Lattice Graph × Andreessen Horowitz (a16z)

Lattice Graph operates a materials knowledge graph spanning millions of compositions, connecting formula to crystal structure, predicted and measured properties, synthesis routes, and patent claims in a single traversable graph. When a16z needs to evaluate an AI-materials or critical-minerals company, that graph is the independent reference layer — not a third party's slides, not a scraped public dataset, but a governed, cross-source structure that resolves disagreements between computational databases and flags where the evidence is thin. Every composition in the graph carries trust and disagreement signals across multiple DFT sources, so a diligence team can distinguish well-corroborated property predictions from cases where the underlying references conflict — a direct, quantitative gauge of whether a startup's model is calibrated or overfit to a favorable subset of the public record. Validation on Lattice Graph does not stop at a single engine. Candidate materials are evaluated using multiple independent physics engines — including machine-learning interatomic potentials such as MACE and CHGNet alongside traditional DFT — with phonon and thermodynamic stability consensus required before a material advances. That multi-engine consensus discipline is directly relevant to diligence: when a target company reports a hit rate or a stability fraction, a16z can test whether those candidates hold up under independent engines or collapse when the evaluation method changes. Freedom-to-operate and patent-whitespace screening across more than 300,000 materials patents closes the IP loop, so novelty assertions in a deal memo can be tested against actual claim coverage rather than asserted on the strength of a lawyer's letter. The single most differentiated instrument in the Lattice Graph stack for an investment firm is what we call the negatives moat: 23,196 labeled failed-experiment and kill edges that capture what does not work, alongside the chemistry and conditions under which it failed. This data is largely internal and absent from the public repositories that most foundation-model decks are trained and evaluated on. A positives-heavy public benchmark can make a materials model look far stronger than it performs in practice — the negatives moat is the instrument that reveals the difference. For a firm writing large checks on AI-for-science and critical-minerals theses, having access to the ground truth most competing diligence cannot reproduce is a structural advantage.

Andreessen Horowitz has made deep-tech and American Dynamism a central strategic thesis, and an increasing share of that pipeline runs through categories where technical truth is genuinely hard to read from a pitch deck: foundation models for chemistry and materials science, generative-discovery platforms, self-driving labs, battery and catalyst startups, and domestic critical-minerals supply chains. A generative-materials company can present compelling benchmark scores, high candidate hit rates, and a defensible-looking IP roadmap that all clear a generalist screen — while the load-bearing questions about model calibration, chemistry novelty, patent freedom, and buyer depth stay open until after the round prices. That is the gap Lattice Graph closes: an independent, computationally grounded technical read that can be delivered at the pace a16z deploys. The strategic pressure on the firm right now is that round sizes and valuations in AI-materials have outrun the tools available to underwrite the underlying science. Most evaluation frameworks for materials models rely on public, positives-heavy datasets — the failed experiments that constitute the honest signal of what actually does not work are largely missing from the public record. That systematic gap means a16z can be shown an eval score that looks like a moat but is simply a reflection of what the model has seen. Simultaneously, the domestic-supply and critical-minerals narratives that recur in American Dynamism deals are frequently asserted rather than grounded in composition, concentration, and conversion-route data. Lattice Graph addresses both failure modes in a single diligence engagement. The fit is not just analytical — it is structural. Lattice Graph does not compete with a16z portfolio companies for discovery assets or commercial deals. We are the independent diligence layer: a scoped, fast technical read built on assets most AI-materials teams cannot reproduce, delivered as a buyer-grade memo that can travel directly into an investment committee. The engagement model is calibrated to the pace and budget of venture due diligence, with per-deal reads at roughly $30–60K and an annual subscription structure for firms running multiple deals in this category per year.

