Part 4: assumptions and sources
Financial figures and definitions
Anthropic’s S-1, printed p.158, Quarterly Consolidated Statements of Operations, reports unaudited interim results in millions. Q1 2026 revenue is directly disclosed as $4.731bn. Q2 revenue is $11.555bn, total operating expenses $10.581bn and operating income $974m. The expense components are compute/infrastructure $5.558bn, personnel $1.857bn, Bartz settlement zero, sales/marketing/partnerships $2.135bn, and other operating/administrative $1.031bn.
H1 revenue is $16.286bn, operating loss $693m, and operating cash inflow $3.929bn. Cash, cash equivalents and short-term investments total $131.651bn at June 30, not at the October article date. The cash-flow statement F-13 includes receivables -$4.982bn, prepaid/other assets +$2.094bn, payables/accruals/other liabilities +$2.679bn, deferred revenue +$1.380bn, and noncash adjustments. These selected items help explain collection/payment timing; they are not a complete reconciliation or evidence that advance payments explain the whole difference.
Revenue is sales recognized during a period. Operating income subtracts operating expenses; it is not net income or cash available for lenders. Deferred revenue represents payments received before the corresponding revenue is recognized. An annualized run-rate extrapolates a shorter period and is not a reported annual total.
Holding Q2 operating expenses fixed, $0.974bn / $11.555bn = 8.4293%: the revenue reduction that eliminates that quarter’s operating cushion. Variable costs and future commitments can change. This is not a default threshold.
Reproducible scenario model
All money below is nominal USD billions. Baseline Q2 2026 recognized revenue: 11.555. Model periods: Q3 2026 through Q2 2029. As of October 27, Q3 has already elapsed; the original model row is a sensitivity, not an observed result. The approximate Q3 anchor below is derived from a rounded narrative, not an exact disclosed result. These paths illustrate a three-year horizon from the selected baseline, not a newly observed October run-rate.
For elapsed years t, revenue = 11.555 × adopting-pool ratio × category paid-conversion ratio × normalized paid-usage ratio × Anthropic realized-price ratio × Anthropic workload-share ratio. Conversion means paying the category, not converting to Anthropic; otherwise including company share again would double-count capture. Price must use the same normalized workload and quality/mix definition as usage.
The pool coordinate is assumed to start at0.60 and approach0.90:
A(t) = 0.90 / (1 + 0.5 × exp(-k × t)); pool ratio = A(t)/0.60; k = 3g.
This is not measured global or Anthropic penetration, nor the tracker’s60.8% adult-use observation. It is a chosen saturation sensitivity with at most1.5× pool expansion. Histories inform slowing shape and broad rate comparisons, not empirically fitted company forecasts.
| Path | Initial pool pace g | Modeled first-year pool growth | Annual conversion factor | Paid-usage factor | Realized-price factor | Share factor |
|---|---|---|---|---|---|---|
| Internet-like diffusion | 10% | 9.4564% | 1.02 | 1.25 | 0.85 | 1.00 |
| Faster AI integration | 20% | 17.7019% | 1.08 | 1.50 | 0.90 | 1.05 |
| Price-compressed growth | 6% | 5.8100% | 1.00 | 1.20 | 0.70 | 0.90 |
All factors are author-chosen assumptions, not historically measured AI price, conversion, share or usage estimates. Initial pool paces sit within a broad comparison with household-internet share growth3.44% to20.35%; they are not fitted estimates from those points. The0.60/0.90 choice is also an assumption. Rates slow as the pool fills. Other factors compound only over this finite scenario horizon, with no assigned probability.
| Revenue measure | Internet-like | Faster | Price-compressed |
|---|---|---|---|
| Calendar2027, four projected quarters | 55.9555 | 90.0691 | 36.0010 |
| Calendar2028, four projected quarters | 65.0712 | 151.6014 | 28.5858 |
| Four quarters ending Q2 2029 | 69.7711 | 193.6647 | 25.4127 |
| Q2 2029 ×4, reference only | 73.3399 | 229.7597 | 23.1971 |
Feasibility: underlying paid-conversion and share fractions cannot exceed100%. Faster requires baseline category paid conversion no greater than79.3832% and baseline Anthropic workload share no greater than86.3838%; these bounds are calculated conditions, not estimates of actual starting shares. It also requires serving about6.8189× normalized paid workload by Q2 2029. Efficiency and product mix prevent translating that into a GPU count. Supply, customers’ ability to pay, mix definitions and recognition timing must hold. There is no capacity-feasibility, future-cost, free-cash or creditor-coverage model.
