AI replaces all white-collar work
- Jobs lost directly
- 87.1M
- Extra jobs lost from spending cuts
- ~13M
- Employee pay lost
- $10.9T
- Share of GDP
- 35%
- Consumer spending, year 1
- −25%
- Labor share of GDP
- 51% → 16%
Twice the Great Depression peak.
Anthropic Form S-1 · filed 2026-10-26 · read 2026-11-18
Gross margin, first half 2026
85%
(Revenue $16,286M − cost of revenue $2,380M) / revenue. The filing prints no gross margin.
Note 8, p. F-39
Gross margin after R&D, first half 2026
26%
(Revenue $16,286M − cost of revenue $2,380M − research and development $9,741M) / revenue. R&D includes training compute.
Note 8, p. F-39
Earned per $1 spent, Q2 2026
$1.09
Revenue $11,555M / total operating expenses $10,581M. First half 2026: $0.96. Full year 2025: $0.36.
MD&A, p. 157; Statements of Operations, p. F-7
Total assets
$147.4B
Of which $112.4B cash and $19.3B short-term investments. Liabilities are $117.5B. At June 30, 2026.
Balance Sheet, p. F-5
Development spending (R&D), first half 2026
$9.7B
5.3 times the first half of 2025 ($1.85B), and 60% of revenue. Includes training compute.
Note 8, p. F-39
Net revenue retention
~900%
Existing customers spend about nine times what they spent a year earlier. Excludes Pro and Max subscribers.
MD&A, p. 142
Compute commitments
~$853B
What Anthropic has promised to pay for compute: purchases, leases not yet started and new deals to Oct 26, 2026.
Notes 7, 13, 17 (sum by LLMeter)
Net loss, first half 2026
−$80.7B
$72.4B of it is the rising value of convertible notes owed in shares, not cash. Operating loss was $0.7B.
Statements of Operations, p. F-7
Revenue divided by the cost named in the row, in dollars. All rows are calculated by LLMeter from the figures in sections 1 and 2.
| Revenue earned per $1 of | 2025 | H1 2026 | Q2 2026 |
|---|---|---|---|
| All operating expenses | $0.36 | $0.96 | $1.09 |
| Compute and infrastructure | $0.63 | $1.82 | $2.08 |
| Cost of revenue (serving customers) | $1.65 | $6.84 | not reported |
| Research and development | $0.72 | $1.67 | not reported |
Example: in the first half of 2026 each dollar of total operating spending brought back $0.96 of revenue, so the business was still $0.04 short per dollar spent. In Q2 2026 it was $0.09 ahead. Serving cost alone is far below the rest: each $1 spent on serving customers earned $6.84. The filing reports cost of revenue and R&D only by half year, not by quarter.
The filing uses several similar words for money that is promised but not yet moved. They point in opposite directions.
Remaining contracts (remaining performance obligations, $10.8B): money customers have agreed to pay Anthropic under non-cancelable contracts, for service it has not delivered yet. It is future revenue that is already signed. It does not include pay-as-you-go usage with no commitment, so it is small next to the $16.3B of revenue earned in the first half of 2026 alone.
Future obligations (purchase commitments, $265.6B, and leases not yet started, $321.4B): money Anthropic has agreed to pay cloud and chip providers such as Google, Amazon, Microsoft, Broadcom and others. They are future costs that are already signed, and most cannot be cancelled. New deals signed from July 1 to October 26, 2026 add about $266B.
Deferred revenue ($1.65B) is a third item: cash customers already paid for service not yet delivered. It is part of the $10.8B.
Read together: the first is what is coming in on signed contracts, the others are what is going out. Today the money going out on contracts is about 79 times the money coming in on contracts, which is why the filing leans on revenue growth to cover it.
Where: remaining performance obligations in Note 2, Revenue (p. F-26); purchase commitments in Note 13 (p. F-58); leases in Note 7 (p. F-35); subsequent commitments in Note 17 (p. F-65). The sum is in section 3 below.
