Wise, Hyperscalers vs Labs, Meta
Here is a selection of interviews published last week. Visit our platform for all research published.
Published Last Week
AI Labs, Hyperscalers, & Open Source Models
In a world of token optimisation, there is a perspective that the LLMs are becoming commoditised. 'Commoditise your complement’ strategy 101:
LLMs are becoming a commodity. The focus is shifting toward providing governance for companies using tens, hundreds, or thousands of agents, or, for those with a single AI SaaS product, ensuring control across all three levels - the data, the workflow and the front door into the application.- Former General Manager, Azure AI Engineering at Microsoft
And in a world of commoditised models, value may accrue to the orchestration layer. Nadella publicly said Microsoft aims to ‘decouple the harness from the models’. Hyperscalers like Microsoft and Amazon have an incentive to neutralise the labs and offer more, and often open source, models. But this doesn’t seem to align with the physics of how GPUs are run efficiently:
Having fleets of GPUs running open source models is inconsistent with running high-revenue, highly profitable businesses. I'm not buying that the hyperscalers will flip to being open source model delivery platforms…It generates fine-tuning revenue — clients fine-tune, customize, and build on top of it — but that is the opposite of what Microsoft is optimizing towards. Microsoft is optimizing towards high-scale inference on frontier models with no model fine-tuning and no custom model development. That is how you run a million GPUs at scale. Providing small pools of GPUs for customization work is a terrible business with terrible gross margins. You need to run base model inference at massive scale. The entire model customization and fine-tuning of open source models is a very low-margin business because you have to isolate small pools of capacity for low-volume models. It undermines fleet economics and fleet efficiency. It is a poor business model. - Former General Manager, Azure AI Engineering at Microsoft
Neutralising OpenAI and Anthropic is important. But how may this impact hyperscaler margins if there is a significant increase in adoption of fine-tuned open source models?
From a hyperscaler's perspective, you need to run one global fleet of roughly three million GPUs, running as few models as possible, without any customization or fine-tuning, fully connected to your agent stack and so forth. That's the only way you can run a profitable business. The world of open-source models is not a profitable business. You can't run it efficiently because there's no fungibility — you end up with small pools of dead capacity in data centers around the world. It simply doesn't work that way. - Former General Manager, Azure AI Engineering at Microsoft
This emphasises why MAI may be so important. Microsoft can route demand to MAI and the frontier models to both neutralise the labs whilst driving fleet efficiency.
The only things that really matter at Microsoft right now in terms of models are OpenAI, Anthropic, and MAI. Those are the three that matter. The rest is window dressing. - Former General Manager, Azure AI Engineering at Microsoft
The interview also explores how labs are becoming full cloud providers:
If you look at any of these companies, it's a journey of abstraction off the silicon. Take OpenAI as an example — they started extremely tightly coupled to Microsoft. Back in 2022, anytime they needed more GPUs, I had someone on my team manually shifting GPUs over to the fleet. They were using the entire Microsoft tooling stack. Compare that to where they are today, managing their own fleets and getting bare metal compute from multiple providers. It's a journey from tight coupling to a cloud provider, to shifting up off the GPU and running their own stack. They're already on that journey. - Former General Manager, Azure AI Engineering at Microsoft
The delta between models commoditising and hyperscalers running a more complex GPU fleet is discussed in more detail in the following interviews:
Wise and Kristo Käärmann Reference Checks
Last month, there was a press report of a Belgian investigation surrounding illegal casinos and other criminal operators using Wise to transfer ~500m EUR of payments. Two former Wise compliance executives believe this could partly be due to compliance debt from early rapid growth.
What Wise had adopted was a threshold-based application of this risk-based approach, which equated risk with transactional volume... That is not what I would consider an orthodox interpretation of risk-based approach guidance... while transaction value can be one factor in assessing risk, it should not be treated as the sole determinant. - Former Head of KYC, Wise plc
We’re also exploring Käärmann’s leadership style in an IP management reference check that shares a perspective into the company's culture:
I think Wise is addicted to chaos. There's a sense there that the company can deal with anything: problems arise and they get solved. That's the whole ethos…He is without a doubt the most analytically intelligent person I’ve ever met. From my observations, his emotional intelligence levels are quite low. He's a very driven person with a particular thinking style…conversations with him are difficult. - Former Head of KYC, Wise plc
The interviews below go on to explore the company’s culture and founder in more detail:
Meta Reality Labs
Since 2019, Meta has racked up cumulative losses of ~$90n on Reality Labs without much to show for it to date. Maybe no one wants to speak up to Zuck?
My take is that no one wanted to bell the cat. No one wanted to say, "We can't do this in this timeframe." The response was always, "Give us more money and we'll get it done." That, in my view, is the core issue. - Former Hardware Finance leader at Meta Reality Labs
My larger point is that a lot of different investments were made, and then the decision was made not to proceed with them. Money was spent, but nothing came of it. For example, a company was acquired for the sensor technology, and those sensors were eventually integrated into the watch — and then the watch was cancelled. - Former Hardware Finance leader at Meta Reality Labs
The interview shares more insight into the history and internal operations at Meta Reality Labs and can be read alongside:
Related Content
Azure AI Foundry: Model Routing, Quota Constraints & Frontier Model Economics
Meta Reality Labs: Hardware Investment Strategy & Product Roadmap Economics
Reddit Advertising: B2B Targeting & Subreddit Contextual Strategy
CCC Intelligent Solutions: AI Adoption, Pricing Pressure & the Orchestration Challenge
© 2026 In Practise. All rights reserved. This material is for informational purposes only and should not be considered as investment advice.