OpenAI vs Anthropic IPO Race: Who's Really Burning Cash? A Deep Dive Into AI Giants' Finances

Recently, the hottest topic in the tech industry has been the IPO rumors surrounding the AI giants. In the latest episode of the Big Technology Podcast, host Alex Kantrowitz and Ranjan Roy of Margins took a deep dive into the financials of OpenAI, Anthropic, and SpaceX, revealing the "true ledger" behind these star companies. When we strip away the narrative cloak of "exponential growth," what we see is a capital game about timing, storytelling, and survival.

The Truth About OpenAI's Finances: For Every $1 Earned, $1.22 Lost

According to a report by The Information, OpenAI generated roughly $5.7 billion in revenue in the first quarter—nearly $1 billion more than Anthropic over the same period—with the Codex product serving as an important growth engine. Yet behind the glossy numbers lies staggering losses—OpenAI disclosed to investors that its adjusted operating margin was negative 122%.

Understanding this number requires some context. Operating margin is a key indicator of a company's core business profitability, calculated as operating profit divided by revenue. Negative 122% means that for every $1 of revenue OpenAI generates, it spends $2.22 at the operational level. For traditional SaaS software companies, once R&D investment is complete, marginal costs approach zero, and scaling up brings continuous improvement in profit margins. AI large-model companies, however, face a fundamentally different cost structure: every user query requires substantial GPU compute for inference—a variable cost that grows linearly with usage, rather than a fixed cost.

The inference cost of large language models is central to understanding AI companies' financial dilemmas. Unlike traditional software, whose marginal costs approach zero once development is complete, every LLM inference requires loading billions of parameters into GPU memory and completing forward-pass computation. A single inference on a GPT-4-class model may require hundreds of A100/H100 GPUs working in concert for hundreds of milliseconds, translating to a monetary cost of roughly $0.03–0.12 per thousand tokens—while user subscription fees often struggle to cover the compute consumption of high-frequency usage scenarios. High-performance GPUs like NVIDIA's H100/H200 cost up to $30,000–40,000 each, and combined with data center power and cooling (AI data centers generally have higher PUE than traditional IDCs) and ongoing model training costs, they form a cost ceiling that AI companies cannot solve by simply "turning off the sales faucet" as traditional software firms can. Industry estimates suggest that mainstream LLM inference costs currently drop about 75% every 18 months, but demand grows even faster, causing total costs to keep rising.

CNBC even ran the question "Is OpenAI the new WeWork?" as a chyron on its broadcast.

Guest Ranjan pointed out that what truly surprised Wall Street wasn't the loss itself, but that the loss appeared at the "operating profit" level. Operating profit already accounts for cost of goods sold, R&D, and selling, general, and administrative expenses—all traditional outlays. A negative figure at this level shows that the AI business model differs fundamentally from the traditional software industry.

"AI companies are more like industrial companies, because you have to pay for compute, and the resources you invest scale linearly with size."

This assessment explains why AI companies can't replicate the Uber-style path—software companies can "turn off the sales and marketing faucet" to achieve profitability, but AI companies' cost structures make that road impassable. A more accurate analogy is an oil refinery or a semiconductor manufacturer: every additional unit of output requires a corresponding increase in raw material inputs (compute, electricity). The emergence of economies of scale takes far longer than for a pure software company and depends on the continuous decline of the hardware cost curve.

Why Now, of All Times? Anthropic's Catch-Up and Shrinking Gulf Money

With financials this ugly, why is OpenAI in such a rush to IPO now? Host Alex offered a fairly persuasive judgment: this is entirely about the pressure created by Anthropic's rapid growth.

According to The Wall Street Journal, Anthropic is experiencing explosive growth—$4.8 billion in sales in Q1, projected to double to $10.9 billion in Q2, and potentially achieving its first-ever operating profit in some quarter. While OpenAI remains the revenue leader, Anthropic's growth curve is steeper.

Alex analyzed that what OpenAI most wants to avoid is Anthropic going public first and seizing the narrative high ground of "rapid growth." If Anthropic tells its story first, then when OpenAI follows, the data comparison would be extremely awkward—at that point it would be hard to convince Wall Street to believe "we'll ultimately overtake our rival through infrastructure bets." Defining the narrative before your rival thus becomes OpenAI's strategic choice.

Deeper pressure comes from shrinking financing channels. The episode noted that although OpenAI recently completed a record-breaking round of funding, the Gulf state sovereign wealth funds that serve as the source of its last-resort "ammunition" are becoming increasingly uncertain due to regional situations (such as the war with Iran and disruptions to Qatari natural gas).

