OpenAI Audited Financial Report: $38.5 Billion Loss in 2025, $19.2 Billion Burned on R&D, Microsoft Takes $17.2 Billion in a Year

marsbit2026-06-17 tarihinde yayınlandı2026-06-17 tarihinde güncellendi

Özet

OpenAI's audited financials reveal staggering losses. In 2025, the company reported a net loss attributable to OpenAI of $38.53 billion, a dramatic increase from a $5.1 billion loss in 2024. While revenue grew from $3.7 billion in 2024 to $13.07 billion in 2025, costs skyrocketed to $34 billion. A massive $41.55 billion loss from the fair value change of convertible equity and warrant liabilities, tied to its shift to a for-profit structure, significantly contributed to the 2025 result. Research and development spending reached $19.18 billion in 2025. The filings also show OpenAI paid Microsoft approximately $17.2 billion in 2025 for services, including $10.59 billion categorized under "research and development," likely for model training costs. Other major payments included $867 million from SoftBank and $303 million from Microsoft to OpenAI. The report raises serious concerns about the AI company's path to sustainability and profitability, highlighting financial figures far exceeding external expectations.

Author: Ed Zitron

Compiled by: Deep Tide TechFlow

Deep Tide Introduction: OpenAI's audited financial report has been exposed for the first time, revealing losses far exceeding external expectations. The loss soared from $5.1 billion the previous year to $38.5 billion in 2025, with R&D expenditure as high as $19.2 billion, while Microsoft received $17.2 billion in fees from OpenAI in a single year. How far is this AI star company, valued at hundreds of billions, from achieving profitability?

Today, I can exclusively report, based on audited financial documents reviewed by this publication (independently verified by the Financial Times), that OpenAI lost approximately $38.5 billion in 2025, along with other key details about the company's financial condition.

Given the seriousness of this report, I will not comment extensively, as the numbers speak for themselves.

OpenAI Lost $5.1 Billion in 2024

2024 – OpenAI revenue was $3.7 billion, costs and expenses were $12.4 billion, and the net loss attributable to the company was $5.1 billion.

OpenAI's financial statements tell the story of a company with staggering losses.

Revenue: $3.7 billion

Cost of Revenue: $2.65 billion

Research and Development: $7.81 billion

Sales and Marketing: $1.11 billion

General and Administrative: $907 million

Total Costs and Expenses: $12.48 billion

Operating Loss: $8.78 billion

Other factors including interest income and interest expense brought its net loss to $8.84 billion. It then marked $3.74 billion of the loss as "Net loss attributable to noncontrolling member capital," making the net loss attributable to the company $5.1 billion.

It is currently unclear what this means or how OpenAI adjusted to remove $3.74 billion in costs. I will not speculate further.

OpenAI Lost $38.5 Billion in 2025

2025 – OpenAI revenue was $13.07 billion, costs and expenses were $34 billion, the loss was $20.92 billion, and the net loss attributable to the company was $38.53 billion.

Revenue: $13.07 billion

Cost of Revenue: $7.5 billion

Research and Development: $19.18 billion

Sales and Marketing: $5.73 billion

General and Administrative: $1.57 billion

Total Costs and Expenses: $34 billion

Operating Loss: $20.92 billion

Note that 2025 was the year OpenAI transitioned from a non-profit to a for-profit entity, resulting in a $41.55 billion loss due to changes in the fair value of convertible equity and warrant liabilities.

Considering other minor factors such as interest income and interest expense, OpenAI's net loss was $60.35 billion. This was reduced to $38.53 billion by removing $17.87 billion via "Net loss attributable to noncontrolling member capital" and an additional $395 million via "Net loss attributable to redeemable noncontrolling interests."

Ultimately, the net loss attributable to OpenAI in 2025 was $38.5 billion.

At year-end, OpenAI had assets just over $50 billion, nearly half of which was cash.

SoftBank Paid OpenAI $867 Million, Microsoft Paid $303 Million in 2025

In 2025, SoftBank paid OpenAI $867 million. Microsoft paid $303 million.

The documents disclose how much OpenAI paid Microsoft in service fees. In the 2025 calendar year, OpenAI paid Microsoft $10.59 billion for "research and development" expenses. We believe this likely refers to the costs of training OpenAI models.

The documents also mention a $6.047 billion fee related to "cost of revenue," $527 million in sales and marketing expenses, and $42 million in "general and administrative expenses." In total, OpenAI's payments to Microsoft amounted to $17.2 billion.

