DARPA's July SBIR Drop: FALCON Wants LLMs That Reason Over Structured Data, and a Memory Chip That Survives −269°C to +600°C. Opens July 22.

DARPA runs its SBIR program on a rhythm most founders never internalize: new topics pre-release on the first Wednesday of each month, open roughly three weeks later, and close a month after that. Miss the pre-release and you have effectively lost a third of your preparation window before you even knew the topic existed. On July 1, 2026, DARPA's Release 4 pre-released two topics that reward very different companies but share the same brutal calendar — they open July 22 and close August 19, 2026 at 12:00 PM ET. That is a compressed window, and the teams that win it are already assembling.

This is the deep dive on both topics — FALCON and Non-Volatile Memory for Extreme Environments — what each is actually asking for beneath the acronym, who is positioned to win, and how to build a proposal that survives DARPA's review. It follows our earlier coverage of DARPA's BTO and DSO drops earlier this summer.

FALCON: Making LLMs Reason Over Structured Data Without Melting the Compute Budget

FALCON — Fusion of Abstract Learning and Context-Optimized Neural-methods (topic DPA26BZ03-DV016, out of the Information Processing Techniques office) — targets one of the most practically important open problems in applied AI. Large language models are extraordinary at unstructured text but notoriously inefficient and unreliable when the underlying data is structured: tables, graphs, sensor streams, relational records. The objective, in DARPA's own framing, is to "combine advanced machine learning (ML) methods that can be computationally efficient in structured data with large language models."

Read that carefully, because the emphasis on computational efficiency is the whole game. DARPA is not asking for a bigger model. It is asking for an architecture that gets the reasoning and language capabilities of an LLM while doing the heavy lifting over structured data with lightweight, specialized ML methods — the kind that can run inside a power-and-thermal budget appropriate to a fielded system rather than a data center. In defense contexts, that constraint is not academic. A model that needs a rack of GPUs is useless on a forward-deployed platform, an unmanned system, or an edge sensor. The winning FALCON proposal will show a credible path to efficiency-per-inference, not just accuracy.

Who is positioned to win FALCON: teams that have already done real work at the boundary between symbolic or graph-based methods and transformer models — retrieval-augmented systems, neuro-symbolic architectures, graph neural networks feeding LLM reasoning, or model-compression and distillation applied to structured-data pipelines. If your only asset is a fine-tuned off-the-shelf model, you are not competitive here. DARPA reviewers will look for a genuine architectural insight about why your fusion approach is more efficient than brute force, backed by preliminary data or a defensible theoretical argument.

Non-Volatile Memory: A Flash System That Survives Space and Deep Cryogenics

The second topic is pure hardware and it is unforgiving. Non-Volatile Memory for Extreme Environments (topic DPA26BZ03-DV017, out of DARPA's Multi X office) asks for a temperature-hard (−269°C to +600°C) and radiation-tolerant NOR Flash memory system. That temperature range is staggering: −269°C is within four degrees of absolute zero, the regime of superconducting and quantum hardware; +600°C is the environment inside a jet engine, a hypersonic vehicle's skin, or a deep-drilling geothermal sensor. A single memory system that functions across that entire span, while also shrugging off radiation, does not exist as a commercial product.

This is a topic for a very specific kind of company — one with real materials-science and semiconductor-device expertise, ideally with prior work in rad-hard electronics, silicon carbide or wide-bandgap devices, or specialized non-volatile memory chemistries. The applications are obvious once you see the temperature envelope: space systems (where radiation and cryogenic cold coexist), hypersonics and propulsion (extreme heat), and quantum computing infrastructure (deep cryogenic). DARPA is signaling that it wants memory that can live where conventional silicon fails, and it is willing to fund the feasibility work to get there.

Who is positioned to win: semiconductor and materials teams with a differentiated device physics story. The proposal needs to name the mechanism — the specific memory technology, the material stack, the packaging approach — that gives you a shot at the temperature and radiation envelope. Vague assertions that you will "explore" solutions lose to teams that arrive with a hypothesis and preliminary characterization data.

The Money and the Mechanics

DARPA SBIR awards run larger than the typical DoD SBIR. Based on DARPA's recent structure, Phase I awards are commonly around $300K for a 6-to-12-month feasibility effort, with Phase II in the $1.5M–$1.8M range. Some topics also offer a Direct-to-Phase-II (DP2) path that skips the feasibility gate and awards up to roughly $1.5M immediately — but DP2 is only available where the topic explicitly allows it, and it demands documented prior feasibility work.

That last point is where most DP2 applications die. As we noted in earlier DARPA coverage, the most common DP2 failure is asserting feasibility without documenting it. If a topic offers DP2 and you want it, you need explicit evidence — prototype data, third-party validation, a prior funding history that demonstrates the feasibility phase is genuinely behind you. Reviewers will not take your word for it.

Building a Proposal on a Three-Week Runway

With the topics opening July 22 and closing August 19, you have roughly four weeks of open submission on top of the three-week pre-release head start. Treat the pre-release period as your real preparation window:

• Days 1–5: Topic match and team formation. Confirm your technology genuinely fits the topic's intent — not the acronym, the objective statement. FALCON is about efficiency; Non-Volatile Memory is about a specific temperature and radiation envelope. If you are stretching to fit, the reviewers will see it.
• Weeks 1–2: Technical narrative and preliminary data. DARPA rewards a crisp technical hypothesis backed by whatever preliminary evidence you can marshal. Assemble it now, not in the final week.
• Register early. SBIR submission requires active registrations (SAM.gov, SBC, and the DoD portal) that can take days to weeks to clear. A lapsed or incomplete registration is the single most preventable way to lose a submission. Verify all of them the day the topic pre-releases.
• Build in a 72-hour buffer. The close is a hard 12:00 PM ET on August 19. Portal congestion in the final hours is real. Plan to submit three days early.

The Strategic Read

These two topics tell you something about DARPA's priorities right now. FALCON is a bet that the near-term defense value of large language models depends less on raw scale and more on efficient, deployable reasoning over the structured data that military systems actually produce. Non-Volatile Memory is a bet that the physical extremes of space, hypersonics, and quantum hardware are outrunning the commercial semiconductor supply chain, and that the government needs to seed the components itself.

Neither topic is winnable by a generalist. Both reward companies that have been working on exactly this problem and can prove it. If you are one of them, the calendar is the enemy — the topics pre-released July 1, open July 22, and close August 19. The window to turn a genuine technical advantage into $300K of non-dilutive Phase I funding is measured in weeks, and it is already open.