NextFin News - Anthropic has told investors it expects to book an operating profit for a second consecutive quarter, a milestone that separates the Claude maker from rivals still burning hundreds of billions of dollars on compute and sharpens the question hanging over a potential $2 trillion initial public offering: can a frontier AI lab stay profitable while the infrastructure arms race accelerates?
The company disclosed the outlook to investors as part of an ongoing financing push, according to people with knowledge of the matter. It marks a rapid reversal for a business founded in 2021 that had been expected to lose money for years.
The numbers behind the shift are stark. In the second quarter, Anthropic is projected to generate $10.9 billion in revenue, more than double the $4.8 billion it posted in the first quarter, and to turn an adjusted operating profit of $559 million - a margin of about 5.1% that includes model-training costs but excludes stock-based compensation. That made the June quarter the company's first profitable period ever. Now the company is telling investors the current quarter will also clear the line.
The profitability milestone arrives as Anthropic's revenue curve steepens. Annualized revenue climbed from roughly $9 billion at the end of 2025 to $14 billion in February, $19 billion in March, $30 billion in April, $44 billion in early May, $47 billion by mid-May and more than $65 billion by the end of July.
Anthropic's annualized revenue run rate climbed from roughly $9 billion at the end of 2025 to more than $65 billion by July 2026 (chart: output/anthropic-runrate.png).
The company also told investors it now has about 6,000 customers spending at least $100,000 a year - up sevenfold in twelve months - and more than 1,000 customers paying at least $1 million annually, a figure that more than doubled in roughly two months. Roughly 80% of Anthropic's business comes from enterprises, its chief executive has said, helped by a viral coding tool, Claude Code, that has reached a $2.5 billion annualized revenue run rate with business subscriptions quadrupling since the start of the year.
The context is a funding market that has rarely been hotter. Anthropic closed a $30 billion Series G round in February at a $380 billion post-money valuation - the second-largest private technology financing on record, behind only OpenAI's round of more than $40 billion the prior year - then raised more at roughly $965 billion in May. It has been in discussions to raise at least another $30 billion at a valuation above $900 billion, and it confidentially filed for an IPO on June 1 with a listing targeted for October. Some investors model a $2 trillion valuation at flotation, with a few, including Gavin Baker of Atreides Management, arguing the company could be worth $3 trillion today.
That ambition would put Anthropic in the company of the year's other mega-listing. SpaceX raised $75 billion in June at an initial valuation of roughly $1.75 trillion. An Anthropic float at $2 trillion would not merely rival it; it would become the largest technology IPO in history.
The Margin Turned Because Revenue Outran a Pre-Committed Cost Base
The first question is mechanical: how does a company that spends tens of billions on GPUs suddenly print profit? The answer is that Anthropic's revenue is growing faster than its cost base can absorb it, at least for now. A research firm estimated in May that gross margin on Anthropic's inference infrastructure had risen from 38% to more than 70% as token-production costs fell, and later pegged inference margins on one of its flagship models above 85%. With roughly three-quarters of annual recurring revenue coming from usage-based APIs rather than flat subscriptions, each incremental token sold carries improving unit economics.
That is the opposite of the textbook AI-lab trajectory. The standard model - spend on training, give the model away, hope monetization catches up - produces widening losses as scale increases. Anthropic's adjusted operating income went positive because revenue more than doubled in a single quarter while a large share of its compute bill was pre-committed under long-term cloud contracts.
But the margin is thin, and it is an adjusted figure. At 5.1%, the profit line is a rounding error against the revenue number. A valuation analysis calculated that even at a $120 billion annualized run rate, a 5.1% margin produces about $6.2 billion in annualized operating profit - meaning a $2 trillion valuation would trade at roughly 325 times that earnings power. Only at a 30% operating margin would the multiple fall to about 56 times. The profitability headline is real; the valuation math still depends on margin expansion that has not yet been demonstrated at scale.
Two Exponentials Are Racing - Revenue and Compute Cost
Is this profitability cyclical or structural? The honest answer is that two different forces are at work, and they point in opposite directions.
The cyclical leg is the training supercycle. Anthropic has committed more than $130 billion to AWS and Microsoft Azure compute alone, including more than $100 billion to Amazon over ten years and $30 billion to Azure, plus a multi-gigawatt agreement with Google and Broadcom. Independent analyses of the company's forecasts put total cloud commitments across Amazon, Google and Microsoft at roughly $80 billion through 2029. Those bills do not disappear. As the company races to train each successive generation of Claude, capital intensity will rise, and a single large training run can erase a quarter of thin profits.
The structural leg is the demand side. Anthropic gets about 80% of its business from enterprises, and its viral coding tool, Claude Code, has reached a $2.5 billion annualized revenue run rate with business subscriptions quadrupling since the start of the year. Enterprise customers lock into usage through workflows - coding assistants, customer-support agents, document processing - that embed the model into daily operations. That kind of embedding is stickier than consumer chat traffic and more resistant to price competition. If agentic workloads consume ten to a hundred times more compute per developer-day than chat, as industry estimates suggest, the revenue per customer can keep compounding even as per-token prices fall.
The mechanism, then, is a race between two exponentials: revenue per enterprise customer compounding on the back of agentic adoption, and compute cost compounding on the back of the training arms race. Profitability persists only if the first outruns the second.
It's a very capital-intensive business to train AI models.
That is not a skeptic's line - it is what Anthropic's own president, Daniela Amodei, said at a technology conference in early June, days after the company confidentially filed IPO paperwork. The admission matters because it frames the central tension: the company is profitable precisely because a wave of enterprise demand is meeting capacity it already committed to, and the next training cycle will require committing to more.
