807B Loss, 657B Take-or-Pay: Computing the Anthropic 335-Page S-1 Into Three Ledgers with Code
807B Loss, 657B Take-or-Pay: Computing the Anthropic 335-Page S-1 Into Three Ledgers with Code
Part One: From “Rumors” to “Formal Filing” — the First Regulator-Reviewed Full Financial
On October 26, 2026 (US Eastern Time), Anthropic filed an S-1 registration statement with the SEC to list on the Nasdaq Global Select Market under the ticker “ANT”, with Goldman Sachs and Morgan Stanley as lead underwriters. Press reports put the planned pricing week at November 16, 2026; the valuation range is $1.5–$2.0 trillion, with up to $100 billion in proceeds — which at the top of the range would surpass Saudi Aramco’s 2019 ~$29.4B record as the largest global IPO.
We need to distinguish this article from the earlier speculative pieces. The past few months were about leaked drafts, fundraising rumors, and circular-deal questions; this one is a 335-page, SEC-reviewed, legally-accountable formal prospectus. For the first time it nails three things to the regulatory page: real operating profit and loss, the revenue channel concentration, and rigid compute commitments for the next decade. This article will avoid sentiment; instead we will compute three ledgers — the accounting ledger, the structural ledger, and the commitment ledger — with code.
The one core question the prospectus asks
+----------------------------------------------------------+
| Revenue growth vs rigid compute-commitment growth: |
| which runs faster? |
| |
| Win: fixed compute diluted -> huge operating leverage |
| (gross margin ~85%) |
| Lose: take-or-pay unchanged -> fixed commitments become |
| "reverse leverage" eating profit |
+----------------------------------------------------------+
Part Two: Ledger One — why the 807B “loss” isn’t “807B in cash burned”
The headline number is loss: 2026H1 net loss $80.748B. But reading it directly as “the business burned $80.748B in half a year” is wrong. The prospectus itself splits the accounting layer from the operating layer:
- Of the $80.748B net loss, roughly $72.388B comes from fair-value changes in convertible notes, derivative liabilities, and SAFE instruments — as the company’s valuation rose, these future-equity liabilities got “more expensive” on the books, causing a huge accounting loss without any cash actually flowing out;
- Operating loss was only $0.693B (operating margin ~ -4%); excluding equity comp and non-cash charitable donations, adjusted operating profit was $1.618B and adjusted net profit $2.671B;
- Q2 revenue was $11.555B (+1368% y/y), operating profit $0.974B (~8%), adjusted operating profit $2.235B (19% margin) — the first time a foundation-model company reported a quarterly operating profit in a formal filing;
- Gross margin went from negative in 2024 to ~85.4% in 2026H1.
Bridging GAAP net loss back to the operating truth with code is a necessary reading skill for this kind of AI financial:
# Unit: $100M (yi), from Anthropic S-1 (six months to 2026-06-30)
def bridge_gaap_to_operating():
net_loss_gaap = -807.48
fair_value_change = -723.88 # convertible/deriv/SAFE FV, non-cash
other_non_cash = -87.57 # FV of purchase-agreement assets/liabilities etc.
# Counterintuitive: these "losses" come mainly from valuation gains; adding back lands on operating
operating_loss = -6.93
sbc_and_donations = 23.11 # stock comp + non-cash donations (reverse-engineered demo)
adjusted_op_profit = 16.18
adjusted_net_profit = 26.71
check = net_loss_gaap - fair_value_change - other_non_cash
print(f"GAAP net loss : {net_loss_gaap:8.2f}")
print(f"Add back FV non-cash : {-fair_value_change:8.2f} + {-other_non_cash:.2f}")
print(f"Stripped, roughly op : {check:8.2f} (cf. operating loss {operating_loss})")
print(f"Adjusted op / net profit : {adjusted_op_profit:6.2f} / {adjusted_net_profit:.2f}")
return adjusted_op_profit
bridge_gaap_to_operating()
Keep this balanced: non-cash loss is not fake loss — convertible notes create real future dilution; adjusted profit is not cash self-sufficiency — the real funding pressure is in the commitment ledger below. But at least “$80.748B” as a headline number and “the business is bleeding” are two different stories.
Part Three: Ledger Two — customers broadening, channels concentrating (48% and 58%)
Revenue is exploding: 2023 $0.0041B → 2025 $4.585B → 2026H1 $16.286B (+1300% y/y). Structurally 80% is pay-per-token consumption, 20% subscription; customers spending >$0.1M in the trailing twelve months top 11,900, >$1M top 1,900, >$10M ~250; net revenue retention is ~900%; no single customer exceeds 10% of revenue.
