Insights · AI & Data Infrastructure · Insight · 30 September 2026

Why AI infrastructure buyers read the practitioner first

AI infrastructure vendors spent the week arguing in public that their own technology needs restraining [5], while an economist questioned whether the spending returns [4]. None of that helps a VP Engineering choose anything. The reads that travel in this category come from named practitioners who publish their method and their defect numbers [1][3].

Why a vendor's own warning carries so little weight

The Verge counted the AI executives who spent recent days calling for the industry to slow down: Sam Altman, Dario Amodei, Demis Hassabis, Satya Nadella, and others who stand to make a great deal of money from the technology they want restrained [5]. Read it as a buyer and the problem is obvious. A warning from someone selling the product tells you nothing about which product to buy, because it is priced into both the warning and the sale. The same week, MIT Technology Review covered Jessica Wachter, a finance professor at the University of Pennsylvania, assessing what AI does to the economy over the next few years [4]. That is the macro question. A Head of Data Platform holding a renewal quote has a narrower one: does this cut our inference bill without costing us reliability, and who has seen it do that. Neither the executive warnings nor the economics answer it.

What Duolingo published that a buyer can use

Sarah Deitke's account of AI adoption at Duolingo is the opposite kind of artefact [1]. It describes internal AI literacy workshops, observability dashboards, and a redesigned code review built around an automated risk-assessment bot on pull requests. The claim attached is narrow and checkable: pairing developer education with guardrails sped up delivery without increasing defect rates. Note what makes that usable. There is a named person, an organisation, a mechanism, and a metric that could have come out the other way. A CTO can take that to their own team and ask whether the same instrumentation exists here. No vendor deck offers that, and not because vendors are dishonest. A vendor cannot publish the case where their product raised the defect rate, so the absence of that case in their material tells you nothing either.

Evidence that survives a question about cost

Irakli Betchvaia's piece takes the same shape on a harder problem [3]. The approach, Typed Domain Grounding, embeds a domain-specific language inside a mainstream typed language so the compiler can reject model output, with generate-compile-repair loops and kUML benchmarks behind the claim [3]. Whether the approach holds is a question for an engineer. What matters to a buyer is that it arrives with benchmarks, so it can be argued with. Set that beside MIT Technology Review on the materials side, where semiconductors and data centres are running into limits on thermal management, electrical efficiency and reliability [2]. FinOps scrutiny of inference cost is downstream of physics now, and a committee under that scrutiny needs evidence that survives being checked. Benchmarks survive. Positioning does not.

Where trust concentrates in this category

Put the week together and the pattern is not about volume of content. Deitke and Betchvaia both published something that could be wrong, and said how you would tell [1][3]. That is the whole mechanism. In a category where the loudest voices have a position in the outcome, credibility accrues to whoever shows their instrumentation. For a revenue team selling AI infrastructure, that has an uncomfortable implication. The people your buyers read are engineers with their own reputations to protect, and they will not repeat a claim they cannot defend in a thread. A sponsorship changes nothing about that. What changes is whether they have access to a real deployment with numbers they are allowed to publish, which is a product and customer-success question long before it is a marketing one.

What this would look like once measured

One closing note on measurement. Everything above is a claim about where credibility sits, and a revenue team is entitled to ask for the number. Influential Partners builds attribution from a creator's post through to the named accounts that engaged and the pipeline that followed. Two limits belong in the same breath. The trail starts at a click, so a piece read and acted on days later in a private conversation leaves no trace, and measured influence is a floor rather than a total. And audience composition is checkable only where a paid boost returns demographics from the platform. Self-declared follower counts stay self-declared, and they carry that label wherever they appear.