Two financial disclosures landed within 48 hours of each other last week and reshaped the picture of how the world’s most important technology gets funded. On June 15, Ed Zitron’s newsletter published audited OpenAI financial documents, independently verified by the Financial Times, showing the company spent $34 billion and lost $38.5 billion in 2025. The next day, The Information reported that China’s DeepSeek closed its first funding round at more than $7.4 billion, using a deal structure designed to keep founder Liang Wenfeng in complete control. Reuters confirmed the report, adding that Tencent is considering 10 billion yuan ($1.4 billion) and battery maker CATL is looking at 5 billion yuan ($700 million).

The contrast reveals two fundamentally different theories of how to build and control the companies that produce frontier AI.

The DeepSeek Structure: Capital Without Dilution

DeepSeek’s deal is unusual by any standard, and unprecedented in AI. According to Reuters, investors did not put their money directly into DeepSeek. Instead, they committed capital to a limited partnership managed by CEO Liang Wenfeng. Investors are subject to a five-year lock-up period and receive no voting rights. The only exception is China’s National Artificial Intelligence Industry Investment Fund, which invested directly in DeepSeek and retained both voting rights and freedom from the lock-up.

Reuters also reported earlier this month that Liang committed 20 billion yuan (~$2.8 billion) of his own money into this round. If accurate, that means the founder personally contributed roughly 40% of the total raise.

The valuation exceeds $50 billion. For context, DeepSeek has never taken outside capital before. This is its first funding round, and it is already the largest AI funding round ever closed.

The structure accomplishes several things simultaneously. It provides DeepSeek with massive capital for compute, infrastructure, and research. It preserves the founder’s complete strategic authority over the company. And it insulates technical decisions from investor governance pressure, the kind of pressure that contributed to OpenAI’s board crisis in November 2023 and its subsequent restructuring.

OpenAI’s Financials: Growth and Losses at Scale

The picture at OpenAI is different in every dimension. According to audited financial documents obtained by Ed Zitron and verified by the Financial Times, OpenAI generated $13.07 billion in revenue in 2025, up from $3.7 billion in 2024. That is roughly 3.5x revenue growth in a single year.

But spending grew faster. Total costs and expenses reached $34 billion, broken down as follows according to Zitron’s reporting:

  • Cost of revenue: $7.5 billion
  • Research and development: $19.18 billion
  • Sales and marketing: $5.73 billion
  • General and administrative: $1.57 billion

The loss from operations was $20.92 billion. After factoring in a $41.55 billion non-cash charge related to OpenAI’s conversion from a nonprofit to a for-profit entity, the gross net loss was $60.35 billion. OpenAI reduced that to $38.53 billion by attributing $17.87 billion to “noncontrolling members capital” and another $3.95 billion to “redeemable noncontrolling interests,” according to Zitron’s reporting.

As Gizmodo noted, the Financial Times cited a source familiar with OpenAI’s finances who estimated the company’s adjusted losses for 2025 were closer to $8 billion once one-time restructuring charges are stripped out. Even at that more favorable interpretation, OpenAI spent more than $2.50 for every $1 it earned.

OpenAI’s valuation on private markets has reached $833.7 billion, and the company has announced plans for a September IPO.

Comparing the Capital Stacks

The numbers alone tell a story about divergent strategies, but the structural differences matter more.

OpenAI has raised capital through a complex web of convertible interests, strategic partnerships (most notably the Microsoft relationship, which accounted for $17.2 billion in expenses to Microsoft in 2025 alone), and equity rounds that have diluted the founding team’s control. The nonprofit-to-for-profit conversion required restructuring that added tens of billions in accounting charges. Multiple rounds of fundraising have brought in investors with governance rights, board seats, and strategic agendas.

DeepSeek’s LP structure is the opposite. No board seats for investors. No voting rights. A five-year lock-up that prevents early exits and the short-term pressure that comes with them. Government investment treated differently from private capital, with China’s national AI fund receiving direct equity and voting rights while all other investors go through the limited partnership.

