The Economics of AI — When the Math Doesn't Work

On March 29, 2026, OpenAI quietly pulled the plug on Sora, its flagship video generation product. No grand farewell. No "pivot to a new vision" blog post. Just a tersely worded announcement that the service would be winding down over the next 30 days. The reason was as blunt as it was damning: Sora was burning approximately one million dollars per day in inference costs, and after eighteen months of availability, it had failed to attract more than 500,000 paying subscribers.

Let that math sink in. A million dollars a day. Three hundred and sixty-five million a year. To serve fewer than half a million users, most of whom were on the $20/month tier. Even if every single user paid the maximum subscription price, the revenue would not have covered a quarter of the operating costs. This was not a product struggling to find product-market fit. This was a product that had found its market and discovered the market was not large enough, not willing to pay enough, and not engaged enough to justify the staggering computational expense of generating video from text.

But the Sora shutdown is not just an OpenAI story. It is the clearest signal yet that the AI industry is entering its reckoning phase — the moment where the difference between a technically impressive demo and a commercially viable product becomes impossible to ignore. And the casualties are piling up.

Disney, which had been in advanced negotiations for a reported billion-dollar licensing and integration deal with OpenAI built around Sora's capabilities, saw that deal evaporate overnight. Runway, the other major player in AI video generation, has been quietly laying off staff and raising emergency bridge rounds. The broader AI video generation space, which attracted over $4 billion in venture capital between 2023 and 2025, is now facing a collective existential crisis.

The core problem is not that AI video generation does not work. It works remarkably well. The problem is that it costs too much to run, and the people who want it do not want to pay what it actually costs. This is the fundamental tension at the heart of AI economics in 2026, and it extends far beyond video. Across every modality — text, code, images, audio, video, 3D — there is a spectrum of economic viability, and the companies that fail to understand where their products fall on that spectrum are going to burn through their capital and vanish.

This article is about that spectrum. It is about which AI products actually make money, which ones are bleeding cash while pretending they are not, and what the economics of inference mean for every developer, founder, and enterprise buyer trying to figure out where to place their bets.

The Inference Cost Spectrum: Not All AI Is Created Equal

To understand why Sora died while ChatGPT prints money, you need to understand the fundamental economics of AI inference — the cost of actually running a trained model to produce output for a user.

Every AI product, at its core, takes some input, runs it through a neural network, and produces some output. The cost of that process depends on three things: the size of the model, the amount of computation required per token or frame or pixel of output, and the volume of output the user expects. And these three variables create an enormous range of costs across different AI modalities.

Text generation — the kind that powers ChatGPT, Claude, and every coding assistant — is astonishingly cheap. A typical GPT-4-class response costs somewhere between one and five cents to generate, depending on length and complexity. At scale, with batching and optimization, that number drops even further. A user who sends fifty messages a day to ChatGPT might cost OpenAI a dollar or two in compute. Against a twenty-dollar monthly subscription, that is a healthy margin.

Code generation is slightly more expensive because it often involves longer context windows and more complex reasoning chains, but it is still firmly in the "affordable at consumer price points" category. Anthropic's Claude Code, GitHub Copilot, and similar products can operate at margins that traditional SaaS companies would envy.

Image generation sits in an interesting middle ground. A single image from a model like DALL-E 3 or Midjourney costs roughly four to ten cents to generate, depending on resolution and model size. That is cheap enough to build a business around, as Midjourney has proven, but expensive enough that you need to be thoughtful about pricing and usage caps. You cannot give away unlimited high-resolution image generation at twenty dollars a month and expect to survive.

Audio generation and speech synthesis are moderately expensive but manageable, particularly for short-form content. The real-time streaming requirements add infrastructure costs, but the underlying inference is not dramatically more expensive than text.

And then there is video. Video generation is where the economics go catastrophic. Generating even a fifteen-second clip at reasonable quality requires running a diffusion model through hundreds of denoising steps across dozens of frames, each of which is essentially a high-resolution image. A single sixty-second HD video can cost ten to twenty dollars in raw compute. When your user is paying twenty dollars a month and expects to generate multiple videos per session, the math does not just fail to work — it fails spectacularly.

