The AI Multiplier: Why the Market is Correct on the Bubble and Still Missing the Discontinuity
Bubble Burst or Breakthrough? Navigating Agentic AI's Value Multiplier Amid Public Market Froth
Perhaps we should declare August “International AI Bubble Month.”
Several weeks ago, OpenAI CEO Sam Altman caused a stir when he told reporters following the troubled launch of GPT-5 that yes, possibly there is such a bubble. “Are we in a phase where investors as a whole are overexcited about AI? My opinion is yes. Is AI the most important thing to happen in a very long time? My opinion is also yes.”
His remarks rebounded around the internet and global markets:
Throw in the widely discussed MIT study that claimed 95% of generative AI projects at companies are unprofitable, Apollo economists warning that “AI bubble now bigger than 1990s tech mania,” and famed dot-com bubble-caller Howard Marks hinting that the “market's current behavior reminds him of the late '90s tech craze,” and there are all the elements for a toxic brew of panic by believers and triumphant I-told-you-so’s from critics.
It is the kind of pessimistic soul searching not seen in the AI sector since…August 2024:
The disconnect between what I saw then as the short-term panic and long-term promise surrounding AI prompted me to write a manifesto last year, outlining my larger view of this moment: “Why GenAI Is Not Disruption — It’s Discontinuity.”
Discontinuity has become the cornerstone for how I think about the scale, scope, and timing of the transformation being catalyzed by generative and agentic AI. As we navigate the 2nd Annual AI Bubble Debate, it seems like the right time to revisit that concept and explore why the real value creation—now driven by agentic AI—hasn't truly started.
By using the lens of discontinuity, we can avoid focusing solely on the short-term bubble narrative, which misses the larger truth: we're witnessing the early stages of a general-purpose technology deployment on par with electricity or the internet. The question isn't whether AI will transform the economy. It's which companies will capture the value. And when.
The Bubble Signs
Let’s acknowledge some of the legitimate critiques. Many intelligent individuals are thinking deeply about this moment and bringing tremendous experience to their analysis. They are bound by a desire to help their businesses and clients navigate a moment when new technologies are evolving at an almost incomprehensible pace.
The parallels to 1999-2000 are unmistakable, such as a handful of stocks dominating index returns. Torsten Sløk, Apollo Chief Economist, notes: “The difference between the IT bubble in the 1990s and the AI bubble today is that the top 10 companies in the S&P 500 today are more overvalued than they were in the 1990s.”
In mid-August, Marks published an extensive missive to clients, “The Calculus of Value.” He took a slightly different tack than Sløk by looking at the non-elite (Apple, Microsoft, Alphabet, Amazon, Meta, Nvidia, and Tesla) and concluding: “I think it’s the average p/e ratio of 22 on the 493 non-Magnificent companies in the index – well above the mid-teens average historical p/e for the S&P 500 – that renders the index’s overall valuation so high and possibly worrisome.”
Value, Not Hype
In contrast to this caution are the recent astonishing AI investment numbers in both the private and public markets.
There was the $13 billion raised by Anthropic at a $183 billion valuation. The Bank of America Institute just reported $40 billion of U.S. data center construction spending in June, up 30% from the previous year and a new all-time high driven by an unquenchable thirst for AI infrastructure. Oracle CEO Larry Ellison became the world’s richest man thanks to the company’s growing footprint in AI. (The $300 billion compute deal with OpenAI that just sent Oracle’s stock booming also points to how AI compute is bifurcating into two distinct paths, a dynamic that I’ll explore in a two-part series starting Wednesday).
Capital expenditure has exploded. Big Tech will spend $364-400 billion in 2025, more than double 2023 levels, driven by AI infrastructure demands. In FY26, Meta implied that spending would hit $100 billion, and Microsoft suggested it could go up to $120 billion in the same fiscal year.
"The depth of the demand, the scale of the demand, the breadth of the demand is overwhelming," CoreWeave co-founder and CEO Michael Intrator said at the recent Goldman Sachs Communacopia conference. "The industry's capacity to deliver the compute that is required by OpenAI, by the hyperscalers, by the enterprise, by the sovereigns, it's just layer after layer of overwhelming demand."
In a report released earlier this year, Goldman Sachs' research staff argued that there is no bubble and that things are far less frothy than during the dot-com era.
Sure, that might seem like a long time ago. But, still, why the disconnect between these two AI bubble camps?
Marks is right to sound the alarm. His prudence is a virtue in a market where psychology and FOMO often outrun fundamentals. Those elevated P/E ratios for the non-Magnificent Seven scream "priced for perfection" if AI adoption hits speed bumps.
But in using such a classic calculus, Marks also highlights one of the central challenges of this moment, one that has become the focus of much of my work:
Are the traditional benchmarks, which track historic performance against linear growth metrics and are still tuned to linear growth norms, giving enough weight to the AI-driven discontinuities that are quietly starting to impact earnings power?
I believe they are not.
Based on what we already see, tangible productivity gains from generative and agentic AI are compressing costs and boosting efficiency in ways that historical comparables cannot capture.
