Built in public ยท iterates most weeks
Dacard.ai Q3 readout, worked example Back to the point of view →
Worked example, not a real readout. The four bets are Vercel products announced in public. Every number, status and decision below is constructed to show the shape.

Product readout · worked example

Q3 product readout to the board.

Subject: Vercel Sixty seconds spoken Four questions answered Evidence attached to every line Three asks, all decidable today

At a glance.

One number per question, then the shape of the quarter. Everything below this is the same story with the evidence attached.

Q1 · chose 4 / 11 bets backed, requests declined
Q2 · shipped 138 merged changes, indexed to 100 a year ago
Q3 · landed 2 of 4 1 on watch, 1 stalled
Cost to serve -34% per unit of agent work, same volume
Q4 · stopping $310k per quarter, reassigned this month
Landed 2 Watch 1 Stalled 1

Shipping rose. Landing did not follow.

Changes shipped What landed
140 120 100 Q3 last year: shipped 100, landed 100 Q4: shipped 112, landed 104 Q1: shipped 121, landed 101 Q2: shipped 130, landed 103 Q3 now: shipped 138, landed 102 138 102 Q3 last yr Q3 now
Changes shipped is merged pull requests. What landed is the count of those tied to a success condition that moved. Both indexed to 100 five quarters ago, so one axis carries both. That gap is why line 02 demotes the shipping number.

The four answers, before anything else.

A board asks four questions and only one is about outcomes, which is the one most readouts route around. These are the answers. Everything after this is how they get delivered and what they rest on.

Q1 · delivered by line 01

Are we building the right things?

Four bets on one thesis, that the agent era needs different primitives. Eleven requests declined, and I made the call on each.

Source: the Q2 portfolio doc, unchanged since it was signed

Q2 · delivered by line 02

Are we moving fast enough?

Highest shipping quarter on record across the framework, the SDK and the compute layer, at 34% less cost per unit of agent work. I am treating that as capacity, not achievement.

Source: release log, public

Q3 · delivered by lines 03 and 04

Is it working?

Two landed, one on watch, one stalled. The counts and the stalled name are volunteered, not extracted. The early read I watch mid-quarter is second-week retention on a new agent workload; it moved first on both bets that landed, and never moved on the one that stalled.

Source: success conditions fixed in Q2, scored against the same definition

Q4 · delivered by line 05

What would you stop?

The marketplace, this month. Its $310k a quarter is already assigned to the agent runtime. The destination is what makes it checkable next quarter. Budget that is only freed tends to quietly not move.

Source: this readout, offered before it was asked for

The same four answers, as sixty seconds spoken.

Nothing new below this line. It is the summary above, sequenced for a room: outcome lands on three, the bad news on four while there is still space for it, the decision on five. Nothing waits to be asked for.

01

Last quarter we backed four bets and declined eleven requests.

the choosing
02

We shipped more than we ever have, and I treat that as an input, not a result.

the reframe
03

Two landed. The SDK and the compute change moved behaviour in most enterprise accounts.

the outcome
04

One stalled. The partner marketplace has been live a quarter and changed nothing measurable.

the uncomfortable part
05

We are stopping it this month and moving that $310k a quarter to the agent runtime.

the teeth
06

Every one of those reads uses the same definition, including the one I championed.

the credibility

58 seconds read aloud 4 questions answered before the first interrupt 1 decision offered, not requested

The portfolio the script rests on.

The board does not see this table unless it asks. Every line in the script has to survive it.

Bet Success condition, set in Q2 Read Cost to run Status
AI SDK agent-building layer Agent workloads named in renewal conversations in 28 of 44 enterprise accounts, the number set in Q2. 31 of 44 accounts $0.11 / call
was $0.19
Landed
Fluid compute billing and cold start Customers stop citing idle cost as a reason to keep AI workloads off the platform. Cited in 2 of 19
losses, was 11 of 23
$41k / qtr
down 34%
Landed
v0 generated UI to production Generated work reaches production without a rewrite on 10 of 15 non-early-adopter teams. Below 10 by 31 Oct it is scored stalled, not extended again. 6 of 15 teams
flips at 10, or stalls
$2.40 / screen
flat
Watch
Partner marketplace constructed for this example Partner-sourced deals reach 8% of new pipeline within two quarters of launch. 1.2% of pipeline
live 1 quarter
$310k / qtr
2 eng + partner ops
Stalled

The stalled row is invented, and only that row. The other three are real public bets. A stall is a claim about a real company's real product, and nothing outside the building has standing to make one, so the example carries a fictional initiative instead. Inside a real board room you would name it, because you hold the evidence and you own the call.

The decision, carried in.

Stopping this month

Partner marketplace → agent runtime

Live one quarter, 1.2% of new pipeline against an 8% condition. Two engineers and the partner-ops headcount move to the agent runtime, where the SDK read says the demand already is. The condition for the new work gets written before the money moves, not after.

Line four and line six are a pair. One names a loss. The other says the naming was not selective. Cut either and this reads as a story with numbers attached.

What I need from you.

Three things, all decidable in this meeting. Everything above is context for these.

Ask 1 · from line 05

Sign off on the reallocation, today.

The $310k and the three people move to the agent runtime this month. If you want the marketplace kept alive instead, I need the success condition rewritten in this meeting rather than after it.

Decision needed: today

Ask 2 · from the watch row

Hold me to 31 October on v0.

It flips to landed at 10 of 15 teams, or it is scored stalled. I am asking you to refuse a second extension, including if I am the one requesting it.

Decision needed: today, enforced 31 Oct

Ask 3 · from Q2

One intro, to a design-system buyer at scale.

The v0 read is thin because the teams we reach are already early adopters. The fastest fix is access to an organisation that is not, and that is a board asset rather than a product one.

Decision needed: names by end of week