Anthropic (the company that makes Claude, a chat assistant you type to) told the public on November 2, 2026 that it sold about $16.96 billion of its product in the three months ending September 30. In the same document it reported a net loss (the bottom-line loss after every accounting entry, including ones that are not cash) of about $23.29 billion, and it wrote down a compute commitments: a promise to spend about $895.4 billion on computers and buildings, mostly between now and 2045.
Figures below were checked against Anthropic’s amended Form S-1/A HTML on sec.gov (accession 0001628280-26-068676) on Nov 9, 2026 Asia/Calcutta.
Figures below were checked against Anthropic’s amended Form S-1/A HTML on sec.gov (accession 0001628280-26-068676) on Nov 9, 2026 Asia/Calcutta. Those three numbers are being passed around as if they were the same kind of fact. They are not. One is sales. One is mostly a paper mark on an old loan. One is a stack of contracts, and only a slice of that stack is due soon. If you remember only one thing, remember that split.
The receipt, before the vocabulary
Related reading on this site: Anthropic cut the internet off its own tests and why OpenAI’s safety writer quit. For agent risk outside the model, see the cage around the tool.
A company is a machine that takes in money from customers and pays money out to suppliers, staff, and landlords. Profit (money left after the costs you count) is just what is left. A loss is the opposite. The trouble is that accountants also write down changes in the value of pieces of paper, and those changes can dwarf the cash.
Here is the quarter, in the order a reader should meet it. Sources are at the bottom. Where two outlets agree, this write-up uses the tighter figure.
| What people repeat | What it actually is | Did cash leave for this line? |
|---|---|---|
| $16.96 billion of revenue (sales) in the quarter | Customers paid, or owed, that much for Claude and related products | This is money coming in, not going out |
| $567 million operating loss (revenue minus the cost of running the business) | $16.956 billion of sales minus $17.523 billion of operating costs | This is the real quarter for the business. It is a loss, and it is not $23 billion |
| $2.46 billion adjusted operating income | The operating loss, with stock pay and a non-cash donation match added back | A management view. Not a bank balance |
| $23.29 billion net loss | The operating result, plus a $21.94 billion paper loss on convertible notes (loans that can turn into shares) | No. About $21.9 billion of it is a revaluation |
| $6.6 billion of operating cash flow over nine months | Cash the operations produced from January through September | This one is cash, and it came in |
| $895.4 billion of commitments | Signed promises to pay for compute (rented thinking-machines) and buildings, out to about 2045 | Not yet. About $69.8 billion is the next 12 months |
The arithmetic on the adjusted line is the part most roundups skip, so here it is in public:
$$
-567 + 1,070 + 1,960 = 2,463
$$
Read it as millions of dollars. Start from the $567 million operating loss. Add back $1,070 million of stock-based compensation (pay delivered as shares, which does not write a check today). Add back $1,960 million of non-cash charitable charges (a donation-matching entry that PPC Land says did not move cash this quarter). You land on $2,463 million, which is the adjusted operating income they report, about 15 cents of that kind of profit per dollar of sales. Last quarter that adjusted margin (profit as a share of sales) was about 19 percent. The business got less efficient even while sales jumped.
What an S-1 is, then what this one changed
A private company (one whose shares are not for sale on a public exchange) can decide to sell shares to the public. That first public sale is an IPO (initial public offering). Before it happens, the company files a Form S-1 with the SEC (the U.S. Securities and Exchange Commission, the agency that polices stock sales). An S-1 is a long confession: how you make money, what can go wrong, and the numbers. An S-1/A is an amendment, a corrected or completed version of that confession.
Anthropic filed the first version on October 26, 2026. Several big blanks were still empty, including the total of its compute promises and the third-quarter numbers. The November 2 amendment filled those blanks. Coverage of the October file said the planned exchange is Nasdaq and the planned ticker (the short trading symbol) is ANT. The November coverage reviewed here did not restate that ticker; treat “ANT” as a plan from the filing, not as a confirmed listing ticker.
The company on the cover is Anthropic, PBC. A PBC (public benefit corporation) is an ordinary for-profit company whose charter also names a public mission. It is not a charity. It can still sell shares, pay staff in stock, and lose money.
Where the $23 billion goes, if it is not cash
A convertible note is a loan with a switch. The lender gives the company cash. Later, instead of only getting the cash back, the lender may turn the loan into shares. If the company becomes more valuable, that switch becomes more valuable too, even when nobody has flipped it.
