64% of VC Money Flows to AI: Only 3 of 8 Directions Are Buildable Today

1. Consensus Is Right, But Technologists Must Hear the Deeper Meaning

The three firms' consensus in one sentence: AI shifts from "tool" to "labor," value moves from selling software to selling outcomes. For technologists, the real implication is that your moat is no longer "knowing how to tune models." Model tuning is being commoditized — GPT-6.1 Terra charges $1 per million input tokens and $5 per million output tokens, several times cheaper than early-2024 flagships; once Qwen4's 27B open weight is released, a single consumer GPU can run it. As models become cheaper and more standardized, leverage shifts from "model capability" to "industry data + outcome delivery + customer relationships."

Therefore, the only filter used here: Can a small team today use off-the-shelf models plus local/cloud APIs to ship a revenue-generating prototype?

2. Scoring the Eight Directions for "Can We Start Today?"

Service-as-Software (sell outcomes) — Direction is correct but slow-burn. Technically, existing models can handle insurance brokerage pre-screening, compliance review — tasks that are "judgment-light, reasoning-heavy." The real blockers are customer acquisition and outcome-based pricing, not technology.

Vertical Agents (multi-party coordination) — Most friendly starting point for technologists. Multi-agent frameworks + Ollama/MLX can run prototypes locally. Pick a narrow slice like one step in real-estate transactions or supply-chain procurement; don't build a general-purpose agent.

Infrastructure for Agents — Bets on "agent volume growth drives demand." Lightweight middleware (payments, permissions, auditing, API auto-adaptation) can be built by two people, but avoid heavy layers; big tech will eventually enter.

Trust Layer (security/compliance/human verification) — Underestimated hard need. As agents gain autonomy, proving they are trustworthy, compliant, and interacting with real humans becomes critical. Red-teaming and evaluation can start with current models, but regulated sectors (finance, healthcare) have high entry barriers.

Data Layer (turn messy data into usable assets) — The steadiest "sell shovels" play. Cheaper models make clean, traceable, industry-specific data more valuable. Structuring, governance, annotation are tasks technologists can do directly; local models handle extraction.

Physical AI (robotics/world models) — Chinese teams have hardware supply-chain advantages, but pure-software squads can only touch edge pieces (scheduling software, data/testing for one subprocess). Heavy assets are out of reach.

Human Needs (silver economy/education/personalization) — Direction is valid, but pure consumer acquisition is expensive (YC 2026 consumer share dropped to 4–5%). Enter with caution.

Heavy Capital (compute/energy/defense) — Ordinary teams can only build light-asset pieces like transaction matching or scheduling software; avoid heavy assets.

Bottom line: The three directions you can prototype on your Mac today are ② Vertical Agents, ③ Agent Infrastructure, ⑤ Data Layer. ① is a long-term slow burn, ④ is an underestimated necessity, ⑥⑦⑧ either require heavy assets or suffer from hard customer acquisition — technical squads should park them for now.

3. Three Core Viewpoints

Viewpoint 1: The technologist's leverage lies "outside the model." When model tuning becomes a cheap, per-token commodity, your real product isn't "I used a bigger model" but "I wrapped the model in a closed loop that owns the outcome." The source says "moats are built on customers"; I'll add: the first asset a small team should accumulate is the structured data from every delivery — that's something model vendors can't casually replicate and big tech can't easily seize.

Viewpoint 2: On-device/local inference is your hidden differentiator. Today most agents run on cloud APIs, metered by token, with data leaving the premises. But you can run a 70B model on a 64 GB MacBook Pro, or a 405B model on an M5 Ultra with 512 GB, enabling medical pre-screening, financial compliance, and other privacy-sensitive, data-cannot-leave scenarios — a selling point cloud solutions can't match. Before selling "outcomes," ask: "Can this outcome run entirely inside the customer's firewall?"

Viewpoint 3: Before selling outcomes, quantify the agent's failure rate. This is where technologists crash most. The source gave "four judgment questions"; I'll reframe: don't be fooled by benchmark scores. In my earlier Memdoor five-agent write-up, the gap between 9.8 and 8.0 came largely from evaluation noise; real long-task success rate is the only metric that pays. What percentage of real tasks does your agent complete in one shot? If you don't know that number, don't rush to shout "sell outcomes."

4. Four "Start Tonight" Actions

Tonight: Spin up a vertical agent prototype locally with Qwen3.8-27B (e.g., "contract first-pass review") and feel the model's capability boundary firsthand.

Pick one narrow scenario where the customer already pays for outsourcing, then write down "AI does 80%, your expert covers the last 20%" — that 20% is where your value lives.

Price by outcome, even if it's just a one-month paid pilot; avoid the old "per-seat" trap.

Turn every delivery into structured data to build your own moat, not to do free labor for model vendors.

5. Three Technologist-Specific Pitfalls

Treating "model tuning" as the product: A model upgrade zeroes you out; that's outsourcing, not entrepreneurship.

Ignoring agent failure rate while selling outcomes: Without a measured real-world long-task success rate, the more you promise, the more you'll pay in penalties.

Over-reliance on cloud: In heavily regulated industries, data leaving the premises is an instant disqualifier; on-device inference isn't optional — it's the entry ticket.

Conclusion

Capital keeps flooding into models and compute, but the technologist's real leverage sits outside the model — in data, in industry depth, in outcome loops. Of the eight directions, ②, ③, ⑤ can be prototyped on your Mac today; ① is a long-term ticket, ④ is an underestimated must-have. Don't chase hype; chase the spot where the customer is already paying, the model handles 80%, and you cover the last 20%.