The Number That Matters Is Not $900 Billion

Anthropic reportedly hit operational profitability in Q2 2026, closed a $30 billion round at a $900 billion valuation, and overtook OpenAI's last private mark of $852 billion in the same week. The reaction has been about the valuation. That is the wrong number to fixate on. The interesting claim, the one that will either reshape how frontier AI gets financed or quietly evaporate when the S-1 lands, is that a lab actively training next-generation models is now making money on operations.

If that claim survives audit, the unit economics of frontier AI have inverted, and most of the analyst frameworks built in 2024 and 2025 are stale. If it doesn't, this round will look in hindsight like the high-water mark of pre-IPO storytelling. Either way, the $900 billion is the least informative figure in the press release.

What "operational profitability" would actually mean

The phrase is doing a lot of work, and it matters which definition is in play.

A frontier lab has three buckets of cost. There is inference, which is the cost of serving customers — GPUs running, tokens flowing, electricity bills arriving. There is training, which is the cost of building the next model — capex on compute, data, and the salaries of the people who push the frontier forward. And there is everything else: research overhead, go-to-market, the lawyers, the safety team, the building.

Operational profitability usually means revenue exceeds operating expenses, and operating expenses usually include current-period training spend. If Anthropic is genuinely operationally profitable while training Claude's next generation, then revenue from the deployed fleet is now large enough to fund both the cost of serving it and the cost of building its replacement. That is a structurally different business from the one OpenAI is reportedly running, where Ed Zitron's analysis of leaked figures puts the Q1 2026 non-GAAP operating margin at negative 122 percent on something like $12.7 billion of annualised revenue.¹

The two companies are not in the same financial regime. If both numbers are accurate, Anthropic is generating roughly 3.5x OpenAI's revenue at positive operating margin while OpenAI is burning more than two dollars for every dollar of revenue. That is not "Anthropic is winning" — that is a different kind of company entirely.

There are reasons to be careful. "Operational profitability" can be defined to exclude training capex, in which case the achievement is real but considerably less dramatic; it would mean Claude pays for Claude, but not for Claude-next. It can also be defined on a recent-quarter run-rate that flatters the trailing twelve months. The S-1 will resolve this. Until then, the right posture is: interested, sceptical, watching the definition.

The $45 billion question

The ARR figure is where the story gets uncomfortable.

Anthropic reportedly went from around $4 billion ARR in early 2025 to $45 billion now. That is roughly 10x in 12–15 months. For context, the fastest enterprise software ramps of the SaaS era, Snowflake, Datadog, the ServiceNow run from $1B to $4B, moved an order of magnitude over three to four years, not one. A 10x in a year is not a software growth rate. It is either a category-defining business event or an accounting artefact.

There are at least three ways the figure could be defensible and still misleading.

First, ARR at most private AI labs is a composite of API consumption (which annualises poorly because it can spike and collapse), enterprise contract value (which can be booked as ARR on signature even if usage hasn't started), and committed minimums (which annualise into ARR but may never be consumed). A clean GAAP revenue number for Anthropic in calendar 2025 would tell us how much of the $45B is actually flowing. We don't have that number.

Second, annualising a strong recent quarter can produce big ARR jumps that don't survive contact with the next quarter. If Anthropic shipped a major model upgrade in Q1 and saw a usage spike, multiplying Q1 by four overstates the durable revenue base.

Third, the largest enterprise contracts at this scale, and KPMG with 276,000 employees plus PwC are very large enterprise contracts, tend to be structured with ramps. Year one is a fraction of the headline TCV. Counting the headline as ARR is common practice and not dishonest; it is also not the same as cash.

The interesting claim is not the valuation. It is that a lab actively training next-generation models is now making money on operations.

None of this means the number is wrong. It means the number is load-bearing, unverified, and about to be audited in front of public-market investors. Analyst scepticism has been mentioned in the reporting but no analyst has put their name to the dispute on record, which itself is telling. Either nobody wants to be wrong in public against a company about to IPO at $900B, or nobody has clean data to be confidently right with.

The inference-economics frame, and why it might be the answer

My standing prior on this kind of question is that inference economics, the actual margin on serving tokens, is where the truth lives, and that headline ARR figures mislead because they don't tell you what the cost of revenue looks like.