Amazon holds most of these notes. PPC Land’s reading of the amendment puts the face amount (the cash originally owed) of Amazon’s notes at about $6.2 billion, and the accounting value of those same notes at about $130.5 billion on September 30. The gap is not new cash Amazon wired this quarter. It is the market’s implied price of the right to become a shareholder.
When that implied price rises, the company records a loss. That sounds backwards. It is not a judgment that the product got worse. It is the rule for this kind of contract: a more valuable conversion right is a more expensive obligation, on paper. In the third quarter that paper loss was $21.94 billion. Reuters’ net-loss figure of $23.29 billion is then mostly that entry, plus the much smaller operating loss, interest, and tax. Reuters also says the net loss shrank from $66.3 billion in the second quarter. The loss got “better” because the paper mark got less violent, not because the product suddenly cost less to run.
Cash on hand is the other side of that sentence. At September 30 the amendment, as read by PPC Land, shows about $95.6 billion of cash and about $36.3 billion of short-term investments (safe IOUs that turn back into cash quickly). Add a revolving credit facility (a pre-agreed loan you can draw, repay, and draw again) and the same reading puts total liquidity (money you can reach) near $146.6 billion. This is not a company that runs out of cash next Tuesday. It is a company that has promised to keep paying landlords and chip suppliers for a generation.
What “compute” is, from the wall socket
Before the dollar figure, the object. A model (the file that answers you) is a giant list of numbers. To produce those numbers, and later to use them, you do an absurd amount of multiplication. The chips that are good at that multiplication are accelerators (specialized processors; Nvidia’s are the famous kind, and Google’s are called TPUs, tensor processing units). They sit on boards next to stacked memory (the scratchpad that has to feed the chip fast enough that the chip is not waiting). The boards sit in racks. The racks sit in a data center (a warehouse built around power and cooling, not around people). The warehouse sits next to a substation (the fenced yard of transformers that steps grid power down to something the building can drink).
So “compute” in this filing is not a metaphor. It is electricity, plus the machines that turn electricity into multiplications, plus the rent on the building, plus the right to keep doing it for years.
!
!
There are two jobs for those machines. Training (building a new model, a one-time construction project that can cost a fortune) and inference (using a finished model to answer one person, over and over). A chat reply is inference. A new Claude is training. Anthropic’s customers pay mostly for inference and for the tools around it. Anthropic itself also burns compute to build the next model. The filing does not split the $895.4 billion into those two jobs. When someone on social media does, they are guessing.
Why this quarter’s computers got more expensive
Sales were $11.56 billion in the second quarter and $16.96 billion in the third. That is a 47 percent rise in three months. Reuters notes the previous quarter had risen 144 percent, so the pace slowed. Slowed from what? From a pace that could not last. A 47 percent quarter is still a company doubling in well under a year, if it repeated, which it may not.
The cost of the machines did not slow with it. TechCrunch, on November 3, reported that compute ate 58 percent of revenue in the new quarter, up from 48 percent the quarter before, and that the compute bill was about $10 billion. PPC Land’s line-item figure is $9.863 billion, which is the same story in tighter digits. Inside that bill, $5.0 billion was short-term compute (capacity rented on contracts you can cancel within a year, or that last a year or less). The company says that short-term capacity “currently costs significantly more” than the same work bought under multi-year deals.
PPC Land does the useful subtraction. Take the $5.0 billion back out and compute would have been about 29 percent of revenue, not 58 percent. The quarter did not reveal a permanent law that Claude eats 58 cents of every dollar. It revealed a company that ran out of cheap, pre-booked capacity and paid the spot price to keep customers served. The filing, in that same reading, says Anthropic expects short-term compute to keep squeezing margins.
One more quiet line. Price cuts from June 30 to September 30 offset about $600 million of revenue. Customers used more, but each unit was cheaper, so the sales number understates the usage jump. That is good for the people paying the bill. It is awkward for a company that has already promised to pay suppliers at yesterday’s capacity plan.
The $895.4 billion, taken apart
The sentence in the amendment, as quoted by PPC Land:
> As of and subsequent to September 30, 2026, we have entered into purchase commitments and lease agreements, including short-term commitments, as well as leases not yet commenced, with an aggregate committed spend of approximately $895.4 billion.