Customers are widely distributed, but sales channels are exceptionally concentrated — the most informative structural numbers in the filing:
- Revenue through AWS / GCP / Azure marketplaces accounted for 48% of total (11% in 2023, 32% in 2024, 47% in 2025);
- Corresponding accounts receivable share was ~58% — nearly six-tenths of collections go through the cloud vendors’ hands;
- Channel-partner payments rose $2.7B y/y in H1, ~17% of revenue;
- In 2025 ~47% of sales ran through AWS and GCP; Anthropic recognizes revenue on a gross basis, then books the ~16% cloud cut as selling expense — OpenAI has privately complained this gross treatment “inflates” revenue.
Concentration and channel drag can be computed directly:
# Channel concentration & effective rate (2026H1, unit $100M)
revenue_h1 = 162.86
csp_share = 0.48 # CSP marketplace share of revenue
ar_csp_share = 0.58 # CSP share of AR
channel_fee = 27.0 # H1 incremental CSP channel fee
csp_revenue = revenue_h1 * csp_share
effective_channel_rate = channel_fee / revenue_h1
print(f"CSP marketplace revenue : {csp_revenue:6.1f} bn")
print(f"Channel fee / revenue : {effective_channel_rate:6.1%}")
print(f"AR via CSP : {ar_csp_share:6.0%} -> payment lifecycle in CSP hands")
# Simplified HHI demo (with disclosed customer tiers as a rough demo)
tiers = {">=$10M": 250, ">=$1M": 1900, ">=$0.1M": 11900}
# Demo only: many customers but each <10% -> low HHI = revenue distributed; channel HHI is high
def hhi_from_shares(shares):
return sum(s * s for s in shares) # shares in %, higher HHI = more concentrated
revenue_side = [8, 8, 7, 7, 6] # assumed top customers (each <10%)
channel_side = [22, 15, 11] # AWS/GCP/Azure approx
print(f"Revenue-side HHI (low=diverse) : {hhi_from_shares(revenue_side)}")
print(f"Channel-side HHI (high=concent) : {hhi_from_shares(channel_side)}")
The conclusion is clear: Anthropic has diversified away the risk of losing any single customer across tens of thousands of enterprises, but pledged the channel to reach them to three cloud giants. Revenue diversification lowers single-customer churn risk; channel concentration simultaneously amplifies pricing-power loss, cash-cycle drag, and “they see our playbook” risk.
Part Four: The “frenemy” with four identities — investor, customer, channel, supplier, competitor
The risk section of the S-1 itself describes the cloud vendors as simultaneously playing the roles of investor, customer, distribution partner, and competitor, and openly states they “are able to learn about our pricing and business terms” and may prioritize their own models. The funding side of this relationship is also striking:
- Amazon: ~$18B invested so far and up to another $15B committed on milestones; convertible notes from 2023/2024 purchase agreements can convert to up to 1.84B shares (~10.9% of fully diluted shares at June), with a 12% holding cap (excess auto-converts into new notes, can be waived);
- Google: ~$23B invested across rounds and another $20B committed on performance milestones; latest round has a 14.99% cap (also waivable);
- The shares of both are excluded from the “5%+ voting holders” list because they carry no voting rights, but after IPO conversion each will hold above 5%;
- More subtly, the cash flows loop: Anthropic is a major tenant on AWS and Google Cloud (using their data centers and chips to train Claude), while Amazon and Google resell Claude via Bedrock and Google Cloud, and each runs Nova / Gemini / Copilot competing directly.
This structure — “funds you, bills you, sells your product, and also competes with you” — is clearer as a diagram than prose:
+-----------------------+
| Anthropic |
| (Claude / ANT) |
+-----------+-----------+
equity/conv. | | channel take (~17%)
+-----------------+ +-----------------+
v v
+--------------+ leases chips/datacenter (charges) +-------------------+
| AWS/Google/ |<-------------------------------------| 48% revenue |
| Microsoft |------------------------------------->| 58% AR |
+--------------+ via Bedrock/GCP resale +-------------------+
| |
| also build Nova/Gemini/Copilot competing w/ Claude
+-----------------------------------------------+
and can see Anthropic's pricing & terms
The S-1 explicitly admits: moving workloads off these clouds would be “costly, time-consuming, and may impair computing efficiency.” That is, the dependency has gone deep from “supplier relationship” into the marrow of infrastructure and business model, and has already drawn antitrust attention.
