Frontier AI Valuation Lab

An interactive valuation and capital-decision studio for illustrative Anthropic and OpenAI scenarios. A discounted cash flow (DCF) model values each company under Bear, Base and Bull cases built on public-data assumptions. The Capital Strategy Studio then compares hypothetical funding rounds: how much capital each one raises, how much of their stake existing holders keep, and how each plan holds up across the cases.

Open first: the worked decision in Capital Strategy Studio. It tests three hypothetical rounds against two goals: raise at least $25B, and let existing holders keep at least 88% of their starting stake. The company model shows the cash flows behind each case.

Run it locally with Node.js 24 (the version in .nvmrc). The test suite also needs Python 3.11 or newer.

git clone https://github.com/ult666666/frontier-ai-valuation-lab.git
cd frontier-ai-valuation-lab
npm test
npm start    # then open http://127.0.0.1:4321/capital.html#worked-decision

The company figures rest on illustrative assumptions and incomplete public disclosures. They are not current valuations of either company or investment advice, and the project is not affiliated with or endorsed by Anthropic or OpenAI. Charles Dai (Carnegie Mellon University) built it with AI assistance.

More: fintech portfolio · engineering case study · walkthrough · worked decision as PDF, HTML or JSON · portable desktop package (ZIP)

Try a complete decision

The worked illustration compares three primary rounds. All figures are USD billions. Quotes are independently entered hypothetical prices. Retention is the fraction of starting ownership kept by an existing holder who does not subscribe for more shares.

Plan	Pre-money quote	Primary capital	Starting stake retained	Meets $25B capital and 88% retention?
A · Smaller round	250	25	90.91%	Yes
B · More capital	300	35	89.55%	Yes
C · Higher retention	300	20	93.75%	No: $5B below the capital objective

1. Define. Open Try the tradeoff. Preview $25B / 90%: only A meets both objectives in the worked example. Preview $30B / 90%: none qualifies. Preview $30B / 88%: B qualifies. Each preview uses your current submitted plans; choose Apply to update the two objectives together.
2. Compare. Apply $25B / 88% again to compare A and B in the worked example. Inspect individual plan markers, exact pricing and the equivalent table. A capital shortfall, an ownership constraint and an invalid input have different meanings.
3. Challenge. Evaluate the same plans against Bear, Base and Bull operating cases, or edit every assumption and continuation horizon. For A and B, the signed modeled existing-cohort allocations cross at $635B of pre-money modeled equity residual, with $600B allocated to the cohort. B allocates more below that switching value; A allocates more above it.
4. Record. Choose a plan deliberately, explain its tradeoff, name information still needed and set a review trigger. Export an exact decision link, JSON record or script-free brief. Imported output is recalculated from the inputs.

Previewing preserves the exact decision, its analyst text and its verification state. Applying a different pair preserves all plan terms and scenarios, recalculates the complete decision, and marks existing reasoning and API verification for review. Editing the draft clears an old preview. The preview helper composes the same validated record and calculation engine used by the rest of the application.

The default hypothetical bridge assumes cash of 20, debt of 5 and no other claims. The default capital objective is an analyst input, not an estimate of either company's cash requirement. See the worked mathematics, numeric conventions and limits.

Four disciplines, one product

Discipline	What the demo makes inspectable	Implementation and evidence
Fintech	Operating EV, modeled equity residual, independently entered financing price, primary proceeds and ownership have separate definitions.	Capital method, financing method, preserved FCFF engine
Quant	Finite-scenario regret, dominance, pairwise switching values and all-scenario model-conditioned pricing intervals. No scenario probabilities are invented.	Pure calculation, independent Python reference, boundary tests
Engineering	Complete state flows through the browser, bounded stateless API, exact share codec, import validation and standalone outputs. Late responses cannot verify newer inputs.	Architecture, API contract, HTTP checks, browser workflow
Business

The selected comparison metric is retention × (modeled pre-money residual + primary proceeds). It allocates a signed modeled residual under equal economic rights, unchanged operations, retained cash and zero fees. It is not an investment return or a limited-liability common-share price.

Only plans meeting the capital and retention objectives enter regret and dominance comparisons. Minimax identifies the smallest maximum gap to the eligible best across the submitted cases and returns all ties. It never changes the analyst's selection. A finite set of cases does not establish global robustness.

