OpenAI, AppLovin, S&P & Moody's Analytics
Here is a selection of interviews published last week. Visit our platform for all research published.
Published Last Week
OpenAI
How much can partners such as Accenture trust OpenAI?
A former Microsoft engineer, who was involved in the early days of the OpenAI relationship, shares a glimpse of OpenAI’s approach to ‘partnership’. In 2024, Microsoft's internal forward-deployed engineering team signed large, long-term contracts with two telecom companies.
I had about 40 people split across two large projects. One was with Korea Telecom and one was with G42. Both involved custom model training taking GPT-4.0 at the time and creating bespoke Korean and Arabic language versions of OpenAI models to ensure it was correctly being delivered as completions with cultural nuances.— Former General Manager, Azure AI Engineering at Microsoft
While OpenAI was informed and agreed to work with Microsoft on these deals, OpenAI added the functionality to its next model and killed the projects:
OpenAI holds IP to any work done... They were aware of it, and they simply went ahead and built those capabilities directly into GPT-5... we all knew it was a waste of time, because those capabilities were inevitably going to be trained into the models…These projects failed miserably because by the time the projects were done, these were 12 to 18 month projects, GPT-5.0 was there, so rendering the 4.0 model was uncool, unsexy, and unnecessary. — Former General Manager, Azure AI Engineering at Microsoft
This seems fair game. OpenAI are building general intelligence into the model for all users. But what does this mean for customers and ‘partners’ of OpenAI?
How much access to internal workflows and proprietary processes should a company provide a company like OpenAI? What can and cannot they build into the model?
Accenture’s relationship with OpenAI comes to mind.
Accenture is a founding member of OpenAI’s Frontier Alliance to move enterprises from experimentation to scaled deployment of OpenAI’s agent platform. Accenture Federal Services is also an OpenAI Implementation Partner for federal agencies. How much has the company opened up its business processes to OpenAI in the hope to build agents that deliver value to enterprises?
Accenture seems to be betting on a shift from selling time and materials to building pre-packaged industry agent solutions that sit above the model layer. Accenture has built decades of what we call process intellectual property or unique business know-how. Its long history and reputation serving enterprises will likely maintain its position as a critical distribution layer for technology companies. But just how much profit Accenture captures in the agentic enterprise value chain will be partly determined by how it protects its proprietary processes and workflows from the labs.
We’ve seen this disintermediation risk play out in various other industries. Amazon scaled its 3P marketplace and then built 1P products to undercut the 3P suppliers. It learned which products converted and why, and then took their business away. Supermarkets and private labels also do this all the time. Is there any reason why the models won’t follow this strategy? This is all the more relevant for OpenAI given its weaker unit economics compared to Anthropic's supposed API margins.
Sergey Brin also made an interesting comment:
when you train for a certain class of problems, let's say you're training for coding, that that actually can help your math reasoning and vice versa. When we were training this model, we didn't expect it to come out nearly as powerful as it did or to have all the capabilities that it does. I honestly don't really understand how a model does that…even the people building these models do not fully understand what they have created - Sergey Brin, DeepMind Build Day, June 2026
This leads to the following dynamics at play:
- OpenAI with a history of disintermediating partners
- Models having some unknown, latent capability to develop new skills
- Engineers training the models not truly understanding how this works
- Increasing token optimisation and the threat of open source models
This combination may encourage OpenAI to increasingly look to disintermediate partners. A question for Accenture and all partners and entrenched customers: if a lab can absorb a bespoke Korean- and Arabic-language GPT-4 build into its next base model, what stops it absorbing process IP Accenture is now feeding into OpenAI's agent platform. Language translation may well be simple. But where is the line in what models can do? And how much of the agentic value chain is left for the distribution layer to keep?
S&P Global
Proprietary data companies (a topic close to our heart) must also take note of the analysis above. AI impacts S&P’s ratings and indices business differently than datasets within Market Intelligence such as CapIQ.