Part Five: Ledger Three — a $657B “compute indenture,” ~80% take-or-pay
What really determines this company’s fate is the commitment ledger. Adding the disclosed main contracts roughly, future compute and infrastructure arrangements exceed $650B (different oral range $542–$657B, or higher if including Google’s additional $125–$150B TPU top-up, terms up to 2045), representing ~3.5 GW of long-horizon dedicated capacity:
Counterparty Amount ($B) Term / notes
Broadcom ~161 chip + equipment finance lease (excl. interest)
SpaceX AI up to 120.5 NVIDIA cluster to 2029, ~$3.1B/mo at full
Google at least 114.5 cloud/compute to 2033, short-fall top-up (take-or-pay)
Amazon at least 110 cloud services to 2036
Microsoft at least 44.9 cloud infra to 2033 (up from draft 31.4)
Nscale ~44.6 newly disclosed
Lambda ~41.3 newly disclosed
AMD initial >20 plus $5B equity-purchase commitment back to Anthropic
The key clause is take-or-pay: by some reported oral ~80% of commitments are irrevocable — Anthropic pays regardless of actual use — “if actual spend falls short, we must make up the difference to Google and Amazon.” The only obvious “escape hatch” is the SpaceX AI deal: after an initial non-cancellable period in Jan–Mar 2027, either party can cancel with 90 days’ written notice.
Stress-testing the commitments against cash flow is the core of this ledger:
# Unit: $100M. Commitments span many years; this is order-of-magnitude annualized stress test
commitments = {
"Broadcom(equip)": 1612, "SpaceX AI(NVIDIA)": 1205,
"Google(cloud/TPU)": 1145, "Amazon(cloud)": 1100,
"Microsoft(cloud)": 449, "Nscale": 446, "Lambda": 413, "AMD(initial)": 200,
}
total = sum(commitments.values())
take_or_pay_ratio = 0.80
years = 10
annual_fixed = total * take_or_pay_ratio / years
spacex_monthly_full = 31.0
print(f"Main commitments total : {total:6.0f} bn ($0.{total/100:.2f}T)")
print(f"Rigid (80%) avg annual : {annual_fixed:6.1f} bn/yr")
print(f"SpaceX full-load monthly: {spacex_monthly_full:6.1f} bn/mo")
# Reverse-leverage scenario: revenue growth slows, fixed compute doesn't fall
def operating_margin(revenue_q, fixed_compute_q, var_ratio):
var_cost = revenue_q * var_ratio # cost scales with revenue
op = revenue_q - fixed_compute_q - var_cost
return op / revenue_q
scenarios = {
"bull (revenue keeps scaling)": (200, 40, 0.35),
"base": (130, 40, 0.35),
"stall (revenue halved)": (65, 40, 0.35), # fixed compute unchanged
}
for name, (r, fx, vr) in scenarios.items():
print(f"{name:20s} op margin = {operating_margin(r, fx, vr):7.1%}")
The output matches the prospectus’s subtext: fixed compute fees get diluted away at high growth (gross margin reaches 85%), but they don’t fall when revenue falls. If demand stalls, the same leverage crushes margins the other way — this is a contract you can only press the accelerator on, and never the brake.
Part Six: Liquidity cushion and “two-way leverage” — what 22% coverage means
As of June 30, 2026, the company had about $112.4B in cash and equivalents, $19.3B in short-term investments, plus the $15B revolving credit expanded in September, for **$146.7B of usable liquidity**; H1 operating cash flow was positive $3.929B, financing inflows were $110.774B, cumulative funding ~$141.1B. But versus the $657B commitment, the liquidity coverage is only ~22%.
What looks really good is the unit-economics trajectory: revenue per $1 of compute & infrastructure expense went from $0.40 in 2025Q1, sub-$0.70 in 2025Q4, to $1.40 in 2026Q1 and $2.08 in 2026Q2 — a 5x improvement in just over a year. The Opus 5.5 released in September runs 40% cheaper than Opus 5. Compute & infrastructure as a share of revenue dropped from 1,056% in 2023 to ~48–55% in 2026H1.
# Operating leverage & liquidity safety margin
liquidity = 1124 + 193 + 150 # cash + ST invest + revolver
commit = sum([1612,1205,1145,1100,449,446,413,200])
coverage = liquidity / commit
print(f"Liquidity / commitment : {coverage:6.1%} (safe near term; long-term needs growth)")
# Trajectory of "revenue per compute dollar"
revenue_per_compute_dollar = {
"2025Q1": 0.40, "2025Q4": 0.70, "2026Q1": 1.40, "2026Q2": 2.08,
}
base = revenue_per_compute_dollar["2025Q1"]
for q, v in revenue_per_compute_dollar.items():
print(f"{q}: $1 compute -> ${v:.2f} revenue, vs 2025Q1 x{v/base:.1f}")
# Simple cash-flow sustainability: OCF vs annualized rigid commitment
ocf_h1 = 39.29
annual_fixed = commit * 0.8 / 10
print(f"H1 operating cash flow : {ocf_h1:6.1f} bn (annualized ~{ocf_h1*2:.0f} bn)")
print(f"Annualized rigid (80% over 10y): {annual_fixed:6.1f} bn")
print(f"OCF covers fixed (excl. financing): {ocf_h1*2/annual_fixed:6.1%} -> still reliant on financing + growth at this stage")
The conclusion is clear: short-term liquidity is absolutely safe, but whether the business can self-sustain long-term depends on whether OCF growth can keep pace with the depreciation/payment peak of commitments that lands in 2027–2030. The 22% is not a danger signal, but it says this is a balance sheet that is “discounting today’s compute against tomorrow’s growth.”
