Why the most expensive Seed deals are the cheapest | E2299

What It Covers

Thomas Tunguz (Theory Ventures), Michael Downing (Castalia Capital), and Paige Doherty (Behind Genius Ventures) analyze the 2026 IPO wave featuring SpaceX, OpenAI, and Anthropic — collectively representing roughly $3.5 trillion in potential liquidity — while examining how AI is reshaping seed valuations, founder leverage, model routing economics, and startup growth benchmarks.

Key Questions Answered

• •Series A Growth Benchmarks: The minimum threshold to raise a competitive Series A has shifted dramatically. Three years ago, 3x annual revenue growth secured strong terms. Today, 10x annual growth is considered mid-pack. AI-native companies in early-stage portfolios are routinely hitting 100x revenue growth within a single year, driven by enterprise AI budget expansion — over 50% of which, per Morgan Stanley analysis, represents net-new spending rather than reallocation.
• •Seed Valuation Paradox: Carta data shows the 95th percentile US seed round now prices at $174M valuation, up from $66M in 2022. While median seed deals appear overpriced, the top 1–5% of seed deals may actually be underpriced given the scale of potential outcomes. Founders should benchmark their entry valuation against realistic exit multiples in markets where AI enables 3–7x higher pricing power than traditional SaaS.
• •Token Spend as Primary Use of Funds: The dominant use of capital in new funding rounds has shifted from headcount and office space to AI token and compute spend. Founders raising $25M rounds now cite token costs as the primary deployment target. This changes how investors should evaluate burn rate and unit economics — token spend is a direct revenue-generating input, not overhead, requiring a different analytical framework.
• •Model Routing to Cut Inference Costs: Startups can dramatically reduce AI infrastructure costs by routing tasks intelligently across model tiers. Use frontier models like Anthropic's Fable 5 only for high-level reasoning and orchestration; deploy smaller, open-source, or local models for repetitive tasks. One practitioner on the panel reduced local model inference from 65% to 91% of total workload using skill distillation in markdown files, materially cutting token spend.
• •Skill Distillation Architecture: The emerging application layer architecture involves frontier models generating reusable "skill files" — markdown-formatted instructions for specific tasks — which are then executed by smaller, cheaper local or open-source models. This approach lets software companies deliver state-of-the-art AI capabilities to customers without passing on state-of-the-art pricing. Expect this pattern to define competitive SaaS products over the next 18–24 months.
• •LP Capital Reallocation Risk for Early-Stage Funds: Family offices and high-net-worth LPs — the primary capital source for pre-seed and seed funds — spent disproportionate VC allocations on late-stage secondaries over the past two to three years. The incoming SpaceX, OpenAI, and Anthropic IPO liquidity events will deliver generational returns on those bets, likely reinforcing late-stage pre-IPO strategies over early-stage fund commitments, potentially constraining seed-stage capital supply.