Whoever sings down Anthropic may be disappointed

The problem isn't in this picture. This picture is very well done, and the data is probably done seriously. The problem is that almost everyone who retweeted it was using it to answer a question it couldn't answer at all.

01
Let's first figure out who made this picture

There is a Claude icon in the upper left corner. The color scheme is Claude's familiar orange. At first glance, it looks like an official product.

It's not.

The author of this picture isTickerTrends, My name is written in the upper right corner. It's a familyThird-party data tracking agenciesIt relies on various external signals (application data, payment panels, recruitment, channel caliber, etc.) to estimate the revenue of an unlisted company. The line in the picture is written very honestly: “tracked allocation” -- the percentage of allocation that has been tracked.

Let's be clear: Anthropic has never publicly disclosed Claude Code's individual ARR numbers, not once. Every point on this curve isEvaluated by outsiders。

What is more important is the caliber itself. ARR's algorithm is”Earnings for the most recent period × 12”. Enterprise software contracts are not executed evenly every day; they are signed batch by batch. Big orders signed at the end of a quarter will jump a week's curve by a large margin; if the next quarter's big orders aren't signed, the curve will go sideways. Weekly ARR tracking is extremely insensitive to this kind of massive landing—it will”The pace of signing” Draw”Changes in demand”.

02
This picture just punches myself in the face

Next, I haven't seen anyone mention this, but it's the most interesting part of the whole thing.

The picture says two numbers: Claude Code is$151.2 billion, accounting for Anthropic's total ARR21.9%。

Do a division:

And the official caliber numbers reported by Bloomberg, Reuters, and CNBC on August 17 were — end of July$650 billion。

Clear: This chart, which is being used to prove “slowing growth,” its own implied total number of companies is 4 billion US dollars higher than the official figure ten days ago.

Push back another ten days until today. If the trend continues,Over 70 billionIt's a reasonable estimate (this sentence is an inference, not data).

03
You're looking at the picture below. The money is in the picture above

That picture has two blocks above and below. Above is the absolute amount (how many billion dollars), and below is the percentage change (how much more than a percent increase from four weeks ago).

The vast majority of people's reasoning is that they saw 85% drop to 5.2% on the picture below, and “it collapsed,” then press this conclusion back to the picture above and say, “Look, the columns are all the same height.”

Count the numbers first. 5.2% × $15.12 billion = additional annualized revenue of approximately $750 million in one month. However, in the 85% month of March 2025, the base figure was less than 100 million US dollars — in the “outbreak” month, the actual increase was only in today's “lackluster” monthtenthsleft and right.

Percentages are fractions, and the denominator keeps getting bigger. An 85% increase of 100 yuan was 85 yuan, and a 5.2% increase of 15,120 yuan was 786 yuan. The latter is more than nine times the former. The “collapse” you see in the picture below was “nine times more profitable” in terms of money.

But Alan wants to be complete; otherwise, he's just kidding:Definitely newIndeed, from the peakRetreatThat's it.

In April and May, Claude Code rushed from 5 billion to 13.5 billion, with an average monthly increase of 4 billion; by July, there was only 300 million left. A decline of more than five times cannot be fooled by “you misread the percentages”.

So the real question isn't “is there any deceleration” — yes. There are really two things to ask:Why slow down?And why is it crooked again?

Please take another look at the far right of the image. The last pillar jumped up, and the line below also started2%It's back5.2%。A curve that is picking up at the right end was read as leveling off.

04
After three weeks of leveling off, the model went offline

If you take out the calendar and check it right, the answer is on the face; it's not mysterious at all.

On June 9, Anthropic made Mythos-level models publicly available and called Fable 5 and Mythos 5. On June 12, in order to comply with the US Department of Commerce'sexport control order, both models are offline. The Ministry of Commerce lifted the control on June 30 and resumed on July 1.

That so-called “”Platform period“, I just got stuck in this hole for three weeks.

What's more fundamental than these three weeksproduction capacity. Anthropic publicly acknowledged in April that demand was putting pressure on infrastructure and impacting reliability and performance during peak periods. And the vast majority of the computing power agreements it has signed with Amazon, Google, Nvidia, and Microsoft will arriveEnd of 2026或 Early 2027It just went online.

05
Percentage of people who can talk: 17% increased to 22%

If Claude Code actually stalls, its share of the company's revenue should drop. The opposite is true.