3D model generation is even worse. Generating a detailed, usable 3D asset from a text prompt requires multiple passes, consistency checks across views, and often iterative refinement. Costs per generation can easily exceed fifteen to twenty dollars, putting it firmly in the territory of "enterprise-only pricing or death."

The chart above shows the general trajectory of inference costs, indexed to Q1 2023 as 100. Costs have fallen dramatically — roughly 80 percent over three years — but this decline has been roughly uniform across modalities. If text inference cost a penny per request in 2023 and now costs a fifth of a penny, that is great. But if video inference cost fifteen dollars per request and now costs three dollars, that is still completely untenable for a consumer product.

The declining cost curve helps everyone, but it helps the already-cheap modalities far more than it helps the expensive ones. Text-based AI products have gone from "good margins" to "incredible margins." Video-based AI products have gone from "catastrophically unprofitable" to "merely ruinously unprofitable."

This is the inference cost spectrum, and it is the single most important factor determining which AI companies survive and which ones die.

The Winners: Text and Code AI Print Money

Let us start with the good news. The AI products built around text and code generation are not just surviving — they are thriving. They represent the clearest product-market fit in the history of enterprise software, and their unit economics are extraordinary.

ChatGPT: The Cash Machine

OpenAI's ChatGPT crossed 400 million weekly active users in early 2026. At a blended average revenue per user that includes free tier users, Plus subscribers at $20/month, and Team/Enterprise contracts, the product is generating an estimated $8-10 billion in annualized revenue. Against inference costs that, at scale, average out to less than $2 per user per month for typical usage, the gross margins are staggering.

ChatGPT works economically because text is cheap, engagement is high, and the willingness to pay is strong. A knowledge worker who uses ChatGPT fifty times a day to draft emails, summarize documents, and answer questions gets enormous value from the product. Twenty dollars a month is a trivial expense against the productivity gains. And serving those fifty requests costs OpenAI somewhere between fifty cents and two dollars, depending on the complexity of the queries and the model used.

The enterprise tier is particularly lucrative. Companies paying custom enterprise rates — often $50 to $100 or more per seat per month — are getting dedicated capacity, compliance features, and data governance. The marginal cost of serving these users is barely higher than consumer users, but the revenue is three to five times greater. This is classic SaaS economics at their best: low marginal cost, high switching costs, increasing returns to scale.

Claude Code and GitHub Copilot: The Developer Tax

If ChatGPT is the consumer success story, coding assistants are the enterprise success story. GitHub Copilot now has over 15 million paying subscribers, and Anthropic's Claude Code has rapidly grown to an estimated 4 million users since its full launch. These products represent something rare in software: a genuine productivity multiplier that developers will pay for out of their own pockets if their employer will not.

The economics work because code generation, while slightly more expensive per request than conversational text, delivers disproportionate value. A developer who saves thirty minutes a day using Copilot or Claude Code is recovering roughly $100-$200 per day in salary equivalent. Against a $10-$40 monthly subscription, the ROI is absurd.

The margins on coding AI products typically range from 60 to 75 percent, which is remarkable for products that are doing real-time neural network inference on every keystroke. The key factors enabling these margins are aggressive model optimization (distilled models for autocomplete, larger models only for complex generation), smart caching (many code patterns repeat, and cached responses are essentially free), and pricing that reflects the genuine value delivered.

Cursor, the AI-native IDE, has been particularly clever about this. By owning the full editor experience, they can make intelligent decisions about when to call a large model versus a small one, when to use cached results, and when to batch requests. This kind of inference orchestration is becoming a core competency for profitable AI companies.

These products also benefit from strong retention. Developer tools are notoriously sticky — once a developer's workflow depends on AI-assisted completion, removing it feels like losing a limb. Churn rates for coding AI products are estimated at less than 3 percent monthly, which is exceptional by any SaaS standard.

Enterprise Search and Knowledge Management

The third major winner in text-based AI is the enterprise search and knowledge management category. Products like Glean, Moveworks, and various RAG-based (Retrieval Augmented Generation) internal search tools are charging enterprise customers $15-30 per seat per month to make their internal knowledge bases searchable and queryable via natural language.

The inference costs for these products are modest — a typical query involves an embedding lookup, a retrieval step, and a text generation step, totaling perhaps two to five cents per query. An employee who makes twenty queries a day costs the provider less than a dollar. Against enterprise pricing, the margins are excellent.

