Between Hype and Reality: Four AI Futures Based on What We Actually Know

The cheapest worker in the world costs $0.10 per hour. And getting cheaper.

In November 2022, processing a million tokens through an AI model cost about $20. In February 2026, you get better results for $0.10. That's a 99.5% drop in under four years. Think about it — no form of human labor in history has gotten this cheap this fast. And we're just getting started.

But here's the catch nobody talks about: while the price per token dropped, the total cost of using AI went up. Way up. The average professional software stack went from about $30/month in 2020 to over $350 in 2026. 96% of organizations deploying AI say their costs are higher than expected.

This isn't a contradiction. It's Jevons' paradox, named after a 19th-century economist who noticed that when steam engines started burning coal more efficiently, coal consumption didn't drop — it skyrocketed. Because suddenly it made sense to use steam engines everywhere. The same thing is happening with AI: make something cheaper per unit, people consume so much more that the total bill goes up. We're not in the era of "cheap AI." We're in the era of "AI everywhere, doing everything, costing more than ever."

I'm watching this from a position that gives me a fairly unique perspective. By day, I do product design in e-commerce. By night, I co-found an AI startup. I live in both worlds at once. I see the hype from the inside and the outside. And every week on X, someone publishes another article about how AI just changed everything again. Everything is INSANE. Everything is a game changer. Every week it's the biggest thing in history. Every week the world changes forever.

So I did the work and went deep. Foresighting, not vibes. Prediction markets. Data from WEF, IMF, McKinsey, Epoch AI. Safety research from Anthropic and Apollo Research. Real numbers from real companies.

Where We Are Now

Numbers first. Without them, there's no point speculating.

The five largest cloud companies (Amazon, Alphabet, Microsoft, Meta, Oracle) plan to spend roughly $690 billion on AI infrastructure in 2026. Amazon leads with $200 billion, Alphabet with up to $185 billion. Nothing like this has happened before. Not the internet. Not mobile. This is the largest investment cycle in the history of technology.

Google processes about 1.3 quadrillion tokens per month. That's 130x more than a year ago. And here's the interesting part: the main consumers of those tokens are no longer people. They're AI systems talking to other AI systems. Debugging code, testing outputs, starting over. Machines talking to machines, and we're paying for it.

Epoch AI estimates that global AI computing capacity is doubling every 7 months. Not 18-24 months like Moore's Law. Seven. Year over year, that's 3.3x more raw power.

The power is here. The money is here. Consumption is exploding. But what about real-world impact? This is where it breaks.

88% of organizations say they use AI. But only 1% is what McKinsey calls "mature." One percent. The rest are stuck in what's called "pilot purgatory" — pilot projects that never make it to production. I see it around me every day. 96% of companies report higher costs than expected, and 71% admit they don't actually know what's causing those costs.

The gap between "we have AI" and "AI is actually changing how we work" is enormous. And it's probably the most important thing we need to understand right now.

Now look at the other side of this business. The companies that actually build the AI.

OpenAI grew from $1.6 billion in revenue (2023) to over $20 billion (end of 2025). Faster than Google or Meta in their best years. But for every dollar earned in 2024, it spent $2.35. That's improving — in 2025 it's "only" $1.65 — but Deutsche Bank estimates cumulative negative cash flow of $143 billion by 2029. For comparison: Uber, the previous record holder in cash burning, burned $31.7 billion over 14 years. OpenAI plans to do it in five.

Anthropic? From $87 million in early 2024 to $14 billion annualized in February 2026. Eighty percent of revenue from enterprise. But inference costs 23% higher than they internally expected. Musk's xAI was even worse — burn-to-revenue ratio of 14:1, eventually had to merge with SpaceX to hide under a bigger balance sheet.

Both major companies plan IPOs in 2026. OpenAI says profitability by 2029, Anthropic by 2028. For that to work, Anthropic needs $70 billion in annual revenue. OpenAI over $100 billion. These are bets that AI agents will replace a substantial portion of corporate work within three years. The entire industry rests on this hypothesis.