The model-conditioned pricing interval is a separate diagnostic. Every submitted scenario must meet the chosen modeled-value-to-quote ratio, while the quote must also preserve the ownership floor. The default Bear residual makes the interval empty for all three entered raises. That conflict remains visible even though A and B meet the capital and ownership objectives.

Valuation under uncertainty

Three cases hide the spread of outcomes between and around them. The Model workspace now has a Monte Carlo panel (open it) and the API a matching simulate action. Both reuse the unchanged DCF engine for every draw.

• Drivers with ranges. Revenue growth path, terminal operating margin, discount rate, terminal growth and, with a hypothetical equity bridge, uncompensated dilution. Each takes a bounded uniform, triangular or truncated normal distribution.
• Correlated draws. A Gaussian copula with a Cholesky factor of your correlation matrix. A matrix that is not positive semidefinite is rejected with its smallest eigenvalue.
• Reproducible. A seeded xoshiro128** generator. The same request returns byte-identical JSON in the browser and the API, which CI checks in Chromium.
• Outputs. P5, P25, P50, P75 and P95, the share of draws above a hurdle with a 95% interval, a tornado of one-way P10 to P90 swings, and convergence diagnostics: the standard error of the mean and the median across five seeds.

With the server running, this request gives the default panel ranges around the Anthropic Base case:

curl --fail-with-body http://127.0.0.1:4321/api/valuation \
  -H 'Content-Type: application/json' \
  -d '{"action":"simulate","assumptions":{"rev2026":53.26,"g1":0.45,"g2":0.24,"g3":0.12,"m27":-0.02,"m30":0.18,"m36":0.22,"turns":2.5,"wacc":0.125,"tg":0.03,"roic":0.2,"stub":-3,"tax":0.25},"bridge":null,
       "drivers":[{"key":"growth","distribution":{"type":"triangular","min":0.7,"mode":1,"max":1.3}},
                  {"key":"m36","distribution":{"type":"triangular","min":0.16,"mode":0.22,"max":0.28}},
                  {"key":"wacc","distribution":{"type":"normal","mean":0.125,"sd":0.015,"min":0.095,"max":0.155}},
                  {"key":"tg","distribution":{"type":"triangular","min":0.02,"mode":0.03,"max":0.04}}],
       "correlation":[[1,0.4,0,0],[0.4,1,0,0],[0,0,1,0.3],[0,0,0.3,1]],"draws":10000,"seed":20261018,"hurdle":300}'

It returns P5 $197.2B, P50 $364.2B and P95 $694.6B around the deterministic Base value of $367.8B, with 69.5% of draws above the $300B hurdle. The growth path and the discount rate swing value most (about $240B and $236B between their own P10 and P90). These figures describe the illustrative ranges in the request; they are not a forecast or a probability that either company is worth any amount. See the method, references and limits and the API contract.

Run the complete stack locally

The commands at the top run the tests and start the server. Node.js 24 is the supported version for serving, building and testing. On an older Node.js, npm test stops with a message asking for Node.js 24; nvm use switches to the version in .nvmrc. The test suite also needs Python 3.11+ for desktop-package verification. The application has zero npm runtime dependencies, and needs no API key, account, database or paid data feed.

To choose another port, use PORT=8080 npm start. The same Web-standard HTTP handler runs locally and in the Vercel function. npm run build creates the public application in dist/.

Python 3.11+ and Node 24 reproduce the independent references:

python3 -B scripts/check_capital_reference.py --output /tmp/capital-reference.json
python3 -B tests/reference_parity.py
python3 -B scripts/check_horizon_reference.py --output /tmp/horizon-reference.json
python3 -B scripts/check_financing_reference.py --output /tmp/financing-reference.json
python3 -B scripts/check_ownership_reference.py --output /tmp/ownership-reference.json
python3 -B scripts/check_thesis_reference.py --output /tmp/thesis-reference.json
python3 -B scripts/check_uncertainty_reference.py --output /tmp/uncertainty-reference.json

The capital reference uses separate arithmetic and declares its tolerances and tested cases. Browser/API agreement is a deployment and state-transport check, not an independent proof of the formulas. Current and historical evidence are distinguished in the verification record.

