The embed-and-build model goes mainstream.
AWS put $1bn behind embedding its own engineers inside customers to build their AI. The embed-and-build model has gone mainstream, and it is the clearest signal yet that the value is in rebuilding the workflow, not buying another licence.
Hi folks,
This week AWS, the biggest cloud provider, put a billion dollars behind the exact way we work: sending engineers to sit inside a business and build the AI with them, rather than sell them a licence.
The one thing
The embed-and-build model continues to go mainstream
- AWS committed $1bn (30 June) to a new Forward Deployed Engineering unit: teams of five to six engineers embedded inside customers for roughly 45-day builds, shipping agentic AI alongside their own staff. Early partners include the NBA, NFL and Southwest (Amazon).
- OpenAI, Anthropic and Google have all stood up the same kind of unit this year. Palantir coined the term a decade ago; it has now reached the largest AI vendors (CNBC).
What this means in plain terms:
- "Buy a licence and roll it out" is giving way to "have someone build it with you, inside your systems." The vendors concluded that software alone does not move adoption; embedded delivery does.
- The honest read on why internal AI stalls: the gap is not tools, it is someone senior sitting in the workflow and rebuilding it. It's the education and everything else we discuss in our weekly newsletter.
In the mix
- Claude Sonnet 5 is the new default (Anthropic)
- Near-flagship quality at a fraction of the price, a 1M-token context, built to run multi-step work on its own. Intro pricing $2 in / $10 out per million tokens to 31 Aug.
- Why it matters: pointing a capable model at analyst-length work (a data room, a model roll, an IC memo) got cheaper again. More of the workflow is now economic to automate.
- The switched-off models are back, but gated (CNBC)
- Fable and Mythos returned (1 July) after June's export controls were lifted; the most powerful tier now comes back only to government-approved firms. OpenAI shipped its newest models the same way and reportedly offered Washington a 5% stake to ease pressure (The Next Web).
- Why it matters: model access is now a policy variable. If your stack leans on one US frontier lab, "who controls access" belongs in diligence next to price and uptime.
From my week
- The three questions we get in almost every fund conversation
- "We know we need to do something with AI, we just don't know where to start or who to trust." The most common opener, from MDs and analysts alike.
- "Where does our data go, and will it train on our data?" Usually the IT-side blocker, and often the deciding one.
- "We're a Microsoft shop, won't IT just wait for Copilot?" What lands: the work is implementing AI behind the Microsoft stack, permissions and SharePoint included, not around it.
- Building the monitoring, learning to trust the output
- Most of the week went into loan-monitoring tooling for a finance team: ingest the monthly pack, roll the model, flag covenant breaches, with fail-closed checks so nothing passes on a guess.
- The unlock is evals: take a run you have checked and know is right, and make it the bar every future run must clear. If a person cannot verify the number without redoing it by hand, the tool is not done (Hamel Husain).
Levercon is the AI team for investment funds. We embed alongside your team to automate workflows, build custom tools, and compound operational gains. Ultimately we help Australian based investment funds leverage AI.
Book a strategy call.
See where AI fits in your fund.
We'll show you where AI lifts efficiency and performance across your organisation.