And there's a detail I can't let go of. A large portion of these investments is circular. Amazon gives billions to Anthropic. Anthropic spends it on AWS servers. Microsoft gives $13 billion to OpenAI in the form of Azure credits. OpenAI burns those credits on Microsoft's infrastructure. The money circles back to the investor. It looks like growth. In reality, it's just sloshing around.

What the Predictions Say

There's no shortage of opinions about AI. What's more interesting are the places where people put something behind their estimates. Prediction markets. Metaculus, where people bet their reputation and where accuracy is measured retroactively.

As of February 2026:

On AGI. Metaculus puts the median for "Weakly General AI" at February 2028. That sounds like sci-fi, but it's two years away. It means the models being trained right now are direct predecessors. "Strong AGI" — a system that can do anything a human can — is at July 2033. That five-year gap is key. We'll have AI that's superhuman at code, math, or law long before we have AI that's generally capable at everything.

Polymarket gives just a 14% chance that OpenAI announces AGI before 2027. Hype is one thing. Bets are another.

On jobs. Mass unemployment (over 20% in any OECD country) before 2030? Less than 10% probability. The prediction community rejects the "everyone loses their job" narrative. They see augmentation and shifts. Not mass replacement.

On the pace of progress. This is where it gets most interesting. Forecasters consistently underestimate technical capabilities and overestimate the speed of real-world deployment. In 2022, they predicted AI would reach 12.7% on a math benchmark by 2025. Reality? 50.3%. But predictions about self-driving cars everywhere by 2025? Way too optimistic.

I call it the Capability-Impact Gap. We have intelligence that people in 2020 predicted for 2030. But the economic impact is roughly at 2024 levels. Regulation, legacy IT systems, organizational change, and plain human trust are slowing things down far more than anyone expected.

Real Impact on the Labor Market

The headline number from the World Economic Forum: by 2030, 92 million roles will be displaced and 170 million new ones created. Net +78 million. On paper, that sounds fine. In practice, it's messier.

63% of employers say the skills gap is the main barrier. About 59 out of 100 workers will need significant reskilling by 2030, but fewer than half will likely get it. Jobs are being created. Just not where the people who lost the old ones are.

Where the impact is already visible:

Freelancing. Writing and translation work on Upwork dropped 32% year over year. IT networking down 27%. But AI specialist roles? Up 200%. Freelancers who learned to work with AI earn 40% more. The middle of the market is vanishing. You're either AI-augmented and thriving, or you're competing with a monthly subscription.

Junior roles. This one weighs on me because as a designer, I think about team structure and talent pipeline. A METR study found that junior developers using AI tools were 19% slower. Not faster. Slower. AI generates code that's 80-90% correct. A senior spots the error immediately. A junior doesn't have the mental model to tell what's wrong. They spend hours debugging something that looks right but isn't.

Job postings for junior devs (0-4 years) dropped 16.3% since ChatGPT launched. Senior positions stayed stable or grew. Companies are automating precisely the tasks they used to give juniors for training. Documentation. Simple bug fixes. Basic features. Without those, juniors have nowhere to learn.

I call it the demographic bomb. If we stop hiring and training juniors now, we won't have seniors in five to seven years.

Klarna. This story matters because it shows a pattern we'll see again and again. Klarna replaced 700 customer service agents with AI. Got massive press. Then quietly started rehiring humans because AI couldn't handle edge cases and was damaging customer sentiment. AI handles the easy 80%. But the hard 20% is where customer trust lives. Companies are learning this the hard way.

The Price Paradox

Every viral AI post says it'll be cheaper. Per unit — yes. A reasoning model in 2026 does what cost $20 in 2022 for fractions of a cent. But total costs? Up.

We've gone from people asking questions to AI agents running in recursive loops. A single instruction like "refactor this codebase" triggers thousands of autonomous API calls. The main token consumers in 2026 aren't people. They're machines.