In what is known as”Platform period“In that window, Claude Code's share of the company's revenue increased from 17% to 22%. And in the same window, the total number of companies increased by about$180 billion。

06
Compared to the internet? The direction is completely reversed

Some people in the group said, “Internet users weren't that slow back then.” There are three mistakes in this sentence; they all go in the opposite direction.

First, the speed ratio is reversed

There were around 16 million internet users worldwide in 1995 and 360 million in 2000. Twenty-three times in five years, annualized about 86%. Claude Code isSix times in six months. Put it on the same ruler, it's more than an order of magnitude faster than the internet became popular.

Another harsher reference: Salesforce spent about 20 years generating $30 billion in annual revenue. Anthropic started with the first dollar in less than three years.

Second, the shape of the S curve is reversed

The percentage growth rate of internet user growth declined from day one. From 1 million subscribers in 1994 to 23 million in 2000, AOL's monthly growth rate has been declining, while net monthly growth has continued to rise. All S curves look like this:Percentage decrease, absolute value increase。

Third, the dimensions are wrong

The number of users is a variable. Revenue isSeats × usage × unit priceThe three variables are multiplied.

07
What you can't see in the picture: it's actually getting stronger

Anthropic hasn't just been “not that bad” this half year. It actually got stronger everywhere where it couldn't be taken in this picture at all.

1. Coming out of the “money burning” narrative

Q2 After adjustmentOperating profit is positive. In cutting-edge model labs, this is rare. The core doubt about AI companies in the past two years was “revenue was bought with money, and the unit economy was not established”. A positive operating profit figure reverses the burden of proof on the skeptical party. Moreover, this was done in half a year when computing power was tight and the most expensive.

2. Stop standing on one leg

Claude Code accounts for only 22%, and another 78%APIs and enterprise business. The market refers to Anthropic as “the company that makes programming tools,” but three-quarters of its revenue doesn't come from that tool. This means that even if one day competition among programming agents cuts Claude Code's share in half, the company won't collapse. However, today's pricing in the market is largely still tied to the “programming” story.

3. Customers are moving forward

Annual expensesOver $1 millionCustomers, from a dozen two years ago to more than 500, toMore than 1000 houses; Annual costOver $10 millionThe number of customers increased by about a year7 times. Major customers don't rely on sales to renew their fees, and they can't be pulled out once embedded in a workflow. This is a net revenue retention machine, and it is also the root reason why such companies can sustain high multiples even after the growth rate naturally declines.

4. Relative position changed

Over the same period of time: OpenAI doubled from about 20 billion to about 40 billion; Anthropic from 9 billion to 65 billion, 7.2 times. The two companies may have different caliber, but there is no disagreement about the direction of the gap.

08
Where is the next wave and when will it come

The following is Alan's judgment and is not a company disclosure. Four points, with time window.

Tap 1 | Computing power online — Q4 2026 to Q1 2027

This is the most straightforward and computable one. If the June and July platform is indeed due to production capacity, thenProduction capacity launchedIt's the mechanical unlocking of revenue: multi-gigawatt Amazon deployments, Google and Broadcom's TPU production capacity, and the computing power deal with xAI/SpaceX all point to this window. The observation method is rudimentary but effective: see if the speed limit has been relaxed, and whether the new-tier models are no longer queuing up. Relaxation on the product side will tell you the answer two months earlier than any financial report.

Point 2 | Moving out of the developer — Starting in the second half of 2026, volume in 2027

There are approximately 30 million professional developers worldwide. Knowledge workers are counted in billions. There are two orders of magnitude apart in the middle. Claude Code completed a proof question:agentThere are really people willing to pay for this stuff, and they are willing to pay a lot. Next, the @Claude set in Cowork, Excel, PowerPoint, Word, Outlook, Chrome, and Slack sells this verified willingness to pay to anyone who doesn't write code. This curve must have a slower slope than Claude Code — adoption by non-developers is slower and the workflow is more fragmented. But its ceiling isn't on the same floor. This is also the only serious answer to “what to do if the developer penetrates to the top”.

Point 3 | Mythos tier pricing — the timing is yet to be determined, but it must be a step

presentlyMythos PreviewOpen to only a few trusted agencies (Project Glasswing), Fable 5 is a public version with safety measures added. When this tier is actually used on a large scale, each customer's revenue jumps one level because it doesn't sell the same thing. This level of schedule has already been interrupted once by export controls (that is, the three weeks of June), so it is alsoPolicy risksOne of the most concentrated points.

