Calls Ledger

The index reports what the industry did. This page is where we stake what we think it means — dated, falsifiable, and kept on the record when we're wrong.

This index reports what the industry did — the solves / alternative_to / competes_with edges are observable facts. But we also make editorial calls about where things are heading, and those deserve a different standard: dated, falsifiable, and publicly scored. This page is that ledger. Every call states what would change our mind. When we're wrong, it stays here with the correction.

Writing registers used across this site:

  • Observed — supported by relationship edges and primary sources; stated plainly.
  • Reported — a claim from a named source we haven't independently verified; attributed inline.
  • Our read — our synthesis; always dated, always with revision conditions. Everything below is this register.

Open calls

2026-07-26 — Semantic forgetting and cache eviction collide; the first casualty is compliance

The call: Within three quarters: (a) a public incident or regulatory action involves personal data that was "deleted" from an agent's memory bank but persisted in a KV-cache tier (DRAM/NVMe/object-storage prefix cache) — the GDPR-Article-17-through-the-cache problem no benchmark currently tests and no provider documents; (b) cache-aware financial routing ships in at least one production framework — routing on which provider holds the warm prefix rather than list price, because a 90% cache discount makes the "expensive" provider functionally cheaper mid-session. Both follow from one fact pair: deletion is what memory systems measurably do worst (Graphiti 4.4–7.0% on ForgetEval-Adv) while the cache layer's business model is retention (Tensormesh bills cached tokens at zero).

Signals: ForgetEval-Adv's deterministic-substring purge protocol exposing near-zero intent-aware deletion in graph systems; Tensormesh $24.5M with AMD/CoreWeave/NVentures; DeepSeek's 91%-discount Foundry endpoint vs zero-discount native API (same weights, 10× effective cost spread); llm-d shipping cache-hit-rate as an SLO with epsilon-greedy affinity routing.

We revise if: providers publish documented cache-purge semantics with compliance guarantees before an incident forces it; ForgetEval-class control-plane scores improve to parity with recall scores across two consecutive independent evaluations (the gap closes architecturally); or cache-state APIs stay too opaque for cross-provider routers to act on, keeping routing static through mid-2027. Extends the 2026-07-21 consolidation call below — AMB's emergence as a neutral evaluation gym is an early positive signal on that call's clause (a).

2026-07-21 — The AI data stack consolidates persistence onto object storage across three fronts

The call: By mid-2027: (a) a credible cross-vendor agent-memory benchmark harness exists and at least the top three vendors (Mem0, Letta, Zep-class) publish against it, because the July 2026 numbers chaos (92.5/93.4/94.4/94.8 on overlapping benchmarks, plus a third-party +34.4% counter-claim) is unsustainable for buyers; (b) the first independent production case studies for post-Parquet formats (Lance, Vortex, Nimble) surface, converting vendor-benchmark claims into deployment evidence; (c) at least one managed cloud product ships KV-cache-on-object-storage as a billable feature, moving the ObjectCache/LMCache pattern from research to product. Underneath all three, the NAND supply crisis keeps flash rationed to hot paths and makes S3-compatible tiering the persistence default.

Signals: Two critical Mem0 CVEs in one quarter marking the production-hardening gap; TurboQuant's first independent reproduction (AMD RDNA4); ObjectCache holding ~5.6% of DRAM latency at 64K over RoCE; enterprise SSD escalation of roughly 3.5–4× with vendor pass-through (Pure ~70%, Wasabi, Backblaze) — originally cited here as "472%"; corrected August 28, 2026 to the TrendForce contract-price trail.

We revise if: flash pricing normalizes materially before mid-2027 (fab capacity coming online), keeping local-first economical; the memory-layer market consolidates by acquisition before any neutral harness forces comparability; or the post-Parquet formats stall without production adoption beyond their sponsoring vendors.

2026-07-15 — The general-purpose open Iceberg catalog consolidates on Apache Polaris

The call: Polaris becomes the default open catalog choice; Project Nessie's Git-for-data semantics get absorbed (into Polaris or the Iceberg REST spec) rather than winning as a standalone catalog.