Ratings seem more protected given the issuer provides proprietary information to the agency to issue the rating:
About the rating agencies, what I say to people is they have a methodology, and the methodology for every industry is quantitative and qualitative. People have said, can't I just replicate that? My answer is that the analyst working with an issuer on that credit is getting proprietary or private information from bankers of the companies that you will never get; you will never have. - Former Senior Vice President, Data Solutions at Moody's Analytics
Ratings are set by committee-level debate of a cross-sectional group of analysts:
Let's say you're rating Cargill; it won't only be the Cargill analyst, it will be someone from the sovereign or banking or different areas to come in and sit on that committee for that rating. They vote by committee in the end; that is how it works. There is no machine to do that. I don't know how you could get a machine to have different analysts debate it out, but every person gets a vote, and even the chairperson does not get to overrule the committee. They have one vote, and that's how it goes by committee. - Former Senior Vice President, Data Solutions at Moody's Analytics
Market Intelligence and CapIQ specifically seem less protected. A former Moody’s Analytics Director estimates ~30% of Moody’s Analytics portfolio can be scraped by LLMs:
I am saying maybe 30% could be scraped; I am starting out conservative on that. - Former Senior Vice President, Data Solutions at Moody's Analytics
The proprietary nature of data is nuanced. Public data is commoditised. ‘Branded data’ not so much:
S&P's specialty is clearly rooted in ratings. The S&P rating is extremely valuable, and its distribution is irreplaceable — there is only one source for an S&P rating, which is S&P itself. Even if you access it through Bloomberg, you are purchasing it under a redistribution license. Similarly, S&P index data is irreplaceable; there is only one source for the S&P 500, which is S&P. This is where you distinguish between branded data and non-branded data. Some things can only be obtained from S&P — ratings, index data, and GICS, their market sector classification, are prime examples. - Former Chief Data Officer, S&P Global
Distribution is also changing. Is the terminal dead? Why use someone else’s UI when you can design your own and pipe third-party data in to combine with your own internal data? This is where the market seems to be heading:
People are looking at whether they want to use a workstation like a Cap IQ workstation with all the data... or whether they want to build their own tools with the data. AI makes it more possible. I can take my own internal proprietary data and mix it with their data... and I can build agents or chatbots or my own version. — Former Senior Vice President, Data Solutions at Moody's Analytics
they don't want to lose these workstations, but the day is coming that they're not going to be able to hold onto them. When that is exactly, I've been in meetings for years with people sitting at the banks and financial institutions saying they don't want everything a Refinitiv terminal has, they don't want everything that a Bloomberg has. They want to build their own. Now with the technologies and components for the web and AI and all this, you're getting into that world. - Former Senior Vice President, Data Solutions at Moody's Analytics
Due to customers more willing to access data via feeds rather than through a terminal, management recently claimed customers are willing to 35-45% more for S&P’s AI-ready data:
Perhaps maybe one other example I would provide is, in the quarter two financial clients who are just subscribing to our data at renewal, were opting to get that data available in an AI-ready format. We're willing to pay in the range of 35%-45% on the renewal increase to get the AI access. Again, early days, but some very strong signal here around the monetization from an enterprise value standpoint. - S&P CEO, Q1 26
But piping data via API still needs to be protected from public LLMs.
One thing all these companies have to be very careful of is which large language models they are feeding. The advantage these companies have is their data and their IP, and they need to protect it. They have to be careful not to push all of that data into the public LLM domain, because then they lose their advantage. - Former Chief Data Officer, S&P Global
Whether AI-ready pricing is a durable premium or a bridge to disintermediation is the question the interview works through for various S&P product lines.
AppLovin
AppLovin has ~1bn active daily gaming users on mobile games globally. These users are reached via the MAX mediation SDK which is on nine out of 10 games. AppLovin controls the supply.
MAX is simply mediation for publishers earning money, while direct Ads Manager campaigns run on the Axon platform serve into those publishers. Some of the competitors mentioned, like Liftoff and Moloco, are also using it because nine out of ten gaming apps have AppLovin's SDK integrated. Those competitors are actually buying AppLovin supply to run their campaigns on. - Former Sr. Growth Director, AppLovin
Controlling supply and seeing every advertising bid seems difficult for entrants to replicate. AXON matches supply and demand and drives the highest ROI for advertisers and payouts to app developers:
Today the majority of ads go through the AppLovin pipes on Google Android and Apple iOS, giving AppLovin view on every bid... Combine that visibility with their data advantage, it becomes difficult for new entrants to the market to compete. - Former Director at AppLovin
AppLovin is now aiming to funnel non-gaming advertisers to its controlled supply. Those 1bn active gamers don’t only want to play new games; they want to buy cookware, beauty products, or (maybe) insurance:
The billion+ daily active users are not just gamers… They're also not just gamers and shoppers for D2C products… being able to service them with financial services offers, health insurance, auto insurance, is a big part of our strategy - Applovin CEO, Q1 2026
This marketing executive at a beauty e-commerce company explains how Applovin fits in the ad budget:
What's interesting about AppLovin is that the ads are non-skippable, typically 15 to 30 seconds, and someone has to watch the entire thing. When we took our Meta ads and put them on AppLovin as a test, I think that's why they worked so well - Former VP of Growth at leading Beauty Brand
The advertiser suggests AppLovin inventory and ROI is incremental to existing channels. It’s not currently eating into Meta’s budget, for example.
If AppLovin went away tomorrow, it would probably feel like it did before we had them. As I mentioned, it felt very incremental to our ad spend, so we probably wouldn't reallocate that budget anywhere…The number of sites that have AppLovin's pixel on them, and the data they can capture even without serving ads, is just a fraction of what Meta has - Former VP of Growth at leading Beauty Brand
Meta has pixels on ~10m websites, AppLovin is still in the thousands. There is a long road ahead. Whether AppLovin’s supply advantage reaches non-gaming budgets depends on outcome, intent and identity data AppLovin does not yet have for e-commerce. The following interviews explore how quickly it can close that gap.
Related Content
© 2026 In Practise. All rights reserved. This material is for informational purposes only and should not be considered as investment advice.