AI Inference Portability, Amrize, Salesforce Churn, Copart Share Loss
A collection of interviews published last week. Visit our platform for full coverage
Published Last Week
AWS: Inference Portability
Last month, we explored the forecasted revenue ramp for 1GW of AI capacity at AWS:
My benchmarks per gigawatt, based on 32 business cases I've built over the last 18 months, are as follows. Fully loaded CapEx is approximately $30 billion per gigawatt. Annual revenue at today's rates is about $6 to $7 billion. EBITDA comes in somewhere between 45 to 50% over the contract period. Annual cash contribution is roughly $3 to $3.5 billion. Cash payback is five to six years. Those are the numbers I use specifically when building a business case for Amazon or any hyperscaler. - Former Global Executive, AWS
The model considers the cash payback on gross capex over Year 1-6 considering:
- Utilisation ramping from ~50% to 90%+.
- Workloads shifting from ~60% training / 40% inference to ~70% inference
- Trainium ramping from 40% to 55% of capacity
- Bedrock and managed services attach rate ramping from 5-15% to 50-60%
A major variable driving AWS ROIC is the attach rate of high-margin managed services serving inference for customers.
Data storage is the biggest driver [of the attach rate]. Database search performs equally — both are about 25% of attach revenue. The AI platform is about 20%, networking is about 15%, security is about 10%, and observability is about 10%. Lambda and the serverless applications account for another 10%. - Former AWS Executive
But what if the inference can be ported elsewhere and AWS loses all this high-margin revenue?
There is definitely a scenario where inference becomes portable and begins to look like nothing more than an API. If that application is using Claude and running through standard Claude APIs, the inference could run anywhere. - Former AWS Executive
We explore which workloads and applications could be at risk from inference portability.
Banks like Lloyds or JP Morgan are going to keep S3, identity, application and security on AWS. If they send the inference to Nebius, without moving the data permanently, the application can get the context from AWS. That way I can send only what I need to an external inference provider. AWS continues to own the workload but potentially loses the compute component. - Former AWS Executive
We discuss the three main risks of AWS losing workloads. An interesting company posing a risk to AWS is Fireworks AI:
Fireworks AI with optimized model inference. Fireworks has a compelling approach. The key distinction is that you do not move the entire application — you move or redirect the model execution. For instance, suppose an enterprise has AWS S3, Amazon Bedrock and an AWS application. With Fireworks, the system of record can stay in AWS. What moves is the prompt and the context for the inference request. Fireworks is not a broker — they host, operate and run the inference and serving engine themselves. - Former AWS Executive
This comment was also interesting comparing the quality of the traditional data center business to new AI compute considering switching costs, ROIC, and the structure of the value chain:
At the end of the day, inference is compute, but it is economically and architecturally very different. You are using GPUs and ASICs versus CPUs, and the workloads are completely different. In the old world, you were using VMs and applications. Here, you are generating tokens. I would argue it is easier [for customers to switch]. There is lower overhead compared to CPUs. You get better leverage of memory. It is easier to optimize your inference. You get better economics per megawatt. You get better performance gains from the accelerators, and on top of all the improvements already being made, your software will likely also perform better. Those gains compound over time, which is why inference per megawatt will become dramatically faster and cheaper than traditional CPU compute. - Former AWS Executive
The interview goes on to explore the role of data gravity and risks to AWS losing inference workloads and can be read alongside:
Amrize & US Cement Pricing Power
After years of pricing increases above inflation, US cement customers seem to be pushing back on Amrize:
Significant price increases were passed to the market, especially from 2023 to 2024, and then again from 2024 to 2025… the price increases we passed through were well above what we had in terms of cost increases. It was more because we could — because we had the leverage and the demand was there. But in 2025, we started seeing problems passing price increases, especially because at that point you start hitting import parity pricing. - Former Senior Executive at Amrize
The growth in customer-owned terminals seem to provide the scale for ready-mix producers to import directly:
Ozinga basically calls Amrize and says, "I'm not going to buy from you anymore. I've made the investment and the decision that I'm better off — no matter what you promised me for this next year, I already decided that I've lived through so many years of price increases. I've made the investment, and I'm going to bring my own cement." - Former Senior Executive at Amrize
The interview goes on to explore import pricing parity economics and competition and can be read alongside:
Salesforce Marketing Cloud: Churn Drivers & Share Gains for Braze and Klaviyo
Agentforce Marketing, previously Salesforce Marketing Cloud, has an estimated 8-12% annual churn:
Roughly 30% of their Marketing Cloud account base comes up for renewal each year... Within that 30%, we are generally seeing anywhere from 8% to 12% churn depending on the year and the segment. - Former Strategic Director, Salesforce
Of clients that churn, ~70% are moving to a point solution like Klaviyo or Braze:
Within that 8% to 12% churn figure, roughly 70% are moving to a single-point solution, while about 20% to 30% are moving to an entirely new platform. The reason the platform replacement figure is smaller is that a full rip-and-replace is far more arduous for an organization — in terms of headcount allocation, spend, and consultants required. It is much easier to simply tell the marketing team they will be using a different tool in three months, since only the marketing team is affected. That is why those percentages are distributed the way they are. - Former Strategic Director, Salesforce
Also, this is interesting regarding Agentforce ARR accounting…
if you were a large customer spending over $1 million, over 90% of those upsells to Agentforce were being given away for free in the first 12 months of a three-year deal... What you would see in those years 2 or 3 is that inside of our books, we would transition the revenue from that legacy platform to the new Agentforce platform. It looked like those were growing from a revenue standpoint, when in reality we had simply transitioned them to a new platform. - Former Strategic Director, Salesforce
This interview is best read alongside:
Copart vs IAA
A Former State Farm Manager who led the salvage team explains how it chooses salvage providers:
What corporate looked at was different criteria: how quickly they pick the cars up, how quickly they get the car ready to be sold, and how quickly they process the title and get it back from the DMV. Some DMVs get backed up significantly, and that's out of the hands of both the vendor and the insurance company. On the insurance side, Copart, IAA, and I formed a committee and sat down with some state departments to see what we could do to improve the process. We made improvements in Virginia, North Carolina, West Virginia, and South Carolina. Florida was also very important. - Former Manager at State Farm
And how and why IAA may continue to win share from Copart:
I'd say you may see some moves; not dramatic shifts like Progressive made, where they gave all their business to IAA. I don't foresee State Farm doing that. I think we prefer the competitive environment. The split is roughly 52-48 in favor of IAA right now, and I think IAA may gain another 5% to 10% based on their performance. - Former Manager at State Farm
This can be read alongside prior Copart research:
Related Content
© 2026 In Practise. All rights reserved. This material is for informational purposes only and should not be considered as investment advice.