Signals: ASF Top-Level graduation (Feb 2026), monthly release train through 1.6.0, Cloudera adoption + contributed Ranger plugin, CVE cluster handled with fast disclosure/patching; Nessie lacks credential vending and RBAC, and two independent practitioner assessments advise against greenfield use.

We revise if: Nessie ships native credential vending/RBAC; significant new greenfield Nessie deployments surface; a third catalog (Lakekeeper, Gravitino) takes the general-purpose slot instead.

2026-07-15 — Kafka becomes optional in single-consumer lakehouse CDC

The call: For pipelines whose only consumer is the lakehouse, direct CDC→Iceberg (Flink CDC, Supermetal-class single-process tools, streaming SQL databases) displaces Debezium→Kafka→processor as the default architecture for new builds by end of 2027. Where the change stream genuinely fans out to many consumers, Kafka stays.

Signals: Normalized benchmarks showing 7× snapshot gaps untuned; Flink CDC's own no-Kafka positioning; the "Kafka-less" tooling wave (RisingWave, BladePipe, Estuary) documented across independent sources.

We revise if: new-deployment surveys/case studies through 2027 show Kafka-based CDC holding share for lakehouse-only pipelines; the single-process tools stall pre-GA or accumulate correctness incidents.

2026-07-15 — The two-engine (Spark + Flink) pattern erodes where feature parity matters

The call: Spark Real-Time Mode erodes Flink's share of workloads whose pain is training/serving logic drift — not Flink's latency crown. Consolidation is a 2027 question; the burden of proof has flipped, adoption hasn't yet.

Signals: RTM GA (March 2026) with sub-100ms production references (Coinbase, DraftKings); the architectural mechanism (long epochs, concurrent stages, non-blocking operators) is published, not just benchmarked.

We revise if: RTM production regressions surface at scale; Flink's response closes the single-engine gap; the launch references stay the only references by mid-2027.

2026-07-15 — Vector search dissolves into the lakehouse; the standalone RAM-first vector DB niche narrows

The call: The default home for vectors becomes lake-resident formats and object-storage-native engines (External Collections, S3 Vectors, Turbopuffer-class, Lance/Vortex); standalone RAM-first vector databases narrow to latency-critical serving tiers.

Signals: Milvus 3.0's zero-copy lake queries and Zilliz's Lakebase pivot; every 2026 format release (Paimon 1.4, Lance, Vortex) adding vector primitives; S3 Vectors' price cuts; Pinecone responding with lakehouse integration (OneLake) and a price-floor tier.

We revise if: lake-resident query latency stalls above serving requirements; the RAM-first vendors' growth re-accelerates on workloads that were supposed to dissolve.

2026-08-24 — Object storage reclassifies from cold persistence to AI's active memory tier

The call: Object storage's role in AI systems shifts from cold archive to active memory tier — KV-cache and agent state land in S3-compatible tiers as a default architecture, and I/O topology (not capacity) becomes the axis storage products compete on for AI workloads.

Signals: Tensormesh's published break-even framework (~one cache reuse per 94 hours justifies S3-tier persistence; vendor-published); Mooncake in production at 100B+ tokens/day treating pooled SSD/remote storage as an active tier; Wasabi Fire shipping NVMe object storage at $19.99/TB positioned for GPU-adjacent inference; Redpanda Cloud Topics writing payloads straight to S3 with Raft kept local; DuckDB 2.0 rebuilding its entire I/O layer as asynchronous for network-attached storage.

We revise if: S3-tier cache-hit latency proves too slow for production TTFT budgets outside long-context niches; NAND pricing normalizes enough that local NVMe undercuts the remote-tier arithmetic; the 94-hour-class break-even frameworks fail independent replication.