Part Seven: Verdict — a “frontier-AI call option + heavy-industry balance sheet”
Bringing the three ledgers together, a few judgments.
First, the IPO story is fundamentally a growth-rate race, not “will AI make money.” Q2 operating profit already proves foundation models can be profitable; the real suspense is whether revenue (mostly volatile, token-consumption) can consistently outrun rigid commitments (take-or-pay) over the next decade. Win, and operating leverage is enormous; lose, and fixed commitments turn into reverse leverage.
Second, this is a “very noisy accounting, very small governance, very heavy commitment” complex note. The $80.748B loss is mainly non-cash revaluation, but convertible dilution is real; Amazon’s 12% and Google’s 14.99% caps stacked on no-voting-rights arrangements leave external public shareholders with limited governance voice. Many key issuance terms (price, share count) are still blank.
Third, circular dependency and regulation are the sword overhead. Three cloud vendors are simultaneously shareholders, suppliers, channels, customers, and competitors; circular-deal criticism has been leveled, and antitrust scrutiny has arrived. CSPs see the playbook and can prioritize their own models or play hardball on capacity allocation. 48% channel concentration is a growth flywheel in good times and a chokehold in bad.
Fourth, the timing itself is a signal. Q3 revenue-growth-slowdown reports already hit this month and triggered a tech selloff, with markets worried AI spending is outrunning real demand. Anthropic chose to sprint through an unfriendly mood, partly to ride the fundraising window, partly because it’s an “advance-compute-mortgaged capital race” where standing still means losing ground. Compared with OpenAI still widening losses and Zhipu/MiniMax still losing money in H1, first-to-file earned Anthropic the narrative of “first pure foundation-model company to go public.”
Fifth, the employee leverage is another implicit optimistic assumption. With roughly 5,300 employees and TTM revenue per capita of about $6.6M, far above the average for mature tech firms, the bet is that every one of them keeps pulling AI models through the long chain from paper, training, tuning, alignment, to deployment. If a training bottleneck appears (say Opus 6 lags, alignment quality drops, or a key researcher is poached), those 5,300 instantly become 5,300 lines of fixed high-compensation burn. The prospectus adds equity comp back to operating profit as a non-cash item, which essentially says “these people are taking a bet on the future” — a bet that pays off is profit; one that doesn’t is the double hit of cash spend plus talent loss.
Sixth, revenue structure and cash quality deserve another layer of dissection. The 80% token consumption / 20% subscription mix looks stable, but token consumption is “pay per invocation” — client invocation behavior is spiky and heavily tied to model quality, price, and the launch cadence of competing products. That means Anthropic’s quarterly revenue curve will be steeper than a subscription-led company’s, and whether the Q1–Q2 blow-out sustains into Q4 and next year depends on whether Claude keeps pulling away from OpenAI and Google in agent capability, long context, and enterprise API stability. The 20% subscription piece is more like “floor,” but it’s small and cannot offset the volatility of token consumption.
For investors and practitioners, the pragmatic read is not to fixate on the headline $80.748B or the $2T valuation, but to keep tracking three measurable leading indicators: ① revenue per unit of compute (already $2.08 — can it keep rising?); ② cloud-channel revenue share (can 48% come down to cut dependency?); ③ OCF coverage of rigid commitments (when can the business self-sustain without financing?) Those three numbers, more than any roadshow deck, will decide whether this ends up as “the largest IPO in history” or “the most expensive IOU for compute.”
Sources: Anthropic’s S-1 prospectus filed with the SEC (2026-10-26, ticker ANT); CaiLianShe/Sina Finance “Anthropic: Google, Amazon, Microsoft — frenemies”; 36Kr “Anthropic unveils 335-page prospectus”; Huxiu 42-Channel-Transmission; CaiJing Magazine “Dissecting the Anthropic prospectus”; Guoxin Securities “Anthropic prospectus review: 2026H1 revenue 16.286 bn, 657 bn compute contracts disclosed.” Commitment totals vary by reported oral range ($542–$657B, or higher with additional TPU); all are from public reporting and research notes. Python estimates are order-of-magnitude demonstrations; annualizing and the 80% take-or-pay share are assumed parameters, not company forecasts.