The operational counterpart to the Distribution Thesis — who we pursue, in what order, on what evidence. Companion docs: Neocloud Lexicon · Reference Architecture · Vendor Dossiers. Synthesized from 19 primary-source vendor dossiers.
Tiering
Tiers reflect strategic fit, not company size. Drivers, in priority order: (1) are they actually moving on agentic services? (2) do they have customer mass that needs memory? (3) is the integration topology clean? (4) is the inference economics aligned (do they want more tokens through their pipe)? (5) is leadership AI-fluent or operator-DNA (sets pitch register)?
Tier A — pursue aggressively, in parallel, now
The five we run as a parallel BD process. Each gets a dedicated owner, a tailored deck, and a 90-day pilot proposal.
| Company | Why Tier A | First node | Status |
|---|---|---|---|
| CoreWeave | Most complete agentic stack (Sandboxes + Serverless RL + W&B + RAG). Memory is the explicit last gap. Open Managed Databases PE role — window before architecture freezes. | Lukas Biewald (ex-W&B) and Kyle Corbitt (ex-OpenPipe), both inside; CoreWeave Ventures as a parallel wedge. | Cold |
| Fireworks | Most agent-fluent public POV. Lin Qiao’s blog tells customers to add “vector-based memory with scope/expiration” — Honcho’s problem space in their words. Computer-use/code-agent customer book. MongoDB strategic investor. | Lin Qiao (CEO, ex-PyTorch lead). Warm via Sarah Guo (Conviction). | Approach pending |
| Together | Only neocloud whose research bench could have built Honcho (Tri Dao, Chris Ré, Percy Liang). Sandbox shipped without a memory layer. Nous Research is a customer (Plastic adjacency). | Charles Zedlewski (CPO, ex-Cloudera); Arielle Fidel (VP Partnerships); Tri Dao academic channel. | Approach pending |
| Baseten | Most agent-dense customer book (Decagon, Bland, EliseAI, Hebbia, Mercor). NVIDIA $150M strategic. Published MongoDB Atlas + Chains reference pattern. Chains is the orchestration substrate; Honcho is the missing memory primitive. | Tuhin Srivastava (CEO). Warm via Sarah Guo (also Series E investor). | Intro in motion |
| Crusoe | Already says “agentic AI era.” No memory primitive, no agent runtime — structural gap. Ex-MongoDB CFO Michael Gordon now COO/CFO. Critically, Baseten/Fireworks/Together/Modal are Crusoe GPU customers — Crusoe is the substrate under the others. | Patrick McGregor (CPO) or Nadav Eiron (SVP Cloud Eng). Warm via Michael Gordon (MongoDB alumni). | Initial conversation underway |
Cross-tier insight: Baseten, Fireworks, and Together all buy raw GPU from Crusoe. MongoDB shows up adjacent to four of the five Tier-A vendors — but MongoDB is a co-marketing channel, not a storage backend. Honcho runs on Postgres + pgvector; it coexists alongside Atlas in customer stacks, it doesn’t run on it. A MongoDB partner-directory listing amplifies reach; the integration story is alongside Atlas, not on top of it.
Tier B — pursue deliberately, stagger behind Tier A
| Company | Why Tier B | First node |
|---|---|---|
| Anyscale | High fit — explicit agentic surface (Agent Skills, MCP, A2A), but Ray-centric stack constrains generality. No memory product. | Robert Nishihara (co-founder); warm via Ion Stoica. |
| Modal | First-class “agentic” product surface, strong dev-experience brand, Modal Workflows. Smaller but architecturally tight; Honcho deploys naturally as a Modal function. | Erik Bernhardsson (CEO, creator of Annoy — knows retrieval). |
| Groq | Most agent-forward product (Compound, 100k+ devs, MCP) but NVIDIA acquired the LPU IP + senior leadership Dec 2025 — org clarity murky. Timing-sensitive. | Simon Edwards (new CEO, ex-CFO). Confirm post-deal product ownership first. |
| Cerebras | Only silicon-vertical to publicly recognize the agent ecosystem (LangChain, LlamaIndex, AgentOps, Weaviate launch partners). IPO’d May 2026. | Alan Chhabra (EVP Worldwide Partners). |
| Featherless | Highest strategic alignment despite smallest size. Pricing tiers literally “Agent Standard/Pro.” Hosts Hermes Agent (Nous) one-click. CEO Eugene Cheah publishes memory research. | Eugene Cheah; warmest intro via Nous Research overlap. |
| RunPod | Dev-cloud whose own articles tell customers to BYO vector DB — a philosophical fit. $120M ARR, 750K devs, agent-IDE skills distribution. | Pardeep Singh (CTO); the incoming partnerships hire. Warm via Daniel Docter (Dell Technologies Capital). |