Update (2026-08-28): first production evidence against the latency revision condition — Cohere on LMCache + CoreWeave AI Object Storage reports 22–32% TTFT reduction and 41% higher decode throughput, with object-tier cache reads 1.2×–3× faster than S3 Express One Zone (vendor-published, named customer). The call stands, strengthened. Two supporting numbers from the original post were corrected the same day (KV/request 2.6→1.3 GB after the TopKV paper's v2; the "~472%" SSD figure retired for TrendForce's ~3.5–4× contract trail) — neither correction weakens the call's mechanism. Details in blog #29.

Update (2026-09-03): a second named vendor deployment — Dell integrated GPUDirect RDMA KV-cache offloading into ObjectScale, merged upstream into both vLLM and LMCache (vendor-published release notes). Two independent vendors now ship the S3-tier-as-memory pattern as product. On the NAND revision condition: TrendForce's Q3 survey shows contract prices moderating to +10–15% QoQ — a plateau at a permanently elevated baseline, not normalization, so the remote-tier arithmetic holds. The call stands, further strengthened.

2026-08-28 — Verification becomes citation practice

The call: Within two quarters: (a) at least one more named production deployment publishes S3-tier KV-cache latency numbers, converting the Cohere case from anecdote to pattern; (b) "verify against the paper's current revision" becomes standard citation practice for AI-infrastructure references, because answer engines quoting stale v1 preprint figures will force it — the way pinned dependencies became standard for builds.

Signals: TopKV's v2 revision silently halving a headline figure 17 days before we quoted v1; our own 472% correction tracing to a secondary source outliving the primary trail; Cohere/CoreWeave publishing real TTFT deltas; this index now being cited by answer engines, which converts our errors into propagated errors.

We revise if: Cohere-class case studies stay singular through Q1 2027; or arXiv-revision drift proves rare enough in practice that no one else adopts revision-checking.

Update (2026-09-03): partial progress on clause (a) — Dell ObjectScale is now the second named production platform shipping S3-tier KV-cache offload (GPUDirect RDMA, merged upstream into vLLM and LMCache). The Cohere case is no longer singular as a deployment; what Dell has not yet published is latency numbers, which is what clause (a) literally asks for. Counting this as progress, not satisfaction. The same verification sweep also confirmed the TopKV v2 figures remain the best public numbers (no v3, no rebuttal as of early September) — the revision-checking discipline this call bets on caught nothing new this cycle, which is what success looks like.

2026-09-03 — The fabric choice becomes an object-storage product feature

The call: By end of 2027, the "which network fabric" decision migrates from cluster design into storage procurement: Ethernet-family fabrics (UEC/RoCEv2) carry the majority of new S3-to-GPU deployments — the Dell'Oro crossover trajectory holds — and S3-compatible storage vendors ship fabric-specific tuning as named product capabilities rather than reference-architecture PDFs. Buying object storage for AI will mean choosing its fabric the way buying a database means choosing its consistency model.

Signals: UEC 1.0 shipped (June 2025) with Broadcom Tomahawk 6 silicon shipping now; Quantum-X800 XDR live in production at CoreWeave (Vera Rubin NVL72) and Azure (GB200); Dell ObjectScale wiring S3-over-RDMA KV-cache offload upstream into vLLM/LMCache; DDN posting MLPerf-audited 120.68 GB/s from 2RU; networking cost and power rivaling GPU spend, which forces the fabric decision up to the CFO level where storage TCO already lives.

We revise if: UEC-capable hardware fails to reach named production storage deployments by mid-2027 (silicon shipping ≠ storage tiers using it); InfiniBand holds or grows share in new AI back-end deployments through 2027, breaking the Dell'Oro crossover; or S3-over-RDMA stays a two-vendor story without a second round of audited MLPerf-class results from competing platforms.

