Two years ago, calling AI a bubble got you laughed out of the room. Today, saying it's not a bubble might get you the same treatment. But both camps are confusing asset class with individual company valuations — and that confusion is costing investors real money.
The Bears Are Not Wrong — They're Just Pointing at the Right Problems
The bear case is grounded in real data. OpenAI projected $11.6B in revenue for 2025 while burning through $5B+ in operating losses. The top 10 generative AI application companies raised at a combined $80B+ in valuations, most generating less than $100M in ARR. That math only works if everything goes perfectly — and in venture, nothing goes perfectly.
McKinsey's 2025 enterprise AI survey found that 67% of AI pilots stall before reaching production. That's not a technology problem — it's a deployment, integration, and organizational change management problem. Enterprises are paying for AI licenses they cannot fully operationalize. Net revenue retention at many AI SaaS companies is masking churn behind new logo growth.
Valuation multiples are also stretched beyond precedent. The average AI SaaS company trades at 25-40x ARR in private markets versus 8-12x for traditional SaaS with equivalent growth rates. That delta only makes sense if AI companies have structurally superior retention, margin expansion, and TAM. Some do. Most don't — and the market hasn't priced that distinction yet.
But the Infrastructure Bet Is Not Going Anywhere
Here's what the bears consistently underweight: the scale of committed capital at the infrastructure layer. Microsoft, Google, Amazon, and Meta have collectively committed over $500B in AI-related capex for 2025 and 2026 combined. This is not speculative spend — it's contracted data center construction, GPU procurement, and network buildout tied to existing revenue streams.
NVIDIA generated $22.1B in data center revenue in a single quarter in 2025. Blackwell chip orders are sold out through 2026. These are not hype numbers — they are purchasing orders from the most financially disciplined technology buyers on the planet. The Fortune 500 does not spend $100M+ on infrastructure without a signed business case.
I've seen this pattern in my portfolio. The companies getting the most durable AI revenue are not selling AI to early adopters — they're embedding AI into workflows that existing enterprise customers already depend on. That is structurally different from the 2021 crypto cycle, where infrastructure spend never connected to sustainable use case demand.
Bubble vs. Structural Shift: How to Tell the Difference
- •Dot-com 1999: 80% of capital went to companies with no revenue model. AI 2026: the top five cloud providers have $1T+ in combined AI-related infrastructure revenue — real and growing.
- •Crypto 2021: speculative tokens with no real-world utility commanded $3T in total market cap. AI 2026: productivity studies show 14-40% efficiency gains in software development, legal research, and financial modeling — verifiable with production data.
- •The at-risk layer: AI wrappers with under $5M ARR raising at $50M+ valuations on GPT API access and a clean UI. These companies have no moat, no proprietary data, and no switching cost. If OpenAI cuts API pricing — which it will — the margin structure collapses.
- •The durable layer: verticalized AI trained on proprietary domain data, embedded in workflows with real switching costs, serving industries where regulatory complexity creates natural barriers. These companies will compound through any correction.
- •Key tell: if an AI company's value proposition disappears the moment a foundation model provider adds a native feature, it was never a real company. If it gets stronger when models improve because the moat is in data and workflow — that's real.
The Three Signals That Will Tell Us Who's Right
I track three metrics as leading indicators of where this resolves.
First: GPU demand signals. If NVIDIA data center revenue drops more than 20% in two consecutive quarters, it means cloud providers are pulling back capex — which is a demand problem, not a supply problem. That would validate the bear case. So far, every quarter has beaten the prior. Blackwell orders suggest that continues through at least Q3 2026.
Second: Enterprise AI renewal rates. Net revenue retention is the cleanest signal of real value delivery. If AI SaaS companies show NRR below 100% at scale — meaning customers are contracting rather than expanding — it is a usage problem masquerading as a product. Watch the Q4 2026 earnings calls for enterprise software companies that have AI as a primary revenue line. The NRR numbers will be telling.
Third: Application-layer M&A pricing. If large technology companies begin acquiring AI application startups at discounts to their last private valuation — say, $200M acquisitions for companies that raised at $500M — that is the bubble deflating in real time. It is not a bet; it is confirmation. Strategic acquirers are the most ruthless valuators because they have the most information about what customers actually pay.
The bubble is not in AI. It's in the specific stratum of AI applications with no proprietary data, no vertical depth, and a valuation built entirely on vibes. That layer will compress 60-80%. Everything else will compound.
Stay current with VC and startup trends at Value Add VC. Originally published in the Trace Cohen newsletter.