Look at the fascinating and simultaneously alarming picture unfolding right before our eyes. On one hand, news feeds are literally bursting with victory dispatches: Nvidia is breaking every conceivable market capitalization record, tech giant CEOs with glowing eyes are declaring the dawn of a new technological era, and AI data centers are being built at a pace as if electricity and silicon had become free. On the other hand, if you open the financial statements of companies selling real applied AI software to end businesses, a giant gaping hole between compute spending and actual revenue jumps right out at you.
In this endless stream of euphoria, one simple truth always puts me on guard: the fundamental laws of economics cannot be repealed by a pretty chatbot interface. When infrastructure spending grows exponentially while user revenue grows linearly, the air inevitably starts to smell like a classic market bubble. Let's take this colossal mechanism apart together, bolt by bolt, and see why the familiar twenty-dollar-a-month model for accessing neural networks is hitting a dead end and what will actually happen when venture subsidies dry up.
The Infrastructure Boom and the Scale of Real Costs
Let's start with dry numbers that would make any CFO gasp. The combined capital expenditures (CapEx) of the "Big Four" hyperscalers—Microsoft, Alphabet, Meta, and Amazon—have crossed the $200 billion per year mark. To give you a sense of the scale of that figure: it is larger than the annual budget of most developed nations in the world. And the lion's share of that money goes not toward abstract scientific research, but toward buying physical silicon and bringing in gigawatts of power.
A single modern cluster for training and inference of next-generation foundation models requires anywhere from 50,000 to 100,000 GPUs of the Nvidia H100 class or the latest Blackwell B200. Yet the chip itself is only the tip of the iceberg. To make this armada of silicon operate as a single giant supercomputer, you need ultra-high-speed InfiniBand networking fabrics or specialized RoCE v2 switches, optical transceivers, direct-to-chip liquid cooling, and dedicated transformer substations. In engineering practice, a strict empirical rule applies: every dollar spent on buying Nvidia accelerators requires at least one more dollar for supporting data center infrastructure—from concrete and cable trays to heat rejection systems.
Now add the energy crisis to the mix. A data center with a capacity of 500 megawatts or a full gigawatt consumes electricity on the scale of a major steel mill or a city of half a million people. Hyperscalers are already buying up capacity from decommissioned nuclear power plants, signing direct Power Purchase Agreements (PPAs), and investing in small modular reactors (SMRs). The question is who ultimately has to pay for this extravagant banquet. Sequoia Capital partner David Cahn formulated the famous $600 billion question: if you add up the projected revenue of semiconductor fabs, multiply it by two to account for supporting infrastructure costs, and factor in a normal 50% gross margin for software companies, the AI software industry needs to generate $600 billion in annual revenue just to pay back the installed servers. Meanwhile, the actual combined revenue of all applied generative services today barely scrapes together $50–$60 billion. That's a tenfold gap.
The Economic Dead End of the Twenty-Dollar Subscription
You're probably used to a subscription for top-tier neural networks costing a flat twenty dollars a month. Pay a modest sum—and use the latest models from OpenAI, Anthropic, or Google without limits. It's incredibly convenient for you and me, but from the perspective of service unit economics, it is pure financial absurdity sustained solely by investor subsidies.
Conceptual cross-section of reasoning model thinking process: a massive hidden bulk of dark compute blocks and a tiny glowing cube on the surface
Look at what's happening under the hood here, especially with the arrival of reasoning models like Claude Opus 5.5, GPT-o1, or Gemini Ultra. In classic conversational models, the relationship was simple: you sent a short prompt, the model generated a couple hundred response tokens, and the server spent a fixed fraction of a second computing. Reasoning architectures, however, work completely differently. Before outputting even a single visible word to you, the model runs a hidden Chain of Thought, generating thousands of internal tokens. For every visible token of a quality answer today, there can be anywhere from 5 to 20 hidden tokens of deliberation, hypothesis double-checking, and pruning dead-end branches.
If an ordinary user opens the chat a couple of times a day to rewrite an email to a colleague or generate a birthday greeting, they cost the platform mere cents. But an active developer, researcher, or content creator ("power user") who runs terminal agents, triggers reviews of entire repositories, or feeds heavy documents into the context generates a real server inference cost for the provider of around $80–$150 per month. A paradox emerges: the more useful the tool becomes for a professional and the more tasks it solves, the larger the direct losses it inflicts on the platform under a flat-rate subscription.
Providers are already desperately trying to plug these financial leaks: they introduce floating hidden message limits, trim the length of reasoning chains during peak hours, and gently nudge users toward pay-per-token plans. The unlimited twenty-dollar all-you-can-eat buffet is living out its final months—inevitably, it will be replaced by strict metering of every kilowatt-hour of compute used.
The Scaling Plateau: Why Chips No Longer Work Miracles
Remember the intoxicating feeling of the first years of the generative boom? It seemed you just had to feed a neural network ten times more text from the internet, increase the model's parameter count, and we would automatically get artificial superintelligence. This empirical Scaling Laws principle, formulated by researchers, worked flawlessly for several years. But by the end of 2026, the industry slammed at full speed into a solid physical wall.
First, humanity has simply run out of quality data. All publicly available digitized books in the world, arXiv paper archives, years of discussions on specialized forums, the code of every open repository, and terabytes of subtitles have already been vacuumed up and baked into the weights of existing base models. There is nowhere left to pull information from. Attempting to feed models synthetic text generated by themselves in previous iterations leads to the well-known phenomenon of "model collapse": the algorithm starts stewing in its own juices, accumulating statistical errors, losing rare linguistic constructs, and churning out bizarre hallucinations.