2026-09-03 — Quota opacity becomes a product category; the two-tier exit becomes the documented default

The call: By end of 2027: (a) effective-quota observability becomes a named product category — gateway/observability tools (Helicone/LiteLLM-class) ship subscription-quota reconstruction as a first-class feature, because no US frontier provider exposes an auditable quota API; (b) at least one major US provider publishes a machine-readable usage-accounting surface for subscription plans under competitive pressure; (c) the two-tier architecture already documented on this index for the data plane — open-weight floor models for token-heavy volume, metered closed frontier for reasoning peaks — becomes the documented default in enterprise AI architecture guidance, driven by subscription-quota risk at least as much as by capability or price.

Signals: Fourteen months of dated evidence on the new AI Subscription Quota Opacity pain point: Cursor's June 2025 re-denomination (apology + refunds), Anthropic's user-discovered March 2026 peak-hour throttle and announced net −17% effective September 14, GitHub Copilot's second unit re-denomination in eighteen months (AI Credits, June 1), xAI's deprecation-with-auto-redirect-billing at ~6.25× (May 15), our own unresolved first-party Ollama Cloud session-yield measurement (~2M → ~700K tokens on an unchanged $20 plan); against it, the counter-trend — OpenAI's August 6 text uncap, Anthropic's May 6 doubling, DeepSeek's permanent price floor — showing bifurcation, not uniform squeeze.

We revise if: US providers converge back to transparent flat quotas (headline unit = independently measurable unit) and hold them for two consecutive quarters; the open-weight floor reprices upward (DeepSeek-class permanent cuts reversed, open-weight API prices rising broadly); or adoption evidence shows users absorb quota opacity without countermeasures — observability, routing, and local-exit categories stalling through 2027.

Explicitly NOT called

The post-MinIO storage race

We have no winner call. RustFS leads mindshare and our own traffic, but it is pre-GA (1.0.0-beta series), independent benchmarks show MinIO still ~3× faster on reads, and clustering is the stated GA blocker. The "RustFS won" narrative circulating in 2026 runs ahead of the software. We index the race; we don't score it yet.

Corrections

Corrections are part of the record. Silent fixes don't build authority; visible ones do.

  • 2026-07-15 — RustFS benchmark framing. We carried a vendor-side "~2.3× faster than MinIO" claim for two content waves. Independent benchmarks (rustfs#2154; a Milvus community evaluation) show the advantage is write-path only — MinIO leads pure reads by ~3× with ~10× better TTFB. The node now carries the full picture and an operator verdict.
  • 2026-07-15 — Apache Paimon vector index. We briefly paired Paimon 1.4's Lumina vector index with "DiskANN" based on a secondary summary. Primary documentation doesn't support the pairing; the claim was removed the same day.
  • 2026-07-15 — Duplicate Vortex node. The July wave briefly created a second Vortex entry alongside an existing richer one; merged same-day, count corrected 417→416.
  • 2026-08-12 — Apache Gravitino incubation status. The node's "What it is" line said "originally developed by Datastrato" without noting graduation; corrected to state Gravitino graduated from Apache Incubator to a full Apache Top-Level Project in June 2025 — a claim that had sat stale on the page for over a year.
  • 2026-08-12 — IDrive e2 feature framing. The top-line summary said "No object lock, no versioning, no event-notification surface." Two of those three claims were already contradicted by evidence sitting in the same node's Recent-developments section (Event Notifications since March 2025, Veeam-certified Object Lock since 2026) but the summary line hadn't been reconciled. Corrected to reflect both features as shipped.
  • 2026-09-03 — DuckDB 2.0 has not shipped. Our September research brief carried the working assumption that DuckDB 2.0 was due to ship in September 2026, and blog #29's fine-print summary described the Quack server protocol as "going stable" — forward-looking phrasing that read as nearer-term than reality. The verification pull is unambiguous: as of early September 2026, DuckDB 2.0 remains in Preview ("A Preview of DuckDB v2.0," August 17); the current stable release is v1.5.5 (July 2026) and the LTS is v1.4.5 (June 2026), per the project's own release history. The duckdb node now states the not-shipped status explicitly. The published site never claimed 2.0 had shipped — the node said "tracked for Fall 2026," which stands — but the expectation deserved a public check because we set it.