Two companies took 43% of all venture funding in 2026. Seed fell 27%.
Golden Hour invests in vertical AI, defense technology, and business infrastructure at the stage everyone else just left.
The headline is a boom. The seed market is a recession.
Global venture funding hit $510B in the first half of 2026, more than all of 2025. OpenAI and Anthropic alone took $217B of it, about 43% of every startup dollar raised on earth.
In the same period North American seed funding fell to $4.9B in the second quarter, down 27% year over year, with early-stage deal count at a five-quarter low.
Those two facts are not in conflict, and the gap between them is the whole opportunity.
Four things we believe.
Underwrite 50% gross margins, not 80%.
Observed gross margin on AI products was 45% in 2025, projected to 52 or 53% in 2026. Application-layer companies post the lowest margins of any group, around 49%, because model inference alone is 23% of the cost stack. At 50% margin instead of 80%, a company needs roughly 1.6 times the revenue to fund the same engineering and sales. Rule of 40 comparisons against the SaaS comp set are not valid. Anyone modeling 80% margins at exit is treating a 2027 projection as observed fact.
Defensibility in 2026 is purchased, not engineered, and it costs gross margin.
Half of AI companies now plan to scale forward-deployed engineering as a permanent go-to-market motion, projecting coverage of 34% of enterprise customers by 2027. The moat being funded is implementation depth and switching cost, not model or data advantage. It is a real moat and a margin tax at the same time. We think that is usually the right trade. We just say which trade we are making.
In defense, capital is the least binding constraint. Underwrite the budget line, not the technology.
Defense startups captured 1.3% of Pentagon contract obligations in 2025, up from 0.6%, while absorbing $9.6B of venture capital. The gating question is never whether the thing works. It is which program element, in which appropriation, in which fiscal year. Diligence that does not produce that answer is not diligence, and the sector is currently pricing seed companies with no program of record at multiples several times what the category's best asset trades at.
Infrastructure is a small, fast-growing category, not a large one, and we would rather say so.
Of $37B in enterprise generative AI spend in 2025, foundation model APIs took $12.5B. Storage, retrieval and orchestration, the entire layer most funds mean when they say agent infrastructure, took $1.5B. That is 4% of the spend. It is growing fast and it is worth investing in. It is not where the value is accruing, and a fund that tells you otherwise has not read the survey.
Five places we look first.
Two passes we make every time.
AI rollups from a minority position. Seventy nine percent of surveyed investors expect these to exit to a private equity consolidator and only 13% believe the returns are venture scale. Without a control stake a fund supplies 100% of the acquisition cash for a minority of the assets, and the technology's addressable market is limited to the companies it bought. Take control or do not play. The middle is structurally unreturnable.
Anything whose pitch turns on a foundation model capability that improved thirty points last year.
Name the program element and we will move fast.
Or the appropriation, or the switching decision you are attacking. Specificity gets a fast answer from us.
Say hello