mirror of
https://github.com/hanzoai/papers.git
synced 2026-10-29 07:34:31 +00:00
* paper: unified sovereign-tenant cloud — measured per-tenant memory + linear scaling Measured (host evo, AMD Ryzen AI Max+ 395, Go 1.26.4) marginal RSS per open per-tenant SQLite handle via the exact production open path (dbx.Open + sqlite. PragmaDSN): ~0.25 MB/active tenant (cgo/mattn, warm), ~8x below the assumed 2 MB; cache_size is a ceiling not a footprint; idle tenants ~0 live RAM. Capacity model puts 1e9 registered tenants on ~103 32GB nodes at 1% concurrency. One-binary vs 56-subsystem mesh fixed-memory multiple: 11-32x. All numbers measured or marked modeled/cited; honest limitations section. Cites HIP-0106/0116/0125/0302/0107/0032/0118 and the cloud-unified-binary + pq-object-store companion papers. * paper: expand — 128GB nodes, GPU inference axis, cloud-in-a-box, hanzo.network fabric - Capacity retabulated for the standard 128GB unified-memory node (U=125GB): 1B registered @1% concurrency = ~20 nodes (measured 0.25MB constant), 3-node HA floor up to 1e8 @1%. - New section: the two scaling axes — tenant state (CPU/RAM, O(active), shared-nothing) vs GPU inference (VRAM, O(distinct models) — weights shared across all tenants, batched). hanzo-engine FFI-embeddable (hanzo_engine_ffi.h). - New section: the unified-memory box as atomic self-hostable unit — one $3,200 128GB box holds per-tenant SQLite + model weights in one address space; ~47M registered users/box @1% (measured-m-grounded); ~$0.00007/user capex. - Culminating section: hanzo.network — leaderless PQ BFT fabric (luxfi/consensus) as control plane scheduling PLACEMENT not writes; per-tenant single-writer data plane; TEE trust tiers (H100 CC/TDX/SEV-SNP); PQ end-to-end (pq_mode + ML-DSA/ ML-KEM auth, Quasar); auto-provisioned MPC wallet; AI-coin incentive (HIP-0001); viral hanzod-up deployment (hanzod:cloud :: bitcoind:ledger). - Explicit BUILT (cited to source files: writerpin, role, luxfi/consensus PQ, hanzod, wallets, capauth, lux/crypto mldsa/mlkem) vs ROADMAP (fabric-VM, on-chain attestation/scheduling, auto-join orchestration) split. Adds HIP-0001/0005/0117, luxconsensus, quasar, TEE, cross-backend-inference cites. 14pp, compiles clean.
255 KiB
255 KiB