The frame is worth testing here. If Anthropic is operationally profitable at $45B claimed ARR, there are two stories that fit the data.

The first: inference margin has improved dramatically. Claude's serving costs per token have come down through some combination of model distillation, hardware efficiency, and Anthropic's deep AWS Trainium partnership (Amazon has committed up to $4 billion, mostly in compute credits and infrastructure). At sufficient scale, gross margin on inference can plausibly exceed 60 percent — high enough that the revenue base funds training spend.

The second: enterprise contract structure is front-loading cash. Large committed deals with KPMG, PwC and others may carry significant upfront payments, multi-year prepayments, or aggressive minimum commits that show up as revenue or deferred revenue and look operationally healthy in the current period even if the underlying consumption is lower.

Both stories can be partially true. The S-1 will disclose deferred revenue, RPO (remaining performance obligations), and the split between consumption and committed revenue. That disclosure will tell us which story dominates. My honest read, holding the prior loosely: it is probably more story one than most sceptics assume, because the Trainium relationship is real and the inference cost curve has bent harder than the public narrative gives it credit for, but story two is doing meaningful work and the audited gross margin will be lower than the operational-profitability claim implies.

Valuation inversion is a positioning move, not a financial fact

Anthropic at $900B versus OpenAI at $852B has been written about as if it were a market verdict. It is not. Private valuations at this stage are set by the lead investor of the most recent round, on terms that include liquidation preferences, ratchets, and information rights that public-market comparables do not carry. The $900B figure is the price someone agreed to pay for a specific slice of equity with specific protections. It is not a market clearing price.

This matters because the inversion will be read by three different audiences in three different ways.

Enterprise buyers will mostly not care. The CIO at a Fortune 100 bank choosing between Claude and GPT for a fraud detection workflow is not running a comparison of private valuations; she is running a comparison of model performance on her data, latency, governance posture, and contractual liability terms. The valuation inversion is irrelevant to that decision, and the KPMG and PwC wins suggest Anthropic is competing, and winning, on the dimensions that matter to that buyer. The valuation is a vanity datum from her perspective.

Regulators will care, but slowly. A frontier lab approaching trillion-dollar private valuation creates pressure on the FTC, on the AI Office in Brussels, on the UK's CMA. None of those bodies move at the speed of a funding round, but the political surface area Anthropic now carries is meaningfully larger than it was a month ago, and the safety-focused founding narrative becomes harder to sustain when you are the most valuable private company in the world.

IPO underwriters will care a great deal, and they will discount the $900B aggressively. Public-market comparables — the closest being Nvidia on the supply side and the hyperscalers on the demand side — trade on multiples of audited GAAP revenue, free cash flow, and forward growth. The $900B will not survive the underwriter's spreadsheet intact unless the audited financials are remarkable. They might be. The point is that the private mark is a starting position in a negotiation, not a finishing one.

Private valuations set in late-stage rounds reflect negotiated terms and investor strategy, not independent market pricing. The $900B is a starting position, not a finishing one.

The interesting question is not whether Anthropic is worth more than OpenAI. It is whether either company is worth what their last round implies, and on that question the honest answer is: we will find out, in public, within twelve months.

Why KPMG and PwC are the actually important news

If I had to pick one item from this week's reporting that will matter most in three years, it would not be the valuation, the ARR, or even the operational profitability claim. It would be the bookending of Anthropic's enterprise strategy with KPMG and PwC.

The professional services firms are the distribution layer for enterprise AI in the way that systems integrators were the distribution layer for enterprise software in the 1990s and the cloud consultancies were for IaaS in the 2010s. They are how the Global 2000 actually buys, deploys, and absorbs new technology. Together, the Big Four employ more than 1.5 million people, most of them in some flavour of knowledge work that current-generation AI can plausibly transform within the next deployment cycle.

Embedding Claude into KPMG's audit workflows, PwC's advisory practice, and the consulting work that sits beneath both is not a feature sale. It is a structural claim on the workflow itself. If Claude becomes the model that KPMG audit partners use to review work papers, and the model that PwC consultants use to draft client deliverables, then every Fortune 1000 client of those firms is touching Claude indirectly — and the procurement path for direct adoption gets dramatically shorter.