The pattern across all these winners is clear: text-based inference is cheap enough that you can build real businesses with real margins at price points that customers are willing to pay. The value delivered exceeds the cost of delivery by a wide margin, and the products are sticky enough to sustain low churn. This is what a viable AI business looks like.

The Losers: Video and 3D Generation Collapse Under Their Own Weight

Now for the bad news. If text-based AI is printing money, video and 3D generation are incinerating it at a pace that should terrify every investor in the space.

Sora: A Post-Mortem

Sora's failure is the defining cautionary tale of the AI hype cycle. When OpenAI first demonstrated Sora in February 2024, the tech world lost its collective mind. The quality of the generated videos was unprecedented. Hollywood studios panicked. VCs threw money at every video generation startup they could find. The narrative was set: AI was going to revolutionize video production, and OpenAI was going to lead the charge.

The reality was far more sobering. When Sora launched as a product in late 2024, it faced immediate challenges. Generation times were slow — often minutes for a short clip. Quality, while impressive in cherry-picked demos, was inconsistent in real-world usage. And the costs were astronomical.

User growth plateaued at around 490,000 paying subscribers by mid-2025 and then began a slow decline. The core problem was that Sora sat in an awkward no-man's-land: too expensive for casual users who just wanted to make fun clips, and not reliable or controllable enough for professional video production. The professional users who might have justified higher pricing needed frame-level control, consistent characters, and predictable outputs — capabilities that Sora could not consistently deliver.

Meanwhile, the costs were relentless. Each active user was generating an average of 8-12 videos per month. At an average cost of $5-8 per video (blending short low-res clips with longer high-quality ones), the per-user cost was $40-96 per month. The $20 subscription tier was underwater by a factor of two to five. Even the $200/month Pro tier, which attracted fewer than 20,000 subscribers, was barely breaking even per user — and nowhere close to covering the fixed infrastructure costs.

The Disney deal was supposed to change the calculus. A billion-dollar partnership would have given OpenAI a marquee customer, a massive revenue injection, and a showcase for Sora's capabilities in professional production. But Disney's due diligence team reportedly ran their own cost analysis and concluded that the per-minute generation costs made Sora uneconomical even at studio budgets. When your product is too expensive for Disney, you have a problem.

OpenAI's internal projections, according to people familiar with the matter, showed that Sora would need to either reduce inference costs by 90 percent or increase its user base to over 5 million paying subscribers to reach break-even. Neither was achievable on any reasonable timeline. The decision to shut down was painful but rational.

Runway and the Broader Video AI Graveyard

Runway, the other major player in AI video generation, has fared slightly better than Sora — but only because it was spending slightly less while also generating slightly less revenue. Runway's Gen-3 model, while technically impressive, faces the same fundamental cost structure problem. The company has burned through the majority of its $680 million in raised capital and was reportedly seeking additional funding at a significantly lower valuation in early 2026.

The broader video AI startup graveyard is growing. Pika, which raised $135 million at a $530 million valuation in 2024, has pivoted to focus on short-form social video effects — a less compute-intensive use case that might actually work economically. Stability AI's video efforts have been essentially abandoned as the company continues its broader restructuring. HeyGen, which focused on AI avatar videos for marketing, has survived by keeping generation times short and pricing aggressively for enterprise customers, but its margins remain thin.

The aggregate numbers are damning. The AI video generation space has attracted over $4 billion in venture capital since 2023 but generated estimated total revenue of only $340 million in 2025. That is a ratio that would make even the most patient venture capitalists uncomfortable. For context, the AI coding assistant space attracted roughly $2 billion in funding and generated over $3 billion in revenue over the same period.

3D Generation: The Most Expensive Modality

If video AI economics are bad, 3D generation economics are worse. Generating a usable 3D model — one that a game studio or product designer could actually work with — requires not just generating geometry but also textures, UV mapping, rigging, and often multiple iterations with human feedback. The cost per usable output can exceed $20-50, and the quality is still far below what a skilled 3D artist produces.

Companies like Meshy, Luma AI, and Kaedim have found small niches in rapid prototyping and concept visualization, but none has found a path to the kind of scale that venture investors expected. The total addressable market for AI 3D generation is smaller than video, the costs are higher, and the quality bar set by existing professional tools is punishingly high.