Second, a period of sharply diminishing marginal returns has arrived. Whereas scaling up a compute cluster tenfold used to yield a qualitative breakthrough in context comprehension and logic, today a tenfold increase in spending yields a modest 1.5–2% gain on standard benchmarks like MMLU-Pro or SWE-bench. To take a tiny step forward, you have to burn millions of dollars on electricity. Extensive pre-training scaling has exhausted itself, and that's what scares investors most: the golden age of easy progress through brute-force silicon has officially ended.
The Threat of Open-Weight Models and Commoditization
While the creators of closed cloud monopolies were burning billions of dollars, a quiet but devastating revolution took place on the open market. If two years ago open-weight models lagged light-years behind the flagships from OpenAI and Anthropic, today the gap has narrowed to a matter of weeks—and in some scenarios, open code has even pulled ahead.
The emergence of architectures like DeepSeek V3 and DeepSeek R1, the Llama 3.3 family, and recent releases of Qwen and Mistral proved that fundamental reasoning capabilities can be packed into elegant, mathematically sound structures without astronomical budgets. The introduction of Multi-Head Latent Attention (MLA), adaptive sparse attention, and high-quality distillation techniques made it possible to run powerhouse models on relatively modest servers.
For closed providers, this became a real catastrophe. When a developer can deploy an open model of comparable caliber on their own rented server or via an independent endpoint, closed labs lose the ability to dictate pricing terms. Over the past two years, the market price of one million tokens has plummeted by more than 80%. Intelligence is turning into a widely available commodity—a basic utility just like tap water or electricity from a wall outlet. And you cannot sustain 80% profit margins on a commodity unless you hold an exclusive patent or a monopoly on raw materials.
Corporate Inertia and Adoption Barriers
You might ask: "But big business has unlimited budgets! Won't corporations buy up all this capacity to automate their departments?" It's a logical question, and venture analysts pinned their main hopes on the corporate sector. Yet reality collided with ironclad corporate inertia.
Executives at major banks, insurance giants, and retailers gladly pose for photos at conferences and launch pilot projects in internal marketing departments. But when it comes to integrating AI agents into core business processes—processing payments, managing warehouse supply chains, calculating taxes, or assessing credit risk—legal departments and security teams dig in their heels to the death.
There are several reasons for this, and all of them are completely pragmatic:
- Non-determinism and hallucinations: traditional enterprise software may be slow and clunky, but it guarantees that given the same input, the output will always be identical. A neural network, by contrast, is probabilistic by nature and capable of producing an absurd decision one time in a thousand—at a cost measured in millions of rubles in damages.
- Regulatory pressure and compliance: the requirements of the European EU AI Act, personal data protection standards, and local laws demand full transparency in decision-making logic. Try explaining to a government auditor which hidden weights led a transformer to deny a customer's insurance payout.
- Lack of personal accountability: if a staff accountant makes a mistake, the internal investigation procedure is clear. If an autonomous agent makes a fatal error in a financial ledger, it's unclear whom to hold liable for the losses—the model developer, the system integrator, or the system administrator.
As a result, billions of corporate dollars that were supposed to flood into the software industry get stuck in months-long security approvals and endless pilots with no commercial follow-through.
The Scenario of an Inevitable Market Correction
So what will happen when the market finally realizes that trillion-dollar expectations are detached from reality? Let's draw a parallel that many seasoned engineers and financiers are recalling today. What we are witnessing now bears a painful resemblance to the dot-com boom and the telecom crisis of 1999–2001.
In the late nineties, telecom operators invested hundreds of billions of dollars in laying fiber-optic cables across ocean floors and entire continents. They were certain that the explosive growth of the internet would instantly fill those pipes with traffic. When it turned out that actual demand was years behind, most operators went bankrupt and tech stocks crashed. Yet the fiber didn't vanish. The physical infrastructure stayed in the ground, dropped in price by orders of magnitude, and a few years later became the foundation for video hosting platforms, streaming services, cloud platforms, and the modern digital world.
Exactly the same thing awaits the AI industry. When the venture hype subsides, we will see a wave of bankruptcies among overvalued wrapper startups that simply resold access to someone else's API at a markup. But the giant data centers, power substations, and millions of installed accelerators will remain. The cost of compute will inevitably drop to the marginal cost of generating electricity, cleansing the industry of meaningless hype. In practice, the Jevons paradox in AI economics works brilliantly here: as soon as access to compute becomes cheap and utilitarian, total demand for it skyrockets, opening the door to solutions that previously seemed economically senseless.
Survival Strategy: From Hype to Unit Economics
For you and me as practitioners, developers, and researchers, this valuation crisis is no reason to panic—it's a clear compass. Those who survive won't be the ones shouting loudest about a revolution, but the ones who know how to count money and build reliable engineering pipelines.
If you are designing the architecture of your projects today, build three foundational rules into it:
- The Model Routing architectural pattern: drop the habit of sending every user request to a flagship model at maximum price. Build an intelligent router: simple classification, entity extraction, and text formatting tasks should go to compact specialized models or local inference, and only in the 5% of genuinely complex scenarios should a request fly off to an expensive reasoning network.
- Uncompromising context optimization: using Prompt Caching, semantic prompt compression, and local vector search lets you cut the volume of transmitted tokens by 70–80%. Every token not sent is money saved in your pocket.
- Reliance on open weights and independent infrastructure: services completely tied to a single corporation's proprietary cloud API risk dying the very day the provider decides to raise prices or change its terms of service. Invest in a local stack—frameworks like vLLM, SGLang, Ollama, and open models guarantee your business stability through any market storm.


