DARPA Pre-Released Three FY26 SBIR Topics on July 1 — Two Are AI Forecasting Problems, One Is a Memory Chip That Survives 600°C. The Real Deadline Is Sooner Than August 19.

On July 1, 2026, DARPA pre-released three new SBIR topics as part of Release 4 of the FY2026 Department of War SBIR Broad Agency Announcement. They open for proposals on July 22 and close on August 19, 2026 at 12:00 PM ET. That is the schedule every headline will repeat. It is also the schedule that quietly misleads first-time applicants into losing before they start.

The date that actually decides who wins is July 22 — the open date, not the close. Between the July 1 pre-release and the July 22 open, applicants can email the topic author directly with technical questions. Once the topic opens, that channel goes dark: no contact with the government is permitted until after the selection is made. For a DARPA SBIR, the pre-release Q&A window is the single most valuable and most underused asset in the entire process. You have roughly three weeks. This is the deep dive on the three topics and how to use those weeks.

The three topics

Art of Novel Signals: Predicting and Forecasting with High Confidence (topic DPA26BZ03-DV015) asks for a predictive/forecasting model that leverages Temporal Knowledge Graph Forecasting using In-Context Learning from novel, multilingual, and multimodal data. Strip the jargon and the intent is clear: DARPA wants to forecast events — geopolitical, logistical, adversarial — by reasoning over a graph of entities and relationships that changes through time, and it wants that reasoning to come with a calibrated confidence estimate rather than a single point guess. The phrase "with high confidence" is doing real work. Anyone can produce a forecast; DARPA is funding the ability to say how much to trust it.

Fusion of Abstract Learning and Context-Optimized Neural-methods (FALCON) (topic DPA26BZ03-DV016) is the sister problem. It asks applicants to combine machine-learning methods that are computationally efficient on structured data with large language models that are general and can extract context from unstructured data — the goal being interactive statistical analysis at scale. This is the "analyst copilot" pattern rendered as a research problem: fast, cheap models do the numerical heavy lifting on tables and time series; an LLM handles the messy human-language framing, the "what am I actually looking at" layer. FALCON is a bet that the future of defense analytics is neither pure LLM nor pure classical ML, but a disciplined fusion of the two.

Non-Volatile Memory for Extreme Environments (topic DPA26BZ03-DV017) is the outlier — and the most concrete. The objective is a co-packaged, temperature-hard, and radiation-tolerant NOR Flash memory system that survives from −269°C to +600°C. That temperature floor is a few degrees above absolute zero; the ceiling is hot enough to melt lead twice over. This is memory for spacecraft, for down-hole sensors, for hypersonic vehicles, for anywhere a conventional chip would either freeze into unreliability or cook itself. There is no LLM anywhere near this problem. It is a materials-and-packaging problem, and it will be won by a semiconductor team, not a software team.

Taken together, the three topics tell you something about DARPA's FY26 posture: two-thirds of this release is AI, but the AI it wants is decision-grade — forecasting with confidence intervals, analysis a human can interrogate — and the remaining third is the unglamorous hardware that keeps the whole edifice running where it's cold, hot, or bathed in radiation.

Why the pre-release window is the whole game

DARPA SBIR topics are written by a program manager who already has a mental picture of the technical approach that would excite them. The one-paragraph topic description you can read publicly is the compressed, declassified surface of a much richer intent. The pre-release Q&A period exists so you can probe that intent before you commit weeks to writing.

Use it deliberately. Good pre-release questions are specific and reveal that you've done the reading: Is the confidence estimate expected to be calibrated against a held-out event set, or is a self-reported uncertainty acceptable? For FALCON, does "structured data" presume tabular inputs, or should the fusion layer handle graph-structured inputs as well? For the NOR Flash topic, is the −269°C to +600°C range a single-device requirement, or can the co-packaged system partition across the range? Questions like these do two things: they get you an answer that sharpens your proposal, and — because topic authors remember who engaged thoughtfully — they establish you as a serious team before a single evaluation score is assigned.

Vague questions ("Can we apply?" "What's the budget?") waste the one contact you get and signal that you didn't read the solicitation. Once July 22 arrives, the topic author cannot speak to you at all. Every question you didn't ask becomes an assumption you're gambling on.

The Phase I math and the Direct-to-Phase-II question

DARPA runs its SBIR through the DoD SBIR/STTR Innovation Portal (DSIP), and proposals for these topics submit there. Under the FY2026 program structure — reshaped by the SBIR reauthorization enacted earlier this year — Phase I feasibility awards fund the proof-of-concept stage, with Phase II carrying the substantial development money. For the fast-moving deep-tech firm, the more important question is whether any of these three topics offer a Direct-to-Phase-II (D2P2) path, which lets a company with prior, documented feasibility skip Phase I entirely and compete for the larger award.

The published topic pages for DV015, DV016, and DV017 do not, as of the pre-release, indicate a D2P2 option — which is itself a question worth asking the topic author before July 22. DARPA has used D2P2 aggressively in recent releases (its Biological Technologies and Defense Sciences offices leaned on it heavily in the June 3 tranche), so its absence here is worth confirming rather than assuming. If you already have relevant feasibility data — a working forecasting prototype, a fusion-analytics demo, a rad-hard memory cell characterized across temperature — you want to know now whether there's a lane that lets that history count.

Who should actually apply

Art of Novel Signals favors teams that live at the intersection of temporal graph learning and calibrated uncertainty — the difference between a group that can build a forecaster and one that can defend its confidence numbers to a skeptical analyst. If your team's prior work is all point-estimate prediction with no calibration story, this topic will expose that gap fast.

FALCON rewards teams that have already shipped hybrid ML-plus-LLM systems and can speak concretely about the handoff between the two — latency budgets, when the cheap model defers to the expensive one, how the LLM's outputs get grounded in the structured layer so it can't hallucinate a statistic. A proposal that is really just "we'll wrap an LLM around a database" will not survive review.

Non-Volatile Memory for Extreme Environments is a semiconductor play, full stop. It suits fabless design houses, national-lab spinouts, and specialty foundries with a credible packaging partner. A software shop cannot pivot into this topic in three weeks, and shouldn't try.

The clock

Here is the sequence that matters. Now through July 22: read the full topic text on DSIP, register your firm's SBC and SAM.gov credentials if you haven't (this alone can take longer than you expect and cannot be rushed at the deadline), and send your sharpened questions to the topic author. July 22: topics open, government contact ends, and you begin writing against a fixed target. August 19, 12:00 PM ET: proposals are due, and DSIP will not accept a late upload for any reason.

The teams that win DARPA SBIRs are rarely the ones with the most impressive résumé. They are the ones who understood the program manager's intent well enough to write a proposal that reads as if it were commissioned. That understanding is available for the next three weeks and then it is gone. For related DARPA analysis, see our breakdowns of the Defense Sciences Office FY26 SBIR XL topics and the Biological Technologies Office FY26 pre-release.