“Substantially all” of it is compute and infrastructure. “Not yet commenced” means the lease is signed and the building or the cluster may not be switched on. You still owe it when it starts. This is not one check for $895.4 billion. It is a pile of contracts. PPC Land splits the pile into the five pieces the liquidity section describes.
| Piece | Amount | Plain meaning |
|---|---|---|
| Contractual obligations already on the books at September 30 | $320.0 billion | Leases and purchase promises the accountants already list. Most of this is purchase commitments, about $314 billion |
| Short-term commitments at September 30 | $17.4 billion | Cancelable within a year, or lasting a year or less. The expensive kind from this quarter lives near here |
| Leases signed but not yet started | $451.2 billion | The largest piece. Expected to start in 2026 to 2028, with a weighted average life of about 10 years |
| New commitments after September 30 | $100.2 billion | Signed in the weeks after the quarter ended, mostly more buildings and cloud hosting |
| New short-term deals after September 30 | $6.6 billion | More of the pricey, short kind, signed after the quarter closed |
| Total they name | $895.4 billion | About $447.5 billion of the cash is expected before December 31, 2030. Most of the rest sits in 2031 to 2045. About $69.8 billion falls in the 12 months after September 30, 2026 |
Two doors are hidden in that total, and they are the reason a serious reader does not treat $895.4 billion as a single due date.
First, about $213.8 billion is subject to termination rights or conditions (financing, buildings, approvals). Some of the promise can die if a condition fails. It is still a promise. It is not the same as rent you already owe a landlord who has handed you the keys.
Second, the $895.4 billion leaves out deals that have no fixed dollar amount. PPC Land puts that excluded range at about $150 billion to $175 billion, including a Google TPU arrangement it reads as $125 billion to $150 billion and an AMD Helios arrangement above $20 billion. Add them and the loose total is about $1.05 trillion to $1.07 trillion. This excluded range is treated as single-source and left out of the main argument. Reuters, TechCrunch, CNBC, and Bloomberg’s reported totals all repeat $895.4 billion. They do not all repeat the extra trillion. Use $895.4 billion as the company’s own aggregate. Mention the extra only as “deals with no fixed price, which one close reading says could add $150 billion to $175 billion.”
Google’s named cloud commitment, in that same close reading, moved from about $114.5 billion in the October prospectus to about $121.3 billion. That is a single-source delta. The direction is what matters: the blank that was filled in November was filled upward.
The trap in one sentence the company already wrote
A consumption contract (you pay for what you use this month, the way you pay for electricity) is not a lease (you pay whether you fill the building or not). EO Magazine, reading the October prospectus, quoted Anthropic on the mismatch:
> our revenue is primarily consumption-based, and much of it is not subject to long-term contractual commitments, while our compute and infrastructure commitments generally extend over multiple years.
The November numbers make that sentence concrete. Of the $33.24 billion of revenue in the first nine months of 2026, about 79 percent ($26.31 billion) was consumption and about 21 percent ($6.94 billion) was subscription (a set fee for a period of access). The nine-month sales total is up from roughly $2.5 billion in the same stretch of 2025. The year-ago quarter that PPC Land uses is $1.382 billion, against $16.956 billion this year, about twelve times. The growth is real. The shape of the growth is the risk: customers can shrink a usage bill in a month, and Anthropic cannot shrink a 10-year lease in a month.
How big are the customers who would have to keep growing? At September 30, more than 11,900 customers had trailing-twelve-month spend (the last year of their payments, not a forecast) above $100,000. More than 1,900 were above $1 million. About 250 were above $10 million. The top five customers were about a quarter of nine-month revenue. Concentration (a few buyers being a large share of sales) means two procurement teams can move the quarter.
Net revenue retention (what last year’s customers spend now, divided by what those same customers spent then, ignoring brand-new logos) was about 900 percent. Work a toy example before you trust the percent. If those customers paid $1 a year ago and pay $9 now, retention is 900 percent. It does not mean the company grew 900 percent, and it does not mean the next year will 9x again. It means the accounts that already existed got enormous, which is what happens when a tool goes from a pilot to the default. Pilots do not 9x twice.
Headcount was about 5,300. Trailing-twelve-month revenue was about $35.3 billion. Divide, and you get about $6.6 million of revenue per employee. That number is not evidence that each person is a miracle. The factory is the rented power plant. The people steer it.
