If AI is smarter than ever, when’s the revolution coming?

I recently attended an artificial intelligence (AI) venture capital presentation where AI’s accelerating speed and capabilities were on full display.

Between 2025 and 2026, AI shifted dramatically from conversational chatbots to autonomous execution, driven by two compounding breakthroughs: exponential gains in agent task endurance and unprecedented inference speeds. Standardised benchmarks show that the time an AI agent can perform complex, multi-step tasks without human intervention roughly doubled every three to four months, while hardware innovations – such as wafer-scale engines pushing open-weights past 2,600 tokens per second – slashed response times to near-zero latency. Together, sustained, multi-hour problem-solving and lightning-fast execution transformed AI from a passive assistant into a real-time, autonomous workforce capable of handling end-to-end software engineering, scientific research, and complex workflows in seconds.

But despite this, two questions are emerging and dominating Wall Street conversations and commentary: If AI capabilities are accelerating exponentially, where is the economic revolution? And if “everything” is changing inside research labs, why does “nothing” seem to be changing in broader macroeconomic data?

In a recent podcast, The Atlantic staff writer Derek Thompson sat down with Wharton professor and leading AI scholar Ethan Mollick to try to unpack this puzzle.

For stock market and venture capital (vc) investors trying to separate real utility from the capital expenditure hype, I thought the discussion offers a potentially helpful framework for evaluating AI’s true trajectory and commerciality.

The great disconnect: capability vs. implementation

Investor scepticism stems from a divide between frontier AI performance metrics, which are off the charts, and broader economic indicators that remain surprisingly subdued.

In terms of performance, capabilities once thought to be years away – such as solving Millennium Prize math problems – are being checked off in months. Model capabilities are following a logarithmic line, and previous state-of-the-art tools are becoming cheap, run-on-your-phone utilities in a matter of months.

On the other side, unemployment remains at historic baseline norms rather than spiking from mass automation, productivity growth shows modest bumps, but nothing resembling a structural paradigm shift and corporate Return On Investments (ROIs) – outside of semiconductor manufacturers (e.g., chip and memory makers) – remain indistinguishable from returns being generated prior to the advent of AI.

So, why haven’t improving model capabilities immediately shown up in corporate earnings and returns across the broader S&P 500? In fact, at this juncture it’s worth noting just how many stocks in the S&P500 are trading at significant discounts to their previous highs.

Figure 1. Percentage of stocks below 52-week highs.

Why the lag?

Mollick argues that markets are making a fundamental mistake by assuming human organisations can adapt as quickly as digital networks.

The idea is that two organisational frameworks explain why implementation lags capability:

1. The O-ring principle of jobs

Named for the single faulty rubber seal that caused the Space Shuttle Challenger disaster, the O-Ring model holds that a process fails when any one step breaks down.

While AI can now handle 90 per cent or 95 per cent of the tasks in a specific role, such as software engineering or legal analysis, human professionals are still needed to manage, audit, and fix the remaining 5 per cent where AI falls short. Because jobs are non-additive systems, value continues to accrue to human oversight until the entire task chain can be completed end-to-end without breaking down.

1. The garbage can model of organisations

In organisational theory, corporations are rarely streamlined, perfectly orchestrated machines. Instead, they’re like “garbage cans” – disorganised bundles of legacy problems, competing political solutions, unmapped processes, and redundant workflows.

Deploying an AI model within a ‘garbage can’ corporate structure means redesigning its messy, informal workflows from scratch. For now, AI is mostly a single-player tool that boosts individual workers’ output, rather than a multiplayer system that can automatically streamline entire corporate architectures.

The next structural shift: proactive agents and swarms

We’ve seen this recently. It took a swarm of 10,000 agents working together for 88 hours to solve the Navier–Stokes existence and smoothness problem, one of the seven famous Clay Mathematics Institute Millennium Prize Problems. That was in September 2026.

Following the Navier-Stokes announcement, OpenAI expanded its scope and pointed its unreleased AI model at roughly 4,000 open problems in mathematics. On October 6, 2026, the company posted the results to GitHub, claiming it had solved nearly 400 math problems humans had not yet been able to solve.

For investors looking for the catalyst that bridges single-player software gains to enterprise-wide ROI, Mollick reckons the rapid emergence of autonomous agents and agentic swarms (such as Meta’s Muse or OpenAI’s internal tools) is that bridge.

Shifting from prompting to proactive execution

Another catalyst is the shift from prompts in a box to autonomous agents, executing tasks continuously through apps on handheld devices.

Until recently, the dominant interface for AI was a blank chat box – requiring humans to know what to ask.

But next-generation personal and corporate agents will access data streams, such as emails, calendars and workflows, to anticipate friction points, execute multi-step tasks, cancel subscriptions, audit tax forms, and negotiate customer support without explicit human prompting.

Then there’s the disruption business models reliant on consumer friction will experience (e.g., store credits, gift cards, low-cost airlines, low-yield savings accounts, rebates and refund processes, warranty claims, and subscriptions). As proactive agents navigate friction points on consumers’ behalf, corporate margin structures across retail, finance, and services will face unexpected pressure.

The power of swarm architecture

Furthermore, when autonomous agents coordinate in multi-agent swarms, they bypass classical human organisational limits. Unlike humans, AI agents face no principal-agent misalignment (when one party (the agent) is hired to make decisions or act on behalf of another party (the principal), but the agent’s personal incentives don’t align with the principal’s best interests), internal office politics, or communication friction. They pass data telepathically at machine speed, solving massive, long-running tasks previously impossible for single models or human teams.

Implications for investors

Forecasters regularly underestimate AI acceleration simply because the human mind struggles with exponential curves. So don’t assume today’s roadblocks are long-term barriers.

Obviously, building agentic systems requires substantial compute, constant architectural updates, and massive efficiency gains. So, small, simple apps that just plug into existing AI technology have little protection against competitors, while massive tech companies with billions of users already have a huge natural advantage.

The metrics worth following are those that track an AI’s ability to run autonomously on complex, multi-week objectives without human intervention and without error.

Meanwhile, Derek Thompson and Ethan Mollick expect a turbulent economic transition period as enterprise software, customer service models, and hiring pipelines are rebuilt around AI capabilities.

While the macroeconomic revolution is almost always mentioned but hasn’t fully registered in quarterly earnings calls outside of AI infrastructure providers, the underlying capabilities are advancing rapidly. Thompson and Mollick argue the lag isn’t a sign of an AI bubble bursting, but rather a quiet prelude to a massive structural reconfiguration of corporate operations.