Are You Sure We Are Not in a Bubble?

A basic principle in business is that your revenue must cover your costs, and ideally leave some margin for profit.

One of the keys of understanding Venture Capital-funded startups is the time-frame. Startups raise money at different stages in exchange for equity, and use that money to build their offering and grow their user base to a point where they can achieve “escape velocity” and start to actually gain profits.

Once funding is raised, the goal for investors is that the startup either gets bought or goes public, so they get a nice multiple of their investment. If that happens–VC is a game of probability– the startup founders, employees, and other stakeholders who own shares also benefit.

On thing to notice is that the startup may get bought or go public without actually being profitable. For example, for many years Amazon didn’t turn a profit because it reinvested every penny to build their logistics infrastructure.

Of course, sometimes things are not that simple.

OpenAI recently announced their latest funding round: $122 billion of commited capital at a post-money valuation of $852 billion. Like OpenAI’s prior funding rounds, only around $50 billion of the $122 billion are money from VCs, retail investors, or SoftBank. Amazon, NVIDIA, and Microsoft participation is not cash-in-hand. It is conditioned to OpenAI spending on those partners’ products or services.

For example, let’s take a look at NVIDIA’s $30 billion dollar commitment:

The $30 billion are part of a prior $100 billion non-binding announcement by NVIDIA and is what would be actually commited for the moment. In exchange, OpenAI is expected to build 5GW of dedicated AI capacity based on NVIDIA’s Vera Rubin systems.

How does NVIDIA benefit from this?

According to CNBC¹, NVIDIA’s CEO Jensen Huang told investors on an earnings call in August 2025 that building one gigawatt of datacenter capacity costs between $50-60 billion, of which about $35 billion NVIDIA’s chips and systems. This means that for NVIDIA’s $30 billion investment, OpenAI is expected to spend around 5 gigawatts x 35 billion, or around $175 billion in NVIDIA hardware. NVIDIA has an estimated 70% gross margin on its datacenter products. So NVIDIA gets $122 billion from a $30 billion investment, a 4x multiple.

Besides that, NVIDIA also gets $30 billion of OpenAI’s equity in the form of non-voting shares, deployed progressively as gigawatts come online.

From OpenAI’s perspective, deals like these are for now one of the few ways it has to build the massive infrastructure OpenAI needs and is an important motivation for the IPO. Private companies with massive losses can’t access the bond market at reasonable rates. There are no audited public financials, no credit rating, no market cap to anchor against. Going public gives OpenAI access to convertible bonds and other financial instruments at a significant lower cost.

But NVIDIA is not the only company OpenAI has commitments with. According to investor Brad Gerstner, OpenAI spend commitments in November 2025 where $1.4 trillion².

Profitability

OpenAI cannot escape the fact that it needs to be profitable at some point in the future. John Gruber writes that OpenAI’s own estimate is that the company will lose $111 billion between 2024 and 2029. By comparison–also from Gruber’s article–the accumulated net income for Berkshire Hathaway, Walmart, Samsung, and Eli Lilly for 2025 was $139 billion dollars. OpenAI losses will exceed those companies combined profits.

However, according to its own projections, OpenAI expects to be profitable on 2030.

OpenAI’s CFO said that the company’s annualized revenue crossed $20 billion in 2025³. The figure comes not from the year’s revenue, but from multiplying December 2025 revenue by twelve. The actual booked 2025 revenue is estimated to be circa $13 billion⁴.

Valuation multiples should be applied to booked trailing 12-month revenue, not run rate. At the current valuation of $825 billion, that means a 66x multiple for OpenAI. Such a multiple is not only incredibly high by any current or historical standard. It’s an indicator that OpenAI’s current business model has no future as is.

How about marginal contribution? OpenAI gross margin is 33%⁴. Also, it has a revenue share agreement where Microsoft gets 20% of OpenAI’s revenue.

For every $1 dollar of revenue, OpenAI gets:

Available for everything else is the contribution margin. Those $0.13 per dollar of revenue must cover R&D, sales and marketing, general and administrative expenses, and leave something for profit.

Software companies usually bear high development costs up front, but once the product sells, they enjoy near zero marginal costs. That’s why Microsoft enjoys a gross margin close to 70%, and many mature SaaS companies run at 70%-85%. The get a gross margin of ~$0.70 on the dollar.

AI companies selling inference, however, consume tokens each time the customers use their products. While cost per-FLOP fall 30%-40% per year, new frontier AI models get bigger. Reasoning consumes more tokens. Agents make more calls, so per-task token consumption is raising sharply. So, essentially, AI companies gross margin remains the same, even if the face of compute getting cheaper.

AI Companies, specially OpenAI and Anthropic, are trying to win the AI race by growing usage and building better and faster models. The cost of running those models doesn’t dilute with more user. They are part of the marginal cost. And the scale needed to cover operating expenses is simply unreal.

Future scenarios may include the cost of using AI going up, monetization through advertising, devising new techniques that reduces cost inference. What is certain, however, is that the current pace will not be sustainable in the future years.