A collection of interviews published last week. Visit our platform for full coverage
Linde & Industrial Gas Competition
It’s interesting to compare Linde and industrial gas plant to AI data center capex. A neocloud or AWS AI customer will typically have a 3-6 year contract for a minimum level of capacity. And AWS sources power from utility companies on 10-20 year contracts. GPUs, which are ~70% of the data center capex, also only have 5-6 year life. It seems much more favourable for Linde, Air Liquide and industrial gas companies. Anchor customers limit the risk of building new capacity and are locked in for 15-20 years. The plant can also be used for ~30 years.
The anchor customer contracts are 15 to 20 years. It's also nearly impossible to dislodge a supplier in the anchor space — even if the 15-year contract expires, the supplier already has a completely paid-for plant. There's a golden period after that for another 15 years, because these plants last forever. It's very tough for someone else to come in, so it's a very profitable business. - Former Senior Vice President at Linde plc
The industrial gas competitive game also seems fairly rational between three players that own a fairly even combined ~80%:
Competitive bidding between major producers does compress returns during the negotiation phase. Major gas companies rarely approve a combined project where the projected after-tax return for the anchor and merchant business falls below 11%. - Former Senior Vice President at Linde plc
If all players are bidding on similar projects with similar pricing, operational efficiencies become a key driver in driving differences in unlevered returns:
The old Linde could ever manage to get to that Praxair-type operational philosophy before the merger — and it took that merger to do it. With a European company, the shareholder is often not the most important stakeholder. You almost feel like employees, the community, and others take precedence. It is very difficult to create that model... I am very skeptical that Air Liquide has the culture to deliver it. I would almost rule it out entirely. - Former Senior Vice President at Linde plc
Read more on the economics of industrial gases and competition below:
S3 & Data Gravity in AI Workloads
An attraction of the traditional cloud computing business was the concept of data gravity. Once a workload runs through EC2 and S3, it’s uneconomical for customers to move the data and compute to another provider.
But how does data gravity compare for AI workloads run on GPUs or custom ASICs?
S3 and AWS’ storage products are typically 25-30% of a customer’s cloud bill of which AWS reportedly earns >90% gross margins:
In the past, servers were attached to hard disk drives. Today, cloud providers place SSDs directly next to the CPU and GPU within the servers as Direct Attached Storage, bypassing the separate block storage rack to maximize speed. Normalizing the pricing shows that tape costs $1, Infrequent Access is $12.50, standard S3 is $23, and EBS averages $100. Direct Attached Storage costs roughly $200 or more per terabyte per month because it delivers top-tier performance. - Former Head of Storage & Memory at AWS
Porting data from S3 to run workloads on a neocloud or inference provider like Fireworks AI is a difficult technical challenge:
Kubernetes handles compute workloads because compute is stateless. Storage is stateful, which introduces different constraints. The challenge relates to data synchronization and transactional consistency. If an employee creates a document summarizing sales figures and stores it in S3 within the AWS ecosystem, that data must be replicated to a separate object store, like Backblaze, utilized by CoreWeave. If another team member queries the neocloud instance before synchronization completes, they will receive stale or inaccurate information. Achieving real-time synchronous replication across two separate, independent object storage providers is difficult. Most cross-cloud systems rely on batch replication at the end of the day, meaning the data states are inconsistent during operations. - Former Head of Storage & Memory at AWS
As in traditional workloads, data gravity seems to drive stickiness for incumbent enterprises using AWS for AI workloads:
Data gravity remains a strong retention mechanism. Enterprises that have spent years accumulating databases, EBS volumes, and object storage within AWS face high technical and financial hurdles to migrate that data out. The battleground is centered on new users and pure AI startups who do not have legacy data dependencies.- Former Head of Storage & Memory at AWS
Hyperscalers with object storage also have an advantage today over Neoclouds without a complete offering:
Neocloud providers like CoreWeave, Nebius, or Crusoe do not possess native object storage architectures like S3. They rely on SSD configurations similar to Express One Zone deployed directly on racks. This is eight to 10 times more expensive. As costs escalate, these providers must differentiate between hot data needed immediately for active sessions and cold data like old PDFs or presentation decks, which should move to object storage. This creates an opportunity. When a user accesses a five-month-old session, the data must be retrieved from hard disk drives, moved onto the SSDs attached to the GPUs, then loaded into HBM to recalculate the context. - Former Head of Storage & Memory at AWS
This can be read alongside:
NVIDIA vs Hyperscalers
Even if supply wasn’t an issue, Anthropic seems to prefer using ASICs vs GPUs:
Even if we were not compute constrained, I think the focus would be more on hyperscaler AI ASICs, because they are now so much more performant. I think there will be three categories of compute chips that really persist. The first is the existing incumbent GPUs, NVIDIA and AMD. Second are the hyperscaler AI ASICs; Microsoft just announced its Maia 300 chip yesterday, Meta has MTIA, and everyone is building. The third category will be frontier labs building their own chips, like Jalapeño and the one Anthropic is building - Former Senior Executive, AI Accelerators & GPUs at Anthropic
Google's TPU price per performance seems to be becoming more and more attractive relative to GPUs:
When you come to serving inference, hyperscaler AI ASICs like TPUs are really good, because even if you get 80% of the performance at 50% of the cost, it is still a very good deal. Google can fundamentally rent out TPUs much cheaper than it can rent out GPUs; if you go to Google Cloud right now, you will see that difference — Former Senior Executive, AI Accelerators & GPUs at Anthropic
The price performance seems to be better because ASICs are specifically made for math needed for AI workloads:
That is why AI ASICs from hyperscalers are the custom solution for the very specific mathematics needed for AI workloads, matmuls or matrix multiplications. AI workloads are fundamentally multiple dimensions of numbers that need to be multiplied with each other. - Former Senior Executive, AI Accelerators & GPUs at Anthropic
The interview goes on to explore how NVIDIAs hardware and software stack is under threat from hyperscalers:
Klarna Replacing Salesforce
Klarna switched out Salesforce, Workday and various other tools with internal AI-built solutions:
Most tools were replaced. Workday was simply replaced with something else. In my six years at Klarna, I think we had five different HRIS systems. I don't think that was AI-driven; they were just never happy with anything. But Salesforce was 100% AI-driven. Our customer service tools were also replaced, or at least partly replaced. We built a whole AI infrastructure for customer service instead. I'm sure it's much better now, because back then it wasn't the same. - Former Director at Klarna
The interview explores how Klarna replaced SaaS tools, the impact on the business, and how the CEO is focused less on building a complete neobank offering compared to secret non-core projects:
I don't think he feels it's the priority. Similar to the whole GenAI initiative he drove, he has other ideas on how to win. He runs these secret projects. - Former Director at Klarna
© 2026 In Practise. All rights reserved. This material is for informational purposes only and should not be considered as investment advice.