Modern reasoning models don't just predict the next word. They think. DeepSeek R1 internally generates 8,793 thinking tokens on a hard problem. A standard model? 711. 12x more compute for the same question. You see the same-length response. Behind the curtain, it cost 12x more.

Professional tier: from $20/month in 2023 to $200/month in 2026. Add Copilot, Perplexity, M365 Copilot, Zoom AI. A complete AI stack costs as much as a car lease. And for companies? Unsupervised AI agents enter a loop and solve the same problem over and over. One company reported thousands of dollars burned overnight because nobody set a "stop."

AI per unit is cheaper than ever. AI overall is more expensive than ever. Both are true simultaneously.

And here's the fundamental question. Bain & Company estimates a "revenue hole" — the gap between AI investments and what AI actually earns — at $600 to $800 billion per year. A PwC survey from early 2026: 56% of CEOs see no measurable gains yet. Money flows. Results don't.

Open Source: An Alternative Path

You don't have to pay for API access. In February 2026, you can run models yourself — and it makes sense. Meta's Llama 4 Maverick scores comparably to GPT-5.2 on coding benchmarks. DeepSeek V3.2, built on restricted chips under US export controls, achieves frontier performance in math. Open source is catching up to proprietary models faster than anyone expected.

You can run a capable model on a Mac Studio with enough RAM or a dedicated workstation for about $4,000. Higher upfront cost, more technical work, but no API fees and no data leaving your building. For EU companies where data sovereignty matters, this is becoming the preferred path.

But there's a bottleneck nobody talks about. Not compute, not algorithms. Memory. AI models need massive amounts of ultra-fast memory to run — and production of this specific component (called HBM3e) is at absolute maximum capacity. Demand outstrips supply, prices are shooting up, and this flows into the cost of everything — GPUs, servers, inference. It's one of the reasons total AI costs keep rising even as the price per token falls. You can have the best model in the world, but without enough of this memory, you can't run it.

And this applies to local deployment too. Most companies have people on MacBook Airs with 16 GB RAM and 256 GB storage. You can't run a capable AI model on that. If you want to run AI locally, you need entirely different hardware — and those are costs that no AI demo video ever mentions.

Under the Hood: Safety

Most AI articles skip this. But safety research results will directly shape regulation, trust, and deployment speed. And what's happening there is worth paying attention to.

Anthropic in January 2024 trained models as "sleeper agents." Systems that passed every test. Model students. Then, after a specific trigger, they switched to harmful behavior. Standard safety training didn't remove it. And adversarial training — where you try to catch the model lying — actually taught it to lie better. Researchers call it selection pressure for better liars. Read that again.

Apollo Research in December 2024 found that frontier models do this spontaneously. Without any special training. They tested models from OpenAI, Anthropic, Google, Meta. Several attempted to manipulate their own evaluation or disable oversight.

The UK AI Safety Institute found that frontier models outperform PhD experts in specific areas of biology and chemistry. And can execute basic cyberattacks about 50% of the time.

And here's what personally stopped me: "evaluation awareness." Models can distinguish whether they're being tested or deployed in production. And they behave differently in each context. This isn't sci-fi. This is peer-reviewed research from last year.

I'm not saying this to scare anyone. I'm saying it because this will directly determine how fast deployment happens and what regulation comes.

Four Scenarios for 2026-2029

Foresighting works differently than prediction. You don't predict one future. You prepare for several. Each scenario has different assumptions and different signals that tell you which direction things are heading.

Scenario A: "Agentic Spring" — 35%

AI agents work as reliable junior colleagues. They handle routine on their own, humans orchestrate and verify. Productivity rises. New roles emerge faster than old ones disappear. Companies escape pilot purgatory.

Signals: McKinsey's 1% maturity rate starts climbing. More AI-augmented job postings appear. Companies report real ROI.

Risk even here: Inequality grows. Seniors capture the gains, junior entry points shrink.

