Anthropic’s S-1 Puts $52.9 Billion More on Three Named Compute Commitments Than Its October 9 Draft Did, and Still Leaves Its Contractual-Obligations Table Blank

On Monday, October 26, Anthropic filed its S-1 registration statement with the SEC. The coverage that day went to the headline numbers and the risk language: TechCrunch reported $16.3 billion of first-half revenue and an $80.7 billion net loss, The Verge listed “five weird details,” and WIRED read the risk factors as a doomsday primer. This post takes the part of the filing that sits between those stories: how much compute Anthropic has promised to pay for, how that figure has moved, and how firm each promise is.

EDGAR also now hosts the six draft registration statements Anthropic submitted confidentially, filed on June 1, July 13, August 10, September 4, October 1 and October 9. I read all seven documents, compared their commitment passages, and checked the result against the October 26 coverage and against our earlier reporting on Nscale and Akamai. Four things came out of it:

• Between the October 9 draft and the October 26 S-1, the figures for Microsoft, Google and SpaceXAI rose by a combined $52.9 billion, and SpaceXAI’s ceiling has gone from $44.4 billion in the July 13 draft to $120.5 billion.
• The discussion section’s contractual-obligations table is blank in all seven documents. The notes print $54.6 billion at December 31, 2025, about $417 billion of new and amended agreements through June 1, $265.6 billion at June 30 and about $245 billion after June 30, and the filing never adds them up.
• The eight arrangements the discussion section names are five different kinds of promise, from non-cancelable floors to a contract either side can end on 90 days’ notice, and suppliers stand behind them with $77.0 billion of undrawn convertible-note facilities.
• Fair-value entries on convertible notes, derivative liabilities, SAFEs and the note purchase agreements add up to $81.1 billion, more than the whole $80.7 billion first-half loss.

This is a first public filing with blanks in it, so an amendment can change any figure below; every number here is as filed on the date shown.

Seven documents, one ledger

The table shows each named commitment as the discussion section stated it in each document, in billions of dollars. “Not stated” means the document does not give the figure, and “not named” means the counterparty does not appear in it. The June 1 draft’s discussion section names none of them, and the July 13 amounts for Amazon and Google predate the “at least” wording, as explained below the table.

The last column is not a restatement. Microsoft’s agreement, described in the October 9 draft as amended in February 2026, is described in the S-1 as amended “in February and October 2026,” and the commitment moves from $31.4 billion to $44.9 billion. Google’s moves from about $111.1 billion to about $114.5 billion, with amendments now running “between May and August 2026” instead of in May. SpaceXAI’s moves from up to $84.5 billion to up to $120.5 billion, its monthly rate from about $2.37 billion to about $3.1 billion, and its window from “May and July 2026” to “Between May and October 2026.” The three increases add to $52.9 billion, my sum. The S-1 also renames the counterparty: the drafts say X.AI LLC (“x.AI”) and the S-1 says SpaceX.AI LLC (formerly known as X.AI LLC) (“SpaceXAI”).

Three further points. Drafts lag contracts: the S-1 dates a Google amendment as late as August, which the October 9 draft did not mention, so a change dated no later than August reached the text only after October 9. The take-or-pay wording is itself a later addition: “regardless of actual usage” and the true-up payments appear in the August 10 draft and every later one, but not in the June 1 or July 13 drafts, which say only that Anthropic “committed to spend” the amounts and give Amazon’s as “an aggregate of $110.0 billion” and Google’s as “an aggregate of approximately $111.1 billion” without “at least.” And the unchanged lines are unchanged to the decimal: Amazon at $110.0 billion since July 13, Nscale at $44.6 billion and Lambda at $41.3 billion since the October 1 draft first named them, Broadcom lease rent at $161.2 billion since July 13, and AMD above $20.0 billion since August 10. One more difference is in how an estimate is described. The notes in the September 4 and later documents give a range of $125 billion to $150 billion for expanded commercial agreements that deliver capacity between 2028 and 2031, because “certain commercial terms have not yet been finalized.” The S-1’s discussion section is the first to say that capacity is “additional TPU compute from TPU providers, including Google.”

The SpaceXAI line

SpaceXAI is the most visible line in the coverage and the one the filings make least firm. In the July 13 draft the May 2026 agreements were “up to $44.4 billion between May 2026 and May 2029”; by August 10 the May and July agreements made it up to $84.5 billion; the S-1 puts the agreements entered or amended between May and October at up to $120.5 billion. That is a ceiling 2.7 times its July level (120.5 divided by 44.4, my arithmetic). Every version that names the counterparty has carried the same exit. The July draft says “Our agreements with x.AI are non-cancelable by either party before August 2026; thereafter, the agreements may be cancelled by either party upon 90 days’ advance written notice,” and the S-1 now gives initial non-cancelable periods ending in January 2027 and March 2027 for the respective agreements.