Millions of US dollars, unaudited, from the filing's quarterly table (p. 157). Rows in italics are calculated by LLMeter.
| Q1 2025 | Q2 2025 | Q3 2025 | Q4 2025 | Q1 2026 | Q2 2026 | |
|---|---|---|---|---|---|---|
| Revenue | 376 | 787 | 1,382 | 2,040 | 4,731 | 11,555 |
| Compute and infrastructure | 944 | 1,390 | 2,047 | 2,945 | 3,387 | 5,558 |
| Operating income (loss) | (935) | (1,200) | (3,089) | (2,836) | (1,667) | 974 |
| Compute as a share of revenue | 251% | 177% | 148% | 144% | 72% | 48% |
| Revenue per dollar of compute | $0.40 | $0.57 | $0.68 | $0.69 | $1.40 | $2.08 |
The Q3 2025 loss includes a $1.5 billion copyright settlement. Compute is contracted 12 to 18 months ahead and expensed when it comes online, so part of the improvement is demand growing into capacity already bought. The company warns that margins may fall again when the next large blocks arrive.
Q3 2026 revenue is blank in this filing. It states that the quarter was 86 times Q4 2024 revenue of $197 million, which gives about $16.9 billion, up about 47% on Q2 2026 after 132% and 144% in the two quarters before.
Millions of US dollars. The filing reports no gross margin, but Note 8 (p. F-39) regroups expenses by function and says cost of revenue is mostly compute for customer-serving workloads.
| 2023 | 2024 | 2025 | H1 2025 | H1 2026 | |
|---|---|---|---|---|---|
| Revenue | 41 | 386 | 4,585 | 1,163 | 16,286 |
| Compute and infrastructure | 433 | 2,522 | 7,326 | 2,334 | 8,945 |
| Cost of revenue | 26 | 752 | 2,773 | 1,019 | 2,380 |
| Research and development | 422 | 2,275 | 6,338 | 1,850 | 9,741 |
| Gross margin | 37% | -95% | 40% | 12% | 85% |
| Compute outside cost of revenue, at least | 407 | 1,770 | 4,553 | 1,315 | 6,565 |
| ...as a share of all compute | 94% | 70% | 62% | 56% | 73% |
2023 to 2025 are audited; the half years are not. "Compute outside cost of revenue" covers research, training the next models, Anthropic's own use of Claude, and free access, so it is not a pure training figure. It is a lower bound because cost of revenue also holds some staff and overhead costs.
Serving is the small part: at most $2.4 billion to serve $16.3 billion of revenue in the first half of 2026. The training cycle is the large part, and it recurs: model versions released in the previous six months produced over 80% of Q3 2026 revenue (p. 143).
The filing leaves its own total blank. These are the parts it discloses.
| Item | Amount | Where in the filing |
|---|---|---|
| Non-cancelable purchase commitments at 2026-06-30 | $265.6 billion | Note 13, p. F-58 |
| Leases signed but not yet started at 2026-06-30 | $321.4 billion | Note 7, p. F-35 |
| New commitments, July 1 to October 26, 2026 | about $245 billion | Note 17, p. F-65 |
| New short-term commitments, same period | about $21 billion | Note 17, p. F-65 |
| Sum | about $853 billion | LLMeter's addition |
Not in the sum: further Google TPU compute for 2028 to 2031, estimated at $125 billion to $150 billion, and an AMD agreement expected to exceed $20 billion. At the end of 2025 the first two lines were $54.6 billion and $50.2 billion. The sum is about twelve and a half years of revenue at the Q3 2026 rate.
| Counterparty | Amount | Period | Hardware |
|---|---|---|---|
| Broadcom (equipment leases) | $161.2 billion, before interest | five-year leases | Google TPUs |
| SpaceXAI | up to $120.5 billion | into 2029 | NVIDIA GPUs |
| at least $114.5 billion | April 2026 to July 2033 | TPUs and cloud | |
| Amazon | at least $110.0 billion | May 2026 to April 2036 | Trainium and cloud |
| Microsoft | at least $44.9 billion | November 2026 to May 2033 | NVIDIA GPUs |
| Nscale | up to $44.6 billion | into 2033 | NVIDIA GPUs |
| Lambda | about $41.3 billion | into 2034 | NVIDIA GPUs |
Reported on pp. 170 to 171. The SpaceXAI agreement runs at about $3.1 billion a month once all capacity is delivered, expected in December 2026. Against all of this the company held $131.7 billion of cash and short-term investments at June 30, 2026.
