Practice · graded on a filing
Build it from NVDA’s latest annual filing. Every line you enter is checked by code against the filing.
Compound a base unlevered free-cash-flow figure at a stated growth rate, discount each year at a stated discount rate, take a Gordon terminal value, and sum to an implied enterprise value. Every cell is arithmetic; the model stops at enterprise value and never reaches a per-share number. Every figure you are given comes from NVDA’s latest annual filing and is cited below; the assumptions are handed to you and are never graded. Nothing here is a view on the company.
Match the AI
An AI model worked this exact drill and got 9 of 9 lines right. Match it.
Claude Sonnet 5, given the same filed figures and stated assumptions you see, scored by the same grader. One run, on 2026-09-28, not retried. It does arithmetic on this filing only and never forms a view on the company.
Reconcile NVDA's model from the filed figures below (as of 2026-01-25). Derive each linked line to the dollar — the model must reconcile end to end.
NVIDIA CORP 10-K (FY2026, period ended 2026-01-25), accession 0001045810-26-000021, via SEC EDGAR
Educational use only — not investment advice.
Project the company's unlevered free cash flow for a few years, discount each year back at the weighted average cost of capital, add a terminal value for everything after the projection, discounted the same way, and the sum is enterprise value. Unlevered, because the cash flow belongs to debt and equity holders together, which is why the rate is the WACC.
Comes up inIB first-round technicalsIB superdayModeling testEquity research case
Operating cash flow, interest expense, the effective tax rate and capital expenditures, as filed in the latest annual report, bridged to unlevered free cash flow (CFO + after-tax interest − CapEx). Each cell is checked against the value the same pipeline derives from that filing, inside a small tolerance band, and the model has to reconcile end to end. There is no language model in the grade — it is arithmetic over filed figures, so it returns the same verdict every time. The correct values are not in this page: they are recomputed on the server when you submit, and your entries are read as a submission, never as the answer.
Every attempt ends in a review. Each line is marked Best, Inaccuracy, Mistake or Blunder by how far it is off, and the first line that went wrong is flagged, so an error carried through the rest of the model is fixed once, where it started.
Signed out, nothing is recorded: work it as many times as you like. Signed in, each graded attempt is saved to your account so your dashboard can track it.
Each one is built from that company's own annual filing and graded the same way.
Educational use only — not investment advice. Figures come from public SEC filings; Echelon teaches you to analyze data, it never recommends buying or selling any security.