Sensitivity: comparing the price-compressed case with the faster case at Q2 2029, pool/conversion/usage/price/share ratios are0.838370/0.793832/0.512000/0.470508/0.629738. Their product0.100963 equals the two quarterly revenue ratio. Price is the largest individual chosen contrast; conversion, usage and share together have a larger effect. This decomposition is arithmetic, not estimated causal elasticity. A1% multiplicative improvement in an annual factor for3years produces1.01³=1.030301, not a universal one-percentage-point growth adjustment.
Historical evidence and latest-data cutoff
- Rail, Historical Statistics Q274 to Q283, printed p.727/PDF p.25:1882 to1890 ton-miles39,302m to79,193m(+101.4987%); freight revenue$485m to$734m nominal(+51.3402%); realized price1.236 to0.927 cents per ton-mile(-25%). Rounded volume/price products need not exactly reproduce the separately published revenue column. This teaches usage×price, not an AI growth estimate.
- Census household internet, Figure1, printed/PDF pp.1 to2:18.0%1997,54.7%2003,71.7%2011. Annualized adoption-share growth20.3518% then3.4407%; households, not individuals/company revenue.
- 1930 Census manufacturing power, Table2, printed p.112/PDF p.2: electric-motor horsepower shares5,12,26,40,56,67,73,78,82% in1899,1904,1909,1914,1919,1923,1925,1927,1929. The denominator is the table’s aggregate rated manufacturing horsepower, not establishments, output or GDP. Table1 footnote1 changes the value-of-products coverage threshold from$500 in1919/earlier to$5,000 in1923/later. No unmarked continuous fitted curve.
- Bick/Blandin/Deming, November13,2025: revised August2024 adult18to64 use44.6%, August2025 54.6%; matched-launch internet30.1%; work use37.4%; work-hours share5.7%. Overall use is not paid/daily/work use. Do not insert the unrevised39.4% baseline.
- Authors’ tracker, dated cards: August2026 adult use60.8%, employed work use40.5%, work-hours share5.3%. Definitions/revision history behind the chart are not fully audited. Separate vintage, not an unmarked continuous curve with the2025 article.
- Fed April3,2026 and Census May26,2026 explain the November2025 BTOS question broadening. More current National.xlsx is dated October 22, 2026: October22 release, collected October5to18, reference September21toOctober4; Response Estimates E16=24.6% current use, E82=28.6% expected next-six-month use. Current workbook labels say SI/formerlyAI; application text says wording updated starting the October19 collection. Thus labels are not proven identical across every historical response column. These values stay in this note rather than the main narrative. This is the evidence cutoff, not a claim that every survey uses comparable definitions.
Fresh company pages also support May28 run-rate above$47bn, April20 AWS commitment above$100bn over ten years, November2025 Azure$30bn commitment, and Google/Broadcom capacity expected from2027. They do not supply a complete reconciled annual cash schedule. Do not sum potentially overlapping deals, treat an arithmetic average as annual installments, deduct the$63bn financing headline from quarterly revenue, or deduct already-expensed leases twice.
Approximate Q3 reanchoring sensitivity
The filing’s rounded 86-times narrative implies approximately $16.94bn for Q3. This is a derived approximation, not an exact disclosed quarterly result. It is used here only as a sensitivity anchor. Original Q3 2026 projections have elapsed and are not October forecasts.
Method: for each original path, multiply all Q4 2026 and later revenues by 16.94 / that path’s original modeled Q3 revenue. Equivalently R_new(q) = 16.94 × R_old(q)/R_old(Q3). Keep calendar dates and subsequent relative growth unchanged; do not reset the diffusion clock or fit new growth rates. This is a level-shift sensitivity, not evidence that the driver assumptions are now calibrated. No probability or lender-coverage inference. Future cost/cash schedules remain absent.
| Path | Old modeled Q3 | Scale | Q4 2026 | FY2027 sum | FY2028 sum | Four quarters to Q2 2029 |
|---|---|---|---|---|---|---|
| Internet-like diffusion | 12.08 | 1.4022 | 17.69 | 78.46 | 91.25 | 97.84 |
| Faster AI integration | 13.48 | 1.2568 | 19.67 | 113.20 | 190.53 | 243.39 |
| Price-compressed growth | 10.93 | 1.5492 | 16.02 | 55.77 | 44.28 | 39.37 |
All dollar amounts in billions, nominal. Full-year sums use four future quarters, not annualized exit revenue. No FY2029 total is claimed. Baseline approximation uncertainty carries proportionally into every output. The higher starting revenue does not prove more payers, higher prices, capacity utilization or debt-service ability.
Limits of the exercise
The paths are chosen sensitivities, not probability-weighted forecasts. Historical adoption informs the slowing shape, not measured company-specific growth rates. Empirical calibration, delivered capacity, customers’ ability to pay, future costs and a full matched cash-payment schedule remain absent. Reanchoring changes the revenue level, not those missing facts. Neither version establishes lender repayment capacity.