The AI Bubble Nobody Wants to Talk About

Something is wrong with the math. Not the technology — the technology is extraordinary. The math. The economics underneath the most aggressively funded industry in human history don't add up, and almost nobody in a position to say so is willing to say it.

This is not a bearish take on artificial intelligence. It is an honest look at what's actually being built, what it costs, who's paying for it, and what happens when the correction comes. Because it will come. It always does.

What Five Companies Are Spending to Own the Future

In 2024, Microsoft, Google, Amazon, Meta, and Apple spent a combined $400 billion on AI infrastructure. Not research. Not salaries. Infrastructure — data centers, chips, cooling systems, and the land to put them on (company earnings reports, Q4 2024).

To put that in context: $400 billion is larger than the GDP of Denmark. It is more than the United States spent on the entire Apollo program, adjusted for inflation, over thirteen years — spent in a single calendar year by five private companies. Just let that sink in. In only 1 YEAR, only 5 COMPANIES, spent more than the APOLLO SPACE PROGRAM over 13 years.

NVIDIA, which makes the GPUs that power most AI systems, watched its data center revenue jump from $15 billion in 2023 to over $47 billion in 2024. Its market capitalization briefly surpassed $3 trillion. The demand for its chips has been so overwhelming that delivery times stretched to nine to twelve months for the largest orders.

The buildout is real. The infrastructure is going into the ground.

Here is what is also real: the revenue these AI systems are generating is a fraction of what it costs to run them. OpenAI, the most recognized AI company on earth, was projected to LOSE approximately $5 billion in 2024 on roughly $3.7 billion in revenue, according to reporting by The Information. That is a company spending nearly three dollars to make one. Microsoft has invested over $13 billion in OpenAI and has yet to see a profitable AI product line. Google's AI overhaul of search has cost billions and faces a structural crisis: every AI-generated answer it serves is one fewer ad click, which is where Google's actual money comes from.

The global AI market is routinely projected to reach $1.8 trillion by 2030. That projection assumes adoption curves that haven't materialized in previous technology cycles, a revenue model nobody has proven at scale, and something nobody is discussing publicly: that the energy required to run these systems can actually be sourced.

The Energy Problem Nobody Is Solving

A single ChatGPT query uses approximately 10 times the electricity of a standard Google search, according to the International Energy Agency. That gap compounds fast at scale. AI data centers currently consume an estimated 1-2 percent of all global electricity. The IEA projects that figure reaching an 8 to 10 percent by 2030, and are conservative assumptions.

By the end of 2026, data centers are projected to consume more electricity than Japan's entire national grid (IEA, 2024). The response from tech companies has been extraordinary in its ambition and alarming in what it implies. Microsoft has signed agreements to restart Three Mile Island — the nuclear plant synonymous with America's WORST nuclear fears — because conventional power cannot keep up. Google is funding next-generation geothermal projects that have never been built at commercial scale. Amazon is purchasing small modular nuclear reactors that don't yet exist in commercial form.

These are not long-term hedges. These are emergency responses to a problem happening right now, as infrastructure goes in faster than the grid can support it.

The cost of that energy doesn't disappear. It gets built into every API call, every query, every product built on top of these systems. Current AI tool pricing does not reflect actual delivery costs. It reflects what the market will bear while investors remain patient.

The Valuation Gap Nobody Wants to Read

The dot-com crash of 2000 is the standard reference point for technology speculation. At the peak, companies valued at 100 times revenue were called overvalued. The conventional wisdom was that the market had lost its mind.

Several leading AI companies today are valued at multiples that make 100 times revenue look disciplined — with the added structural problem that their core product loses money to deliver. In 2000, overvalued companies were at least not destroying value with every sale.

OpenAI's last private valuation: $157 billion. Anthropic: $61 billion. Neither is a public company with audited financials available for scrutiny. These are private valuations set in funding rounds by investors who need to believe the number to justify the check they just wrote.