According to the data, at the end of the calendar year, OpenAI had liabilities to Microsoft of $3.64 billion, plus another $21 million in "accrued expenses and other current liabilities." The documents also mention an additional $58 million in non-current liabilities.

Follow-up Note

I intend to follow up on this story next month with a more in-depth report related to these documents. The documents are quite detailed, and I need time to fully parse them. You will know once it's done.

OpenAI's financial condition is deeply concerning. A loss of $38.53 billion is an astronomical figure, far higher than most anticipated. The losses also appear to be growing at an alarming rate year over year. I am unsure how this company finds a path to any form of sustainability or profitability.

As mentioned, I have not commented much today. I believe the best thing I can do for the public is to deliver this news as concisely as possible.

İlgili Sorular

QAccording to the article, what were the major cost drivers for OpenAI's massive losses in 2025?

AThe major cost drivers for OpenAI's massive $38.5 billion loss in 2025 were Research & Development (R&D) expenses at $19.18 billion and a significant payment to Microsoft of $17.2 billion. The R&D itself was heavily comprised of a $10.59 billion payment to Microsoft for R&D costs, likely for model training.

QHow much did Microsoft receive from OpenAI in 2025 according to the audited financials, and what was the primary purpose of this payment?

AIn 2025, Microsoft received a total of $17.2 billion from OpenAI. The primary purpose of the payment was for services, with the largest component being $10.59 billion designated for "research and development" fees, which the article suggests likely refers to the costs of training OpenAI's AI models.

QWhat financial impact did OpenAI's transition to a for-profit entity have on its 2025 losses?

AOpenAI's transition to a for-profit entity in 2025 led to a significant accounting loss of $41.55 billion due to "changes in the fair value of convertible equity and warrant liabilities." This non-cash charge was a major contributor to the company's reported net loss.

QWhat was OpenAI's revenue in 2024 and 2025, and how did its losses change between these years?

AOpenAI's revenue was $3.7 billion in 2024 and $13.07 billion in 2025. Despite this revenue growth, its net loss (attributable to the company) increased dramatically from $5.1 billion in 2024 to $38.53 billion in 2025.

QWho were the other major disclosed payors to OpenAI in 2025 besides Microsoft, and what was the amount?

ABesides Microsoft, the other major disclosed payor to OpenAI in 2025 was SoftBank, which paid $867 million. Microsoft was also listed as a payor with a payment of $303 million, separate from the massive $17.2 billion in service fees.

İlgili Okumalar

Beyond APR: After Staking, Who Really Owns Your ETH?

Beyond APR: Who Actually Controls Your ETH After Staking? The article challenges the common focus on Annual Percentage Rate (APR) when choosing ETH staking services. It argues that as staking yields converge, the more critical question is who controls the staked assets. The core of non-custodial staking lies in Ethereum's key separation design. A validator uses two keys: a **Signing Key** (for online consensus tasks) and a **Withdrawal Credential** (the ultimate control over funds). Service providers can manage the Signing Key to run the validator but cannot access the staked ETH. The user retains the Withdrawal Credential, which can be kept offline. Post-EIP-7002, users can even force a validator exit directly from their wallet if a service provider disappears. Four roles are involved: 1. **User:** Retains ultimate fund control via the Withdrawal Credential. 2. **Wallet:** Acts as an interface for managing permissions, not an asset owner. 3. **Node Service Provider:** Manages validator operation (Signing Key), carrying risks related to performance and slashing, but not fund theft. 4. **Ethereum Protocol:** Governs the immutable rules for activation, exit, and withdrawal. The piece contrasts this with Liquid Staking like Lido. Users receive a liquid token (e.g., stETH) representing a claim on pooled staked ETH, enabling DeFi composability. However, they do not control individual validator withdrawal credentials; redemption relies on the protocol's withdrawal queue. This trades direct control for liquidity and lower entry barriers. The conclusion is pragmatic: Liquid staking suits users with smaller amounts or high liquidity needs. For long-term holders with 32+ ETH, where yield differentials are minimal, the security model and direct control of a non-custodial, native stake become paramount. The essential question shifts from "What's the APR?" to "Who holds the keys to my staked ETH?"

marsbit8 dk önce

Beyond APR: After Staking, Who Really Owns Your ETH?