The result is a funding model that looks more like a sovereign wealth fund or an endowment than a traditional venture-backed startup. Liang Wenfeng retains the kind of absolute founder control that most Western AI companies lost years ago.

The Efficiency Question

DeepSeek first gained global attention in January 2025, when its V3 and R1 models drew praise from researchers at OpenAI, Google, and Anthropic. The models achieved competitive performance benchmarks while reportedly requiring a fraction of the training compute used by Western labs. That efficiency advantage has been a recurring theme in DeepSeek’s narrative, though precise training cost comparisons remain difficult to verify independently.

The funding context amplifies this question. DeepSeek’s total fundraise of $7.4 billion, combined with Liang’s personal commitment, gives the company roughly $10 billion in capital. OpenAI spent $34 billion in a single year. If DeepSeek can continue producing competitive models at a fraction of the cost, the capital efficiency gap becomes a strategic advantage that compounds over time.

But there are caveats. DeepSeek operates within China’s regulatory and hardware access constraints. U.S. export controls restrict Chinese companies’ access to the most advanced NVIDIA chips, forcing workarounds in training infrastructure. DeepSeek’s cost advantage may partly reflect lower labor costs and government subsidies rather than pure technical efficiency. And the company’s models, while competitive on benchmarks, have not been tested at the inference scale that drives OpenAI’s $7.5 billion cost of revenue.

What the IPO Dynamic Changes

OpenAI’s planned September IPO introduces a variable that DeepSeek’s structure explicitly avoids: public market accountability.

Once OpenAI goes public at an $833.7 billion valuation, it will face quarterly earnings scrutiny, analyst expectations, and the constant pressure to demonstrate a credible path from $13 billion in revenue to profitability on a $34 billion cost base. Any quarter where spending growth outpaces revenue growth will draw market punishment.

The Anthropic situation adds another dimension. CNBC reported that prediction market traders are pricing 58% odds that Anthropic will restore access to its Fable 5 model by July 1, after the Trump administration directed the company to limit access. Anthropic, which has raised over $12 billion in funding, now faces a regulatory constraint that neither OpenAI nor DeepSeek currently confronts in their primary markets.

The three largest frontier AI companies are now subject to three entirely different capital and regulatory regimes. OpenAI will answer to public markets. Anthropic is navigating government export controls. DeepSeek is insulated by a founder-controlled LP structure backed by Chinese state capital.

The Infrastructure Bet

Both DeepSeek and OpenAI are making enormous bets on compute infrastructure. OpenAI’s $19.18 billion in R&D spending and $7.5 billion in cost of revenue reflect the raw expense of training frontier models and running inference at scale. The company has pledged, though not yet spent, $1 trillion on data center buildouts.

DeepSeek’s infrastructure strategy is less transparent. The company has benefited from China’s aggressive AI infrastructure buildout, with government programs at the national and provincial level subsidizing data center construction and chip procurement. Liang’s own fortune comes from his hedge fund High-Flyer, which generated its returns through quantitative trading strategies that required significant compute infrastructure, giving DeepSeek a pre-existing hardware base.

The compute infrastructure race is the real story beneath both disclosures. Models improve with more compute. Inference at scale requires more compute. And whoever controls more compute, at lower cost, for longer, wins the race for AI capability leadership.

Two Theories of Control

The most consequential difference between these capital structures may be philosophical rather than financial.

OpenAI’s journey from a nonprofit research lab to a for-profit company seeking an $833.7 billion public market valuation represents one theory: that AI development requires massive, diversified capital sources and the accountability structures that come with them. The tradeoff is loss of founder control, complex governance, and the risk that financial pressures distort research priorities.

DeepSeek’s LP structure represents the opposite theory: that AI development benefits from concentrating decision-making authority in a single technical founder, insulated from investor governance and short-term market pressure. The tradeoff is less transparency, less external oversight, and the risk that a single point of strategic failure goes unchecked.

Both theories are being tested simultaneously, at unprecedented scale, with the most consequential technology of the decade at stake. The financial disclosures of the past 48 hours made the terms of that test explicit.