Enterprise adoption tells a quieter story. Despite years of headlines and billions in investment, most Fortune 500 companies are still running pilots. Not deployments. Pilots. The gap between "we're exploring AI" and "AI is generating measurable ROI" is wider than the industry projected in 2022, and wider still than it projected in 2023. The case studies being published right now are marketing. The actual internal data at large organizations, when you can find it, is considerably more cautious.

The Commodity Problem the Incumbents Don't Want to Mention

The thing that makes AI infrastructure valuable is scarcity of capability. When only a handful of companies can run large language models, those companies have pricing power. That scarcity is eroding — deliberately, by some of the biggest players in the market.

Meta has released a series of open-source AI models, including Llama 3, that perform comparably to paid alternatives for a wide range of tasks. Running them requires hardware you can own rather than rent. The moment a meaningful percentage of AI workloads shift from proprietary APIs to open-source models on owned infrastructure, the revenue projections for major AI platforms compress dramatically.

This is not a future risk. It is happening now. The companies spending the most to build proprietary model advantage are racing against their own industry's gravitational pull toward commoditization — accelerated by one of the industry's largest players.

What a Correction Actually Looks Like

The dot-com crash did not end the internet. It ended the version that assumed unlimited capital and no requirement to ever turn a profit. What followed was a decade of consolidation and failure — and then Amazon, Google, and Facebook built the most profitable businesses in history on the wreckage.

The AI correction will follow a similar shape. Not a crash of the technology, but a crash of the economics that currently surround it. Some companies will disappear. Valuations will compress. Pricing for AI products will have to reflect actual costs rather than subsidized access.

What comes after is probably more durable and more interesting than what we have now. The hype burns off and what's left is the technology that actually works.

What Small Creators Can Do Right Now — Before the Window Closes

The tools available today at current price points represent one of the most asymmetric opportunities in the history of small business. A solo creator in 2026 has access to capabilities that would have required a team of twelve and a six-figure budget five years ago. That access exists because the companies providing it are subsidizing your entry with capital that is not yet profitable. They need users. They need adoption data. They need to demonstrate scale to justify the next funding round.

That subsidy will not last indefinitely. Here is how to use it while it does.

Build real skills, not just tool habits. There is a difference between knowing how to use ChatGPT and understanding how to direct AI systems to produce work that reflects your specific taste, voice, and standards. The first skill is temporary — the tool will change. The second is permanent and compounds. The creators who will be fine after the correction are the ones who used the current window to develop genuine AI fluency, not just a subscription.

Stack the capabilities that cost the most to replace. Right now you can access research tools, writing tools, image generation, video assistance, SEO analysis, and analytics — often for under $200 a month combined. Use that stack to build assets that will outlast the pricing: a content library, a brand system, a newsletter audience, a body of work that ranks. The subscriptions may get more expensive. The assets you built with them are yours.

Price yourself as if the tools already cost what they will cost. The biggest strategic mistake small creators are making right now is passing the AI cost savings directly to clients or keeping their own rates low because "it's faster now." If your deliverable quality went up and your delivery time went down, your price should go up — not down. Protect your margin. The correction will hit your costs before it hits your clients' expectations.

Build redundancy before you need it. If your entire workflow runs through one platform, you are one terms-of-service update away from a crisis. Learn two tools that can do the same job. Export your data regularly. Document your processes somewhere you own. The creators who survived every previous platform shift had one thing in common: they never let a single company hold everything.

Own an audience outside the algorithm. A newsletter, a small community, a mailing list — anything that lets you reach the people who care about your work without paying a platform for the privilege. Every piece of AI-generated content you create right now should have one job beyond the content itself: moving someone from a platform you rent to a channel you own. That is the only durable asset in the creator economy, and AI makes building it faster than it has ever been.

The window is open. The question is what you build while it is.

Angel is the founder of Loftiideas, a Toronto-based Creative Studio exploring the intersection of creativity, technology and psychology. She's been inside the AI economy long enough to find it genuinely remarkable — and long enough to read the footnotes in the earnings reports. She writes about brand building, the creator economy, and the very human experience of building something that lasts.

If you're thinking about how to build with AI in a way that actually holds up — let's talk.