DARPA Wants AI That Invents Algorithms to Become Ordinary Software. SPEED DIAL Is a $2M Direct-to-Phase-II STTR Closing September 23.

Most AI-for-science funding in 2026 is aimed at discovery. DARPA's SPEED DIAL is aimed at the part that comes after discovery and almost never gets funded: making the thing usable by someone who was not in the room when it was invented.

DPA26TZ05-DV003 — Scalable Platform for Enterprise Engineering and Deployment towards Mathematics for the Discovery of Algorithms and Architectures (SPEED DIAL) is the single STTR topic in DARPA's FY26 SBIR Release 5. It pre-released August 5, opened August 26, and closes September 23, 2026 at 12:00 PM ET, with technical questions cut off September 16.

It is also, structurally, one of the more demanding topics DARPA has posted this year — and the demands are almost entirely in the entry requirements rather than the technical work.

The parent program: DIAL

SPEED DIAL is downstream of DIAL — Mathematics for the Discovery of Algorithms and Architectures, a Defense Sciences Office "Disruption Opportunity" run by program manager Yannis Kevrekidis. DIAL's premise is a direct assault on how numerical methods have historically come into existence.

The Kalman filter exists because Rudolf Kálmán had an insight in 1960. Wavelets exist because a sequence of mathematicians over decades converged on a representation. Every workhorse numerical algorithm in engineering arrived the same way: human inspiration, plus a great deal of serendipity, plus time measured in careers.

DIAL replaced that with computer-aided algorithm discovery via optimization — treating the search for a good algorithm as a search problem a machine can run. The program targeted defense-relevant domains that break conventional methods: high-speed flow dynamics, advanced materials design, and supply chain risk analysis, all characterized by multiple spatial and temporal scales with tangled physics.

The validation results are what make SPEED DIAL credible rather than speculative. DIAL systems rediscovered the Kalman filter using Transformers. They rediscovered wavelets using genetic programming. They generated optimal meta-solvers for physics simulations. DARPA's own topic materials cite a 1000x improvement from AI-discovered wavelets as a reference point for the magnitude of gain being claimed.

Rediscovery is the right validation strategy, incidentally, and it is worth appreciating why. If your system invents something nobody has seen, you cannot easily tell whether it is brilliant or broken. If your system independently arrives at the Kalman filter — a method the field spent sixty years confirming is correct — you have evidence the search process finds real mathematics rather than overfitted noise.

So the science is demonstrated. The problem SPEED DIAL exists to solve is that essentially none of it is usable by a working engineer.

What SPEED DIAL actually funds

The topic asks for a framework and tool suite that turns algorithm discovery from a research artifact into "enterprise-grade" infrastructure — with the explicit goal of making algorithm generation "as routine and indispensable as simulation or data analysis."

The required platform components are specific:

• A unified discovery and deployment framework — one system that both finds algorithms and ships them.
• Embedded discovery engines using Transformers, genetic programming, or reinforcement learning.
• A version-controlled algorithm library organized by problem class. Algorithms as managed artifacts with provenance, not one-off outputs.
• In-context integration tools for deploying discovered algorithms into MATLAB, Simulink, COMSOL, or C++.
• Interpretability, mandatory. DARPA states it flatly: "Discovered algorithms cannot be black boxes."

That last requirement deserves emphasis, because it quietly excludes a large swath of the current AI-for-science field. A learned surrogate model that predicts the right answer without exposing why is not responsive to this topic. DARPA wants discovered algorithms — inspectable procedures an engineer can read, reason about, bound the error on, and defend in a design review. In a defense context that is not a philosophical preference. An algorithm that goes into a flight control loop or a materials qualification pipeline has to be auditable, and nobody certifies a black box.

The MATLAB/Simulink/COMSOL integration list tells you the intended user just as precisely. This is not built for machine learning researchers. It is built for the mechanical, aerospace, and structural engineers who live inside those environments — the industrial base that would never adopt a discovery engine that required them to learn a new Python framework first.

The money and the milestone clock

Note the unusual shape: the option is larger than the base. Most Phase II structures front-load. Here DARPA funds a smaller proof year and reserves the larger tranche for scale-up, which means year one is effectively an extended audition and the milestone schedule is how you pass it.

The deliverables are fixed and unusually granular for an STTR:

Two of these are worth reading closely. The month-14 gate demands a quantified >10% efficiency gain on a defense problem — a hard number on a real workload, not a benchmark suite. And the month-24 gate requires a second application, which is DARPA testing generality directly. A platform that works beautifully on the one problem you tuned it for is a consulting engagement, not infrastructure. The second-application requirement is the difference.