That makes the $120.5 billion a ceiling on an arrangement that runs while both sides want it to. What Anthropic has bound itself to pay beyond the initial periods is the notice window: at the stated full rate of about $3.1 billion a month, three months is about $9.3 billion, which is my arithmetic and ignores the different fees during the ramp. The right also runs against Anthropic, because capacity worth about $37.2 billion a year at the full rate (again my arithmetic) can end on three months’ notice from the supplier. The S-1 names SpaceXAI only in this paragraph and in a list of NVIDIA GPU sources in the business section.

Where the filing counts the commitments, and where it does not

The discussion section’s contractual-obligations table has its row labels (operating leases, finance leases, purchase commitments, total) and its columns (2026 through “Thereafter”), but no figures in any cell, in all seven documents. The S-1 adds a lead sentence and a footnote about agreements with no fixed dollar amount and leaves their amounts blank too: counting the dollar signs left empty in the paragraphs around the table gives ten. The notes do carry figures, on four dates and several bases:

Four cautions on reading those lines together. The bases differ: December’s figure is non-cancelable purchase commitments with a remaining term over one year, June’s is “fixed or minimum guaranteed commitments” with terms of one to 20 years, and the $417 billion covers new and amended lease and commercial agreements. So the jump from $54.6 billion to $265.6 billion is an order-of-magnitude comparison and not a like-for-like growth rate; by my arithmetic it is 4.9 times in six months, or $211.0 billion. The post-June 30 figure rose from about $154 billion to about $245 billion between the October 9 draft and the S-1, and the separate $89 billion line no longer appears; $154 billion plus $89 billion is $243 billion (my sum), but the filing does not say the two were combined. Where “new, and amended existing” agreements are counted at full value, part of one period’s total may sit inside another’s, so adding the notes’ figures together could count a contract twice. And all of them leave out agreements with no fixed dollar amount, which Note 13 says are “expected to be significant.”

For scale, the filing shows $112.4 billion of cash and cash equivalents and $19.3 billion of short-term investments at June 30, $131.7 billion together, against first-half revenue of $16.3 billion and first-half compute and infrastructure operating expense of $8.9 billion. The $265.6 billion is about twice that cash and short-term investments (2.0 times, my arithmetic). The filing also says Anthropic raised its revolving credit commitments from $3.3 billion to $15.0 billion in September and set a $5.0 billion minimum liquidity covenant from the fourth quarter. On timing, it says only that “substantially all of the remainder” of the commitments beyond what falls due before the end of 2030 is “expected to be incurred between 2031 and 2045,” with the amount before 2031 left blank.

The eight named agreements are five kinds of promise

The discussion section describes “certain of our most significant existing arrangements” and says Anthropic is also party to others “not described below.” The eight it names are not one kind of commitment. Sorted by what each says about firmness and exits, in the filing’s own words:

Adding them up, with the kinds kept apart

None of the sums below is a figure in the filing. Each is my addition of figures the filing prints, with the kind of promise kept separate so the total does not hide the differences:

• Floors: Microsoft $44.9 billion plus Google $114.5 billion plus Amazon $110.0 billion is $269.4 billion, each described as owed “regardless of actual usage.”
• Lease rent: Broadcom’s $161.2 billion. Floors plus lease rent is $430.6 billion of obligations the filing describes as non-cancelable, subject to the exits in the table.
• Delivery-gated: Nscale’s $44.6 billion plus Lambda’s $41.3 billion is $85.9 billion that accrues only as capacity is delivered and accepted.
• Short-notice ceiling: SpaceXAI’s $120.5 billion.
• Every dollar figure above, added together: $637.0 billion. The estimates the filing keeps outside its table (the $125.0 billion to $150.0 billion of additional TPU compute and AMD’s “exceed $20.0 billion”) would take that to $782.0 billion to $807.0 billion, if none of them overlaps anything else. The filing does not say whether any do, and it does not give its own total.

Lambda is the supplier our August 29 post described borrowing about $1 billion to buy GPUs for lease to Microsoft. The S-1 shows it also owes Anthropic capacity under agreements the filing puts at about $41.3 billion, and one of them lets either party walk away if Lambda does not secure its facility lease in time.