The Middle Ground: Image Generation's Precarious Balance

Image generation occupies a fascinating middle position in the inference cost spectrum. It is expensive enough that careless economics will kill you, but cheap enough that a well-run operation can build a real business.

Midjourney: The Quiet Profit Machine

Midjourney is arguably the most impressive business story in generative AI. Founded by David Holz with a small team, the company reached profitability without taking a single dollar of venture capital funding. It has an estimated 16 million paid subscribers and generates over $300 million in annual revenue. And it does this while offering genuinely excellent image generation at price points that customers are willing to pay.

The secret to Midjourney's success is disciplined cost management combined with smart product decisions. Images are generated at optimized resolutions and then upscaled using less expensive methods. Generation is batched efficiently. And the pricing tiers — $10, $30, $60, and $120 per month — are calibrated to ensure that even the heaviest users remain profitable.

Midjourney also benefits from a user base that skews toward creative professionals and enthusiasts who value quality and are willing to pay for it. The average Midjourney user generates 200-400 images per month. At a blended cost of three to six cents per image, the monthly compute cost per user is $6-24 — well within the margin envelope for all but the cheapest subscription tier.

DALL-E and Others: The Subsidy Model

Not everyone in image generation is as disciplined as Midjourney. OpenAI's DALL-E has been offered as part of the ChatGPT Plus bundle, effectively subsidizing image generation with the profits from text. This is a viable strategy if you are OpenAI and have ChatGPT's revenue to lean on, but it has distorted market expectations about what image generation should cost.

Adobe Firefly has taken a similar approach, bundling AI image generation into Creative Cloud subscriptions. This makes economic sense for Adobe — the incremental cost of image generation is modest relative to the existing subscription price, and it helps retain subscribers in an increasingly competitive market. But it means Adobe is not building a standalone image generation business; it is using image generation as a feature to support its existing business.

The companies struggling in image generation are the ones that tried to compete on price without Midjourney's cost discipline or OpenAI's cross-subsidy capability. Stability AI's consumer-facing image products have been plagued by thin margins and high churn. Leonardo AI, while popular, has struggled to push users beyond its free tier.

The lesson from image generation is that the modality is viable but unforgiving. You need to be obsessive about cost optimization, smart about pricing, and disciplined about feature scope. The margin for error is thin. This is a meaningful contrast with text-based AI, where the margins are so generous that even moderately inefficient operations can be profitable.

Why Agentic AI Might Be the Sweet Spot

Here is where the analysis gets interesting. If text AI is cheap and high-value, and video AI is expensive and (for most use cases) low-value, where does agentic AI fit? The answer, increasingly, is that agentic AI might represent the optimal point on the entire spectrum — text-level inference costs delivering enterprise-level value.

Agentic AI systems — AI agents that can reason, plan, use tools, and take actions across multiple steps — are fundamentally text-based at their core. An agent that books your travel, manages your calendar, processes your invoices, or deploys your code is running text inference at each step. It reads, it thinks, it acts. The inference cost per step is the same as any other text generation: pennies.

But the value delivered is not "I got a well-written paragraph." The value is "I completed a complex business process that would have taken a human worker thirty minutes." The gap between cost of delivery and value received is enormous, and that gap is what makes a great business.

I wrote about this dynamic in our agentic AI predictions, and the data is bearing it out. Companies building agentic AI products are reporting some of the healthiest unit economics in the industry. An agent that processes insurance claims, for example, might execute twenty inference calls per claim at a total cost of less than a dollar. If it replaces even fifteen minutes of a claims adjuster's time, the value is $10-15 per claim. That is a 10-15x return on compute cost, which supports aggressive pricing while maintaining strong margins.

The production AI stack article we published details the technical architecture required to make this work reliably, and the key insight is that production agentic systems need careful orchestration — not just to ensure reliability, but to manage costs. Smart routing (using smaller models for simple steps, larger models only for complex reasoning), caching (storing results of common sub-tasks), and batching (processing multiple items in parallel) can reduce agent operating costs by 60-80 percent compared to naive implementations.

The Agent Zero article explored how the industry's pivot to autonomous agents was rewriting the rules of software development. What is becoming clear now is that it is also rewriting the rules of AI economics. Agents are the first AI product category where the economics work at every level — for startups, for enterprises, and for the infrastructure providers powering them.