This growing disconnect may seem small now, but it points to the larger discontinuity that is emerging.
Revisiting Discontinuity
You can read the full Discontinuity manifesto here. But let me highlight a couple of key parts:
“I want to pull the lens back to focus on discontinuity versus disruption.
When disruption comes along, it takes a linear trend and changes the curve, increasing or decreasing it by let’s say +10%, +20%, 50%, and so on. The curve becomes steeper or less steep, but it is based on the same fundamentals that created the curves in the first place. That’s not to say that navigating disruption is easy, but leaders and investors can deploy existing tools to analyze the risks and opportunities and understand how the game has changed.
Discontinuity stops the curve and creates a new curve that has an entirely different profile. The curve can continue going in the same direction, and at the same speed, but it can also collapse or accelerate in an unexpected way. In geology, discontinuity is a term that is used to designate structural breaks which are usually unhealed. But when they heal, they exhibit high tensile strength. The same applies in the business world.
All the previous strategies, tools, and historic lessons are ripped away. In this dynamic, discontinuity introduces a break from the past which can lead to something very positive, or it can lead to failure. But it’s philosophically and fundamentally very different than disruption, which does not introduce such clean breaking points.
Such moments present a fundamental challenge that goes beyond the tech: People’s brains are not framed for discontinuity.”
Also:
“Though critics like to mock Silicon Valley when bubbles emerge, tech leaders have proved to be remarkably prescient about what will happen. What they tend to get wrong is when it will happen...Though it may have taken longer than initially predicted, today we are living in that digital world many envisioned in 2000 thanks to smartphones, 4G wireless, broadband, and a host of business model innovations enabled by these waves of infrastructure.”
This is where we are with generative and agentic AI. Once again, fears of a bubble are blurring the larger picture of discontinuity.
Generative and agentic AI represent fundamental breaks because their impact has the potential to extend far beyond tech disruptions such as the cloud, smartphones, or the social Web. The impact of these technologies is likely to rival that of the internet, and most certainly surpass it, because their ability to leverage both structured and unstructured data removes previous limitations on the type of data that can be utilized.
Seeing that is not easy. It requires envisioning scenarios that, in many ways, are unimaginable. Who could have imagined in January 1879, when Thomas Edison built his first high-resistance, incandescent electric light, that it would lead to such sweeping economic and social changes? And, it requires patience as it is easy to conclude that things are not where they should be. Yet.
Today, generative and agentic AI are still in their infancy. Because I focused on generative AI last year, I’ll zoom in on the discontinuity of agentic AI.
The Untapped Productivity Revolution
While it may be impossible to imagine the unimaginable, there are ways to start grasping the massive scope of the looming valuation revolution.
Markets obsess over GPU allocation and the Magnificent Seven's capex arms race. But they're overlooking where the real value creation will occur. Let’s go back to those 493 companies in the S&P that Marks targeted in his analysis. While it’s inarguable that they may be trading above historic norms, we can also say with a high degree of certainty that they haven't yet begun to tap agentic AI's potential.
Consider what's already happening beneath the headlines.
The gains from generative and agentic AI aren't vaporware. They're already hitting balance sheets, particularly through skyrocketing ARR per FTE ratios in AI-native firms (many of them still in the Private Markets). For instance, some domain-specific applications companies are hitting $3M ARR per full-time employee. That’s 10x the typical SaaS benchmark of an elite public company trading at 10x +.
CEO Alex Karp recently projected that Palantir – the highest valued SaaS company – would increase its revenues 10x while decreasing headcount from 4,100 to 3,600. That would increase annual revenue per FTE from ~$1M in Q2 2025 to ~$11M!
That’s just early days. According to a study this year from Model Evaluation & Threat Research (METR), AI’s ability to complete long tasks “has been consistently exponentially increasing over the past 6 years, with a doubling time of around 7 months."
Most importantly, these gains represent primitive, single-agent implementations. They're the equivalent of using electricity solely for light bulbs while missing the entire industrial revolution it would enable.
Of course, counterarguments persist. Skepticism is always rightfully warranted. MIT and Harvard Business Review studies note that 95% of generative AI projects fail due to implementation challenges, and productivity gains may not scale evenly across industries.
Still, these failures often stem from poor organizational readiness rather than technological limits, and emerging evidence suggests agentic AI could address this by enabling more adaptive deployments. By the same measure, the dot-bust seemed like a catastrophic repudiation of online commerce, as almost 800 online startups failed in 2000 and 2001 due to a mix of frankly misguided ideas and a lack of understanding of the economics and execution of these new businesses.
Today, there’s little doubt that online commerce has triumphed. If you had bought stock at Microsoft’s dot-com bubble peak of $58.38 per share in December 1999, the current price of $515.36 per share would probably have seemed impossible even to the biggest optimists.
The Orchestration Revolution: From AI Agents to Agentic AI
We are on the cusp of another leap, from AI Agents to Agentic AI. The difference may sound semantic, but it's architectural. And it changes everything.
As is often the case, it isn't easy to be precise about the speed and impact of this next step.