Point 4 | The listing itself — September-October 2026

A prospectus was confidentially submitted to the SEC in June, and Morgan Stanley, Goldman Sachs, and J.P. Morgan Chase are underwriting. The listing is not a revenue event; it isCapital eventsHowever, its revenue transmission chain is extremely short: the money raised is converted into computational power, and computing power is exchanged for revenue. Also, Anthropic has an opportunity to get ahead of OpenAI — in an industry where computing power is production capacity, get it firstBig capital poolIt is a pioneer that can be compounded.

09
Is $2 trillion expensive? Let's take a look at Hong Kong first

This is the section in full I want people to read the most.

If you want to judge whether 2 trillion is expensive or not, you can't compare yourself to yourself; you have to findtaxon. And the best of its kind isn't NASDAQ — it's in Hong Kong.

January 8 and 9, 2026,Intellectual spectrum(02513.HK) andMiniMax(00100.HK) rang the bell on the Hong Kong Stock Exchange for two consecutive days and became one of the world's first publicly traded big model companies. For the first time, they let the market see what price the secondary market is willing to pay for a pure model company.

Moreover, the number of intellects is not a wild path. Its market capitalization reached HK$691.5 billion in May 2026, then reached the HK$8800-900 billion range in June, and the Hang Seng Technology Index included it in May. According to its MaaS ARR of 1.7 billion yuan at the end of March, the market sales rate once reached 544 times; even using the 2026 revenue estimates of 7 billion yuan from Goldman Sachs and Morgan Stanley, it was still more than 130 times.

Let's expand on Anthropic's own caliber to see it more clearly:

Two trillion isn't repriced at once; it's the same multiple times a larger base. The May private placement round has already been given20.5 times. If ARR falls within investors' expectations by the end of the year, $2 trillion would be cheaper than the May round.

So to prove the bubble, it is necessary to argue that “the 20 to 30 times ARR multiplier itself is wrong” rather than “growth has stopped” — the latter has been directly rejected by the data. The difficulty of these two matters varies greatly: the former involves gross profit margin, retention, and the competitive landscape; the latter only requires reading the wrong picture. Most people chose the latter.

10
4 things that should really make a noise

None of the questions that can change the conclusion are on that picture.

gross profit margin。This is the only weighty point of contention. Reasoned costs sit in the cost category. If 29 times revenue corresponds to 50% gross profit, it is actually 58 times gross profit, and the location is not cheap right away. The reference system already exists: Zhi Spectrum's comprehensive gross profit margin of 41.0% in 2025, MiniMax 25.4% (of which 69.4% on the enterprise side and only 4.7% on the consumer side). The traditional SaaS structure of 75% to 85% has not been achieved by anyone in the big model industry at present.

Revenue caliber。ARR is the most recent month multiplied by twelve. If it is tightened to a more conservative approach before listing, there will be a “contraction” on the books; this is not a deterioration in business. The real hard figure is recorded revenue: $16.23 billion for the first half of the year.

Customer concentration and retention。6How much of the 5 billion miles comes from the top 20 customers? It will be in the prospectus. The impact of this figure on valuation exceeds all curve fluctuations over the next six months.

Developer ceiling。15.12 billion ÷ 30 million developers = $500 per person per year. This penetration isn't low anymore. Further, we must rely on the unit price, usage, or “getting out of the developer” mentioned in section 8.

11
Write at the end

In September or October, the prospectus will replace all speculation with Audited figures. Gross profit margin, retention, concentration, true caliber — everything that is making noise in the crowd today will become a table that can be checked by then.

Until then, using a weekly curve estimated by a third party to determine the fate of a $65 billion ARR company was using the weakest evidence at hand to answer the most serious questions.

That picture clearly shows only one thing: Claude Code started in six months2.5 billionIt grows until151 billionMeanwhile, due to an export control order, the model went offline for three weeks.

The rest are all mind-reading at the height of the pillar.

The diner arrived. See the prospectus for September.

Verification instructions

I. Verified (multiple mainstream media, August 17-18, 2026)

· Anthropic ARR surpassed $65 billion at the end of July; $47 billion in May; approximately $9 billion at the end of 2025. Bloomberg debut, independently confirmed by Reuters, CNBC, Axios, and TechCrunch.

· Initial revenue for Q2 exceeded $11.5 billion, or $787 million in the same period last year; Q1 was 4.73 billion, up more than 140% month-on-month; Q2 adjusted operating profit was positive (documents seen by Bloomberg). OpenAI's ARR was around $40 billion, and about $20 billion (Bloomberg) at the beginning of the year.