The Willingness-to-Pay Advantage

There is another critical economic factor that favors agentic AI: willingness to pay. When an AI product saves someone time on a creative task — writing an email, generating an image — the perceived value is modest. People have been writing emails and creating images for decades, and they have an intuitive sense of what those tasks are "worth."

But when an AI agent completes a business process — processes an expense report, reconciles an account, triages a support ticket — the value is measured in employee hours saved, and enterprises will pay accordingly. The same enterprise that balks at paying $100/month for an AI writing assistant will happily pay $500/month for an AI agent that processes invoices, because the ROI is obvious and measurable.

This willingness-to-pay asymmetry is why agentic AI companies are able to charge premium prices while maintaining high margins. The customer sees a clear ROI, the provider's costs are modest, and everyone is happy. Compare this to video generation, where the customer sees a cool but non-essential creative tool and the provider's costs are through the roof.

The MCP Effect: Infrastructure for Economically Viable Agents

One of the most significant developments enabling the agentic AI opportunity is the Model Context Protocol (MCP), which has reached 97 million installs across the ecosystem. MCP provides a standardized way for AI agents to connect to external tools, data sources, and APIs, and its explosive adoption has fundamentally changed the economics of building and deploying agents.

Before MCP, building an AI agent that could interact with external systems required custom integration work for every tool and every data source. This was expensive, time-consuming, and fragile. A company building an agent to automate customer support workflows might need to build custom integrations with Zendesk, Salesforce, Slack, and their internal knowledge base. Each integration was a mini-engineering project, and maintaining them was a constant drain on resources.

MCP changed this by providing a universal protocol that tool providers and AI platforms could both implement. Now, an agent built on any MCP-compatible platform can instantly access thousands of tools and data sources without custom integration work. The marginal cost of adding a new capability to an agent dropped from weeks of engineering time to essentially zero.

This has profound economic implications. When the cost of building and deploying an agent drops, the range of use cases that are economically viable expands dramatically. Tasks that were not worth automating at $50,000 in integration costs become obviously worthwhile when the integration cost is near zero and the ongoing inference cost is pennies per execution.

The MCP ecosystem has also created network effects that further improve agent economics. As more tools become MCP-compatible, agents become more capable. As agents become more capable, more users adopt them. As more users adopt them, more tool providers implement MCP. This flywheel is driving rapid expansion of the agentic AI market while keeping costs manageable.

As we discussed in our piece on how agentic AI became the infrastructure layer nobody planned for, this infrastructure build-out is transforming agents from impressive demos into production business tools. And production business tools with good unit economics are exactly what the AI industry needs more of.

What This Means for Builders: Where to Invest, Where to Avoid

If you are building an AI product, planning an AI startup, or deciding where to invest your company's AI budget, the inference cost spectrum should be your north star. Here is a framework for thinking about it.

The Green Zone: Build Here

Text-based business automation. Agents that automate repetitive business processes using text inference and tool calls. The unit economics are excellent, the willingness to pay is high, and the market is massive. Insurance processing, legal document review, financial reconciliation, HR onboarding — any process that involves reading, reasoning, and acting on text-based information is a candidate.

Developer tools. Coding assistants, testing automation, deployment agents, code review bots. Developers are willing to pay, the value is clear and measurable, and the retention is exceptional. The market is not as large as general business automation, but the margins are even better because developers tolerate fewer support calls and generate fewer edge cases.

Enterprise search and knowledge management. RAG-based systems that make internal knowledge accessible via natural language. Low inference costs, high perceived value, strong retention. The key challenge is data integration, but MCP is making this dramatically easier.

The Yellow Zone: Proceed with Caution

Image generation. Viable but requires disciplined cost management and smart pricing. Do not try to compete with Midjourney on quality or with DALL-E on price. Find a niche — architectural visualization, product photography, game asset creation — and serve it well at a price that covers your costs with margin to spare.

Audio and music AI. Moderate inference costs with a market that is willing to pay, but smaller than many realize. The professional music market is conservative and values human creativity. The voiceover and podcast market is more promising but competitive. Proceed if you have a clear niche and efficient infrastructure.

