Dragon 2 · Launch alignment
Two tracks, one lock: GTM and Engineering on the same five weeks
Everything public-facing is subordinate to one gate. M4 — one real community GPU bonded, attached, validated, and paid end to end. Until that passes, recruitment is speculative. So the only question that matters is whether the two tracks line up on the same dates around it.
⚠ Dates are PR #76 anchors (Jul 3 / 10 / 17 / 24) — NOT meeting-ratified. Correct the eng row live.
done / verified
in flight
unverified / blocked
waits on other track
The five weeks
Jun 23 – 29
Unblock now
← we are here
Jun 30 – Jul 6
Config Jul 3
decisions
Jul 7 – 13
Automation Jul 10
build + qualify
Jul 14 – 20
Gate Jul 17
🔒 go / no-go
Jul 21 – 27
Mainnet Jul 24
launch
M1 · door open
Preview sign-off → production cutover
Flip the ecosystem site F50 → Dragon 2; smoke-test intake into the CRM.
in flight
Name the review / onboarding owner
Bren vs Filip vs Martijn — unblocks 6 later items.
to decide
Demand prep
Secure ad-credit tranche; build X campaign (7 angles, EU-GPU, utm=dragon2); brief Steven's warm lane.
not started
M2 · interest qualified
Qualify 3–5 EU GPU owners
Ads live + warm intros. Email worker live after DNS + dark test.
not started
M3 · cohort committed
2–3 committed, onboarding calls booked
Needs bonding economics + GPU setup doc from eng first.
waits on eng
M5 · launched
White-glove calls; 2–3 providers live & earning
Only proceeds if the gate passed.
waits on gate
Approve + merge PR #76
Yahor + Sergey approve, Ulad merges. Unblocks timeline + canonical brief.
not started
8-GPU capacity numbers + Zettabyte terms
Feeds the reward model. (Q2 input)
not started
Lock bonding economics + reward model
Stake amount + funder; what validated work pays, budget, payout signer. (answers Q2)
not started
ExternalInferenceProvider interface
The bond → attach path. Plus GPU hardware + setup doc for the call. (answers Q1)
not started
🔒 M4 · the one hard gate
One real external GPU: bond → attach → serve → validate → pay
End to end on the live cluster. Nothing public-facing proceeds without it.
unverified
Mainnet: providers on the live EU cluster
Serving validated inference, settled via DAC.
after gate
The only two things that can sink this — both engineering, both unverified
Q1 · technical loop
Does the loop close?
Can a community 3090/4090 bond → attach → serve inference → get validated → get paid, end to end? If it doesn't close, there is no product to recruit for.
Owner: Vlad (interface) + Yahor · verify by standing up one real test node
Q2 · unit economics
Do the economics work?
Community-GPU cost-per-inference vs Aether, and what validated work can actually pay a provider. If a 4090 can't earn enough to bother, or costs more than the cloud it replaces, the wedge is dead.
Owner: Vlad (capacity) + Yahor / Fred (reward + signer) · verify by load test + signed model
What I need from you, Vlad — so the tracks lock
- Your real dates for the interface, capacity numbers, and the go/no-go. The
Jul 3 / 10 / 17 / 24 here are PR #76 placeholders, not ratified — every GTM date hangs off yours.
- When does the ExternalInferenceProvider interface realistically land? That sets when I can promise a working bond/attach on the white-glove calls.
- Rough 8-GPU capacity + cost-per-inference vs Aether — even a range. Without it the "earn real CERE" promise has nothing behind it.
- Confirm the EU (DE/NL) geography is right — it drives all the ad targeting and copy.