Nye's Digital Lab is a weekly scribble on creativity in the age of rapid change.

This week, I'm considering the question I always seem to get: "Is it a bubble?"

In the first moments of the universe, everything that would ever exist was compressed into a single point. An almost incomprehensible density of energy and potential.

Then in an instant, it exploded.

The Big Bang didn't just create matter. It decentralized it. Flung energy and plasma and particles across unimaginable distances. From that single, unsustainable point of centralization came everything we know. If you look at the red shifts in deep space, you can still see the universe expanding, still moving outward from that first moment of release.

I've been thinking about that a lot lately.

Because the AI industry is doing exactly the opposite. It's compressing. Centralizing. Insanely Investing. And at some point, maybe sooner than anyone wants to admit, it's going to have to let go.

Boom.

I've spoken to many people who tell me that AI is getting cheaper.
And they're right, well... sort of.

The price per token (the unit of text an AI model reads and writes) has dropped about a thousand times over the last three years. That's remarkable. But your actual AI bill went up 320% over the same period.

This is a real economic phenomenon called the Jevons Paradox.

So.
A typical business AI query in 2020 used fewer than 200 tokens, roughly a paragraph of text. By 2025, a single agentic workflow burns through around 22,000 tokens in one session. The newer AI agents, the ones doing multi-step tasks on your behalf, use five to thirty times more tokens than a basic chatbot. Token counts per query could rise to between 150,000 and 1,500,000 by 2030. That's not a rounding error. That's a different kind of energy industry.

The average enterprise AI budget grew from $1.2 million a year in 2024 to $7 million in 2026. Some Fortune 500 companies are running monthly inference bills in the tens of millions. This means that the companies providing this AI aren't making money on it.

Even the big dogs are in the red.

OpenAI brought in $3.7 billion in revenue in 2025 and lost an estimated $5 billion. They're spending $1.35 for every dollar they earn, and those losses aren't from building the next model. They're from the cost of running the current one.

The prices you're paying right now for AI access aren't real prices. They're subsidized by venture capital and hyperscaler cross-subsidies, a collective bet that the optimization of these models will "eventually work out."

And I haven't even brought up the amount of energy we are consuming yet.

Data centers consumed around 400 terawatt-hours of electricity globally in 2024. By the end of the decade, that number approaches 1,000, with AI responsible for roughly a third.

Meta and Microsoft are literally building nuclear plants to keep their data centers running. These are decade-scale infrastructure commitments.

AI ain't cheap. The floor cost of this industry isn't going anywhere.

So where is the bang, Nye? Why haven't we popped?

Because the cushion is enormous. I mean... Crazy enormous.

Amazon at $200 billion.
Google at $175 to $185 billion.
Microsoft at $145 billion.

In 2025, AI startups raised $203 billion in venture capital, a 75% spike from 2024. OpenAI, Anthropic, and xAI together claim over a trillion dollars in private market valuation before any of them have turned a profit.

The bet being made by every investor writing these checks is that efficiency will catch up to demand before the money runs out. The technology is genuinely transformative and the improvements are real. But the Jevons Paradox keeps showing up. (See above) Every time inference gets cheaper, it unlocks new use cases, which drives more usage, which drives more cost.

The conversation in VC circles has to shift from "is this AI startup exciting?" to "can this business survive when the API subsidies go away?" Those are very different questions, and a lot of companies funded in the last two years were built to answer the first one, not the second.

An recent student of mine tipped me to the work of Ed Zitron, who makes the argument that the AI industry is much like the 2008 subprime mortgage crisis. Companies building their futures on centralized AI could be in for a world of hurt.

Layer on top of this a world going sideways geopolitically.

Trade wars, and um... actual wars, economic instability. Energy, water, semiconductor fabrication capacity are real world, physical limits, that don't respond to a funding round.

So I'll just say it.

I think a collapse might actually be good for all of us. Let's just get it over with.

Not a catastrophe. Not a wipeout. But a Big Bang, a moment where the unsustainable centralization releases and the energy gets flung outward. A moment that forces a real question:

were we actually building this AI thing right?

Stewart J. Russell, an AI expert and professor at UC Berkeley has often used the nuclear industry as an analogy for the crisis of AI we are in. His point is that after Hiroshima and Chernobyl, the industry was forced to stop and confront what it had built. Those were genuine tragedies. But the pause they created gave engineers and policymakers time to ask harder questions before the next wave of expansion.

The newest generation of nuclear reactors, smaller, modular, dramatically safer, came directly out of that period of enforced humility. The industry had to slow down before it could move forward responsibly.