Creative writing and content tools. Text inference is cheap, but the willingness to pay for AI writing assistance is lower than many founders expect. Consumers see writing as a commodity. Enterprises will pay for specialized content tools (marketing copy, technical documentation), but generic writing assistants face intense competition from ChatGPT and Claude, which are free or cheap and "good enough."

The Red Zone: Avoid Unless You Have Deep Pockets

Consumer video generation. The Sora story should be a flashing red warning sign. Unless you have a breakthrough in inference efficiency that reduces costs by an order of magnitude, do not build a consumer video generation product. The costs are too high, the willingness to pay is too low, and the competition from existing video tools (CapCut, Premiere, DaVinci Resolve) is too established.

General-purpose 3D generation. Same story as video but worse. The quality bar is higher, the costs are greater, and the professional tools (Blender, Maya, ZBrush) are deeply entrenched. Narrow niches — furniture visualization, dental implant modeling — might work, but general-purpose 3D generation is a money pit.

Multimodal "do everything" platforms. The temptation to build an AI platform that does text, images, video, audio, and 3D is strong, but the economics are punishing. You end up subsidizing the expensive modalities with the cheap ones, which degrades your overall margin profile and makes pricing extremely difficult. Focus on one modality and do it well.

The Coming Inference Cost Collapse: Will Cheaper Compute Save Video AI?

The obvious counterargument to everything above is: inference costs are falling. They have fallen 80 percent in three years. Will they fall another 80 percent in the next three? And if they do, does that save video and 3D generation?

The answer is: maybe, but probably not fast enough, and not uniformly.

There are three major forces driving inference cost reductions: hardware improvements, model architecture innovations, and deployment optimization.

Hardware: The NVIDIA Dependency

The AI industry remains heavily dependent on NVIDIA GPUs, and NVIDIA's pricing reflects this. The H100 GPUs that power most inference workloads cost $25,000-30,000 each, and the newer B200 GPUs are even more expensive. While AMD and Intel are making competitive inroads, and custom chips from Google (TPUs), Amazon (Trainium/Inferentia), and various startups are emerging, NVIDIA maintains a dominant position that limits how quickly hardware costs can fall.

The most optimistic projections suggest inference costs could fall by 80 percent again by 2029, driven by custom silicon, improved chip architectures, and increased competition in the GPU market. The conservative projections suggest a more modest 50-60 percent decline. In either scenario, video generation remains expensive — just less outrageously expensive.

Architecture Innovations: The Distillation Revolution

More promising than hardware improvements are architectural innovations that reduce the amount of computation required per generation. Distillation — training smaller, more efficient models to replicate the behavior of larger ones — has been transformative for text models and is beginning to show results for image and video models.

Recent work on consistency models, latent video diffusion, and single-step generation has shown that it is possible to dramatically reduce the number of compute steps required for video generation. Some research models can now generate reasonable quality video in 4-8 denoising steps instead of the 50-100 steps that were standard in 2024. If this research translates to production systems, it could reduce video inference costs by 80-90 percent.

But "research translates to production" is doing a lot of heavy lifting in that sentence. Research models often sacrifice quality, controllability, and consistency to achieve efficiency gains. The Sora-quality bar — which, it should be noted, was already below what professionals need — may not be achievable with aggressive optimization. And the professional quality bar remains far beyond current capabilities at any cost.

Deployment Optimization: The Low-Hanging Fruit is Mostly Picked

The third category of cost reduction — better deployment practices — has the least remaining upside. Techniques like model quantization, batching, speculative decoding, and multi-tenant serving have already been widely adopted. The major cloud providers and model serving frameworks have spent three years optimizing these systems, and while incremental improvements continue, the era of easy 2-3x efficiency gains from better deployment is largely over.

The Verdict on Future Cost Reductions

My assessment is that inference costs will continue to fall, but the decline will not be fast enough or steep enough to make consumer video AI viable before 2029 at the earliest. Text and code AI will continue to benefit disproportionately from cost reductions, widening the margin advantage they already enjoy. Image AI will become more comfortably profitable. And video and 3D will remain challenging.

The companies that survive in the expensive modalities will be those that find specific, high-value enterprise niches where the willingness to pay matches the cost of delivery. A pharmaceutical company that needs AI-generated molecular visualization at $50 per model might be a viable customer. A teenager who wants to make AI TikToks for free is not.