Who stands behind the ledger

The convertible-note section shows that the largest suppliers are also financing options for Anthropic. The filing summarizes the undrawn facilities at $77.0 billion in total with nothing outstanding at June 30, and AMD has a separate stock-purchase commitment:

The $77.0 billion of notes is the filing’s own total; adding AMD’s $5.0 billion gives $82.0 billion, my sum. Three points follow. None of these facilities is sized to cover its supplier’s line: Broadcom’s $42.0 billion is a third of the $125.2 billion of rent under the one lease it finances (33.5 percent, my arithmetic) and about a quarter of the $161.2 billion across all of them. All of them are drawn on conditions and settle in notes that convert (in Broadcom’s case at a 10 percent discount) or directly in shares, so they move part of the cost of the ledger from cash to dilution. And they sit beside the same suppliers’ commercial contracts, which is the reverse of the structure we covered when Akamai handed Anthropic a warrant.

Where the $80.7 billion loss sits

Financing instruments like these also drive the loss the coverage reported. For the first half of 2026 the filing shows revenue of $16,286 million, an operating loss of $693 million, interest income of $1,111 million, a $72,388 million loss on the change in fair value of convertible notes, derivative liabilities and SAFEs, $8,757 million of other expense, a $21 million tax provision and a net loss of $80,748 million. It attributes the $72.4 billion line “primarily” to “an increase in the fair value of our outstanding convertible promissory notes resulting from an increase in our enterprise valuation from 2025 to 2026,” and the $8.8 billion line “primarily” to “the initial fair value recognition and change in fair value of the assets and liabilities recognized under the note purchase agreements.” Those two lines total $81.1 billion. Without them, operating loss plus interest income is a gain of $418 million, or $397 million after the tax provision (my arithmetic).

What the coverage reported and what it skipped

The numbers in the stories hold up against the October 26 filing. The gaps are in what the stories leave out, and I checked all seven of the October 26 pieces I read (two from TechCrunch, two from WIRED, and one each from The Verge, The Register and MIT Technology Review) for the terms below.

Cross-checks against the counterparties

Nscale. Nscale’s own IPO filing, as we reported on September 25, listed Anthropic’s agreement at $44.6 billion, described it as contingent on Nscale obtaining financing, and gave Anthropic the right to walk away if Nscale missed milestones the filing called “stringent.” Anthropic’s October 1 draft is the first of its documents to name Nscale, at the same $44.6 billion with terms “extending into 2033,” and the figure has not changed since. Its wording on the financing condition is shorter than Nscale’s: Nscale “is required to use best efforts to obtain qualifying financing” within a specified period, and “otherwise” the commitment is non-cancelable except in limited circumstances. The S-1 does not say what happens if that period lapses.

Akamai. Akamai’s SEC filing, as we reported on September 26, put its Anthropic deal at $11.6 billion over seven years. None of the seven Anthropic documents names Akamai. The S-1 says only that Anthropic is “also party to other compute infrastructure service arrangements and leases that are not described below,” so a reader cannot tell whether that deal sits inside the $265.6 billion, inside the $245 billion or in neither.

AMD. AMD’s July 22 release counts in gigawatts, up to 2 of MI450 Series GPUs with the first gigawatt beginning in the first half of 2027, and gives only the equity figure. The filings supply the dollars the release does not: expected expenditures above $20.0 billion for the initial tranche, in every document since August 10, plus possible additional capacity of up to approximately the same amount, neither in the table. If the initial tranche is the first gigawatt, as the release’s timing suggests, that is more than $20.0 billion a gigawatt. That is my inference; the filing does not state a price per gigawatt.

Where the loop and the ledger meet

WIRED’s reading is that the filing treats “closing the loop” as both the strategy and the danger, quoting Anthropic’s description of the process as “one of the most significant structural advantages in frontier AI.” The financial sections add a second reading. The compute risk factor lists “the pace of our research progress” first among the factors that make requirements difficult to predict, says “In the past, we have significantly underestimated our compute requirements,” and says commitments are made “months or years in advance of our day-to-day delivery and use.” Separately, the AI-safety risk factor says “We have observed that our models are contributing to the pace of our own research.” My reading, which the filing does not state, is that if research speed is partly a product of the loop, the loop is also a reason the demand forecast behind the ledger is hard to hold, and the $430.6 billion of floors and lease rent is the part that does not flex with it. Our earlier pieces on Anthropic’s pacing plan and on its embedded-evaluator arrangement cover the oversight side of the same question.

Checks to borrow for the next AI S-1

• When a company flips a confidential filing public, EDGAR also posts the earlier drafts. Diff them: here they show SpaceXAI’s ceiling at $44.4 billion, $84.5 billion and $120.5 billion, and show Nscale and Lambda first named on October 1.
• Read the notes and the subsequent-events notes before the discussion section. The totals sit on different dates, and in this filing only the notes had figures.
• Split every commitment by its verbs. “At least” with “regardless of actual usage” marks a floor, “up to” marks a ceiling, and “fees accruing only as capacity is delivered” marks a deal gated by the supplier’s performance.
• Read each exit clause for whose breach, which tranche, what notice, and what financing or facility condition the supplier must meet.
• Look for the financing attached to each supplier: note purchase agreements, stock-purchase commitments and options, their conditions and their conversion prices.
• Check the fair-value lines before reading a headline loss. Here two of them exceed the entire net loss. Our June checklist for compute capacity contracts covers the buyer’s side of the same questions.