The Second-Order Effect: A Repricing of the Whole Private AI Cohort
The market's first-order read is simple: profitable AI lab, buy the IPO. The second-order question is different: what happens to the rest of the private AI market when one lab proves profitability is possible?
The immediate transmission channel runs through capital allocation. OpenAI, by comparison, has told investors it is targeting roughly $600 billion in total compute spending by 2030 - a figure it scaled back from the $1.4 trillion in infrastructure commitments its chief executive had previously touted - while projecting 2030 revenue above $280 billion and not expecting to reach profitability until 2030. SpaceX's IPO filing showed its AI division, which includes xAI and Grok, ran a $6.4 billion operating loss in 2025 on $3.2 billion of revenue, consuming 61% of the company's $20.7 billion capital spending that year. In 2024, that same division lost $1.56 billion on $2.62 billion of revenue - the gap between what it earns and spends is widening, not narrowing.
If Anthropic can demonstrate that an API-first, enterprise-weighted model reaches profit first, investors will start pricing the laggards accordingly. That creates a split in the private AI market: enterprise-API labs get rewarded, consumer-and-training-heavy labs face scrutiny. The second-order effect is a repricing of the entire private AI cohort before any of them list.
There is also a pricing trap buried in the numbers. Anthropic's revenue growth has partly ridden a wave of enterprise experimentation - a spending surge that some observers have called "tokenmaxxing," where companies briefly prioritized token volume with few cost controls. If customers start optimizing usage instead of expanding it, revenue growth could slow even as model capability improves. Recent reporting on Anthropic's model mix has shown that its most powerful model is struggling to attract users as cheaper tools thrive - a signal that performance leadership does not automatically convert into usage share.
The risk is not theoretical. Steve Eisman, the investor known for his early bearish bet against subprime mortgages, has warned that cheap Chinese open-weight models could force a price war that wrecks valuations at both OpenAI and Anthropic. His point is structural: if frontier capability becomes commoditized, per-token prices fall toward marginal cost, and the 5.1% margin compresses back into loss territory. Eisman has also noted that OpenAI and Anthropic together underpin roughly 70% of hyperscaler AI revenue - a concentration that makes the two companies' pricing power a systemic variable for the cloud providers that host them.
The Adversarial Case - A Profitable Quarter Is Not a Profitable Company
The strongest argument against the bullish read is straightforward: Anthropic is not a profitable company; it posted a profitable quarter. The company is estimated to have accumulated $10 billion to $15 billion in net losses since its founding, and it carries committed cloud spending through 2029 that analysts put at roughly $80 billion. A single profitable quarter against that backdrop is a milestone, not a business model.
The counter-thesis also has a mechanism. As inference becomes commoditized and open-weight models improve, per-token prices should fall toward marginal cost. Anthropic's margins depend on staying ahead of that curve with each new model generation. If a competitor - OpenAI, a well-funded open-source effort, or a hyperscaler's in-house model - matches capability at lower cost, Anthropic's pricing power erodes and the 5.1% margin compresses back into loss territory.
This counter-argument is not marginal. It attacks the core of the bullish thesis: that the current profitability is durable rather than a transient window created by a revenue spike meeting pre-committed capacity. It is backed by a mainstream voice in Eisman, and it is reinforced by the company's own acknowledgment of how capital-intensive the business remains.
The signal that would falsify the bullish view is concrete: if Anthropic's quarterly adjusted operating margin falls below zero for two consecutive quarters while annualized revenue growth drops below 30% year over year, the structural-profitability thesis is broken. A single bad quarter is noise; two, combined with decelerating top-line growth, would confirm that the profit was cyclical.
Outlook - Three Horizons and a Specific Watchlist
The mechanism cashes out into a clear asymmetry. In the short term, the profitability headlines support a strong IPO window and could let Anthropic price closer to the $2 trillion that some investors model. The beneficiaries are the company's existing backers and the enterprise-software peers that can point to Anthropic as proof that AI monetization works. The exposed are the private labs still running deep losses with consumer-heavy or training-heavy models - their next fundraising rounds will be underwritten against a new benchmark.
Over the medium term, the story turns on execution. The base case is that Anthropic stays profitable through the IPO, lands in October, and uses public currency to keep funding the training race while enterprise usage compounds. The upside case is that agentic adoption pushes annualized revenue toward the $100 billion to $120 billion that investors model for year-end, letting margins expand toward the 30% level that would make a $2 trillion valuation look merely expensive rather than absurd. The downside case is that customer optimization - the end of "tokenmaxxing" - slows revenue growth just as a new training cycle hits, pushing the company back into loss and forcing a valuation reset.
Split by horizon: short-term sentiment favors the bulls on the back of the profitability headlines; medium-term fundamentals depend on whether revenue growth outruns the committed compute bill; long-term structural durability depends on whether enterprise embedding creates a moat that survives commoditization of inference.
The watchlist is specific. First, the Q3 results: confirm the second straight profitable quarter and watch the margin, not just the revenue. Second, the IPO terms: a $2 trillion float would leave little room for a growth miss. Third, customer concentration: with two coding-tool customers having driven close to a quarter of revenue at one milestone, any loss of a major API buyer would show up quickly. Fourth, the price war: if Chinese open-weight models force per-token pricing down faster than Anthropic can cut its own costs, the margin story breaks regardless of revenue growth.
Anthropic has done something the AI industry said was years away. But a profitable quarter is not the same as a profitable company, and the difference between the two is where the next $2 trillion will be made or lost.
Explore more exclusive insights at nextfin.ai.