Recursive self-improvement, from a file of numbers
A model does not “think” in the way a person does. It is a file of weights (the numbers that say how strongly one internal signal should nudge another). Training is the long job of nudging those numbers until the file is good at a task. Research is the job of deciding which nudges to try next: which data, which code, which experiment.
People already use today’s models to write that code and to run those experiments. That is ordinary tool use, the way a compiler (a program that turns human-readable code into something a chip can run) is tool use. Recursive self-improvement (a loop in which the tool gets good enough at making the next tool that progress speeds up by itself) is the claim that this stops being a convenience and becomes the pace-setter. Each new file helps build a stronger file, in less time, and the safety work (figuring out what the new file can do, and how it can fail) has to finish inside that shorter time.
Anthropic’s own name for the guardrail is its Responsible Scaling Policy (a written set of “if the model gets this capable, we will pause or add these protections” rules). The amendment, as quoted by PPC Land from the business section, says:
> Using AI to help build more capable AI could also pose a considerable safety risk, particularly if it sharply accelerates the pace of progress. This is one of the areas covered by our Responsible Scaling Policy, which sets out our plans to monitor and mitigate the risks as we approach what has been termed “recursive self-improvement.”
The November language, again from that reading and matched by TechCrunch’s reporting as summarized by The Neuron, draws a line that headlines erase:
> However, we do not anticipate being able to safely use Claude to fully autonomously design and build successive generations of frontier models without meaningful human supervision or understanding, if and when that becomes technically possible. Accordingly, we are considering updates to our Responsible Scaling Policy to keep humans in control of frontier model development at Anthropic.
A frontier model is simply the most capable model a lab is willing to ship or train, the one at the edge of what anyone has measured. “Fully autonomously” means without people in the loop. Anthropic is saying two things at once. Current models have not crossed its threshold for dramatically accelerated progress. Future ones may get close enough that the threshold matters operationally, not as science fiction. And even if a later Claude could design the one after it, they do not expect to let it do that alone.
That is not “the AI woke up.” It is “the research department is starting to be made of the product, and we are not willing to hand it the keys.” It is also a risk factor in a stock prospectus, which means the lawyers wanted it on the record before strangers buy shares.
Separately, The Information reported on November 3 that Anthropic has told investors a 15-person chip-design team is already using Claude for engineering work on its own chips. That is an anecdote about the loop starting in a narrow, supervised way. It is not a line from the financial tables, and it is not evidence that the threshold has been crossed.
What to watch next
- The next 12 months of the promise, about $69.8 billion, against quarterly revenue. At a $17 billion quarter, a year of sales is on the order of $70 billion only if this pace holds. The compute bill for those same 12 months can eat that entire year. That is the comparison. Not the $23 billion paper loss.
- Whether short-term compute shrinks as the pre-bought buildings turn on. If the 58 percent falls back toward the high 20s or 30s, this quarter was a scramble. If it stays near 58 percent after the long contracts start, the product is structurally expensive.
- Net revenue retention the next time they print it. 900 percent is a one-time stretch from pilots to deployment. The second reading tells you if the land-and-expand story (a small team buys it, then the whole company buys it) is still running.
- Any update to the Responsible Scaling Policy. They said they are considering one, specifically to keep humans in control of building the next frontier model. The text of that update matters more than the adjective “imminent.”
- How much of the $213.8 billion with escape hatches actually remains after the conditions are tested. A cancelled lease is a different company from a company that pays all of it.
Skip the IPO valuation gossip. Posts on November 3 guessed $1.5 trillion to $2 trillion and a November debut. None of that is a number the amendment has to state, and this draft does not use it.
Common questions about Anthropic compute commitments
Is the $895.4 billion a cash bill due this year?
No. It is an aggregate of purchase commitments and leases, including some not yet started. About $69.8 billion falls in the 12 months after September 30, 2026, and about $213.8 billion is subject to termination rights or pre-start conditions.
Did Anthropic “burn” $23 billion in the quarter?
No. Operating loss was $567 million. The roughly $23.29 billion net loss is mostly non-cash fair-value marks on convertible notes and related items. Nine-month operating cash flow was about $6.6 billion.
Where does the $895.4 billion figure come from?
Anthropic’s amended Form S-1 (S-1/A), filed with the SEC, states aggregate committed spend of approximately $895.4 billion as of and subsequent to September 30, 2026, substantially all related to compute and infrastructure.