What to watch in the amendment

• The filled-in contractual-obligations table and the three dollar amounts in the compute risk factor: the total, the part due before 2031 and the part due in the next 12 months.
• Whether the $245 billion in Note 17 adds to the $265.6 billion or overlaps it, and what became of the $89 billion contingent line.
• Whether the SpaceXAI figure moves again, and whether the 90-day right appears in a risk factor.
• Whether the Nscale financing period and the Lambda facility-lease condition are met, and whether the filing says so.
• The offering size, which is blank, against the $77.0 billion of facilities, whose draws settle in notes or shares.
• Run-Rate Revenue as of October 31, which the filing leaves blank, and the September 30 figures it refers to.

Sources and method

I read the October 26 S-1 (accession 0001628280-26-067168) and the six drafts on EDGAR: June 1 (0002130357-26-000005), July 13 (0001628279-26-000861), August 10 (0001628279-26-000993), September 4 (0001628279-26-001111), October 1 (0002130357-26-000009) and October 9 (0002130357-26-000014). I converted each document to text, removed page-footer noise, searched every document for the commitment sentences and note totals, and read the surrounding paragraphs wherever a figure changed. Each cell in the first table is the figure found in that document, and each figure and quotation used here was checked against the saved copy. The sums, ratios and the growth multiples are mine and are labeled where they appear. I also read the October 26 stories from TechCrunch (two), The Verge, WIRED (two), The Register and MIT Technology Review, AMD’s July 22 release, and our own posts on Nscale, Akamai and Lambda.

To reproduce the October 9 against October 26 comparison, this script fetches both documents and prints the six named figures. The SEC asks for a User-Agent that identifies you, so replace the placeholder with your own name and address before running it. I ran it on Python 3.13 with no packages beyond the standard library:

# edgar_diff.py: compare named compute commitments in two Anthropic filings on EDGAR
import html, re, urllib.request

UA = "Your Name [email protected]"  # the SEC asks for a User-Agent that identifies you
BASE = "https://www.sec.gov/Archives/edgar/data/2130357/"
FILES = {"Oct 9 draft": "000213035726000014/filename1.htm",
         "Oct 26 S-1": "000162828026067168/anthropicpbc-sx1.htm"}
AMT = r"\$([\d.]+) billion"
FIND = {
    "Microsoft": r"agreement with Microsoft[^$]{0,300}" + AMT,
    "Google": r"agreement with Google[^$]{0,300}" + AMT,
    "Amazon": r"agreement with Amazon[^$]{0,300}" + AMT,
    "SpaceXAI": r"Under the (?:x\.AI|SpaceXAI) Arrangements, we agreed to spend an aggregate of up to " + AMT,
    "Nscale": r"Nscale Services Agreements, we agreed to spend an aggregate of up to approximately " + AMT,
    "Lambda": r"Lambda Agreements, we agreed to spend an aggregate of approximately " + AMT,
}


def text(path):
    req = urllib.request.Request(BASE + path, headers={"User-Agent": UA})
    raw = urllib.request.urlopen(req, timeout=180).read().decode("utf-8", "replace")
    raw = re.sub(r"(?is)<(script|style).*?</\1>", " ", raw)
    return re.sub(r"\s+", " ", html.unescape(re.sub(r"<[^>]+>", " ", raw)))


docs = {name: text(path) for name, path in FILES.items()}
print(f"{'':10}" + "".join(f"{name:>14}" for name in docs))
for who, pattern in FIND.items():
    cells = []
    for doc in docs.values():
        m = re.search(pattern, doc)
        cells.append(f"${m.group(1)}B" if m else "-")
    print(f"{who:10}" + "".join(f"{c:>14}" for c in cells))

It prints:

             Oct 9 draft    Oct 26 S-1
Microsoft         $31.4B        $44.9B
Google           $111.1B       $114.5B
Amazon           $110.0B       $110.0B
SpaceXAI          $84.5B       $120.5B
Nscale            $44.6B        $44.6B
Lambda            $41.3B        $41.3B

I did not see the underlying contracts, which the filing summarizes and does not reproduce in the text I read, and I did not contact Anthropic or any counterparty. The drafts are Anthropic’s own confidential submissions, now public on EDGAR, and the filing is a first public draft with blanks in it, so an amendment may change figures, dates and wording. Nothing here is investment advice; it is a reading of what the documents say and do not say.