Are we in an AI bubble?
So what is a bubble anyways?
According to the Fed, the term “bubble” is used to describe an asset price that has risen above the level justified by economic fundamentals, as measured by the discounted stream of expected future cash flows that will accrue to the owner of the asset.”
Ok cool! So let’s see if the current valuations of leading AI companies fit in with this description.
OpenAI 2025 stats: $12B ARR, $8B annual burn rate, $500B valuation.
So does this $500B valuation justify economic fundamentals as measured by discounted streams of expected future cash flows that will accrue to the owner of OpenAI? Absolutely fucking not, we are in the biggest bubble of our lifetime!
Let’s do some quick maths.
The OpenAI DCF: Does $500B Make Any Sense?
According to leaked projections from The Information, OpenAI expects revenues to grow from $13B in 2025 to $125B by 2029. That’s a 76% CAGR—more than double Facebook at its hyper-growth phase (~35%) and nearly 3x Google’s long-term trajectory (~27%). In other words, OpenAI is projecting it will grow faster than any of the most successful tech companies in history, despite already starting from a massive $13B base.
So let’s run the numbers. Assume OpenAI somehow operates at Microsoft-like efficiency (35% net margin) and use a 10% discount rate (standard for Big Tech).
Case 1: No Terminal Value (Just 2025–2029 Cash Flows)
2025: $13B → $4.6B profit → PV ≈ $4.6B
2026: $29B → $10.2B profit → PV ≈ $9.2B
2027: $54B → $18.9B profit → PV ≈ $15.6B
2028: $86B → $30.1B profit → PV ≈ $22.6B
2029: $125B → $43.8B profit → PV ≈ $29.9B
Total DCF Value (no terminal): ≈ $82B
→ That’s about 6x lower than OpenAI’s current $500B valuation.
Case 2: With Terminal Value (3% Perpetual Growth Beyond 2029)
Add a perpetual growth assumption: 2029 profit grows at 3% forever.
Terminal value = ~$644B, discounted back to present ≈ $440B.
Total valuation with terminal value: ≈ $522B.
Total DCF Value (with terminal): ≈ $522B
→ That’s actually right in line with OpenAI’s current $500B price tag.
Okay, how about other AI companies?
Anthropic in Numbers:
Revenue Projections & Growth
Current run-rate: Anthropic has soared from around $1 billion ARR in December 2024 to approximately $3 billion by May 2025, and now up to $4 billion as of mid-2025.
End-of-year outlook: Some estimates project it could reach $9 billion ARR by year-end.
Projections for 2027: A “bull” scenario from The Information notes Anthropic could hit $12 billion to $34.5 billion revenue in 2027.
Valuation Context
As of March 2025, the company raised $3.5 billion at a $61.5 billion valuation, implying a forward multiple of ~44x ARR (on $1.4 B ARR).
A new mega-round is in the works. Anthropic is raising up to $5 billion at a valuation around $170 billion.
DCF Analysis: Anthropic's Valuation vs. Fundamentals
Key Assumptions
2025 ARR baseline: $4 billion
Growth path: From $4B in 2025 to $12B (base) or $34.5B (bull) by 2027
Profit margins: Optimistically 30% net (similar to Big Tech at scale)
Discount rate: Standard 10% WACC
Scenario 1: Base Case (to $12B by 2027)
2025: $4B revenue → $1.2B profit → PV ≈ $1.2B
2026: $8B → $2.4B → PV ≈ $2.2B
2027: $12B → $3.6B → PV ≈ $2.7B
Summed PV ≈ $6.1B
Add terminal value (3% perpetual growth): PV ≈ $49B
Total DCF Value: ~$55B
Scenario 2: Bull Case (to $34.5B by 2027)
2025: $4B → $1.2B → PV ≈ $1.2B
2026: $17.25B → $5.2B → PV ≈ $4.7B
2027: $34.5B → $10.35B → PV ≈ $7.7B
Summed PV ≈ $13.6B
Add terminal value: PV ≈ $141B
Total DCF Value: ~$155B
Interpretation
Scenario DCF Valuation (approx.)
Base case $55 billion
Bull case $155 billion
Current ask $170 billion
Let’s also not forget the recent AI labs that have been founded by ex OpenAI leadership. Thinking Machines founded by ex OpenAI CTO Mira Murati raised a $2 billion SEED round (yes seed round) with valuations of $10-12B. No product, revenue and entire operation in stealth.
Safe Super Intelligence Inc. founded by Ilya Sutskever raised $1B at a $5B valuation and only 6 months later raised at a $30B valuation. Again, no product and no revenue.
Dark Fiber → Dark GPUs
Back in 2000, telecoms buried trillions of dollars in fiber cables under our feet. By 2005, 85% of those lines were still dark, never even switched on. It was the perfect symbol of the dot-com bubble: mountains of hardware, no one to use it.
Fast forward to 2025, and the same thing is happening—this time with GPUs. Microsoft, Google, and Amazon are racing to stack millions of NVIDIA H100s, each costing ~$50k a pop. In Q1 2025 alone, data center capex jumped 53% year-over-year to $134 billion. At this pace, hyperscalers will spend over $300 billion in 2025, a record figure that dwarfs almost any prior infrastructure cycle.
Now compare that to demand. The entire global inference market—all the money made from actually using these GPUs—was worth just $76 billion in 2024, projected to grow to $255 billion by 2030. Even at that growth rate, hyperscaler spending in a single year already exceeds the whole inference market today.
Worse, much of the deployed hardware isn’t even working at full tilt. Studies show 30% of enterprise GPU clusters run below 15% utilization, and inference accelerators rarely break 50% usage. In other words: the shelves are full, but the checkouts are empty.
So yes, demand is rising fast, but the buildout is running way ahead of the curve. And unlike fiber, which sat quietly underground until the world caught up, GPUs rot quickly—obsolete within 3–5 years. If this imbalance holds, today’s AI boom could leave us with billions of dollars of “dark GPUs” humming away in half-empty data centers.
In my opinion, this is playing out extremely similar to the internet era bubble. Extremely optimistic infrastructure buildout and insane valuations that’s so far out of touch with fundamentals it’s funny. However, just like the internet after the bubble burst there were a bunch of winners that actually built real businesses. The exciting thing is some of them haven’t even started yet and I suspect it will be the same in this AI era.