The Sovereign Wealth Fund (SWF) is the key variable in this financing game. Gulf sovereign wealth funds—including Saudi Arabia's PIF (Public Investment Fund, managing over $700 billion in assets), Abu Dhabi's ADIA and Mubadala, and Qatar's QIA—collectively manage over $3 trillion in assets and have become a major force in global tech investment. Their interest in AI infrastructure stems from multiple motives: seeking post-oil-era economic diversification (Saudi Arabia's "Vision 2030" explicitly lists tech investment as a core pillar), consolidating geopolitical influence by controlling compute and data leverage, and pursuing high-yield assets uncorrelated with traditional commodities.

Notably, Gulf SWFs' AI investments are not purely financial acts—Saudi Arabia plans to build over 100GW of compute infrastructure domestically, and the UAE's G42 has reached a $1.5 billion strategic investment agreement with Microsoft. This "capital + compute sovereignty" bundling model means that when AI companies accept Gulf money, they face not just valuation negotiations but complex geopolitical considerations such as data sovereignty and technology transfer. Dario once publicly stated a willingness to "go back on some of our values" to accept Gulf money, but that massive round of financing never materialized.

"Either go public now, or wait three or four years—and OpenAI may not have that much time." This was the consensus reached by the two hosts.

Anthropic's Profitability Puzzle: A SpaceX "Assist"?

Regarding the news that Anthropic is about to become profitable, the episode raised a rather controversial point. Ranjan cited Ed Zitron's analysis, noting that SpaceX's S-1 prospectus disclosed that Anthropic committed to spending $1.25 billion per month through May 2029, but the document explicitly states that a "discounted rate" applies for May and June 2026.

The importance of the S-1 document deserves special mention. The S-1 is a registration statement that the U.S. Securities and Exchange Commission (SEC) requires companies to file before an initial public offering (IPO). Its disclosure requirements are extremely stringent, mandating audited financial statements, Management's Discussion and Analysis (MD&A), and details of material contracts (typically involving customer relationships where the contract amount exceeds 10% of revenue), and it carries legal weight. The core difference between an S-1 and the "information memorandum" (IM) commonly used in private financing is that the former is bound by securities law, and false statements carry civil and even criminal liability—so the contract details disclosed within carry very high credibility. It is precisely for this reason that SpaceX's disclosure of the details of Anthropic's committed-spending contract in its S-1 became a credible basis for reverse inference in the episode.

Notably, tech companies increasingly tend to engage in "selective leaking" before formally filing their S-1—disclosing curated data to specific media outlets to shape the market narrative in advance, so that by the time the S-1 is officially disclosed, the market has been fully primed and the impact of negative data has been diluted. This practice operates in the gray area of disclosure compliance.

"This discount rate might be zero," Ranjan boldly speculated: perhaps Anthropic reached some tacit understanding with SpaceX—exchanging a long-term, massive commitment for a huge discount or even free service in the two key months—so that the quarterly financials "just so happen" to achieve profitability, which is then selectively leaked to The Wall Street Journal.

While this speculation hasn't been confirmed, it reveals an important pivot: AI companies are shifting from the "the more we lose, the more glorious" construction narrative to the "profitability" narrative that Wall Street favors. This shift has a profound capital-market backdrop: the low-interest-rate environment of 2021–2022 gave rise to the "burn cash for growth" venture-capital paradigm, while the high-interest-rate environment since 2023 has made institutional investors once again prioritize a "clear path to profitability" as a core consideration. If Anthropic can demonstrate profitability, it can switch the applicable multiple in its IPO valuation negotiations from a "revenue multiple" to an "earnings multiple," fundamentally transforming the valuation logic. Ranjan called Anthropic a "master of communication," capable of carefully weaving numbers together and leveraging media leaks to shape market expectations. Alex remained reserved, believing that "why only May and June get the discount" is still an unsolved mystery.

Signs of a Growth Peak: The Claude Code "Token Bonanza" Recedes

Is the explosive growth of AI peaking? The episode outlined several warning signs.

Slowing user growth. Data from The Information shows OpenAI's average quarterly weekly active users at around 905 million; although the February peak reached 920 million, a declining average means subsequent usage has fallen back somewhat. The original promise to reach 1 billion weekly actives by the end of last year has now turned into telling employees to "prepare for hard times," because Google Gemini is steadily grabbing share.

Enterprise-side "token budget awakening." The episode cited multiple cases: Uber said it burned through its cloud budget in four months; Microsoft, according to The Verge, is canceling most of its Claude Code licenses in favor of pushing developers to use Copilot CLI. Investor Chamath commented: "The problem isn't that the tools are useless, but that without context and oversight, the tools spin endlessly, creating an enormous cost burden."

