Last week OpenAI shut down Sora, its video generation tool, after months of hemorrhaging money at a staggering pace. By most estimates the service was costing around $15 million a day in compute while generating just $2.1 million in total revenue across its entire lifespan. Its user base had collapsed by two-thirds from its peak, and competitors like Runway, Kling, and Google’s Veo had matched or surpassed its output quality at lower prices. OpenAI made the rational call: kill the project, free up the compute, redirect resources toward products that are actually generating revenue. It was, by any reasonable measure, a straightforward business decision.

The internet, predictably, treated it as something much larger. Within hours the commentary had escalated from “Sora failed” to “AI is failing.” If OpenAI, the most well-funded AI company on the planet, couldn’t make a flagship product work, what does that say about the rest of the industry? How long before ChatGPT follows? Before Copilot? Before Claude? The underlying anxiety is clear enough: people have started depending on these tools, and they’re afraid that dependence is built on something that can’t last. That one day they’ll open their laptop and find it all gone, pulled away the moment the venture capital dries up.

I think this fear is understandable, but I also think it’s wrong. And I think you can see why it’s wrong if you look at three things: how fast AI costs are actually falling, what Sora’s failure really tells us, and the bizarre double standard we apply to AI that we’ve never applied to any other technology we depend on.

The cost curve tells a different story

The most important fact about AI economics right now is not how much it costs to run these systems. It’s how fast those costs are dropping. The price of running a large language model has been falling at roughly 10x per year, a rate of decline that has no real precedent in the history of technology. When GPT-3 first became publicly accessible in late 2021, inference cost about $60 per million tokens. Today, models that match or exceed GPT-3’s capabilities are available for about six cents per million tokens. That’s a thousandfold reduction in under five years.

To appreciate how unusual this is, consider some comparisons. High-definition televisions, one of the more dramatic consumer price drops in recent memory, fell from around $8,000 to under $1,000 over the course of roughly a decade. Home broadband prices dropped perhaps 25% over a two-year stretch. The cost of a Microsoft Office license fell about 15% between 2001 and 2023. These were all considered significant price reductions at the time. None of them come remotely close to the trajectory of AI inference costs, which are collapsing faster and deeper than any comparable technology category over a similar timeframe.

And the decline isn’t slowing. Hardware efficiency is improving at about 30% annually. Energy efficiency gains are running at roughly 40% per year. On top of that, algorithmic breakthroughs—things like better model architectures, more efficient training methods, and smarter inference techniques—are compounding the hardware gains. The claim that AI is “too expensive to sustain” doesn’t just seem premature. It runs directly against the single clearest trend in the data.

What Sora actually tells us

There is an important distinction between “AI is too expensive” and “generating high-fidelity video on demand for a small and shrinking user base is too expensive.” Sora’s problem was not that artificial intelligence costs too much in some general sense. Its problem was specific and structural: each video generation request consumed enormous compute resources, users weren’t willing to pay anywhere near the true cost of production, and the audience was already walking away to cheaper alternatives.

Text-based AI tools simply don’t share this cost profile. They’re far cheaper to run per request, their utility is more immediately obvious across a wider range of tasks, and critically, people are paying for them. ChatGPT has over a hundred million users. GitHub Copilot has millions of paying subscribers. Enterprise spending on LLM APIs is growing quarter over quarter. These are not products being propped up by speculative hope. They have actual, growing revenue streams, and the cost to serve each user is falling every quarter.

Sora failed the way products always fail: it cost too much for what it delivered, the competition caught up, and users left. That tells you something about the economics of on-demand video generation in early 2026. It tells you essentially nothing about the viability of AI tools as a category, in the same way that the failure of Quibi told you nothing about the future of streaming video, and the collapse of Pets.com told you nothing about whether people would eventually buy things on the internet.

The infrastructure double standard

Here is what I find genuinely strange about the anxiety around AI costs: we live in a world surrounded by infrastructure that has never paid for itself, and we never give it a second thought.

The United States spends approximately $626 billion a year on transportation and water infrastructure alone. Highway spending runs over $200 billion annually between federal and state governments, and the gas taxes that are supposed to fund roads have never actually covered their costs. The Highway Trust Fund has required regular bailouts from general tax revenue since 2008. The federal government recently committed $65 billion to broadband subsidies through the Infrastructure Investment and Jobs Act, an acknowledgment that the private market alone will not deliver internet access to large parts of the country.