Andreessen Horowitz has made deep-tech and American Dynamism a central strategic thesis, and an increasing share of that pipeline runs through categories where technical truth is genuinely hard to read from a pitch deck: foundation models for chemistry and materials science, generative-discovery platforms, self-driving labs, battery and catalyst startups, and domestic critical-minerals supply chains. A generative-materials company can present compelling benchmark scores, high candidate hit rates, and a defensible-looking IP roadmap that all clear a generalist screen — while the load-bearing questions about model calibration, chemistry novelty, patent freedom, and buyer depth stay open until after the round prices. That is the gap Lattice Graph closes: an independent, computationally grounded technical read that can be delivered at the pace a16z deploys. The strategic pressure on the firm right now is that round sizes and valuations in AI-materials have outrun the tools available to underwrite the underlying science. Most evaluation frameworks for materials models rely on public, positives-heavy datasets — the failed experiments that constitute the honest signal of what actually does not work are largely missing from the public record. That systematic gap means a16z can be shown an eval score that looks like a moat but is simply a reflection of what the model has seen. Simultaneously, the domestic-supply and critical-minerals narratives that recur in American Dynamism deals are frequently asserted rather than grounded in composition, concentration, and conversion-route data. Lattice Graph addresses both failure modes in a single diligence engagement. The fit is not just analytical — it is structural. Lattice Graph does not compete with a16z portfolio companies for discovery assets or commercial deals. We are the independent diligence layer: a scoped, fast technical read built on assets most AI-materials teams cannot reproduce, delivered as a buyer-grade memo that can travel directly into an investment committee. The engagement model is calibrated to the pace and budget of venture due diligence, with per-deal reads at roughly $30–60K and an annual subscription structure for firms running multiple deals in this category per year.

The most useful surfaces for an a16z technical advisor or investing partner are the diligence-oriented dashboards rather than the chemistry-authoring workflows. The opportunity-index and buyer-intelligence views render the ranked prioritization into an interactive screen: select a materials lane, see opportunity scores, the funded-buyer leaderboard, and the golden-finalists shortlist. This lets a partner frame a target's roadmap and go-to-market against independent rankings inside a single session and carry those rankings into an investment memo. The negatives atlas and trust-and-disagreement views let the team submit a target's claimed candidates for a kill-edge check and cross-source disagreement flag, turning the negatives moat into a concrete pass/fail artifact with a clear audit trail. The knowledge-graph explorer adds provenance and evidence-neighborhood views so every property claim in a deal memo can be traced back to its underlying source and corroboration level. The freedom-to-operate and patent-whitespace dashboard supports composition-level and claim-level screening on a target's headline IP — a fast way to pressure-test the IP slide before it underpins a valuation. Batch screening lets a16z evaluate a portfolio of a target's candidate materials at once, which is essential when underwriting a platform company whose value is a discovery pipeline rather than a single compound. Energy-storage, catalyst, synthesis, and supply-chain workflow pages let the team stress-test specific technical claims against the graph and the supply layer without needing to commission a separate domain expert for each.

Because a16z is a diligence partner rather than an asset buyer or co-developer, the natural engagement is a scoped, fixed-scope technical read on a single live deal: an independent assessment of a target's claims using the negatives atlas for a kill-edge check on the company's candidates, the opportunity index and buyer affinity to evaluate the commercial story, and the knowledge-graph and freedom-to-operate layers to test novelty, calibration, and IP coverage. That read is delivered as a buyer-grade memo in a timeframe consistent with venture due diligence. Per-deal engagements are estimated at roughly $30–60K for a single-round technical read; these are planning estimates rather than binding quotes, and actual scope and terms are set per engagement. Given a16z's pace across AI-materials and American Dynamism deals, the more efficient structure is an annual diligence subscription bundling API access to the opportunity-and-buyer-intelligence and negatives-atlas products with a set number of scoped per-deal reads, and making the knowledge-graph, trust, and freedom-to-operate layers available across all engagements — with the option to extend the same tooling to portfolio companies after a check clears. A practical onramp is a single paid pilot read on a live or recently passed deal, chosen so a16z can calibrate our findings against its own technical view before committing to the broader subscription.