· Investors expect $100-120 billion (FT) by the end of 2026; the company's 2028 revenue forecast for investors is $1900-200 billion (Reuters, August 14).

· The S-1 was confidentially submitted in June 2026, underwritten by Morgan Stanley, Goldman Sachs, and J.P. Morgan Chase, and is expected to go public in September or October. Series H was completed on May 28 at a valuation of $965 billion.

· Fable 5 and Mythos 5 went live on June 9, went offline on June 12 due to the Ministry of Commerce's export control order, the controls were lifted on June 30, and resumed on July 1.

· MiniMax was listed on the Hong Kong Stock Exchange on January 8, 2026, and MiniMax on January 9; Smart Spectrum's 2025 revenue was 724 million yuan (+131.9%), consolidated gross profit margin of 41.0%, net loss of 4.718 billion yuan, adjusted net loss of 3.182 billion yuan; MiniMax 2025 revenue of $79.38 million (+158.9%), gross profit margin of 25.4%, and adjusted net loss of US$250.9 million.

· MiniMax reached a high of around HK$1,330 in March, and its market capitalization once exceeded HK$410 billion; on July 10, the market capitalization was about HK$84.242 billion. Smart Spectrum's market capitalization on July 10 was HK$731.18 billion.

II. Single Source/Estimate/ Unverified

· The $15.12 billion and 21.9% share in the chart are from TickerTrends, a third-party estimate (“tracked allocation” on the page), not officially disclosed by Anthropic. Anthropic has never separately disclosed Claude Code's ARR.

· “Approximately 69 billion company total ARR” introduced by 15.12 billion ÷ 21.9% is an implied value under this third party's caliber and is not an official figure. “Probably over 70 billion today” is the author's inference.

· Claude Code is approximately 8 billion in May 2026, from a single second-hand statistics site; in the table, the company's total ARR for November 2025 and February 2026 is estimated based on public nodes, and is an undisclosed value.

· Zhi Spectrum ARR “Approximately US$1 billion (July 2026)” came from several Chinese media quoting people familiar with the matter, and Zhi Spectrum did not respond to confirmation. The public ARR caliber that can be verified by Smart Spectrum was RMB 1.7 billion for the MaaS platform at the end of March.

· MiniMax's “annualized revenue of about 300 million US dollars” comes from media extrapolation of its ARR exceeding $150 million in February 2026, undisclosed by the company.

· HKD/USD is converted at 7.8. The Hong Kong stock market value is the July 10 data. It fluctuates greatly, and the actual figures may be significantly different when readers read it.

· The target valuation of the $2 trillion IPO is a market expectation quoted by the media and is not publicly stated by the company.

· Customers spending more than 1 million dollars a year “more than 1,000 companies” and customers over 100,000 US dollars “7 times a year” came from previous public statements from Sacra and the company, without checking the latest time.

· The amount of the computing power transaction with xAI/ SpaceX comes from media reports, and there is no original announcement from both parties.

3. Historical References (Public Records)

· Global Internet users were around 16 million in 1995 and 360 million in 2000; AOL subscribers were around 1 million in 1994 and 23 million in 2000. There are differences between different source calibers.

· It took about 20 years for Salesforce to reach approximately $30 billion in annual revenue.

4. Author's Inference (Untrue)

· The main reason for the June-July platform period was capacity constraints and export control shutdowns rather than peak demand.

· The four “next wave” points and time windows are all inferences; the computing power window depends on the publicly reported launch time of the agreement, which is not promised by the company.

· The comparison “The multiple given by the Hong Kong market indicates that 2 trillion dollars is not expensive” assumes that the three companies are comparable. In fact, business structures, customer regions, regulatory environments, and liquidity discounts and premiums are all different. Cross-market multiplier comparisons can only be used as a reference level; they cannot be used as a basis for pricing.

· Gross margin is the only weighty point of contention. 29 times revenue corresponds to an estimate of 58 times gross profit. For illustrative purposes, Anthropic did not disclose gross profit margin.

· The penetration estimate of about $500/year per developer uses the median value of “about 30 million professional developers worldwide”. The caliber itself is controversial.

5. Important Reminder

· ARR and recorded revenue are two different things. ARR is an annualized extrapolation of recent earnings, which amplifies both good and bad months.

· This article does not constitute investment advice. Anthropic is not currently listed; the stock prices of the Hong Kong stock companies mentioned in the article fluctuate drastically, and historical performance is not indicative of the future.