AWS, Anthropic, & AI Backlog Risk, LVMH, Uber vs Waymo
A collection of interviews published last week. Visit our platform for full coverage
Published Last Week
AWS, Anthropic, & AI Backlog Risk
In Q2 26, Amazon reported ~$500bn backlog of commitments. A former AWS executive who reported to Matt Garman, the CEO of AWS, and worked on the Anthropic deal estimates ~30% of the backlog is commitments from Tier 1 customers, the top Enterprise Global 2000 companies.
As AWS increasingly signs deals with traditional enterprise customers, the risk that token consumption is below dollars committed may increase due to the difficulty in forecasting demand.
If I am a Frontier model customer, it is straightforward because I can build against a contractual commitment. If I am a Global 2000 customer going from 200 GPUs to 10,000, now you are dealing with sales optimism, production deployments that get delayed, and many utilization assumptions. With the biggest customers, there is lag in the forecast. I have seen repeatedly, around Anthropic contracts, that a customer will think they are going to commit to five gigawatts, it becomes bankable, then the first year comes in at only two gigawatts. - Former Global Executive, AWS Data Center Business at Amazon
At the end of the day, these contracts are tied to consumption. AWS has gotten better at building consumption into the contracts, but you are hitting on the right point. You could build it and commit to it, but there is the risk that the consumption never comes. That is what we are all waiting to see in the next two years with AWS. - Former Global Executive, AWS Data Center Business at Amazon
There is also a more subtle change occurring at AWS that may amplify this risk. Historically, selling EC2 and S3, AWS followed a product and developer-led sales motion. Build it and the developers will come. Now AWS is signing multi-billion dollar deals, it's more like large, traditional enterprise sales:
The success of AWS until 2024 was built on developers — the culture was to find a developer, someone who loved EC2 and liked all the tools. The shift to data centers, AI, and GPUs required transitioning into the lens of how Microsoft sells into the corporate world and the line-of-business world. That was very uncomfortable territory for AWS sellers. - Former Global Executive, AWS Data Center Business at Amazon
The sales team is effectively shifting from taking orders to strategic, project-led sales:
AWS's sales organization, until the end of 2024, were essentially order takers — how many do you want, and when would you like them delivered? One of the things Matt is still trying to fix is taking that organization away from being order takers and turning them into strategic advisors. That is still a work in progress. Former Global Executive, AWS Data Center Business at Amazon
There is also a shift from horizontal to vertical selling:
AWS traditionally sold horizontally. It did not matter if you were a bank, an insurance company or a healthcare company. In early 2025, they pivoted toward industry-driven sales, and the vast majority of AWS sellers have no domain expertise. As a sidebar, that shift drove even greater importance of the channel in everything AWS is doing, because the channel brings capabilities that AWS sellers simply do not have. Former Global Executive, AWS Data Center Business at Amazon
And, more importantly, there are now much stricter requirements of AWS sales people forecasting consumption because it is driving such huge capex.
In 2023, you would be a hero if you walked in and said you had a half-billion-dollar opportunity. Now he would ask: what is the customer committed to? When is the workload going to production? What does the ramp look like? When will we see the consumption we are forecasting? The inspection in 2024 became much more rigorous. Matt put in much better processes for managing demand. The consumption and CapEx forecast were micromanaged, and that was the biggest behavioral change. Bookings used to be the measure in cloud, but now consumption really matters. - Former Global Executive, AWS Data Center Business at Amazon
Consumption seems very difficult to accurately forecast. The sales team forecasts originate from the customers. Do the customers even know how to accurately forecast their own consumption?
The complex operating realities of deploying AI mean it is currently taking 12-15 months to convert enterprises from pilot to production.
I am currently deeply involved with an AWS deal with a very large insurance company. They built a pilot for claims processing using Amazon Bedrock. The agent we built read the claim documents, summarized the files, and was doing a good job retrieving information. Then the CIO said, let's roll it out to 8,000 adjusters. Security asked whether customer PII was reaching the model; legal asked who is responsible when the model hallucinates; compliance wanted an audit trail; the data team wanted to know how data is being moved around the organization; operations said they needed five-nines availability; and finance asked what this would cost at scale with token volume. All of a sudden, this $250,000 pilot that was fairly straightforward got stuck. In the EC2 days, a pilot would convert in 60 to 90 days. These pilots are now taking anywhere from 12 to 15 months to convert. - Former Global Executive, AWS Data Center Business at Amazon
Our research goes on to explore how AWS is transitioning from product, developer-led sales to larger enterprise deals, how AWS struck its deal with Anthropic, Project Rainier, AWS sales team compensation, and Trainum vs GPU fungibility.
Uber and Waymo AV Economics
Can autonomous vehicle unit economics (after depreciation) work?
Excluding vehicle depreciation and looking only at operating costs, my math puts Waymo's break-even at around 12 trips per day at their current pricing. These are not publicly disclosed numbers, but the broad math works out to approximately 12 trips per day in San Francisco. Once you add depreciation, you simply cannot make money regardless of how many trips you complete in a day, because a typical trip generates anywhere between $4 to $6 - even at a $75,000 to $100,000 vehicle value, that is the scale of depreciation you are dealing with. - Former Director of Business Development, Uber
Also, results of the Austin Uber project suggest Uber’s marketplace drives 20-30% higher per-vehicle utilisation:
In all three of these market types, our observation was that per-vehicle utilization was significantly better — anywhere between 20% to 30% better compared to what Waymo would achieve running their own network only. That was the broad rationale with which Uber continued to believe that, over the longer run, Waymo would want to operate on the marketplace across markets. That, of course, hasn't played out the same way. - Former Director of Business Development, Uber
The interview goes on to explore how many AVs are required to match Uber vehicle utilisation, how AV economics may change, and how Uber is positioned. This can be read alongside other research covering AVs:
LVMH Pricing
This interview with a former LVMH executive who spent over a decade in pricing and merchandising provides a counterintuitive insight into the company’s pricing guidelines:
When I arrived at the company as a pricing manager, they explained that Louis Vuitton raises prices in line with inflation every year: between +1% and +3% at most. If, from a marketing perspective, they wanted to restructure the price architecture because a new product didn't fit well relative to existing bags, the maximum increase would be five percent. That was the absolute ceiling. - Former Global Merchandising Manager, Louis Vuitton
But this may overlook how it really works. LVMH product and pricing strategy focuses on adding small changes to products to increase the price. Price goes up because LVMH "added more value" to the product:
The Neverfull model, for example, is around 50 years old, and they would never increase the price without adding something. They would add a small pochette inside and then increase the price by EUR 200, or they would change the lining and add a card holder inside.- Former Global Merchandising Manager, Louis Vuitton
And an interesting insight for those tracking LVMH SKU range:
I think you should look at what they advertise. Louis Vuitton is not a brand that typically runs advertisements, but sometimes they do, and that can give you hints about what their challenges are. - Former Global Merchandising Manager, Louis Vuitton
Related Content
© 2026 In Practise. All rights reserved. This material is for informational purposes only and should not be considered as investment advice.