Quora Answer: If AI Companies Report Billions in Losses While Claiming Record Revenue, Which Number Should You Trust?

The following is my answer to a Quora question: “Why do companies like Anthropic report such huge losses, and are these figures misleading when it comes to their actual financial health?”

When it comes to companies like Anthropic, OpenAI and even SpaceX, their numbers are made up of equal parts bullshit and fantasy.  There is a particular species of financial storytelling that flourishes in Silicon Valley.  It goes as follows: the company is losing money now, but the losses are investments in a future so large that today’s arithmetic is irrelevant.  The valuation is not based on earnings — it is based on potential.  The potential is unbounded.  Therefore, the valuation is whatever the last funding round says it is.  This is not a business model.  It is a creation myth.  And the AI industry has built its entire financial architecture on it.

The Numbers

We start with Anthropic.  Anthropic reported US$4.6 billion in revenue for 2025 — up 12 times from 2024.  That sounds impressive.  Then look at the other side of the ledger.  Net loss for 2025: US$42 billion.  Anthropic lost US$42 billion on US$4.6 billion in revenue.  For every dollar it brought in, it destroyed nine.  The company’s operating expenses reached US$12.6 billion.  Compute and infrastructure spending alone was US$7.3 billion, up from US$2.5 billion in 2024.  The remaining US$34 billion in losses came from write-downs of liabilities tied to previous fundraising rounds.  Anthropic’s latest funding round — the Series H in May 2026 — valued the company at US$965 billion.  Its IPO is targeting a valuation of US$2 trillion.  The company has committed to US$518 billion in future cloud, computing, and infrastructure obligations.  One quarter of its revenue comes from two customers who are not locked into long-term contracts.

Now, OpenAI.  OpenAI generated US$13.07 billion in revenue in 2025.  It spent US$34 billion getting there.  Net loss: US$38.5 billion.  The loss grew nearly eight times in a single year.  Projections suggest OpenAI will not reach profitability before 2030 at the earliest – I doubt it can reach profitability at all.  Between 2025 and 2029, analysts project cumulative infrastructure spending of US$115 billion — mostly on data centres.  The company’s valuation hovers around US$500 billion.  OpenAI’s annualised revenue topped US$20 billion by the end of 2025.  At a US$17 billion annual burn rate, that is not a viable business.  It is a race to see whether revenue can outrun spending before the funding runs out.  Some analysts warn of cash depletion by mid-2027 without further capital injections.  The model might not survive contact with reality.

The Revenue Problem

Revenue is not profit.  This distinction should be obvious.  Both Anthropic and OpenAI report revenue figures that grow impressively year on year.  What they do not report — at least not prominently — is the cost of generating that revenue.  OpenAI reportedly burns US$2 for every US$1 it earns on inference: the core business of running its models.  The compute margin improved to 70 per cent in October 2025, up from 52 per cent the prior year.  That is progress.  It is not profitability.

The fundamental problem is structural.  Every time an AI model runs, it consumes compute.  Compute costs money.  As usage scales, compute costs scale.  Unlike a software product — where the marginal cost of an additional user approaches zero — an AI inference product has a real and substantial marginal cost per query.  The industry is trying to grow its way out of this by betting that compute costs will fall faster than revenue grows.  That may happen.  It has not happened yet.  It will certainly not happen using the brute-force compute they espouse.

Meanwhile, the capital expenditure required to stay competitive is not shrinking.  It is accelerating.  Anthropic plans to spend US$518 billion on infrastructure over the coming years.  Microsoft and OpenAI announced a US$500 billion data centre programme called Stargate.  Google committed over US$75 billion in capital expenditure for 2025 alone.  Amazon Web Services is spending at a similar scale.  The race to build compute infrastructure has become its own arms race, and the prize — market leadership in AI — does not yet have a clear revenue model attached to it.

The Valuation Problem

Anthropic lost US$42 billion in 2025.  Its IPO targets a US$2 trillion valuation.  To justify a US$2 trillion valuation on US$4.6 billion in revenue requires a revenue multiple of 435 times.  For context, Apple — one of the most profitable companies in human history — trades at 7 to 9 times revenue.  Microsoft, which has embedded AI into its entire product suite and is generating real AI-linked revenue, trades at 12 times revenue.  The 435-times multiple requires the market to believe that Anthropic will grow its revenue by several orders of magnitude while simultaneously reducing its cost base to something approaching profitability — all before the competitive landscape consolidates around one or two survivors.  This is not impossible.  It is a bet.  The investors making it are sophisticated enough to know it is a bet.  The retail investors who will buy the IPO may not be.

OpenAI’s US$500 billion valuation on US$13 billion in revenue represents a 38-times multiple — still extraordinary by any conventional measure, though modest by comparison to Anthropic.  The difference reflects OpenAI’s larger revenue base and its more advanced commercial deployment.  It does not reflect profitability, which remains years away.