And this isn’t unique to the United States. Canada is staring down an estimated $270 billion infrastructure deficit, and its 2025 federal budget committed $159 billion in infrastructure spending over the next five years while running an annual deficit of roughly $78 billion. The most ambitious new project on the table is Alto, a high-speed rail line connecting Toronto to Quebec City—approximately 1,000 kilometres of dedicated, electrified track with speeds above 300 km/h. Total estimated cost: somewhere between $60 billion and $90 billion, with a $3.9 billion design phase alone expected to take four to five years before a single rail gets laid. Construction on even the first Ottawa-Montreal segment isn’t expected until 2029. Nobody is writing panicked op-eds about whether Alto is “sustainable.” The conversation is about when it gets built, not whether the concept of passenger rail is viable.

The pattern isn’t new, either. The transcontinental railroads, the great infrastructure achievement of the nineteenth century, were built on federal land grants and below-market government loans. Rural electrification only happened because private utilities looked at the cost of stringing wire to remote farmhouses and decided it wasn’t worth the investment, so the Roosevelt administration stepped in with subsidized lending programs. The early telephone network was a regulated monopoly, given special tax treatment and guaranteed returns on capital, because that was the only way to make universal service economically viable.

None of this infrastructure pays for itself in any strict financial sense. Roads don’t generate a profit. The electrical grid doesn’t generate a profit. Rural broadband will likely never generate a profit. We fund these things because they’re useful, because the economic value they enable vastly exceeds their direct costs, and because we’ve collectively decided that the alternative—not having them—is worse. Nobody writes anxious think pieces about whether the highway system is “sustainable.” Nobody worries that the power grid might “go away” because it depends on public subsidies. We just accept that some things cost money and are worth it.

Yet AI, a technology that is actually on a credible path toward self-sustaining economics, that is getting cheaper at a historically unprecedented rate, and that is already generating billions in direct revenue, is treated as though it might vanish the moment someone looks too closely at the books. The double standard is remarkable.

The real fear isn’t about technology

I suspect that what’s actually driving the anxiety isn’t a considered assessment of AI economics. It’s something simpler: people are afraid of losing free access. Many AI tools currently offer generous free tiers, or are priced well below their actual cost of delivery, as a strategy for acquiring users and building market share. When something like Sora gets shut down, it confirms a lurking suspicion. This was all too good to be true. Eventually they’ll take everything away.

But this is a pricing concern, not a technology concern, and the distinction matters enormously. Free tiers will probably shrink. Prices will settle at sustainable levels. Some tools will get more expensive. That’s the normal maturation of any market, and it’s not the same thing as the technology disappearing. Your electricity bill went up this year. You didn’t conclude that electricity is a fad. Grocery prices have climbed steadily for years. You didn’t stop eating. The price of a thing adjusting to reflect its actual cost is not evidence that the thing is doomed. It’s evidence that it’s becoming real.

What comes next

Every transformative technology passes through a period where the costs look unsustainable and the skeptics seem vindicated. The internet had its dot-com crash, and plenty of smart people declared the entire enterprise a speculative bubble that had finally popped. Cloud computing was dismissed as too expensive and too unreliable for serious workloads for nearly a decade before AWS became the most profitable division at Amazon. Early mobile phones were thousand-dollar bricks that could barely hold a signal, and the conventional wisdom was that they’d remain niche luxuries for executives and stock traders.

In each case, the pattern was the same: costs dropped, use cases crystallized, and what once seemed like an expensive novelty became infrastructure so mundane that we forgot it had ever been controversial. AI is following the same trajectory, arguably at a faster pace than any of those predecessors. The cost curves point unmistakably downward. Adoption is broad and accelerating. Revenue is real and growing.

Sora dying doesn’t signal the beginning of the end. It signals that the market is maturing, that companies are learning to distinguish between expensive experiments and sustainable products, and that the era of throwing money at every possible AI application without regard for unit economics is coming to a close. That’s not a crisis. That’s the thing that’s supposed to happen.

Sora died because it was a bad business. AI is not a bad technology.