Understanding "token consumption" is key to understanding this crisis. In the billing system of large language models, a "token" is the basic unit of text measurement—roughly 1,000 tokens per 750 English words. When enterprises use AI coding assistants like Claude Code, every code generation, review, and debugging conversation consumes large amounts of tokens, with fees accumulating in real time based on usage.

The emergence of the "token maxing" phenomenon stems from enterprises lacking effective usage governance (AI Governance) mechanisms during the early stages of AI tool adoption—without cost awareness, developers tend to let the model "think" and iterate without limit. This phenomenon is especially pronounced in AI coding assistants: when developers use "agentic mode," the model autonomously decomposes tasks, calls tools, and iterates on verification, potentially consuming tens of thousands or even hundreds of thousands of tokens per task. The deeper dilemma lies in quantifying ROI: productivity gains for software engineers are hard to translate directly into financial metrics, and enterprise CFOs cannot calculate "how many engineer hours were saved," so under budget pressure, AI tool subscriptions become the easiest expenditure to cut. A 2024 McKinsey survey showed that about 67% of enterprises said it was difficult to quantify the direct financial returns of generative AI tools. This "ROI black hole" is becoming a core obstacle to slowing enterprise AI spending.

Ranjan called the past few months the era of "token maxing"—everyone using Claude Code without restraint, with usage and revenue both exploding. But when enterprises begin questioning ROI and can't measure whether individual productivity gains match the costs, the party ends abruptly. "The revenue growth curve Claude has enjoyed during this period is one we may never see again."

This also points to the industry's future direction: model routing and provider neutrality. Intelligently routing to different models based on request type and complexity (VO for video, or more cost-effective options like Seedance and Kling) will gradually become mainstream, and the era of "one model to rule them all" is hard to sustain. This trend is fostering the rise of an "AI middleware" layer—routing frameworks like LiteLLM and OpenRouter allow enterprises to dynamically allocate to the optimal model based on task complexity, cost budget, and latency requirements, essentially commoditizing LLM capabilities and compressing the pricing power of any single model provider.

SpaceX Prospectus: Is the $22.7 Trillion AI Market a Fantasy?

If OpenAI's negative 122% margin is still "within reason," SpaceX's prospectus took the two hosts from shock to anger.

According to Bloomberg, SpaceX posted a net loss of $4.28 billion in Q1 on $4.69 billion in revenue, with Musk maintaining absolute control over the company through supervoting shares. What truly stunned them was its claimed "largest addressable market in history"—$28.5 trillion—of which 93% (about $22.7 trillion) supposedly comes from AI enterprise applications.

"Total Addressable Market (TAM)" is one of the most frequently manipulated numbers in a prospectus. The Total Addressable Market is usually calculated in three ways: top-down (starting from the macro market size), bottom-up (starting from a unit economic model), and value theory (based on the maximum value users are willing to pay). Rigorous institutional investors typically require a bottom-up TAM calculation—multiplying the ARPU (average revenue per user) of a single customer by the reachable number of customers—rather than slicing away layer by layer from a macro figure.

SpaceX's claimed $22.7 trillion AI enterprise application market clearly employs extremely aggressive assumptions—lumping in all global enterprise IT spending, software licensing fees, and even future AI service markets that don't yet exist. This "astronomical TAM" strategy is common in tech IPOs and has specific use cases: when there is a huge gap between a company's current business scale and its target valuation, showcasing a massive potential market can justify a high valuation. WeWork once claimed a TAM of $3 trillion (the total value of global commercial real estate); Uber once counted taxis, public transit, and freight all in. Professional investors typically remain skeptical, but for retail investors, a vast market narrative holds powerful psychological appeal—this is precisely the underlying logic of SpaceX choosing to IPO now, relying on retail enthusiasm to support its valuation.

Alex angrily pointed out that this wasn't even SpaceX's core business a year ago. After acquiring xAI, it suddenly claims it wants to do AI enterprise applications, yet Grok lags far behind Google, Anthropic, and OpenAI in market performance and is nowhere near capable of capturing that astronomical market.

As for the "space data center" concept, Alex flatly said he doesn't believe it: "Things going wrong in a data center is routine—on the ground you just send a technician; if something goes wrong in a space data center, you have to send an astronaut." Although Jeff Bezos's Blue Origin has submitted its "Project Sunrise" plan to the FCC to launch 51,600 data center satellites, the two hosts remain strongly skeptical of this vision. The engineering challenges lie not only in maintenance costs—low-orbit satellites face round-trip latency of roughly 250–600 milliseconds (higher than the ideal for terrestrial fiber backbones), and cosmic rays cause bit-flip error rates in DRAM far higher than on the ground, subjecting the large-scale matrix operations that AI inference relies on to additional data-integrity challenges.