The Market Is Sorting Signal from Noise

The shutdown of Sora is not an isolated event. It is a data point on a trendline that has been forming for two years. The AI industry is going through the same economic sorting that every technology wave experiences: the period where the market figures out what actually works as a business, as opposed to what merely works as a technology.

This sorting process is brutal but necessary. The dot-com boom taught us that "technically possible" and "commercially viable" are different questions with different answers. The mobile app boom taught us that downloads do not equal revenue. And the AI boom is teaching us that generating impressive outputs does not mean you can generate impressive profits.

The winners are already clear. Text and code AI products — ChatGPT, Claude, Copilot, and the growing ecosystem of agentic AI tools — have proven that AI can be an extraordinarily profitable business when the inference economics align with customer value and willingness to pay. These products will continue to grow, consolidate, and expand into new use cases.

The losers are also becoming clear, though the full reckoning has not yet arrived. Several major AI video startups are running on fumes, and the next twelve months will likely see more shutdowns, fire-sale acquisitions, and pivots. The 3D generation space will consolidate to a handful of niche players. And the multi-modal platforms that tried to do everything will be forced to focus on the modalities that actually make money.

The middle ground — image generation, audio AI, and specialized vertical applications — will see a bifurcation between well-run companies with disciplined economics (Midjourney, ElevenLabs) and poorly-run companies that burn through their funding and disappear. The technology works; the question is whether the business works, and that depends entirely on execution.

What Happens to All That VC Money?

The venture capital implications of this sorting process are significant. Over $15 billion in venture capital has been deployed into generative AI companies since 2023. A substantial portion of that capital — perhaps 30-40 percent — was invested in companies whose core modality economics do not work. These investments will generate returns ranging from disappointing to catastrophic.

The smart money is already rotating. VC investment in text-based AI applications — particularly agentic AI and vertical-specific tools — increased by 140 percent in the second half of 2025 compared to the first half. Investment in video AI declined by 60 percent over the same period. The market is repricing, and it is repricing fast.

Lessons for the Industry

The Sora shutdown and the broader economic sorting offer several clear lessons for everyone in the AI industry:

For founders: Start with the unit economics. Before you write a single line of code, calculate the inference cost per user request, the expected usage patterns, and the willingness to pay. If the math does not work at scale, no amount of technical brilliance will save you. The AI industry has seen too many companies build incredible technology that cannot sustain itself financially.

For investors: Demand inference cost breakdowns in every pitch deck. Any AI company that cannot articulate its per-unit economics with precision is either hiding bad numbers or has not done the work. The era of investing in AI companies based on demo quality alone is over. Quality of output is table stakes — quality of unit economics is what separates the survivors from the casualties.

For enterprise buyers: Understand what you are actually paying for and whether the economics support your vendor's long-term viability. An AI product that is priced below its inference cost is being subsidized by venture capital, and that subsidy will not last forever. Build your workflows on products with sustainable economics. If your video AI vendor is burning a million dollars a day and has fewer than half a million users, you should have a migration plan ready.

For developers: Build for the modalities with favorable economics. If you are choosing between building a video generation tool and an agentic business automation tool, the agentic tool has a dramatically better economic profile. This does not mean video AI will never work — it means it does not work right now at current costs and price points, and betting your career on a cost curve that may or may not materialize in time is a risky proposition.

The AI industry in 2026 is not in crisis. It is in a correction. The technology continues to improve at a breathtaking pace. The products that align technical capability with economic reality are generating enormous value and building durable businesses. But the products that confused "technically possible" with "commercially viable" are being forced to reckon with a harsh truth that every industry eventually learns: you can not defy gravity forever. When the math does not work, it does not matter how impressive the demo is. Eventually, the bill comes due.

The death of Sora is not the end of AI video generation. It is the end of the fantasy that AI video generation could be a consumer product at current costs. The technology will survive in enterprise niches, in production pipelines where the value justifies the cost, and eventually — perhaps by the end of the decade — in consumer products powered by dramatically cheaper inference. But the path from here to there runs through a valley of shutdowns, write-downs, and painful lessons about the economics that every technology eventually has to learn.

The market is sorting signal from noise. The signal is clear: text-based AI, coding AI, and agentic AI are the economic engines of the AI industry. Everything else is still looking for a business model. And in the meantime, the math does not care about your demo reel.