Six theses. Six futures.

The market is in its hyperscaler era. We are not. The companies we back sit one layer beneath the frontier labs — they sell the picks, shovels, and shipping infrastructure that the labs need but do not build.

What we look for

• Local inference. Quantization, distillation, and runtime systems that move workloads off-cloud without sacrificing capability.
• Vertical agents. Workflow software that earns its keep in a specific industry — legal, claims, supply chain — not a general-purpose chatbot.
• Evaluation and trust. The tooling layer for prompt regression, hallucination detection, and red-teaming. Where compliance meets engineering.
• Model observability. If you can't see what an agent did, you can't sell it into the enterprise.

What we avoid

We do not write checks into foundation model labs, GPU-hungry training startups, or pure compute resellers. The capital intensity is wrong for our model and the moats decay too quickly.

"The dollars to be made in AI over the next decade are not in the models. They are in everything the models will need to be useful, accountable, and embedded." — Thesis Memo, Frontier AI v3 (2025)

Healthspan has overtaken lifespan as the question that matters. The science of slowing biological aging has moved from speculative to clinical in less than ten years — and the consumer surface is still nascent.

What we look for

• Senotherapeutics & geroprotection. Compounds with credible biomarkers and a regulatory path that does not require defining aging as a disease.
• Metabolic health. Continuous glucose, ketone, and hormone monitoring made invisible, useful, and accurate enough for behavioral change.
• Diagnostics-first companies. Earlier signals on cardiovascular, neuro, and oncological risk — particularly multi-omic and at-home.
• Reproductive longevity. Egg-freezing, ovarian aging, and the under-served biology of midlife.

What we avoid

We pass on undifferentiated supplement brands, anti-aging cosmetics, and clinics that monetize attention rather than outcomes.

The energy transition is the largest industrial buildout in human history, and most of it is unglamorous. Substation transformers. Permitting software. Long-duration storage. Geothermal drilling. The transition will be built by companies that solve boring problems on boring timelines.

What we look for

• Grid orchestration. Software that makes aging transmission and distribution infrastructure behave like a modern system.
• Long-duration storage. 8–100 hour discharge — iron-air, flow chemistries, gravity, thermal. The window the lithium-ion grid cannot cover.
• Next-gen geothermal. Drilling, well design, and closed-loop systems with a path to firm baseload power.
• Atmospheric and ocean carbon. Removal with credible MRV, durability, and a defensible cost curve.

What we avoid

Consumer EV adjacencies, voluntary-carbon credit marketplaces, and any company whose unit economics require a permanent subsidy.

Biology is becoming a manufacturing substrate. The cost of writing DNA is falling on a curve familiar to anyone who watched compute. The category will eventually subsume chemistry, materials, and parts of consumer.

What we look for

• Programmable cells. Microbial and mammalian platforms with credible scale paths and end-product economics.
• Biomaterials. Leather, silk, mycelium composites, and lithium-binders — anything that displaces petrochemical inputs at parity cost.
• Bio-foundries. The CMOs and design-build-test infrastructure that make every other bet possible.
• Crop & livestock genetics. Genome-edited resilience traits with regulatory pathways outside GMO-heavy jurisdictions.

What we avoid

Lab-grown meat at consumer price points without a credible cost curve. Cellular agriculture for novelty rather than substitution.

Foundation models have given language a body. The translation from token to torque is the most under-priced bet in venture today — provided you separate the believable from the demoware.

What we look for

• Vertical embodiment. Robots that do one job in one environment with one paying customer. Generalist humanoids are not our timeline.
• Actuators and tactile sensing. The unsexy hardware layer that makes dexterity tractable.
• Simulation-to-real. Synthetic data infrastructure that compresses the iteration cycle from weeks to hours.
• Teleoperation as bridge. Hybrid autonomy is the interim economic model that funds the autonomy roadmap.

What we avoid

Consumer humanoids without a wedge customer. Self-driving cars. Anything that depends on shipping a fully general agent before the unit economics work.

The slow defection from the food and beverage industrial complex is the longest consumer cycle of our lifetime. We invest in brands and software that make health legible, accessible, and unembarrassing to buy.

What we look for

• Functional foods. Honest nutrition with brand DNA and clinical credibility. Boring categories with disciplined unit economics.
• Behavior change software. Apps with retention curves a CMO would believe and a clinician would defend.
• Mental health infrastructure. The plumbing — eligibility, billing, care delivery — not the consumer journals.
• Sleep and circadian. Diagnostics, devices, and pharmaceuticals targeting the most modifiable healthspan input.

What we avoid

Influencer-led brands without retention. Telehealth that sells GLP-1s. Anything that depends on paid acquisition staying cheap forever.