A 1957 language, still underneath NumPy, still underneath climate models, still underneath the machines that rank the world’s fastest supercomputers.
Fortran shipped its first compiler in April 1957, built by John Backus’s team at IBM for the IBM 704. Nearly seventy years later, it hasn’t been replaced so much as buried, one abstraction layer at a time, under languages people reach for first. Python calls it. MATLAB calls it. The supercomputers on the TOP500 list are still ranked using a benchmark descended from a Fortran library. It’s tempting to call this legacy inertia, but the more accurate word is closer to infrastructure: a layer nobody rewrites because everything above it still depends on it working exactly the way it already does.
The Original Bet, and Why It Never Really Left
Backus’s team built what’s generally considered the first optimizing compiler, and the bet it made, that a high-level language could generate code as fast as a human wrote by hand, is the reason Fortran’s numerical libraries never needed replacing on performance grounds. A 2021 Nature retrospective on the ten computer codes that most transformed science named both the original FORTRAN compiler and BLAS, the Basic Linear Algebra Subprograms, among its ten. Jack Dongarra won the 2021 Turing Award specifically for the numerical libraries he built across four decades, EISPACK in 1974, LINPACK, BLAS, LAPACK in 1992, and ScaLAPACK in 1993, all of them Fortran, all of them still the substrate under most of today’s numerical software.
What’s Actually Underneath NumPy and SciPy
NumPy’s linear algebra routines call LAPACK, and LAPACK is still Fortran by default in most builds. SciPy goes further: Pearu Peterson’s f2py tool, built specifically to wrap Fortran subroutines for Python, is what makes libraries like MINPACK, QUADPACK, and FITPACK, all originally written in Fortran 77 and distributed through Netlib decades ago, callable as ordinary Python functions. The scale of this isn’t a guess. SciPy’s own 1.0 paper, published in Nature Methods, reports GitHub’s own code analysis of the project’s source tree directly.
A quarter of SciPy’s own codebase is Fortran, sitting directly underneath a library most of its users have never had a reason to think of as anything but Python. SciPy’s maintainers have an open, active effort to port some of this legacy Fortran 77 code to C or Python where it’s safe to do so, tracked publicly on GitHub, which is itself evidence for how slow and deliberate that process is: individual routines get ported one at a time, over years, precisely because nobody wants to risk reintroducing a bug that took decades to shake out of a numerical kernel.
Climate Models and HPC: the Same Story at a Bigger Scale
The pattern repeats, at greater scale, in climate and weather modeling, where Fortran remains the dominant implementation language for the physics cores of major models. It’s also why the German government’s Sovereign Tech Fund, a program under the Federal Ministry for Economic Affairs and Climate Action, gave the Fortran-lang organization funding equivalent to three full-time positions for six months in late 2022, explicitly citing the language’s importance to climate modeling as the justification. On the hardware side, the benchmark used to rank the TOP500 list of the world’s fastest supercomputers, High-Performance LINPACK, descends directly from Dongarra’s original Fortran LINPACK package. The tool used to measure the world’s fastest machines is itself a descendant of 1970s and 80s Fortran numerical code.
The Comeback Nobody Predicted: TIOBE’s Numbers
None of this shows up as nostalgia in the numbers that are supposed to track programming language popularity. Fortran spent roughly a decade outside the TIOBE index’s top 20 before bottoming out at position 50, its lowest ranking since TIOBE started tracking languages in 2001, in July 2020. What happened next runs the opposite direction.
By July 2023 Fortran had climbed to 11th place. By November 2024 it reached 8th, its best position since TIOBE started measuring in 2001. TIOBE’s own commentary on this trend, alongside COBOL’s and Delphi’s similar climbs, points to a specific cause: the engineers who wrote and understand these codebases are retiring, and the organizations that depend on them are choosing to extend and maintain what already works rather than gamble on a rewrite. That’s not a story about Fortran being rediscovered for its elegance. It’s a story about risk management at institutions that can’t afford to be wrong.
A Living Standard, Not a Fossil
The language underneath all of this has kept moving. The original 1956 Fortran programmer’s manual ran 54 pages; the Fortran 2023 standard, published by ISO that November, runs 681 pages, and a Fortran 202Y standard is already in development, with generics-style templates as its most requested feature. The Fortran-lang community, formed at the end of 2019, built a package manager, fpm, modeled directly on Rust’s Cargo, and an interactive compiler, LFortran, that can run inside a Jupyter notebook. The organization has taken part in Google Summer of Code every year since 2021. None of that reads like a language coasting on inertia; it reads like a maintenance-critical dependency finally getting the tooling investment its actual usage has quietly justified for decades.
Why Fortran keeps getting wrapped instead of replaced
- Rewriting a validated, decades-old numerical kernel risks reintroducing bugs that took years to eliminate the first time
- Array-oriented syntax and decades of vectorization-focused compiler work still make it genuinely fast for dense numerical loops, not merely familiar
- The standard keeps evolving, Fortran 2023 and the work underway on 202Y, while staying compatible with code written in the 1980s
- The real operational risk is a shrinking pool of engineers who understand the original codebase, which is exactly what’s driving renewed tooling and hiring interest, not the language’s technical limits
What We Learned
Fortran’s second life isn’t a revival so much as an admission of what never actually left. NumPy and SciPy call into Fortran-derived LAPACK and Fortran 77 libraries wrapped through f2py, a quarter of SciPy’s own source tree is Fortran by GitHub’s own measurement, and the benchmark used to rank the world’s fastest supercomputers descends from a Fortran package Jack Dongarra wrote decades ago. TIOBE’s numbers, a climb from 50th place in 2020 to 8th in 2024, track something more specific than nostalgia: institutions extending and maintaining code they can’t safely afford to rewrite, while the language itself keeps growing, 681 pages of standard now against 54 in 1956, and picking up the kind of package manager and tooling investment a genuinely maintained platform gets, not a fossil.