This is also where my prior on professional services collides with the evidence and partially loses. My standing position has been that professional services sit at the point of maximum risk and maximum opportunity simultaneously — that the firms whose business model is selling time will either reinvent themselves around agentic delivery or be hollowed out by it. The Anthropic deals suggest the firms have decided which side of that trade they want to be on. They are not waiting to be disrupted; they are buying the disruptor and embedding it. Whether they can execute on that — whether KPMG and PwC can actually rewire their delivery model around AI faster than their juniors get redundant or their clients route around them — is the genuinely open question.

The counter-case is real and worth holding. Professional services firms have historically been slow to deploy enterprise software at scale, and the announced deal is not the deployed deal. Liability concerns around AI-generated audit work papers are nontrivial; the moment a regulator finds a material misstatement that traces back to a Claude output, the rollout pauses. The history of Big Four technology adoption is littered with announced partnerships that produced press releases and not much else. So the deals are necessary but not sufficient, and the question of whether they convert to durable ARR, and to actual workflow transformation, is open.

But the strategic frame is clear. Anthropic has chosen distribution through the professional services layer. OpenAI's strategy, by contrast, has been more direct: consumer first via ChatGPT, then enterprise through Microsoft's distribution. Those are two genuinely different bets on how enterprise AI gets sold. We will know which works inside three years.

The contrarian read

There is one read of this week's news that cuts against everything else, and I want to hold it openly because the rest of this piece can sound celebratory and I do not want that.

Operational profitability at a frontier lab can mean revenue has scaled past costs. It can also mean training spend has plateaued. If Anthropic has decided, explicitly or implicitly, to harvest Claude 4 longer than originally planned, to push the next major model further out, or to scale back the size of the next training run, then operational profitability becomes much easier to achieve. It also means competitive position erodes against any lab that keeps pushing capability.

I do not think this is what is happening. The compute commitments from Amazon and Google are too large, the hiring at Anthropic has not slowed, and the public posture from Dario Amodei on capability progress has not shifted. But the possibility is worth naming, because the IPO incentive structure rewards profitability narratives, and there is a version of this story where Anthropic is optimising for the listing rather than for the frontier.

If the next Claude generation arrives later than the OpenAI and Google equivalents, and falls behind on benchmarks that enterprise buyers care about, the operational profitability claim will age differently than it reads today. That is the scenario worth watching for over the next six months.

What to watch

Four things will resolve most of the ambiguity in this story.

The IPO registration statement, if it comes in late 2026 as signalled, will give us audited GAAP revenue, gross margin, deferred revenue, and RPO. The delta between the audited figures and the $45B ARR claim will be the most important number in AI finance this decade, and I mean that without hyperbole. If audited revenue is within 20 percent of the claim, Anthropic is the real thing and the valuation will hold. If it is 50 percent below, the round will look very different in retrospect.

The KPMG and PwC deployment cadence. Watch for actual seat counts, actual workflow integrations, and actual revenue recognition from those contracts over the next four quarters. Announced deals matter less than deployed deals.

OpenAI's response, which will probably involve some combination of a counter-round, an aggressive enterprise pricing move, and a renewed push on consumer monetisation through ChatGPT. The negative-122-percent operating margin is not sustainable, and OpenAI's path to financial credibility now runs through either a dramatic revenue ramp or a dramatic cost cut, and probably both.

Claude's next major release. If it lands on time and lands ahead of GPT and Gemini on the benchmarks enterprise buyers track, the operational profitability story is structurally sound. If it slips or underperforms, the contrarian read above gets more weight.

For now, the honest summary is this. Anthropic has either built the first financially sustainable frontier AI lab, in which case the entire industry narrative needs rewriting, or it has prepared a very effective pre-IPO story. The S-1 will tell us which. I would not bet against the first reading — the inference economics genuinely look like they have bent in Anthropic's favour, the enterprise distribution is real, and the Trainium relationship matters more than the public narrative gives it credit for — but I would not stake more than I could afford to lose on it either. Both readings remain live. The $900 billion is, in the meantime, the wrong thing to argue about.

---

Footnotes

Footnotes

1. Ed Zitron, "OpenAI Had A Negative 122% Non-GAAP Operating Margin In Q1 2026", Where's Your Ed At, 2026. https://www.wheresyoured.at/news-openai-had-a-negative-122-operating-margin-in-q1-2026-and-chatgpt-growth-has-stalled ↩

¶