Still, the outlines of this new era are taking shape. Current AI agents are task-specific, reactive tools operating in isolation—sophisticated but limited by their inability to reason causally, maintain persistent context, or coordinate with other systems.
Agentic AI represents a paradigm shift, featuring autonomous systems that set goals, adapt, and orchestrate multi-agent collaboration with shared memory and emergent problem-solving capabilities. Think of it as moving from individual skilled workers to entire coordinated teams.
The orchestration mesh enabling this—combining multi-agent systems, persistent memory, and advanced reasoning—is maturing rapidly in 2025, with trends such as self-healing data pipelines, vertical agents in specific industries, and integration with human workflows.
Consider that net margins have risen by 50% since 2000 (from 9% in 2000 to 13.5%) and are expected to reach 15% by 2027. On a margin-adjusted basis, today’s valuation looks far less extreme: the 12-month forward P/E relative to next-twelve-month net margins stands at 1.6×, only modestly above its post-2005 average.
Impressive, yet modest. When this agentic infrastructure matures, we won't see minor incremental improvements. We'll see entire business processes reimagined. The compound effect here is crucial and widely misunderstood.
Let's do the math that Wall Street seems to be missing.
If a company improves productivity by just 10% annually through AI adoption, that compounds to a 61% improvement over five years. Now, apply that to some or all of the 493 companies on Marks’ list. For those who achieve these gains, the current "bubble” prices will seem cheap. For those that don’t, well, it will indeed look like they were priced into a bubble. Such is the benefit of hindsight.
Consider a pure tech player: Salesforce. The company’s stock is down ~27% YTD, giving it a market cap of $231 billion. Using the traditional SaaS Rule of 40, investors are anxious about slowing rates of growth. But at the same time, Salesforce has positioned itself, with the release of Agentforce 3.0 earlier this year, to establish leadership in some of the most valuable parts of the agentic economy.
Former Salesforce Co-CEO Bret Taylor predicted that the world would see the first trillion-dollar “applied enterprise software” company due to “this new world of agents." If that turns out to be Salesforce, then are investors missing out on $750 billion? Next month’s Dreamforce will be a litmus test.
This is ambitious, but it’s not fantastical thinking.
Part of the AI summer panic was caused by signals that enterprise AI adoption might be slowing, according to U.S. Census Data. But there have been plenty of counter signals. According to Ramp, the overall trends remain bullish as enterprises make adjustments, but still demonstrate a clear hunger for more AI.
During Microsoft’s FY 2024 earnings call in July, CEO Satya Nadella highlighted the ferocious rates of adoption the company is experiencing.
“When we look narrowly at just the number of tokens served by Foundry APIs, we processed over 500 trillion this year, up over 7x,” he said. “This is a good indicator of true platform diffusion beyond a few head apps and services...We are going through a generational tech shift with AI.”
This isn't a one-time boost. It's a permanent step-change in earning power. The enterprises buying these tools are still in the early stages of experimenting and deploying. Multiply these gains across hundreds of S&P 500 companies over a decade, and you are describing one of the largest value creation events in economic history.
Then consider that productivity is just the appetizer. The real feast is agentic AI's potential to spawn entirely new revenue streams and unlock growth by cracking open the $60 trillion global labor market, far beyond the estimated $2.2 trillion TAM that software could reach by 2034, according to Precedence Research.
Unlike siloed tools, agentic orchestration enables "labor-as-a-service" models: autonomous swarms that not only automate but also innovate, generating revenue per outcome rather than traditional per-seat/per-user revenue. In that sense, agentic AI isn't eroding moats; it's forging new ones for operators who capture the labor pie first. We will explore this in further detail.
Yes, There's Craziness—But That's Not the Whole Story
Let's be clear: there absolutely is bubble behavior.
Valuation as a headwind is becoming increasingly apparent. NVIDIA is trading at 50x trailing earnings while projecting massive growth amid $500 billion in US AI infrastructure investment over four years. The Stargate project's $500 billion aspiration for AI data centers by 2029, while ambitious, faces launch delays and concerns over execution. Companies with questionable AI strategies are seeing valuation pops simply for mentioning "agentic capabilities" on earnings calls.
The risk dimension is also being ignored. For every company that successfully transforms with AI, others will see their moats evaporate.
With the S&P trading above 22× earnings, the comparisons to the TMT bubble are easy to make. But the backdrop is different.
The disconnect between AI's current reality and future potential creates a unique moment. The technology is improving daily. Each iteration of models shows dramatic capability improvements, while the orchestration infrastructure advances in parallel.
Yet markets price AI as if we've reached steady state, as if today's ChatGPT and GitHub Copilot represent the ceiling rather than the floor.
The productivity gains aren't theoretical anymore. They're showing up in earnings reports. If anything, the pace of improvement is accelerating. And most importantly, the vast majority of companies haven't even started (really) their AI journey.
The Agentic Era is just beginning.
Be prudent, but don’t be distracted by the wild gyrations of sentiment around AI from month to month. Instead, remember that in Altman’s quote about AI bubbles that caused such a frenzy last month, there was far less attention on the second half of his statement: “Is AI the most important thing to happen in a very long time? My opinion is also yes.”