LVP taps OpenClaw, Claude in hybrid cloud AI play

LVP, a sustainability-focussed investment fund with $1.6 billion in assets under management, is re‑engineering its investment process around AI – but only after rebuilding its cloud and data foundations.

This story is part of the 2026 iTnews Cloud Covered report. Read it for free here.

Head of Data & AI Ivan Wong said LVP, which has invested in everything from cleantech startups to affordable housing schemes, is “moving aggressively on AI”.

“We see it as a way to free up our people for higher-value work, and we think organisations that haven’t deeply integrated AI into their workflows over the next two to three years face an existential problem,” he told iTnews.

AI agents are used to support LVP’s investment team by “automating research, surfacing context on past deals, and helping analysts produce first-draft memos,” said Wong.

“This isn't just AI as a coding assistant for engineers. It's first-draft financial models, presentation decks, and legal contracts for every knowledge worker. The underlying architecture uses Claude via Anthropic's Application Programming Interface (API), with an agentic orchestration layer that I've built in-house to connect our data and tools via a self-hosted Model Context Protocol (MCP) server,” Wong added.

That stance also feeds directly into LVP’s investment decisions: “We want to see management teams with a coherent AI strategy, not just AI talking points,” Wong added.

OpenClaw, on‑prem first, then cloud

A central architectural choice at LVP has been how to let AI agents act autonomously without creating new security risks.

“The biggest shift this year was standing up an on-prem sandbox after the OpenClaw release in February,” Wong explained.

“We needed somewhere we could let AI agents act autonomously without data security exposure, and learn what guardrails were actually necessary before deploying agents into our cloud environment.”  

In his view, sequencing “on-prem first to build conviction and cloud second”, has worked better than starting in the cloud and “retrofitting controls”.

This hybrid stance is also how LVP is thinking about sovereignty and confidential workloads.

“This is tricky if everything you run sits on cloud, which is part of why we went on-prem with OpenClaw,” Wong said.

He quickly saw the potential of the open-source AI agent upon its release in February and set about experimenting to explore its potential to automate workflows at LVP.

“I think we’ll see more hybrid setups going forward, where on-premise hardware running local models handles the confidential workloads and cloud is reserved for everything else.”

The difficult part then becomes a skills issue.

“The question that creates is whether your workforce has the capability to actually secure that hardware and your internal network. That’s where most organisations will get stuck,” Wong said.

Standardising on Claude, keeping the stack flexible

On the model side, LVP has narrowed in on a preferred large language model – Anthropic’s Claude.

“We benchmarked it against other frontier models across multiple use cases and Claude consistently came out ahead for our work,” said Wong.

At the same time, LVP is avoiding model lock‑in: “Our orchestration layer and agents are model-agnostic by design, so if another provider pulls ahead, or Claude’s performance doesn’t hold up, we can switch without rebuilding.”

Wong also directly links architecture control to AI economics. 

“I have the advantage of owning both the architecture and the implementation, so decisions get made on output quality rather than procurement cycles,” he noted, contrasting this with larger enterprises where “existing partnerships, risk committees and governance layers slow things down considerably”.

His view on cost is blunt: “If you’re optimising for revenue or outcomes, compromising on model quality to save on inference cost is a false economy.” He added that many failed AI pilots come from “using cheaper, less capable models to hit a budget target, and then wondering why the outputs don’t justify the spend. The cost question is real, but it’s downstream of getting the model choice right first.”

People, security and governance

For LVP, the biggest AI challenge is not data plumbing.

“The hardest part isn’t technical, it’s the human transformation,” Wong said.

“People are naturally hesitant to change how they work, and getting non-technical teams to move from familiar BAU processes to AI-driven ones is harder than any data quality or integration problem I’ve dealt with.”

With enough in‑house AI capability, “the change management work is what’ll actually slow you down.”

Security and vendor risk are also front of mind as cloud platforms and SaaS products add AI features.

“The category that worries me most is AI features being bolted onto existing SaaS products and new AI-native startups shipping fast without mature security practices,” he said.

“The attack surface from these integrations is often wider than the vendor has matured into. We’re spending a lot more time on vendor due diligence as a result, auditing the underlying tech rather than taking marketing claims at face value.”

On governance, Wong’s position is that standing still is not an option. “If you’re not experimenting with AI, you’re falling behind people who are,” he said, even if “Australian-market disruption is slower to land.”

For most roles, “the governance question isn’t whether to experiment, it’s how to do it without creating data exposure or compliance issues.”

LVP’s response is to “make the safe path the easy path, so people aren’t tempted to route around it.”