Something similar might need to happen with AI. Because the alternative to a bubble pop isn't a stable, sustainable ecosystem. It's a race where only the companies with essentially unlimited capital can operate, and everyone else rents access from them forever at whatever price they decide to charge.

I often get "the eye roll" for continuing to consider decentralized infrastructure. But I still think we need a radical redesign.

The seeds of a different model are already here. Projects like EXO, an open-source tool that networks everyday Apple hardware into a distributed AI cluster, are showing that you can run serious, capable models without a data center. One developer networked seven Mac Minis together and got 496 gigabytes of unified memory, enough to run frontier-class models locally, privately, at a cost that pays for itself in under a year compared to cloud API bills.

Tools like Ollama and LM Studio are making it easy for individuals and small organizations to run powerful open-source models on hardware they actually own. I run Claude Code daily, but have started prototyping my starter code with cheaper local models because I regularly hit my daily token limit.

I think this trend to localization isn't just hobbyist experiments. They're a rehearsal for the world after the bubble explodes. One where compute is owned, not rented, and the cost of intelligence doesn't compound on a tab you can likely, never fully pay off.

The universe has been expanding for 13.8 billion years.

The Big Bang wasn't a disaster. It was the condition of everything that came after. All that energy, released from a single unsustainable point of centralization, became stars and planets and eventually... us.

Maybe the AI bubble popping works the same way. Maybe the best outcome is that the bang happens before the centralized actors have so thoroughly locked up the infrastructure that nobody else can play.

Let it pop. And let the energy go to all of us.

Hey! That’s it for this time. I do this every week; if you vibe to the ideas I express, consider subscribing or sharing with friends. If you like tech-detoxing with a book like I do, I crammed some of last year’s best essays into a printed collection.

This was an improvisation that came from a discussion from a student of mine! I rambled on a morning walk, that became a voice note in Otter.ai, took shape in Obsidian, and was finished in collaboration with Claude Sonnet 4.6.

For more info visit: https://nyewarburton.com

We’ll see you next time.

[1] Per-token prices and enterprise spend data: SoftwareSeni, "The AI Inference Market in 2025" — https://www.softwareseni.com/the-ai-inference-market-in-2025-hardware-consolidation-pricing-wars-and-what-it-means-for-buyers/

[2] Gartner agentic token multiplier (5–30x); OpenRouter token growth data: OpenRouter, "State of AI 2025: 100T Token LLM Usage Study" — https://openrouter.ai/state-of-ai

[3] Token-per-query projections and data center energy forecasts: Illuminem, "The Cost of Context: The Exponential Growth in Tokens" — https://illuminem.com/illuminemvoices/the-cost-of-context-the-exponential-growth-in-tokens

[4] Enterprise AI budget growth ($1.2M to $7M): Oplexa, "AI Inference Cost Crisis 2026" — https://oplexa.com/ai-inference-cost-crisis-2026/

[5] OpenAI revenue and loss figures; Turing Award researcher paper on inference bottleneck: Oplexa, "AI Inference Cost Crisis 2026" — https://oplexa.com/ai-inference-cost-crisis-2026/

[6] Hyperscaler CapEx commitments for 2026: SoftwareSeni, "The AI Inference Market in 2025" — https://www.softwareseni.com/the-ai-inference-market-in-2025-hardware-consolidation-pricing-wars-and-what-it-means-for-buyers/

[7] AI startup VC funding, 2025 figures: Crunchbase / Venture Capital Journal, "Funding for AI Dominated in VC in 2025" — https://www.venturecapitaljournal.com/funding-for-ai-dominated-in-vc-in-2025-crunchbase/

[8] OpenAI, Anthropic, xAI combined valuation: Sapphire Ventures, "Our 2026 Outlook: 10 AI Predictions" — https://sapphireventures.com/blog/2026-outlook-10-ai-predictions-shaping-enterprise-infrastructure-the-next-wave-of-innovation/

[9] VC framing shift on AI startups: GeekWire, "Is There an AI Bubble?" — https://www.geekwire.com/2025/is-there-an-ai-bubble-investors-sound-off-on-risks-and-opportunities-for-tech-startups-in-2026/

[10] EXO distributed inference on Mac hardware; payback period vs. cloud APIs: VentureBeat, "You Can Now Run the Most Powerful Open Source AI Models Locally on Mac M4 Computers" — https://venturebeat.com/ai/you-can-now-run-the-most-powerful-open-source-ai-models-locally-on-mac-m4-computers-thanks-to-exo-labs/ and Awesome Agents, "Mac Studio Clusters Now Run Trillion-Parameter Models for $40K" — https://awesomeagents.ai/news/mac-studio-clusters-local-llm-inference-rdma/