The company's own statements. None was independently measured, and the benchmarks are not ones LLMeter tracks, so do not compare them with the other pages.
| Fact | Value | Page |
|---|---|---|
| Customer cost per task, Opus 5.5 against Opus 5 | about 40% lower | 195 |
| Customer cost per task, Fable 5.1 against Fable 5 | about 25% lower, up to 45% for complex agentic work | 195 |
| ExploitBench: Opus 4.8, Mythos 5, Opus 5.5 | 40%, 78%, 91% within about four months | 189 |
| Internal benchmark, open-ended engineering problems | about 90% in September 2026, up about 50 points in nine months | 183 |
| Share of Anthropic's own code written by Claude | over 90% in September 2026 | 183 |
| Chip families used for training and serving | AWS Trainium, Google TPU, NVIDIA GPU, AMD GPU | 135 |
| Time between contracting compute and using it | typically 12 to 18 months | 134 |
| Staff | about 940 at end of 2024, about 5,300 at 2026-09-30 | 63 |
Full write-up: What Anthropic's S-1 says about the cost of the frontier.
Source: Anthropic, PBC, Form S-1, filed 2026-10-26, read 2026-11-18.
US GDP 2025
$30.9T
Paid to employees
$15.7T · 51%
White-collar wages
$7.5T
87.1M people, 69% of wages
Software developer wages
$250B
1.7M people, 3.3% of white-collar pay
White-collar jobs cut by 2029
3.2M
3.7%, about $269B of pay a year. Range 0.7M to 5.0M
Unemployment by 2029
4.6–6.0%
Up from 4.1% today
White-collar
69% of wages
87.1M people
Blue-collar
18% of wages
35.2M people
Pink-collar & service
13% of wages
33.2M people
Other pay
$4.9T
Not split by collar
Office & admin jobs cut
882K
5.0%, the biggest hit
Software developer jobs cut
53K
3.1%
Twice the Great Depression peak.
Only 33M pink-collar and service jobs left.
| White-collar replaced | Jobs | Rate | Like |
|---|---|---|---|
| Software developers only | 1.7M | 5.1% | Recession signal |
| 5% | 4.4M | 6.7% | Mild recession |
| 10% | 8.7M | 9.2% | 2009 Great Recession |
| 20% | 17.4M | 14.4% | COVID peak, 2020 |
| 30% | 26.1M | 19.5% | Between COVID and 1933 |
| 50% | 43.6M | 29.7% | Germany 1932 |
| 100% (Scenario 1) | 87.1M | 55.2% | No precedent |
Source: ModernMBA macro note “Who Gets Paid From GDP” (US data for 2025), figures as printed there. Directional estimates, not forecasts.
Cost per agent-hour
$55.86
Pure $8.58, then 65% use, 2× R&D, 30% margin
Cost per engineer-year of agent work
$116,196
2,080 h
L1 cost vs salary
111%
One agent-year costs more than a $105,000 junior
US developer payroll
~$346B
Replace everyone, doubling rule
~$756B
About 2.2× payroll, before oversight
Failure-based, L1–L3 cost vs salary
50%–61%
L1 61%, L2 50%, L3 50%, before oversight
Price per agent-year is $116,196 (65% use, 2× R&D, 30% vendor margin). Before human oversight. Red = more than the salary.
| Level | Loaded cost | Doubling (1 → 64) | Linear (1, 1, 2, 3, 4, 6, 10) | Failure-based (agent-years) |
|---|---|---|---|---|
| L1 Junior | $105,000 | $116,196 (111%) | $116,196 (111%) | $63,908 (61%) · 0.55 |
| L2 Mid | $155,000 | $232,392 (150%) | $116,196 (75%) | $77,851 (50%) · 0.67 |
| L3 Senior | $215,000 | $464,784 (216%) | $232,392 (108%) | $108,062 (50%) · 0.93 |
| L4 Staff | $300,000 | $929,568 (310%) | $348,588 (116%) | $302,110 (101%) · 2.6 |
| L5 Principal | $480,000 | $1,859,136 (387%) | $464,784 (97%) | $11,619,600 (2421%) · ~100 |
| L6 Distinguished | $700,000 | $3,718,272 (531%) | $697,176 (100%) | not reachable |
| L7 Fellow | $1,100,000 | $7,436,544 (676%) | $1,161,960 (106%) | not reachable |
Cost = engineer-years × $116,196. Under the doubling rule, replacing everyone costs about 2.2 times the payroll.