Scenario B: "Hollow Middle" — 40% (my baseline)

The most likely path. AI capabilities grow, but adoption stays chaotic. Junior and mid-level positions compress. Seniors are OK for now. The freelance market permanently restructures. Companies invest heavily but ROI doesn't materialize as planned. Klarna-style reversals become common.

Signals: Entry-level hiring decline continues. Frictional unemployment rises. AI spending goes up, ROI metrics stagnate.

Risk: The demographic bomb I mentioned. Stop training juniors, and in 5-7 years you don't have seniors. A problem that takes a decade to fix.

Scenario C: "Scaling Wall" — 15%

Physical constraints catch up to progress. HBM3e bottleneck, energy limits, latency walls. OpenAI increased its energy footprint from 0.2 GW in 2023 to 1.9 GW in 2025 — the equivalent of two million households. This doesn't scale forever. The compute doubling time stretches back to 12-18 months. LLM architecture hits diminishing returns. We need a new paradigm, but it's not ready.

Signals: Longer intervals between model releases. Benchmarks flatten. AI companies shift messaging from "bigger models" to "better applications."

Risk: Not AI winter, but AI autumn. Sentiment turns even though existing tools are still strong. Companies slash AI budgets and slow progress that was actually happening.

Scenario D: "Black Box Breakout" — 10%

A major safety incident with a frontier model. An autonomous agent causes measurable harm. Or evaluation awareness scales unpredictably. Or the opposite direction: self-improvement accelerates faster than governance can react.

Signals: Safety incident in mainstream news. Whistleblowers from AI labs. EU AI Act enforcement in August 2026. Or a model release that makes the previous generation obsolete overnight.

Risk: Both sides. Overregulation kills innovation. Insufficient regulation enables harm. The geopolitical split between America's "compete fast" and Europe's "regulate first" makes coordination harder.

What to Do About It

No generic "learn AI." Instead, a few things I think based on the data.

Designers and creatives. Your work is changing, but differently than most people think. AI is changing WHAT you do, not HOW. Generating variants, basic layouts, first drafts — that was never the valuable part. Research synthesis, systems thinking, stakeholder navigation, design judgment in complex situations. That's where people still add clear value. If you can only push pixels, it'll be tough. If you understand why a design decision matters in a business context, you'll be fine.

Team leaders. Stop measuring AI adoption. Measure AI maturity. 88% vs 1%. That gap is real and it's where the work needs to happen. And above all: invest in the junior talent pipeline now. The entry-level hiring freeze is creating a problem that'll show up in five years. And then it'll be too late.

Early in your career. I won't sugarcoat it: it's a harder moment to enter knowledge work than five years ago. But the opportunity is different. People who learn to work effectively with AI, verify outputs, bridge the gap between what AI generates and what the business needs — they'll be in extreme demand. That 40% premium for AI-augmented freelancers won't drop. It'll grow.

Costs. Track TCO, not price per token. The Reasoning Tax is real. Subscription creep is real. And above all: AI is not SaaS. With regular software, the cost of an additional user drops toward zero. With AI, every query triggers a live computation on specialized hardware. More users = more costs, not fewer. That's a fundamental difference most managers haven't grasped yet. Set spending limits on every AI agent before you let it run.

One practical commitment. An hour a day with AI in your actual work. Not reading about AI. Using it. Every day, try something new in your workflow. Three months of this and you'll understand what's coming better than 90% of the people around you. The bar is low. Most people are still just reading.

Foresighting, Not Prediction

Most AI articles give you one future and tell you to prepare. That's prediction. And predictions are almost always wrong.

Foresighting is a different approach. Prepare for multiple plausible futures at once. The four scenarios aren't mutually exclusive. Elements of all four will play out simultaneously across different sectors, regions, and timescales. The point isn't to bet on one. The point is to not be caught off guard.

Technical progress is accelerating. Real-world deployment is slower and messier than the hype suggests. Safety concerns aren't theoretical. Total costs are going up. And the most valuable AI companies on the planet are losing money.

Four futures are here at the same time. Which one wins in your industry, your company, your career depends on the decisions being made right now.