You may not have noticed that SpaceX's truly money-making business, Starlink ($11.7 billion in revenue, $4.4 billion in profit), is not the core of the prospectus story. This maneuver of "using a vague AI narrative to cover a solid cash flow" precisely reflects the current frothiness and bubble in the AI capital market. As the world's largest LEO broadband constellation, Starlink has a near-monopolistic competitive advantage in connecting remote areas—a defensible moat that, in analysts' eyes, is far more credible than a "$22.7 trillion AI market." Yet in today's narrative-driven market, the valuation multiple for a "solid infrastructure business" is far lower than for a "disruptive AI platform." This is the fundamental reason the prospectus narrative must be packaged with an AI story.

The AI PR Crisis: Why Are Graduates "Booing" AI?

The episode closed by touching on a more macro topic—AI's increasingly severe public reputation problem. Former Google CEO Eric Schmidt was booed at the University of Arizona commencement, several speakers who mentioned AI faced the same treatment, and one CEO even angrily retorted to graduates to "deal with it yourselves."

This phenomenon has deep sociological roots. The term "Techlash" first appeared around 2018, referring to the public's systemic crisis of trust in Big Tech companies, triggered by issues such as the Facebook Cambridge Analytica scandal, algorithmic discrimination, and platform monopolies. The current new wave of backlash triggered by AI (which could be called Techlash 2.0) is structurally fundamentally different: the early internet-era Techlash focused mainly on privacy violations and platform monopolies, affecting a relatively specific group; whereas the impact of generative AI is believed to systemically affect white-collar and creative industries, threatening the economic security of the middle class—lawyers, accountants, designers, programmers.

More critically is the issue of benefit distribution: the main beneficiaries of AI productivity gains are business owners and capital holders (through reduced labor costs), not workers (whose wage-bargaining leverage is weakened). This model of "privatized gains, socialized risks" closely resembles the logic that triggered the labor movement during the era of early-20th-century industrial automation. The booing at commencements reflects precisely the deep resistance of the younger generation to "technological elitism"—they face a huge gap between debt burdens, a shrinking job market, and the pervasive optimism of the AI narrative.

The host cited a sharp observation: regarding Big Tech's AI narrative, one of the following three points must be true—either no one can tell the story of how AI benefits society; or no one has even thought about telling it; or they believe AI will cause harm but think fixing it isn't their job.

Dario's remarks about "AI will cause mass unemployment," Marc Andreessen's statement on a podcast that "robots will never get depressed over a breakup and never complain to HR," and Zuckerberg's internal talk during Meta's layoff of 8,000 people that "AI learns from the operations of smart people" all became triggers that intensified public resentment. The common feature of these statements is: describing AI's impact on the labor market as an inevitable natural law, while pushing the responsibility for solutions (retraining, social safety nets) onto governments and individuals—with tech companies posturing as "tool providers" and deliberately dodging the topic of "the social responsibility of tool creators."

The two hosts' conclusion is thought-provoking: public attitudes toward AI have evolved into a political issue. When candidates begin discussing "kill switches" and moratoriums, and when data centers can't be built due to public opposition, these "boos" gain real impact. Sociological research shows that when a technology is perceived to create systemic inequality with highly concentrated beneficiaries, public sentiment often rapidly evolves from "skepticism" to "political mobilization." The mandatory push of the EU's AI Act, the AI employment impact assessment requirements successively issued by several U.S. states, and unions beginning to include "algorithmic management rights" and "AI displacement compensation" in collective bargaining agendas are all concrete manifestations of this mobilization trend. This will keep the pressure for AI regulatory legislation rising in the coming years.

By comparison, Chinese AI leaders such as DeepSeek^Ad's founder keep a low profile, and ByteDance's Doubao is even named after "steamed bun" (baozi) to draw closer to ordinary users—perhaps Silicon Valley's elite really should "tone it down a bit." Behind the differences in narrative strategy lie two starkly different technological-political cultures: one challenges the existing order in the identity of a "disruptor," viewing the social friction of technological change as a necessary cost of progress; the other deliberately downplays the disruptiveness of technology, emphasizing its continuity and gradualism as a productivity tool. Against the backdrop of a deepening crisis of public trust in AI, which strategy is more sustainable? The answer may already be clear.

Key Takeaways