Over two days at SuperAI in Singapore this June, AI founders and researchers, corporate and technology leaders, and investors and advisers discussed a broad array of unresolved tensions around AI, along with some insights that are coming into focus.

We heard enthusiasm for the rapid progress of open-weight models, alongside continued respect for closed frontier models that still define the leading edge of capability. We heard conviction about the future of centralized hyperscale compute, while others argued that trust in the open Internet is eroding and pushing some activity back toward verified, local, or private networks. AI was discussed as a productivity tool, but again and again, the more interesting questions were about workflows, products, customer experience, and business models. Agents promised autonomy but also raised questions of control, accountability, and risk.

Over the past three years, the dominant conversation around AI has focused on access: Which model is best, who has the largest cluster, how quickly is the frontier advancing? What struck us most at SuperAI is how this conversation is changing—not because progress has slowed but because access to capable intelligence is no longer scarce.

Advantage moves up the stack

When a resource becomes abundant and cheap, it stops being a source of advantage. As models commoditize, the durable question is no longer what intelligence you can access; it is what you can do with it that competitors cannot easily replicate.

That answer lives above the model, in proprietary data, workflow integration, domain expertise, trust, distribution, and the ability to turn capability into new sources of growth. These are also the things many organizations have treated as secondary while focusing on the model layer.

In the discussions we had at SuperAI, this didn’t translate into a simple open-vs.-closed debate. The more practical view was that model choice is becoming more contextual. Many organizations will run a portfolio: open-weight models where privacy, sovereignty, cost, and control matter; closed frontier models where performance changes the outcome; faster models where latency and volume matter; and more capable models where judgment is decisive. The question isn’t open vs. closed. It’s where each model belongs in the architecture of value creation.

If the model is a commodity, the application layer is where competitive differentiation is found. But moving up the stack is not only about better integration. It’s about shifting from productivity capture to growth creation.

Many companies are still using AI to do existing work faster: customer service, marketing automation, software development, document review. These applications are valuable, but they are rarely distinctive. The larger opportunity, which came through repeatedly at the conference, is to build AI into strategic workflows, proprietary data loops, and products that competitors cannot easily reproduce. In biotech, that may mean accelerating discovery through specialized data and simulation. In manufacturing, it may mean combining multimodal AI, digital twins, and shop-floor expertise. In asset-heavy sectors, it may mean modeling systems that were previously too complex or slow to optimize.

Agents change the work, not just the speed

One of the clearest shifts we heard about was from copilots to agents, from tools that assist a person with a task to systems that carry out multistep work on their own. This shift is already underway, and the dominant failure mode is not that agents do not work. It’s that organizations deploy them on top of processes designed for humans, accelerating tasks rather than reconsidering whether the workflow should exist in its current form at all.

This mistake is expensive because it feels like progress. Bolting an agent onto a legacy process produces a measurable improvement and a real sense of momentum, while leaving the underlying structure untouched. It doesn’t address the bottlenecks: organizational silos, handoffs, approvals, and sequential tasks that could be done in parallel.

The value of agentic systems comes from redesigning work around them—defining what a machine owns end to end, what a human owns, and where accountability sits. Companies that do this deliberately will pull ahead.

The interface for work is also changing. Based on our conversations at SuperAI, the direction of travel is clear, moving from software that people operate to agents that operate across software and eventually to more ambient, multimodal, and invisible workflows. In some cases, the best experience may be one where the user never opens the app, fills in the form, or touches the phone.

Trust becomes a scarce resource

Another theme that kept surfacing was trust. As generating cheap intelligence becomes effortless, the open web fills with synthetic content, fake identities, and machine-produced noise. The signs are already visible: spam, scams, impersonation, fake recruitment, synthetic profiles, and data deliberately poisoned to confuse bots. The predictable response is a retreat toward trusted networks and verified relationships.

AI both erodes trust and depends on it.

As open information becomes less reliable, provenance becomes increasingly valuable—knowing where something came from, whether it is current, and whether to believe it.

This has a strategic dimension that goes beyond reputation. The freshness and reliability of data degrade quickly under AI pressure, which shortens the useful life of public training data and raises the value of proprietary, verified, current information. AI both erodes trust and depends on it, and that tension does not resolve on its own. Organizations that can offer verified provenance, trusted relationships, and high-quality proprietary data will hold an advantage precisely because the surrounding environment is becoming less trustworthy.

Trust also runs inward. In conversations with enterprise leaders, it was clear that as AI absorbs more of the work, the people currently doing it have to trust that their organizations will treat them fairly and with dignity as their roles change. Fear about “what happens to my job, my status, or my future” is not a soft issue. It directly affects whether people will participate in the redesign.

The question to answer

What we took from SuperAI was not a single prediction about where AI is heading but a sharper version of the question every organization now has to answer: When intelligence is abundant, cheap, and embedded in nearly every business process, what remains distinctively yours?

Companies that answer it well will be specific about what they know, who they serve, and what they can integrate that competitors cannot replicate. They will redesign workflows and jobs; develop governance that allows them to move safely at pace; use AI to create more compelling experiences and products; and build trust with customers, employees, and regulators.

The next few years are likely to disrupt ordinary planning cycles. Leaders should assume that many of the constraints that shaped their business—bandwidth, scale, content production, software development, customer interaction, and global distribution—will change faster than expected.

Leaders who navigate this well will treat AI not as a technology upgrade but as a redesign of how their company creates value. The work is harder than buying a model. But it’s also the only part that lasts.