Here is a selection of interviews published last week. Visit our platform for all research published.

Published Last Week

1. AWS: AI Data Center Economics, CapEx, Trainium & TAAS
2. S&P Global Platts: Price Assessment Benchmark Journey & Competitive Moat
3. AI Data Center ROIC & GPU Pricing Framework
4. CARFAX: Data Moat, Pricing Power & the Banking and Insurance Group
5. Microsoft, CoreWeave & the Economics of Neocloud Compute
6. Microsoft & CoreWeave: Neocloud Strategy, GPU Capacity Planning & Off-Balance-Sheet Dynamics
7. BNSF & Aurora: Autonomous Trucking's Impact on Rail Intermodal
8. Intel 14A vs TSMC A14: Advanced Node Competition & Customer Adoption
9. Veeva & Oracle: Pharma R&D Software Platform Switching Dynamics
10. GE Aerospace & CFM: LEAP LTSA Economics, Contract Renewal & Pricing Architecture
11. Netflix EMEA: Building Out Production & Post-Production Infrastructure
12. ARM vs RISC-V: Long-Term Competitive Dynamics in the Data Center
13. SAP, Palantir & Databricks: Enterprise Data Platform Strategy at a Major Oil Company
14. Shopify: Org Structure & Plus Upmarket Evolution
15. Advantest vs Teradyne: HPC Test Market Share & Power Supply Architecture
16. Cytiva & Purolite: Resin Competition & Workflow Integration Strategy
17. Intuitive Surgical: Hospital Capital Allocation & Robot Utilization Economics
18. DoorDash: Owning the Full Restaurant Operating Stack
19. Evolution: South Korea, Payment Blocks & Stolen Games
20. Constellation Software & Vertical EHR: AI Opportunity in Senior Care Software
21. Copart vs IAA: Competitive Dynamics, Progressive, & Market Share
22. Copart: Jay Adair's Return, Culture & Competitive Moat vs IAA
23. AutoZone vs O'Reilly: Commercial Strategy, Culture & Parts Network
24. CBRE vs Colliers: Competing for Large Occupier Accounts
25. Martin Marietta & Lhoist: Lime Industry Dynamics & Acquisition Rationale
26. Coupang: FLC Supply Chain Strategy & Seller Onboarding
27. FTAI Aviation & Chromalloy: CFM56 PMA Blade Yields & Power Conversion Challenges
28. FICO vs VantageScore: Adoption Barriers & Mortgage Market Dynamics
29. TopBuild & QXO: Distribution, Installation & Building Products Integration
30. Hasbro & Chris Cocks: IP Stewardship, Aggressive Partnerships & Leadership Transformation
31. Domino's Pizza: Joe Jordan's Leadership Style & CEO Transition
32. Rentokil & Terminix: Post-Acquisition Integration Failures & Lessons
33. TopBuild & IBP: Insulation Contracting M&A Multiples and Deal Dynamics

AWS, Nebius, & AI Data Center ROIC

A Former AWS Executive, responsible for building out AWS' AI data centers, helps us model out the ROIC of building a 1GW AI Data center.

The model considers the cash payback on gross capex over Year 1-6 including:

1. Utilisation ramping from ~50% to 90%+.
2. Workloads shifting from ~60% training / 40% inference to 70%+ inference
3. Custom silicon; ramp up in Trainium vs NVIDIA
4. Customer commitments, contracted capacity, and asset lifecycle

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

From the operator's experience, the total revenue per GW will cap out at ~$20bn per year as customers productise AI, inference ramps, and AWS can attach Bedrock, Sagemaker, and other higher margin products around the compute. This is lower than some other estimates in the market:

People you're talking about who are not operators run around the industry citing $25 to $30 billion a year. You can get there mathematically with peak throughput and list pricing, but I live in the real world as an operator. The lens I'm giving you reflects real-world experience. In real life, you have to account for utilization, input and output mix, batching, latency, and reserved-tier discounts. Once you get to Trainium 3, the numbers will absolutely go higher. There's no question about that, but they're not going to be higher than $20 billion. - Former Global Executive, AWS

Over the six years, the ramp from ~$3-4bn to ~$10bn revenue per year is driven by the following:

First, utilization rises. If you walked into the Anthropic data center in Virginia today, it may be running at 30% to 35%. In two years, it's going to be running at 70% to 85%. That alone doubles the revenue from the same physical gigawatt. - Former Global Executive, AWS

Second, Trainium expands the workload — you get more compute per megawatt. By lowering the price of compute, you can generate more revenue per physical megawatt because you simply have more compute available to sell. - Former Global Executive, AWS

Third, and this is a significant one, customers are beginning to move from experimentation to production. We're still in a world where most of the spend was on training and pilots. When that switch flips to inference next year, you get much larger, recurring workloads. Training is episodic; inference runs every day. - Former Global Executive, AWS

Fourth — and if I leave you with one thing I hope this makes sense today — revenue moves up the stack. AWS isn't just selling accelerators anymore. It's databases, networking, security, Amazon SageMaker, and Amazon Bedrock. Each dollar of accelerator consumption can drive additional AWS revenue elsewhere. That's the way I'd ask you to start thinking about it. It's simply not about price. - Former Global Executive, AWS

Inference at scale and Trainium 3 adoption will drive higher returns:

If we get to 2028, I'm fairly bullish that all this AI infrastructure, as it shifts to inference and you account for the ramp and accelerator depreciation, will get into a 35% to 40% range. These are AWS numbers. - Former Global Executive, AWS

Interestingly, the price difference of a GB300 NVL72 cluster vs Trainium 2 is only 10-15% given the steep discounts AWS receives from NVIDIA:

It's probably going to be 10% to 15% higher. AWS has some significant discounts in place with NVIDIA. You'd probably be surprised by how aggressive NVIDIA's pricing is for AWS. So in reality, there isn't a particularly large cost difference. If you wanted to be generous, I'd say it's maybe around 17%. - Former Global Executive, AWS

But this may change with Trainium 3.