Also note month 4. Five foundational algorithms in the library within a third of a year is only achievable if you arrive with an existing corpus. This schedule is not written for a team starting from a blank repository.

The feasibility gate is where most teams fail

SPEED DIAL is Direct to Phase II only. There is no Phase I, and DARPA requires feasibility documentation showing three accomplishments already achieved outside the STTR program before proposal submission.

To be responsive, that documentation must demonstrate:

1. A track record using algorithmic discovery methods — you have actually run these techniques, not read about them.
2. Quantitative performance gains, benchmarked against the 1000x-scale improvements DARPA references.
3. Evidence that discovered algorithms are interpretable, not opaque.

This is a narrow filter. The population of organizations holding all three is roughly: current and former DIAL performers, a handful of academic groups in scientific machine learning and symbolic regression, a few national-lab-adjacent teams, and the small number of companies commercializing automated scientific discovery.

If you do not have prior results in hand, no amount of proposal quality closes the gap. A Direct-to-Phase-II feasibility gate is a screening step, and reviewers apply it before they engage with the technical merit of your approach.

The university partnership is the leverage point

STTR's defining requirement is a formal partnership with a research institution — at least 30 percent of the work must be performed by a university or federal lab, against SBIR's rule that the small business perform the majority. The small business must have fewer than 500 employees; the topic is coded to NAICS 541512 (Computer Systems Design Services).

For most STTR topics the academic partnership is friction to be managed. For SPEED DIAL it is the entire team design, and DARPA says so — the topic explicitly calls for "a university-company partnership" pairing algorithmic discovery expertise from academia with industrial deployment capability from the small business.

That split is not arbitrary. It maps precisely onto why DIAL's results have not reached engineers. Universities produce the discovery methods and have zero incentive to build deployment infrastructure. Nobody gets tenure for writing a COMSOL plugin or maintaining a version-controlled algorithm library. Companies can build enterprise software but do not have the applied-mathematics depth to make novel discovery engines work. The gap between DIAL's results and a usable product is structural, and DARPA is using STTR's mandatory-partnership mechanic as the instrument to close it.

The practical implication for team-building: the university partner should own the discovery engines and the interpretability guarantees; the small business should own the framework, the integration layer, the library infrastructure, and the commercialization path. A proposal where the small business is a pass-through for university work will read as non-responsive to the deployment half of the topic, and one where the academic partner is a nominal 30 percent letterhead contributor will read as non-responsive to the discovery half.

Proposal mechanics

The technical volume splits in two, which is characteristic of Direct-to-Phase-II topics:

• Part One — Feasibility: 10 pages maximum. This is where the three prior accomplishments live.
• Part Two — Technical Proposal: 20 pages maximum.
• Commercialization Strategy: 5 pages, separate, not counted against the limits.

Seven volumes are required in total: cover sheet, technical volume, cost volume, Company Commercialization Report, supporting documents, Fraud Waste and Abuse training, and the Disclosures of Foreign Affiliations webform. Submission runs through DSIP. TABA is available at up to $6,500 for Phase I and up to $25,000 for an initial Phase II. Cost sharing is permitted but not required — and on a topic this competitive, voluntary cost share signals commitment without buying you evaluation credit, so weigh it against your runway rather than treating it as a scoring lever.

Allocate your page budget accordingly. Ten pages to establish feasibility is tight when you need three documented accomplishments with quantitative evidence, and a weak Part One makes Part Two irrelevant.

Where this sits in the 2026 landscape

SPEED DIAL is one node in a federal push toward automating the machinery of science itself. The DOE Genesis Mission SBIR call funds AI-driven autonomous laboratories and AI-enabled materials design on a September 10 deadline. NSF is standing up programmable cloud laboratories. DARPA's own MATHBAC topic targeted agentic AI for mathematics earlier this year.

The distinction worth holding onto: MATHBAC and DIAL are about making the discovery work. SPEED DIAL is about making the discovery deployable. Those are different companies, different skill sets, and different proposals — and the deployment problem is chronically underfunded relative to how much it determines whether any of the discovery investment ever produces value.

For a small software firm with genuine scientific-computing depth and an academic partner in symbolic regression or scientific ML, this is an unusually clean fit: $2 million, a defined milestone path, a named Phase III transition requirement, and a customer base of engineers who already have the target software installed.

For everyone else, the three-accomplishment feasibility gate is the answer, and it is worth being honest about it before spending September writing.

Proposals close September 23 at 12:00 PM ET. Technical questions close September 16.

Want STTR and SBIR topics screened against your team's actual publication and results history — including the feasibility gates that decide eligibility before merit review begins? Granted tracks federal solicitations across every agency and flags the ones you can genuinely win.