What would it have taken to build Econvitals before AI?
At its launch in October 2026, Econvitals ran on about 484,000 lines of code, all of it written with AI. So Jason asked a simple question: what would it have cost a company to build the same thing the old way — with people writing every line of code, though still free to use AI to draft the prose?
What was in it
At launch the code fell into six pieces: a data layer that pulls about 110 sources straight from U.S. and foreign statistical agencies, central banks and household surveys; a library of about 790 charts, with a system that redraws them as soon as new data come out; the daily, weekly and monthly briefings; the site itself, including the Lab and the tools; Fedspeak, which tracks and scores what Fed officials say; and the GDP and PCE estimates and fiscal projections. About a fifth of the code was automated tests.
Two ways to count
By lines. Averaged over a whole project — design, meetings, debugging, testing — a developer produces only about 10 to 50 finished lines a day. At 25 a day, 484,000 lines is about 84 person-years. Real projects also change direction: of the 566,000 lines written for Econvitals before launch, about one in eight had already been deleted, and the time spent working out what to build shows up in no count at all. Adding 20–50% for that gives about 115 person-years.
By task. As a cross-check that ignores lines, the table below estimates how long a competent professional would take to build each piece, knowing exactly what to build.
| Piece | Lines of code | By lines | By task |
|---|---|---|---|
| Data layer | 164,104 | 28.5 | ~10 |
| Charts and release updates | 156,011 | 27.1 | ~12.5 |
| Briefings | 72,863 | 12.7 | ~6.5 |
| Fedspeak | 44,755 | 7.8 | ~4.5 |
| The site, Lab and tools | 32,651 | 5.7 | ~6 |
| Estimates and projections | 13,580 | 2.4 | ~3 |
| Hosting, monitoring and alerts | — | — | ~2 |
| Total | 483,964 | 84 | ~45 |
With the same allowance for changing direction, that comes to about 60 person-years. The first method probably runs high, because AI writes wordier code than a person would. The second probably runs low, because people always underestimate how long things take.
Best guess: about 85 person-years, with a plausible range of 50 to 150.
What that means for a company
Spread over a normal three-year build, that is about 28 engineers, data scientists and analysts. Add economists to decide what matters and check what the AI writes, plus managers and administrative staff, and the result is a company of 45 to 55 people. At about $250,000 a person-year for experienced Boston-area staff, the build labor alone comes to roughly $20 million (range $12–38 million), and the whole company for three years to $30–35 million. One person working alone would need about 85 years.
For scale
| Software | Lines of code |
|---|---|
| A simple mobile app | 20,000–50,000 |
| Photoshop 1.0 (1990) | ~128,000 |
| Linux kernel 1.0 (1994) | ~176,000 |
| Space Shuttle flight software | ~400,000 |
| Econvitals at launch | ~484,000 |
Instead, Jason built Econvitals with AI, mostly in the five months before launch, joined in September by two Harvard College students, Sriyutha Morishetty and Amar Besic. Lines of code are a weak yardstick, and the two methods differ by a factor of two. But the order of magnitude seems right: a funded startup's three years of work, done in a few months.
How this was counted: code files only (Python, JavaScript, shell scripts, SQL and the like) kept in version control at launch; web pages, settings files and downloaded data are excluded. The deleted-code figure comes from the full version history. The rates, the task estimates and the cost per person-year are rules of thumb, so treat every number as an order of magnitude.