in the release window on Thursday, March 19. On Thursday, March 19, Rafael Quintero, with Adaeze Whitfield, approved a one-year No Upfront Compute Savings Plan at $50 an hour, bought that afternoon in cw-cicd, a member of the sharing group Copperwick limited to the 48 in-scope accounts. It saves $148,682 a year. The existing three-year plan stays fully used after every change, deferred ones included, which is what sized the new one; a three-year plan waits for the 2027 renewal, and the p4d plan was declined at the readout.

No savings were counted this week, by design, because every counted change was already in place. What went in was the nine guardrails, GR-1 to GR-9, each with one owner and built as code: a VPC template with S3 gateway endpoints, a service control policy that refuses new gp2 volumes, Spot-by-default Batch environments, log retention in the account baseline, a throughput review before any new Lustre file system, idle-GPU alarms, notebook auto-stop as an account default, and budgets and anomaly monitors for each account family. One more document set the rule for data: the data-class lifecycle standard says where each kind of data may be stored and who may change it, co-signed by Dr. Osei for its GxP class. No guardrail shortens the retention of a GxP record.

The guardrails could not go everywhere on day one, and the week shows why. They went into 32 accounts first, six of them qualified, where each went in as a GxP standard change with notice to QA. The GMP freeze ran until Friday, March 27, so the nine GMP and quality accounts get theirs on Monday, March 30, as a standard change with notice to QA; the seven clinical accounts wait for topline and get theirs on Monday, April 27. The deferred validated-system changes went to the board on Tuesday, March 24, each with its validation impact assessment.

Every change was measured on the Cost and Usage Report the same way: the cost of the changed resources over the seven days before the change against the seven days after it, annualized, at public On-Demand prices, so the Savings Plans are never counted twice, and then net of the private pricing discount, applied once. Spot is measured at the Spot price paid, and the two volume-driven changes are measured per genome processed, so a slow lab week cannot read as a saving. 22 changes measured $2,401,346 a year net, 95.1% of the $2,526,309 estimated for them, and every miss is published at its measured value. Six measured at their estimate, ten came within 91% to 100% of it, and four came in below 90%: S-1, S-3a, W-4, and EF-1. Scientists overrode workstation auto-stop during a structure campaign (S-1), an executed qualification protocol held a validation environment on for several nights (S-3a), snapshots share blocks so deleting one freed less than the estimate used (W-4), and less data moved to Infrequent Access than estimated (EF-1). Two beat their estimates, C-1 and W-1: the Savings Plan covered more usage than its estimate priced, and the snapshots owners kept bill written blocks only.

The handover gave every remaining change an owner and a date, and set the first monthly cost review for Thursday, May 14. The second invoice, $17,500, was issued with the measurement report, so Copperwick paid $35,000 in total.