$350B vs. $190B by 2027: AI inference is projected to reach about $350B by 2027, passing the $190B database market by nearly 2x. Companies paid roughly $25B to run AI models in 2025.

Customers Pay for Usage: Tunguz expects every software application to become a reseller of AI, so customer bills follow usage and could far exceed the seat license or base subscription.

Margins Shrink: Traditional software keeps about 72 cents of every revenue dollar after direct costs. AI usage could take more than half of revenue, so margins fall unless companies cut AI use per task or prices drop.

2x the Outlier Profiles: Analyzing the data, Gray finds that the average emerging manager backs more than twice as many outlier profiles per investment as a megafund.

Pooling Widens the Range: Combining five emerging manager portfolios widens the range of investments significantly, while the same does not hold for the more homogeneous megafunds, so a basket of small funds buys broader exposure.

Megafunds Cluster on AI: AI makes up 82.7% to 89.5% of five megafunds' technology-labeled Seed rounds in 2023 to 2025, while 11 emerging managers range from 15.4% to 100%. Gray sees megafunds crowding into the most legible category, which leaves discovery of companies outside it to emerging managers.

AI Sped Up the Numbers: AI accelerated innovation so much that some companies now reach $10M in 18 months and $1B in under two years. When that keeps happening, investors lose the ability to tell normal from exceptional.

Metrics Stopped Predicting Rounds: Rounds worth hundreds of millions have closed at very small revenue, while companies with far more revenue have failed to raise. With copied technology no longer differentiating, nearly every round now rests on the founder and team.

Three Elements Make the Moment: A decisive moment is when three things line up: the right business (mostly in the founder's control), the right reason to raise (partly controlled), and an investor with a prepared mind (almost entirely outside the founder's control).

Acts Without Being Asked: A fixed loop runs every ten minutes around the clock, about 52,560 times a year, on a Mac mini with 107 active jobs and 47 databases. It reads mail, messages, and calendar, a drafter writes the replies, and no step can send anything.

A Memory Built Around People: The world model holds what the system has worked out about people from Simon's own correspondence, such as their tastes, role changes, and what he promised them. It covers 20,589 people, about 1,300 of whom carry 5,730 facts, each stored with its source sentence. Simon would build it first if he started again.

Boundaries Took the Quarter: A useful agent took an afternoon. The rest of the quarter went on deciding what it may touch and proving that boundary holds, and only three of eight agents can write anything. Repair verification is not built yet, so the morning report can only say a repair was "tried".

Onboarding and Servicing Gain Most: BCG's chart estimates the share of work AI could take on at 45% to 55% in onboarding and KYC (know your customer checks) and 40% to 55% in account servicing, against 20% to 25% in client acquisition and in mid- and back-office work.

Tailored Advice Stays Personal: Relationship-based referrals, life-stage coaching and goals, multigenerational strategy, investment decisions, and the trusted advisor relationship are marked human-led, with AI handling only 10% to 20%, so the more tailored the service, the less exposed it is.

Redesigned Workflows Win: Adding AI to existing processes delivers small gains. The post argues firms that redesign their work around AI should cut costs, widen margins, and charge more, while firms with scattered, disconnected data rarely get past early trials.

Efficiency First, Then Effectiveness: The move from old-school to productivity VC is a push for efficiency through off-the-shelf tools. The next step, into Workflow and Fullstack builders, is a push for effectiveness through custom AI workflows and in-house data.

Most of the Market Is Still Early: The report places most of the global VC market in the old-school stage, run on spreadsheets and gut feel, while fully autonomous quant VC remains rare, emerging, and contested.

Two Builder Types Look Different: The median Fullstack firm has 23 employees, 2 engineers, 9 investors, and $800M AUM. The median Workflow firm has 7 employees, no engineers, 5 investors, and $151M AUM, so its investors and operators take on the engineering role and build the workflows themselves.