On a cost-per-FLOP basis, AWS Trainium 3 is going to change the economics significantly on the IT compute side…If you look at a Trainium 3 campus, you're at around $30 billion, whereas an NVIDIA Frontier campus for AWS is going to be closer to $40 billion. A fully loaded NVIDIA GPU, comparable to Trainium 3, will be $40 billion per gigawatt…This is the first time AWS fully controls the silicon, the server, and the cloud economics. The other thing to consider is effective capacity. AWS Trainium 3 can deliver dramatically more useful training and inference than a prior Trainium campus. The potential ROI is much better because if you're driving three to four times more throughput through the same secured megawatt, the economic value of that megawatt rises accordingly. - Former Global Executive, AWS

Trainium 3 lowers the unit cost of compute:

what Trainium 3 does, is it changes what you get for your $40 billion [capex]. The compute per gigawatt can really move the needle dramatically. If AWS gets this right — and I think they will — they'll compress the cost per token and the cost per unit of training, and improve power utilization. That's where you materially begin to improve the return. - Former Global Executive, AWS

On the capex side, a former Nebius CFO claims that the company built 1GW of H100s for ~$20bn that generates an estimated $50bn over five years as a bare metal service.

Including server racks, if you buy the land and design the building for the right cooling, you get to a maximum of $8 billion. You can add $12 billion for the H100s which makes it $20 billion. With that 1 gigawatt you can make $50 billion revenue without additional specific revenues. - Former Group CFO, Nebius Group

The CFO claims Nebius spends ~$6-8bn in non-compute capex per GW compared to ~$20bn for hyperscalers. This seems to be partly due to:

1. Nebius’ strategy of buying and designing the buildings, and using cheaper contractors and fit-outs
2. Hyperscaler FOMO and being willing to pay up to put up capacity fast
3. Differences in data center locations

We were building in Finland with H100s, excluding GPUs, for $6 billion, including land and building. The only thing not included in that $6 billion was server racks and GPUs. In an H100 calculation, $27,000 per GPU is very different from $65,000 for a Blackwell Ultra. Location matters greatly as the cost differs significantly. If you build near New York or New Jersey, as Nebius is now doing for Microsoft, it is much more expensive than in rural Ohio. If you do a lot yourself, you can save significant amounts...We built with a local construction company — not the most expensive one with a fancy name. - Former Group CFO, Nebius Group

From a pure data center point of view — and when I say data center, I mean the land, the site, the shell, the electrical, the mechanical, the liquid cooling, the generators, and the internal networks — you're looking at about $12 to $15 billion in CapEx. Then your power and transmission — outside the fence — is probably another $6 to $8 billion per gigawatt. - Former Global Executive, AWS

You can read more on the cost and revenue per GW in the first four interviews listed below and more about neocloud vs hyperscaler data center operating models in other interviews listed:

1. AI Data Center ROIC & GPU Pricing Framework — Former Group CFO, Nebius Group
2. Microsoft, CoreWeave & the Economics of Neocloud Compute — Former Vice President at Oracle Corporation
3. AWS: AI Data Center Economics, CapEx, Trainium & TAAS — Former Global Executive, AWS Data Center Business at Amazon
4. Microsoft & CoreWeave: Neocloud Strategy, GPU Capacity Planning & Off-Balance-Sheet Dynamics — Former General Manager, Azure AI Engineering at Microsoft
5. Azure AI Foundry: Model Routing, Quota Constraints & Frontier Model Economics — Former General Manager, Azure AI Engineering at Microsoft
6. Google TPU vs NVIDIA GPU: Training, Inference & Total Cost of Ownership — Former Technical Program Manager, AI Infrastructure at Google
7. Nvidia, Intel & the GPU-to-CPU Ratio in Agentic AI Infrastructure — Former Global Director of Intel AI Products and Strategy, Intel Corporation
8. Microsoft Azure AI: Capacity Planning, GPU Allocation & Internal Prioritization — Former General Manager, Azure AI Engineering at Microsoft
9. AWS Bedrock: Infrastructure Moat vs Model Provider Competition — Former Product and Strategy Lead, AWS Agentic AI at Amazon
10. AWS: Unit Economics, Infrastructure Scaling, and the AI Transition — Former Finance Manager, AWS
11. Microsoft Azure: AI CapEx Strategy, GPU Economics & OpenAI Contract Structure — Former VP, Azure AI Infrastructure, Optimized Workloads and Storage at Microsoft
12. AWS: Securing Grid Power for Hyperscale AI Data Centers — Former Executive Leader, Energy Strategy and AWS Infrastructure at Amazon