GEICO’s repatriation, in the words they gave reporters

In October 2024 two trade outlets published GEICO’s own operators on why a large US auto insurer was repatriating workloads from public cloud onto infrastructure they were rebuilding. The quotations and figures below are from The Stack (Edward Targett, interviewing Rebecca Weekly) and Futuriom (Mary Jander, reporting Sahid Jaffa’s Open Compute Project Summit remarks). GEICO did not, in those pieces, name a hyperscaler. This post does not invent one.

What Weekly told The Stack

Rebecca Weekly, then GEICO’s VP of platform and infrastructure engineering, confirmed to The Stack that GEICO was repatriating many workloads as part of a wider architectural overhaul. The Stack identifies GEICO as a Berkshire Hathaway insurer and, in that article, the United States’ third-largest automotive insurer by volume, with 2023 revenue of $39.6 billion and pre-tax profit of $3.6 billion.

Weekly’s cost line, as printed:

we have a lot of data – and it turns out that storage in the cloud is one of the most expensive things you can do in the cloud, followed by AI in the cloud…

The Stack says the company had started moving to the cloud in 2013 for 600+ applications, and that the journey was “sub-optimal.” Weekly, on the ten-year mark: GEICO still had not migrated everything; bills went up 2.5×; reliability challenges went up quite a lot too — because spreading data and methodology across “so many different vendors” meant spending time “recollecting that data to actually serve customers.”

She described the earlier on-prem estate as expensive in its own way: many vendors, fixed-function appliances, separate SANs, explicit L2–L3 layers. The cloud move was meant to exit that. What they got, in her phrasing, was to be “at the cumulative mercy of our clouds,” with “no consistent data strategy, no consistent hybrid stack.” Lift-and-shift, she said, took stacks to the cloud instead of refactoring for what the cloud does well:

Just running legacy applications in the cloud is prohibitively expensive. Our use case just highlights that…

Storage and latency sit together in her account. Insurers hold a lot of data and run predictive analytics. State regulators can demand production of records to show the firm did not violate policy terms or apply bias. Ease of access with minimal latency and at minimal cost was a major repatriation driver. The Stack notes egress (“exfiltration”) pricing as the familiar public-cloud complaint she is pointing at.

On capital: she rejected the cartoon that on-prem must be a single crushing capex event. “There’s ways of doing ‘CapEx-heavy’ in an OpEx-fashion on-prem,” and she expected financial-services firms to be willing to make capex over time. She also said reserved cloud capacity for advanced ML/AI — “three year reserve pricing upfront… and it’s nine months to GET that capacity?” — was starting to look like capex while still wearing an OpEx label.

The rebuild The Stack describes: a large OpenStack private cloud, heavy Kubernetes, open-source observability (Prometheus, Grafana, OpenTelemetry exporters), and a homegrown placement path using CAPI. Ajit Jain, in a May 2023 Berkshire remark The Stack reprints, had said GEICO’s technology needed more work than he thought, and that the firm still had more than 600 legacy systems it wanted cut toward 15 or 16.

The Stack’s aside that “public documentation suggests” a heavy Azure history is the reporter’s suggestion. Weekly, in that interview, does not name the provider. We leave it unnamed.

What Jaffa said at OCP, via Futuriom

Futuriom’s 22 October 2024 piece reports Sahid Jaffa, GEICO’s head of hardware and storage engineering, at the Open Compute Project Global Summit.

Jaffa’s frame, as Jander quotes it: they went “all in on the public cloud” to mitigate aging, unreliable infrastructure and to “modernize and innovate at scale.”

Figures Futuriom attributes to that talk:

• By 2020, 80% of workloads in public cloud.
• By 2021, eight cloud providers, spending over $300 million annually on those services — half of that on a single provider.
• Estate: over 200,000 cores of cloud compute and over 30,000 container and VM instances (Jaffa).

Futuriom adds that GEICO “has not acknowledged a particular hyperscaler,” then relays industry rumour. Rumour is not a source. We do not repeat a vendor name.

Jander: frustrated with high cloud costs and limited ability to adapt software, the team looked at open-source options and, after the 2023 OCP Summit, at ODM hardware on open standards. She reports Wiwynn delivering a single platform for storage and compute that also supported HPC and GPU-driven AI.

What these two pieces do not say

They do not publish a completed savings percentage for the whole book of business. They do not name the primary cloud in Weekly’s or Jaffa’s own sentences. They do not claim every insurer should repatriate. They describe one regulated, storage-heavy book of work whose operators say cloud storage, cloud AI, egress, latency, and reliability went the wrong way, and who are building an on-prem / private-cloud alternative with OpenStack, Kubernetes, and (in the OCP telling) open hardware.

Custody for a regulated pile of bytes

An insurer that must produce a file for a state examiner is in the same structural position as a bank that must produce a return: the location, latency, and unit cost of the store are not lifestyle choices. Weekly’s “storage in the cloud is one of the most expensive things you can do” is her sentence about GEICO’s data. Jaffa’s $300 million and eight providers are his OCP figures for 2021.

OpenXperiments is not on that programme. The reason we file the story here is the one we already put on the Approach page: customer data stays in the customer data centre by default — not because a hyperscaler cannot be competent, but because cost, custody, and blast radius are easier to explain when the computers are the institution’s. We are not claiming GEICO’s 2.5× bill or $300 million cloud year as anyone else’s.