Happy Sunday and welcome to Investing in AI. Be sure to check out our AI in NYC podcast for our latest takes on what is happening in the AI world. And if you are into models, the Neurometric Leaderboard for systems 2 thinking AI was just updated with NVIDIA’s Nemotron models. Our leaderboard covers how models perform on a per-task basis rather than a full benchmark, which better matches how you actually deploy AI in agentic systems.
Today I want to talk about the AI bubble. Is there one? Or is it just the early stages of a revolution that requires a lot of investment? We don’t know yet, but here is what to watch to determine the answer.
Nvidia posted $57 billion in revenue last quarter, a 62% year-over-year surge that would suggest insatiable demand for AI infrastructure. But examine the customer base and a more troubling picture emerges: four customers contributed 61% of that total. This isn’t broad-based demand. It’s concentrated dependency masquerading as a market.
The critical question facing investors in 2026 isn’t whether AI is real—it demonstrably is. The question is whether AI revenue is real in the economic sense: distributed across productive enterprises generating measurable returns, or trapped in a circular loop where hyperscalers fund infrastructure, sell capacity to each other and to AI labs they partially own, and call the resulting transactions “growth.” If the $1.15 trillion in hyperscaler capex committed through 2027 doesn’t eventually flow beyond this closed system, we’re witnessing tech’s most sophisticated capex echo chamber—not a platform shift.
The 2015 SaaS Benchmark: What Diffusion Looks Like
To understand what healthy technology adoption looks like, consider the enterprise SaaS boom of 2015. When cloud software spending reached $31 billion that year, the market structure was fundamentally different from today’s AI infrastructure economy. Revenue distributed across thousands of mid-market companies serving businesses with 50 to 5,000 employees. The top ten vendors captured roughly 35% of total market revenue; the long tail of smaller providers held the remaining 65%. Crucially, buyers experienced clear, measurable ROI: lower total cost of ownership than on-premise alternatives, subscription flexibility, and deployment cycles measured in weeks rather than years.
This was genuine productivity diffusion. A plumbing company in Ohio using field service software and a mid-sized law firm in Dallas deploying practice management tools both extracted operational value from the same technological wave. The customer base was diverse, the use cases were concrete, and revenue flowed from actual efficiency gains—not from vendors purchasing each other’s services.
Today’s AI infrastructure economy inverts this model. Revenue flows almost entirely up the value chain to Nvidia from a handful of downstream buyers. The mid-market remains largely a spectator, watching from the pilot-program sidelines.
The Fortune 2000 Litmus Test
Enterprise AI spending reached $37 billion in 2025, up from $1.7 billion just two years prior—growth that would seem to validate the bull case. But examine who is actually deploying, and the diffusion thesis weakens considerably. While 92% of Fortune 500 companies now use ChatGPT in some capacity and over 70% have adopted Microsoft Copilot, these figures obscure a stark reality: only 8.6% of enterprises report having AI agents deployed in production. Nearly two-thirds have no formalized AI initiative at all. The majority remain trapped in what analysts now call “pilot purgatory”—experimenting without operationalizing.
This creates a dangerous capex-to-revenue lag. Hyperscalers now spend 45% to 57% of revenue on capital expenditures, historically unthinkable levels when the SaaS-era norm ran 11% to 16%. The implicit assumption is that inference demand will eventually justify this buildout. But if the Fortune 2000 doesn’t convert from “AI-curious” to “AI-deployed” by 2027, the infrastructure surplus becomes a balance-sheet problem rather than a growth opportunity. Revenue concentration risk compounds when your four largest customers are simultaneously your four largest competitors in the downstream market.
The Bull and Bear Cases
The bearish interpretation is straightforward: hyperscalers are funding their own future demand. Microsoft invests in OpenAI, which rents Azure compute, generating Microsoft revenue that funds additional Azure capex. The “customer” is partially an accounting entry. Nvidia’s revenue concentration mirrors Cisco circa 1999—the indispensable backbone of a technology revolution, trading at impossible multiples, until enterprise spending froze and the stock lost 90% of its value. With 74% of companies reporting no tangible value from AI initiatives and failure rates running 70% to 85%, the circular economy looks fragile.
The bull case rests on early signals of genuine diffusion into non-tech sectors. Healthcare organizations deploying AI report 3.2x ROI, with purchasing cycles compressing from eighteen months to under six. Manufacturing executives show 61% reporting decreased costs from AI-enabled supply chain optimization. Financial services firms achieve 4.2x returns on generative AI initiatives. The $37 billion enterprise AI market grew 3.2x year-over-year—faster than any software category in history. And unlike the dot-com era, the companies writing these infrastructure checks are the most cash-rich and profitable entities in corporate history. This isn’t speculative capex from unprofitable startups burning venture capital.
The Watchlist
Three indicators would signal the bubble is bursting: hyperscaler capex decelerating more than 20% year-over-year as spending reverts toward historical norms; enterprise AI production deployment rates stalling below 15%; and Nvidia’s revenue concentration from its top four customers exceeding 70%, indicating the loop is tightening rather than expanding.
Conversely, three indicators would confirm the floor is solidifying: healthcare and manufacturing AI spending exceeding $20 billion annually; enterprise AI deal conversion rates sustaining above 45%; and Fortune 2000 AI budgets growing 30% or more with documented productivity gains.
The 2026 AI market isn’t a bubble in the 1999 sense—the companies are real, the profits are real, the technology works. The question is whether the customers become real. Until AI revenue flows beyond the circular loop of hyperscalers and their subsidiaries into the productive economy of the Fortune 2000 and below, the trillion-dollar infrastructure bet remains exactly that: a bet on diffusion that hasn’t yet occurred.
Thanks for reading.