The SpaceX Question

SpaceX is a different animal.  SpaceX reported US$18.7 billion in revenue for 2025, with a 49.4 per cent gross margin.  Its Starlink satellite internet division generated US$11.4 billion in revenue at a 63 per cent adjusted EBITDA margin.  Starlink had 10.3 million subscribers as of the first quarter of 2026.  These are real numbers from a business that sells a product people pay for every month.  The problem is the IPO valuation.  SpaceX initially targeted US$1.5 trillion.  Reports subsequently suggested it had reduced that target to US$1.75 trillion to US$1.8 trillion — which is a strange definition of reduction.  At US$18.7 billion in revenue, a US$1.75 trillion valuation represents a revenue multiple of 94 times.  SpaceX’s 2025 GAAP net result was a loss of US$4.9 billion, against operating cash flow of positive US$6.8 billion.  The GAAP loss was driven by US$8.6 billion in research and development spending — 46 per cent of revenue.  The AI segment alone produced an operating loss of US$6.4 billion.  SpaceX is not like Anthropic or OpenAI.  It has a profitable core business in Starlink.  It has real revenue, real cash flow, and a launch business with real competitive barriers.  The IPO valuation is still absurd.  The company carries US$29.1 billion in total debt, including a US$20 billion bridge loan due in September 2027.  The IPO is partly a mechanism to pay that back.

The anecdote that cuts through the valuation theatre is this: in 2022, the Wall Street Journal reported that SpaceX had suffered losses in every year of its existence except one.  Elon Reeve Musk, its founder, had subsidised it for two decades on the proceeds of his other ventures.  The company’s commercial viability rests almost entirely on Starlink — a business that did not exist until 2019 and whose average revenue per user has been falling, from US$99 per month in 2023 to US$66 per month by the first quarter of 2026.  The valuation extrapolates a future that may not arrive at the speed the price requires.

The Bond Market Knows

Equity market investors are optimists by disposition.  The bond market is not.  When a company issues bonds, the yield reflects what the market thinks about the probability of repayment.  Investment-grade bonds from companies with stable earnings and manageable debt trade at modest spreads over government rates.  High-yield bonds — junk bonds — trade at substantially higher spreads, reflecting the higher risk that the issuer will not be able to service or repay the debt.

Anthropic’s bonds, issued in connection with its infrastructure commitments, trade at yields that reflect the company’s underlying economics: burning US$42 billion in a year, with US$518 billion in future obligations and no path to profitability on a conventional timeline.  The bond market is not interested in potential.  It is interested in cash flows.  Anthropic does not have them.  Bond pricing is a real-time referendum on financial viability.  When the bonds of a company valued at US$2 trillion trade at junk spreads, the bond market is saying something the equity investors are choosing to ignore.

The Investor Base

Who funds these losses?  Anthropic’s investors include Google, which has committed US$2 billion, and Amazon, which has committed US$4 billion — with a total commitment of US$8 billion.  OpenAI received US$50 billion from Amazon, US$30 billion from Nvidia, and US$40 billion in a funding round at a US$300 billion valuation.  SoftBank Group Corporation has made large commitments to OpenAI.  These are not passive venture capital investors betting on a long shot.  They are strategic investors ensuring that the leading AI platforms run on their cloud infrastructure, use their chips, and remain dependent on their distribution.  Amazon’s investment in Anthropic comes with a commitment to run Anthropic’s workloads on Amazon Web Services.  Google’s investment comes with similar cloud commitments.  The investment is partly a customer acquisition strategy dressed as venture capital.  This dynamic sustains the AI companies in the short term.  It also creates a dependency that the AI companies may find difficult to escape.  The infrastructure providers hold the leverage.  The AI model companies need the compute.  The compute providers set the price.

The Fundamental Question

The 2000 dot-com bubble destroyed US$5 trillion in market capitalisation.  The companies that failed were not frauds — most of them.  They were companies that had useful products, real revenue growth, and real potential, but valuations that could not survive contact with the requirement to eventually turn a profit.  Pets.com raised US$82.5 million in its February 2000 IPO at a valuation of US$290 million.  It was bankrupt by November 2000.  Webvan raised US$375 million, spent it building grocery distribution infrastructure, and collapsed in July 2001 with US$830 million in losses.  These were not stupid ideas.  Online pet supplies and grocery delivery are both large, profitable industries today.  The businesses failed because the economics of the execution did not match the ambition of the valuation.

The AI industry is not building grocery delivery.  It is building something transformative.  The question is not whether the technology is real.  The question is whether the revenue model is real, whether the competitive advantages are durable, whether the capital expenditure is recoverable, and whether any of the current market leaders will still be standing when the consolidation happens.  A US$42 billion loss on US$4.6 billion in revenue is not a growth phase.  It is a capital destruction machine that requires an uninterrupted supply of investors willing to fund the next phase before the last phase’s money runs out.  That machine keeps running as long as the narrative holds.  Narratives do not hold forever.  The last time the market funded this kind of loss at this kind of scale, it took three years for the reckoning to arrive.  We are well on the way there.

---

Terence Nunis | Executive Chairman, Equinox Zenith | Author, The 1% Playbook: The Billionaire Cheat Code