5 Things About AI That Might Surprise You

1. AI is expensive, but you probably don't notice because it's massively subsidized^[1][2][3]

2. AI is probabilistic, not deterministic — you can't expect the same exact answer twice^[4][5]

3. Most AI projects never make it into production^[6][7]

4. AI doesn't know when it's wrong^[8]

5. AI isn't just software — it's an industrial build-out^[9][10][11]

Here's the twist: none of these are reasons to be afraid of AI. Once you actually understand these five things, you can see how each one is part of what makes AI a superpower. Let's take them one at a time.

1. AI is expensive — you just don't feel it yet

AI is expensive. Genuinely, seriously expensive to run. You just don't notice, because it's massively subsidized. A $20/month subscription feels cheap because it is cheap — for you. It doesn't reflect what it actually costs to run the model underneath it. Under heavy use, a single subscriber can cost a provider thousands of dollars a month in compute, far more than the subscription collects^[1]. That gap is covered by venture capital and corporate balance sheets, not by your subscription fee.

This isn't a temporary rounding error. OpenAI alone is projected to burn through $143 billion in cash as it races to build out capacity^[3], and reporting on the major labs makes clear that today's pricing is a land-grab, not a stable business model^[2]. The bill comes due eventually — either through higher prices, more aggressive monetization, or providers that don't survive the transition.

2. AI is probabilistic, not deterministic

Give a calculator the same input twice and you get the same answer twice. Give an LLM the same prompt twice — same model, same settings, even temperature set to zero — and you can still get a different answer. This isn't a bug that better engineering will eventually fix; it's baked into how the math runs. Floating-point operations on a GPU aren't strictly associative, and the order those operations get batched in shifts with server load, which shifts the result^[4][5]. Determinism is a binary property, not a matter of degree — a system that can produce different outputs from identical inputs is, by definition, not deterministic.

Here's the part that turns this from a limitation into a superpower: not every problem wants one correct answer. Ask an image model for "a medieval city built inside the ribcage of a colossal fossilized dragon" and you don't want a single canonical result — you want a hundred different interpretations to pick from. The variation isn't a defect to engineer away. It's the entire feature, and it's exactly why AI is so good at brainstorming, design exploration, and any task that means searching a space of possibilities instead of executing one known procedure.

3. Most AI projects never make it into production

Getting a model to produce an impressive demo is the easy part. RAND Corporation interviewed dozens of data scientists and engineers across more than 50 organizations and found that more than 80% of AI projects fail — roughly double the failure rate of ordinary, non-AI IT projects^[6]. MIT's 2025 study of enterprise generative AI put the number even higher: 95% of pilots never produced a measurable profit-and-loss impact, despite tens of billions of dollars in enterprise spending^[7].

"Failure" here is doing some quiet work, though. A prototype that never scales might be a failure to a CFO looking for ROI and a success to the engineers who learned what wouldn't work. The honest version of this surprise isn't "AI doesn't work" — it's that turning a working model into something that reliably creates value is a much harder problem than building the model in the first place.

4. AI doesn't know when it's wrong

This is probably the biggest conceptual surprise for people who haven't worked closely with these systems. An LLM has no internal flag that lights up when it's making something up versus when it's stating a well-supported fact. Research from OpenAI and Georgia Tech traces this back to how these models are trained and evaluated: the process rewards confident, fluent answers over honest uncertainty, so a model that says "I don't know" gets penalized relative to one that guesses convincingly^[8].

The dangerous output isn't the one that's obviously wrong — you catch that one immediately. It's the answer that sounds exactly as confident and well-reasoned as a correct one, because fluency and accuracy are produced by the same mechanism and the model can't tell them apart any better than you can, at a glance.

This is why you shouldn't trust AI blindly — but it's not a reason to avoid it either. Let it be creative. Let it brainstorm. Let it take swings it wouldn't take if it were worried about being wrong. Just always make it show its work: ask it to research, cite sources, and back up its claims. And whatever it hands you, the final call — the thing you ship, publish, or act on — stays a human decision.

5. AI isn't just software — it's an industrial build-out

It's tempting to think of AI the way we think of an app: write the code, ship it, done. Frontier AI doesn't work that way. Analysts project roughly $7.6 trillion in cumulative capital spending on AI compute, data centers, and power between 2026 and 2031, with annual hyperscaler AI capital expenditure alone projected to grow from around $765 billion in 2026 to $1.6 trillion by 2031^[9]. Global data-center electricity consumption is on track to more than double, from about 415 terawatt-hours in 2024 to roughly 945 terawatt-hours by 2030^[10] — and the binding constraint on how fast AI can scale is increasingly the power grid, not chip supply^[11].

That reframes what "AI capability" actually means. It isn't purely a software or algorithms race anymore — it's also a contest over chips, electricity, land, and capital. A model isn't something you simply download; running one at frontier scale requires physical infrastructure on an industrial scale.

The takeaway

AI is a superpower once you understand these five things — not despite them, but because of them. Here's what that actually buys you:

1. It's subsidized — so you can get amazing value out of frontier AI right now, while someone else's balance sheet is picking up the difference.

2. It's probabilistic — so you can harness that unpredictability to brainstorm and see a problem from a totally different angle than the one you started with.

3. Most projects fail — so once you know why, you know exactly what to avoid and what actually works, instead of learning it the expensive way.

4. It doesn't know when it's wrong — so it will boldly take you to new places, whether you're coding, writing a poem, or generating an image, without the hesitation a "safe" tool would build in.

5. It's an industrial build-out — so this whole ride is running on borrowed time and someone else's capital. Take advantage of the subsidy while it lasts.

The people who get the most value out of AI won't be the ones who trust it blindly. They'll be the ones who know exactly when to verify, when to lean on something else instead, and when to let the uncertainty work in their favor.

References

1. Cybernews (2025). New analysis shows ChatGPT and Claude subscriptions may cost AI firms thousands per user. Link

2. Axios (2026). AI companies like OpenAI, Google cover costs. But not forever. Link

3. eMarketer. OpenAI's forecast $143 billion cash outflow raises stakes for AI monetization. Link

4. Thinking Machines Lab (2025). Defeating Nondeterminism in LLM Inference. Link

5. Unstract. Why is deterministic output from LLMs nearly impossible? Link

6. Ryseff, J., De Bruhl, B. F., & Newberry, S. J. (2024). The Root Causes of Failure for Artificial Intelligence Projects and How They Can Succeed. RAND Corporation, RR-A2680-1. Link

7. MIT Media Lab / Project NANDA (2025). The GenAI Divide: State of AI in Business 2025, as reported by Yahoo Finance. Link

8. Kalai, A. T., Nachum, O., Vempala, S. S., & Zhang, E. (2025). Why Language Models Hallucinate. OpenAI & Georgia Tech. arXiv

9. Goldman Sachs Insights. Tracking Trillions: The Assumptions Shaping the Scale of the AI Build-Out. Link

10. International Energy Agency. Energy and AI — Executive Summary. Link

11. Inflect. Data Center Power Shortage 2026: Why Grid Capacity Is Now the Bigger Constraint Than GPUs. Link