Make an exact API request

With the local server running, prepare the public worked example's numeric snapshot:

node --input-type=module <<'JS'
import {readFile,writeFile} from 'node:fs/promises';
const record=JSON.parse(await readFile('public/downloads/Frontier_AI_Capital_Decision.json','utf8'));
await writeFile('/tmp/frontier-capital-request.json',JSON.stringify({action:'capital',snapshot:record.snapshot}));
JS
curl --fail-with-body http://127.0.0.1:4321/api/valuation \
  -H 'Content-Type: application/json' \
  --data-binary @/tmp/frontier-capital-request.json

A successful response includes appVersion, modelVersion, valuationDate and the complete capital analysis. The numeric request excludes analyst prose. The endpoint accepts 1–9 complete scenarios and 1–6 plans, rejects unsupported fields, validates every input, and retains the existing 16 KiB streamed request limit. See all endpoints, boundaries and status meanings.

The Studio distinguishes exact verification, a qualified rounding comparison, and a mismatch. Browser and server math libraries can approximate finite-continuation functions differently. The qualified state permits a small relative difference only in named modeled outputs; every input, quoted term, eligibility result and comparison status must remain exact. It does not change the current record or exports. See the comparison policy, pure comparator and captured-runtime regression tests.

Continue into the financial foundations

The original research workflow remains connected through seven navigation destinations: Capital Studio, Overview, Model, Thesis Lab, Value Horizon, Casebook and Evidence.

Workspace	Question	Open
Overview and Model	What operating assumptions and annual cash flows support the valuation?	Company comparison, complete model
Financing Lens and Ownership Frontier	How does one hypothetical primary round translate price into cash and ownership?	Worked ownership example, method
Valuation under uncertainty	What range of value do ranges on the main drivers imply, and which driver matters most?	Monte Carlo panel, method
Value Horizon	How much value depends on cash flows after the explicit forecast?	Finite continuation tool, method
Thesis Lab	Which declared assumption groups account for the difference between two operating cases?	Exact attribution and target landscape, method
Research Casebook

Value Horizon keeps the explicit forecast fixed and retains 0–100 annual continuation cash flows after 2036, with zero residual after the chosen cutoff. Thesis Lab uses all 64 group combinations for exact grouped Shapley attribution across six declared groups, equivalent to averaging 720 orders. The Casebook preserves existing operating-case and comparison schemas; capital decisions have their own record type.

Research scope and method sources

The company data has a valuation date of 18 October 2026. The capital decision tools were added on top of the financial engine, the financing and ownership engines, the six company presets, the dated research register, the original Excel workbook and the original review PDF without changing them. Versions and dated evidence are in the GitHub releases.

• RAND: Robust Decision Making informs the separation of decision levers, explicit futures and performance measures, and the caution against interpreting scenario counts as probabilities. This demo is a finite comparison, not a complete RDM implementation.
• Aswath Damodaran: Probabilistic Approaches informs coherent scenario assumptions and the distinction between scenario-specific values and probability-weighted expectations.
• HM Treasury: The Green Book (2026) informs explicit objectives, alternative options, sensitivity and documented judgment. Its public-sector appraisal conventions are not used as private-company financing rules.

Financing examples do not establish either company's balance sheet, cap table, funding need or available offers. The simple primary issuance excludes transaction fees, preferences, options, conversions, secondary sales, spending-related changes to operations and later rounds. Economic FCFF is not a treasury cash schedule. Negative modeled values remain signed.

Review and deliverables

• Verification and limits lists the local test runs, Python reference checks, browser checks and AI-assisted reviews behind each release.
• Architecture and trust boundaries explains exact state, validation, API parity, portable records and deployment.
• Desktop guide explains the versioned package and local reproduction.
• Reviewed Excel model and original financial review retain the reviewed baseline. The workbook is a committed file. Its generator, scripts/build_workbook.mjs, needs @oai/artifact-tool, which comes from the Codex agent environment used to build this project and is not on the public npm registry, so visitors cannot rerun the generator.
• GitHub releases preserve dated handoffs and downloadable evidence.

Code is MIT licensed; linked third-party research remains with its publishers.