marsbit8 dk önce

Riemann Hypothesis Pushed to 99.55% of Theoretical Boundary: Meta-Architecture AI Reshapes Its 'Brain' Through Thinking

When we marvel at large language models writing poems or coding, scientists face a more challenging question: how can AI tackle complex, long-term scientific problems that require months or years of sustained, uncertain reasoning? Traditional AI approaches use a fixed architecture—like forcing a mathematician, physicist, and engineer to share the same brain. While versatile, such a system often lacks specialization for specific, complex tasks. Recently, an international academic team introduced the Eureka framework, a groundbreaking solution. Instead of a fixed brain, Eureka allows AI to dynamically grow task-specific architectures during problem-solving—creating the optimal dedicated brain for each challenge. In rigorous testing, Eureka completed all 170 complex recursive long-horizon tasks, generating 3,948 verifiable certificates without a single instance of erroneous self-approval or false completion. Eureka demonstrated remarkable adaptability across two distinct scientific challenges: 1. Advancing the Riemann Hypothesis: For this mathematical milestone, Eureka constructed a specialized agent that expanded the scope of a key proof path—local Weil quadratic form positivity—from a ≤ 1/4 to a ≤ 69/200 (0.345). This reaches 99.55% of the theoretical first barrier, representing a significant, verifiable step forward. 2. Reconstructing Physical Theory: In exploring quantum processes and spacetime theory, Eureka generated a theory-discovery agent. It produced five progressive structural findings, rigorously proving that behavioral equivalence does not imply essential equivalence—providing a stricter standard for comparing physical theories. Why Fixed Architectures Fall Short: Predefined AI structures—whether single- or multi-agent—are efficient for known problems but struggle with the long-term uncertainty and structural variability of scientific discovery. They face issues like architecture mismatch and planning failures when early assumptions are invalidated. The Eureka Approach: A Meta-Architecture for Self-Evolution Eureka integrates task execution and brain shaping. Its workflow includes: - Dynamic task compilation into a flexible "obligation graph." - Identifying "architectural hotspots" where specific tools, memory, or collaboration patterns are repeatedly needed. - "Architectural promotion": compiling dedicated "macro-agents" with their own memory, tools, verification standards, and workflows to handle specialized subproblems. - Controlled self-evolution: upgrading macro-agents only when benefits clearly outweigh costs. Results and Outlook: Eureka reduced core contextual overhead by 57.8% and avoided 65.38% of redundant computations. In concurrent testing with 16,000 tasks, it ensured perfect consistency with sequential execution, eliminating data conflicts. This approach paves the way for AI in high-difficulty scientific domains—accelerating mathematical research, aiding theoretical physics discovery, and managing large-scale engineering problems like chip design or drug development. Eureka represents a paradigm shift: future AI capability may depend less on model scale and more on dynamic, task-aware brain shaping.

marsbit13 dk önce

Riemann Hypothesis Pushed to 99.55% of Theoretical Boundary: Meta-Architecture AI Reshapes Its 'Brain' Through Thinking

marsbit13 dk önce

Want to Buy a Dog? Buy NVIDIA First: Are Equity Memes the Engine of This Bull Market?

The article explores the emergence of "stock-meme" coins on the Robinhood Chain, a new trend where meme tokens like Artificial Inu are paired directly with tokenized stocks like Nvidia (NVDA). This mechanism forces trading activity for otherwise static tokenized equities. The narrative suggests this fusion could be a catalyst for the next crypto bull market, akin to how ETFs brought traditional finance capital last cycle. However, the model faces critical challenges. A case study involving HIMS stock tokens revealed a 460% premium over the real stock price on a weekend, exposing a fundamental flaw: token supply is controlled by a single authorized participant who can only mint new tokens when the underlying stock market is open. This leaves the 24/7 crypto market vulnerable to disconnection from real-world pricing and liquidity. Data analysis shows frenetic but shallow activity: 61% of wallets are one-day users, bots generate over half the trading volume, and median transaction sizes are small. While tokenized stock holders increased 155% in 30 days, regulatory uncertainty and geographic restrictions (blocking US, Canada, UK users) limit new capital inflow. The article concludes that while "stock-memes" can ignite short-term hype, their sustainability as a "bull market engine" depends on attracting genuine external capital rather than just recycling existing crypto funds.

marsbit13 dk önce

Want to Buy a Dog? Buy NVIDIA First: Are Equity Memes the Engine of This Bull Market?

marsbit13 dk önce

İşlemler

Spot
活动图片