Tier C — track, open door if easy
| Company | Why Tier C |
|---|---|
| Lambda | Long-tenured GPU rental, slower stack-up cadence, wound down its managed Inference API. Reasonable inference-backend listing; low-priority co-sell. |
| Replicate | Strong dev brand but model-share focused — and acquired by Cloudflare (Nov 2025); conversation now belongs to Cloudflare’s AI Cloud org. |
| Nebius | Yandex carve-out; strong storage primitives (Token Factory ships pgvector-powered managed vector store). Worth tracking as a European sovereign play; dossier argues “pitch aggressively.” |
| Lepton | Acquired by NVIDIA (Mar 2025) — now NVIDIA DGX Cloud Lepton. Relationship routes through NVIDIA, not a sovereign partner. |
| DeepInfra | Pure inference-as-a-service, no agentic ambition. Better positioned as a Honcho commercial backend (SOC2 + ISO 27001 + zero retention) than a co-sell partner. |
Tier D — low fit, honestly
| Company | Why low fit |
|---|---|
| Vast.ai | Pure marketplace, no managed-services layer; founder ideology (Jake Cannell) is explicitly anti-vertical-integration. No state plane for Honcho to attach to. |
| Hyperbolic | Smaller web3-flavored Vast analog (~$20M raised). Verifiable inference is a real niche for on-chain agents, but the niche is small. |
| SambaNova | Custom-silicon lock-in plus genuine financial distress. Vocabulary clash: their “memory” means HBM/SRAM/DDR silicon. Attention is on survival. |
Strategic patterns across the corpus
1. The stack-up is underway, and agentic services is the universal next layer. Every neocloud above Tier C has announced or shipped a stack-up motion in 2025–26: Power/GPU → Managed Inference → Compound AI / Orchestration / Sandboxes → Memory / Agent State. Memory is the missing primitive on every Tier-A roadmap. We’re pitching the next defined slot, not speculation.
2. MongoDB is the adjacent ecosystem, not a pre-wired backend. Four of five Tier-A neoclouds have MongoDB adjacency (Baseten reference arch, Together partner page, Fireworks investor, Crusoe’s ex-Mongo COO). But Honcho has no MongoDB driver — Honcho-on-Atlas would require unbuilt engineering. The right framing: “complementary memory layer alongside Atlas.” The ecosystem value is a discovery/co-marketing channel; a single well-cultivated MongoDB BD relationship is worth a parallel track.
3. NVIDIA is everywhere in (and above) the cap stack. Strategic investor in Baseten, Fireworks, Crusoe, DeepInfra; acquired Lepton and Groq’s LPU IP; preferred partner of Together; CoreWeave exemplar. Implications: (a) Honcho’s NVIDIA-neutrality is an asset worth a deck line; (b) NVIDIA AI Enterprise / NIM is the meta-competitor at the memory layer — track every GTC for memory/state primitives.
4. Customer pull is visible in specific logos. Cursor (on Crusoe, Fireworks, Together, Baseten), Cognition, Decagon, UiPath, EliseAI, Bland, Hebbia, Mercor, Nous Research, plus voice/health clusters (Cresta, Ambience, Abridge). These are the Honcho design-partner pipeline — already on the clouds we’re pitching. One Tier-A co-sell surfaces 5–10 highest-fit candidates. We don’t need to find customers; we ride neocloud co-sell into customers their BD already calls weekly.
5. Leadership AI-fluency splits the corpus, which forks the pitch deck.
- AI-native / research-credible (pitch technical depth + benchmarks): Together, Fireworks, Anyscale, Baseten, Featherless.
- Operator / financial-engineering DNA (pitch revenue, attach, unit economics): CoreWeave, Crusoe, RunPod, Lambda, Replicate, Modal (middle ground).
- Silicon-vertical (pitch inference-volume per chip): Cerebras, SambaNova, Groq. Shortest deck — they get “more agent volume = more chip demand” without convincing.
What this means for the process
- Five Tier-A conversations in motion simultaneously (two already: Crusoe, Baseten); three to activate. Each gets a named owner and a tailored deck variant.
- A MongoDB BD relationship is critical parallel-track infrastructure (it amplifies four Tier-A pitches at once), not a Tier-A conversation.
- An NVIDIA-watching function tracks NIM / AI Enterprise announcements.
- Tier B is reactive — take the warm intro when it shows (Featherless via Nous is the easiest, activate quickly).
- Tier C/D are listings, not pitches — supported as inference backends, no BD bandwidth.