$112B Hole — Big Tech's Hidden Bankruptcy

Project Stargate — hundreds of billions of dollars to build giant data centers. This is not a typo: we are talking about amounts comparable to the GDP of small countries. Microsoft, Google, and Meta annually spend tens of billions on GPU infrastructure: Microsoft alone invested over $50 billion in capital expenditures in 2025, most of it for AI.

The problem is hidden in accounting. H100 generation GPUs become obsolete in 2 years with the release of H200, B100, B200 — each subsequent generation is 50-100% faster than the previous one. But corporations record depreciation over 5-6 years, creating an accounting illusion. Example: a company bought GPUs for $20 billion. In the accounting books, after 2 years, they still “cost” $13 billion (with linear depreciation over 6 years). In reality — they cost ~$5 billion, because the new generation does the same work twice as fast and cheaper.

This creates a hidden deficit: the difference between the book value of assets and their real market value — trillions of dollars across the industry. When (not “if,” but “when”) auditors demand a revaluation — this could lead to massive write-offs, collapse AI company stocks, and provoke a crisis of confidence in the entire industry.

Analysts forecast that OpenAI will accumulate about $112 billion in losses by 2030. This figure is not pulled out of thin air: it reflects a fundamental problem with the centralized AI business model.

On one hand, revenues are growing impressively: billions of dollars annually from ChatGPT and API subscriptions. On the other, expenses are rising even faster. Every new model generation requires exponentially more resources:

• GPT-3 → GPT-4: training costs increased approximately 10-fold
• GPT-4 → GPT-5: another exponential increase — a classic exponential curve
• Inference: millions of users = billions of tokens per day = billions of dollars per year in GPU power

This model only works with an endless influx of venture capital. OpenAI has raised tens of billions in investments, including rounds from Microsoft and SoftBank. But investors are not philanthropists. Sooner or later, they will demand profit. The question is not “if” but “when” — and what will happen to the millions of businesses built on the OpenAI API at that moment?

For comparison: Gonka has raised $80M and is already processing real AI requests through a network of ~4,648 GPUs. Inference cost is $0.0047/1M tokens. This is possible because the decentralized model has no need to recoup trillion-dollar investments in data centers.

Gonka does not build data centers — it aggregates existing GPUs around the world. This is not just an alternative business model — it is a fundamentally different economic architecture that eliminates the root cause of the bubble.

No capital expenditures: the Gonka network does not raise hundreds of billions for construction. The protocol, blockchain, and software are all the team creates. GPUs are provided by independent hosts around the world — each at their own expense.

No 6-year depreciation: when an H100 becomes obsolete, the host simply replaces it with an H200 or the next generation. The decision is made by the equipment owner based on market conditions, not by a corporate CFO trying to hide write-offs.

No accounting tricks: all transactions on the Gonka blockchain are transparent. Rewards are distributed according to a protocol audited by CertiK. There are no «hidden» costs that would surface 5 years later during asset revaluation.

Distributed risk: every host bears its own risk. If one host fails due to a bad GPU investment, it is their problem, not the network's. In a centralized model, one $10B mistake can collapse an entire company. In Gonka, such a mistake is impossible by definition — because there is no single participant capable of making a $10B decision.

The result: the cost of inference via Gonka is $0.0047 per million tokens. This is ~400 times cheaper than OpenAI. And this price is stable — because it is not backed by a trillion-dollar infrastructure that needs to be amortized.

Let's compare the two AI infrastructure models:

Gonka operates about 4,648 GPUs across ~113 participants (~582 MLNodes). The project raised $80M — thousands of times less than what Stargate spends alone. But the network does the same thing: processes AI requests via the Kimi K2.6 neural network, available through an OpenAI-compatible API.

Analogy: imagine that in the 2000s, someone suggested: “Instead of building giant servers for the internet, let every homeowner install a mini-server and get rewarded for participating.” It sounds utopian — but that is exactly how Airbnb works for housing, Uber for transportation, and that is how Gonka works for AI computing. Decentralization is not a utopia — it is the next stage of infrastructure evolution.