| Level | Share (assumed) | Headcount | Payroll | Engineer-years of agent work | Agent cost | Payroll vs cost |
|---|---|---|---|---|---|---|
| L1 Junior | 27% | 514,458 | $54.0B | 514,458 | $59.8B | |
| L2 Mid | 40% | 762,160 | $118.1B | 1,524,320 | $177.1B | |
| L3 Senior | 21% | 400,134 | $86.0B | 1,600,536 | $186.0B | |
| L4 Staff | 8% | 152,432 | $45.7B | 1,219,456 | $141.7B | |
| L5 Principal | 3% | 57,162 | $27.4B | 914,592 | $106.3B | |
| L6 Distinguished | 0.8% | 15,243 | $10.7B | 487,776 | $56.7B | |
| L7 Fellow | 0.2% | 3,811 | $4.2B | 243,904 | $28.3B | |
| Total | 100% | 1,905,400 | $346.2B | 6.51M | $755.9B |
| Requirement | Doubling rule, all levels | Failure-based, L1–L4 only |
|---|---|---|
| Engineer-years of agent work | 6.51M | 1.56M |
| Agents running 24/7 | 1.54M | 0.37M |
| GB300 NVL72 racks (36 agents each) | ~42,900 | ~10,300 |
| GPUs (72 per rack) | ~3.1M | ~0.7M |
| Capital at ~$5.7M per rack | ~$245B | ~$59B |
| Power (140 kW × 1.3 overhead) | ~7.8 GW | ~1.9 GW |
| New tokens processed per month | ~0.87 quadrillion | ~0.21 quadrillion |
| Line | Per rack / yr | Per agent-hour |
|---|---|---|
| Rack write-off ($5.7M deployed ÷ 3 yr) | $1,900,000 | $6.02 |
| Power (140 kW × 1.3 × $0.10/kWh) | $159,432 | $0.51 |
| Colocation and space (140 kW × ~$200 per kW-month) | $336,000 | $1.07 |
| Rack operations staff | $100,000 | $0.32 |
| Rack subtotal (÷ 36 agents) | $2,495,432 | $7.91 |
| Sandbox VM and CI minutes | — | $0.30 |
| Platform and infrastructure support (1 engineer per 2 racks, $230K) | $115,000 | $0.36 |
| Pure cost per agent-hour | $8.58 | |
| Use at 65%, not 100% (÷ 0.65) | $13.03 | |
| Frontier R&D, 2× extra | $39.10 | |
| Cost per agent-hour (30% vendor margin: ÷ 0.7) | $55.86 |
| Level | 100% of salary | 75% | 50% | 25% |
|---|---|---|---|---|
| L1 Junior | $87,159 | $60,909 | $34,659 | $8,409 |
| L2 Mid | $119,317 | $80,567 | $41,817 | $3,067 |
| L3 Senior | $143,634 | $89,884 | $36,134 | ($17,616) |
| L4 Staff | $157,268 | $82,268 | $7,268 | ($67,732) |
| L5 Principal | $194,537 | $74,537 | ($45,463) | ($165,463) |
| L6 Distinguished | $129,073 | ($45,927) | ($220,927) | ($395,927) |
| L7 Fellow | ($41,854) | ($316,854) | ($591,854) | ($866,854) |
| Agents per rack (at 2X) | 1X compute | 2X compute (base) | 4X compute | 10X compute |
|---|---|---|---|---|
| 15 | $21,133 | $40,884 | $80,386 | $198,890 |
| 30 (no gain over GB200) | $11,258 | $21,133 | $40,884 | $100,136 |
| 36 (base) | $11,258 | $17,841 | $34,300 | $83,677 |
| 45 (full 1.5× compute) | $11,258 | $14,550 | $27,717 | $67,218 |
| 60 (bandwidth cap) | $11,258 | $11,258 | $21,133 | $50,759 |
| Event | Impact |
|---|---|
| Base case economic profit | $12,291,841 |
| Price cut 25% to give clients a reason to switch | ($5,377,980) |
| 20% of rack capacity unsold | ($3,188,945) |
| New GPU generation; racks written off over 2 years, not 3 | ($1,900,000) |
| One agent-caused production incident | ($150,000) |
| Adverse case economic profit | $1,674,916 |
Source: “How Many Agents Make an Engineer” teardown memo (US data, BLS 2025), figures as printed there. Level split and several cost lines are the memo’s assumptions.