Neocloud Vendor Dossiers

Consolidated per-vendor profiles for the Honcho in Space initiative — positioning, Honcho fit, contact nodes, and notes. Each vendor is one section, grouped by tier. For the tier rationale see Landscape Map; for pitch language see Neocloud Lexicon; for deployment topologies see Reference Architecture.

Tier A — pursue aggressively, in parallel

Crusoe

• Positioning: “The AI factory company” — vertically integrated power → DC → IaaS → managed inference. Just GA’d Crusoe Managed Inference and Intelligence Foundry. Series E $1.375Bat>$10B (Oct 2025).
• Fit: Only Tier-A vendor whose marketing already anchors on “agentic AI workflows,” but zero memory layer, no agent runtime, no managed Postgres/vector DB — clean gap. Honcho rides on compute; every reasoning call burns Managed Inference tokens.
• Contact: Michael Gordon (ex-MongoDB CFO, now COO/CFO — grasps the host-the-data-layer model) via MongoDB-alumni route; product owner ambiguous between Patrick McGregor (CPO) and Nadav Eiron (SVP Cloud Eng).
• Status: Initial conversation underway. Tactical ask: 90-day co-sell pilot into shared customers (Cursor, Cognition, Odyssey, Figure).
• Note: Structurally the substrate under the other neoclouds — Baseten, Fireworks, Together, Modal all buy GPU from Crusoe. Microsoft anchor customer (Abilene 900MW).

Baseten

• Positioning: “The Inference Cloud for the multi-model era” / “Inference is everything.” Dedicated Inference, Model APIs, Chains (compound-AI/agent orchestration), Loops (RL). Series E $300Mat$5B (Jan 2026) with NVIDIA $150Mstrategic;reportedlyraising$1B at $11B.
• Fit: Most agent-dense customer book (Decagon, Bland, EliseAI, Parallel, Hebbia, Mercor, Scaled Cognition). Chains is the orchestration substrate; Honcho is the missing memory primitive, slotting into their existing Baseten + MongoDB Atlas reference pattern (Honcho coexists, doesn’t displace Atlas).
• Contact: Sarah Guo / Conviction (investor + No Priors host) is the warm path; CEO Tuhin Srivastava is the public voice.
• Status: Intro in motion. Tactical ask: co-build a “Stateful Agents on Baseten” reference arch with 2–3 design partners (Decagon, EliseAI).
• Note: NVIDIA strategic check may bring NIM co-sell pressure; some top accounts may already build memory in-house.

Fireworks

• Positioning: “From Inference to Intelligence” / fastest inference. Serverless 2.0, fine-tuning/RL, Compound AI System f1, transparent per-token pricing. ~$280M+ARR,10T+tokens/day.SeriesC$250M at $4B (Oct 2025). Investors: NVIDIA, AMD, MongoDB, Databricks.
• Fit: Most agent-fluent public POV — Lin Qiao’s blog prescribes “memory + planning + reasoning + tool use” and “vector-based memory with scope/expiration.” They write about memory but sell none. Code/computer-use agent customers (Cursor Composer 2, Cognition, UiPath, Vercel). Cheap tokens make high-volume reasoning-memory economically viable.
• Contact: CEO Lin Qiao (ex-Head of PyTorch at Meta — will instantly grok reasoning-based memory). MongoDB strategic-investor channel for a co-authored reference arch.
• Status: Pre-engagement. Tactical ask: co-author “Stateful Agents on Fireworks” blog + repo; UiPath Computer Use is the most differentiated design-partner target.
• Note: Cognition runs on both Fireworks and Crusoe and is also a Honcho design-partner candidate — watch for duplicated BD. Databricks investment may create agent-product conflict.

Together

• Positioning: “The AI Native Cloud” — research-first (FlashAttention-4, ATLAS, Mamba-3). Four inference SKUs, GPU Clusters, Sandbox (agent runtime, git-versioned FS), Managed Storage, Model Shaping. ~$1Bat$7.5B (late 2025). NVIDIA “Preferred Partner.”
• Fit: “Agents” is a first-class research category; Sandbox ships agent execution/process state but there’s no semantic memory layer — “half a product.” Honcho fills it and amplifies their fastest-inference flywheel.
• Contact: Charles Zedlewski (CPO, ex-Cloudera — gets data-platform-as-leverage) + Arielle Fidel (VP Strategic Partnerships). Research-credibility audience via Tri Dao / Chris Ré (possible joint paper).
• Status: Pre-engagement. Tactical ask: 90-day Topology A pilot, co-target Cursor, Cognition, Decagon (all already on Together).
• Note: Buys GPU from Crusoe. Risk: open “AI Data Products” / “Data Platform” roles could mean they build their own primitive (coopetition). Nous Research is a customer (Plastic adjacency). MongoDB is co-marketing only.

CoreWeave

• Positioning: “The Essential Cloud for AI.” Most complete agentic stack: Sandboxes (execution), Serverless RL (via OpenPipe acq), W&B / Mission Control (observability), RAG (retrieval). AI Object Storage, SUNK, LOTA. Public (Nasdaq: CRWV) since Mar 2025, ~$5B 2025 revenue. Acquired W&B, OpenPipe, Monolith, Marimo, Core Scientific.
• Fit: Has shipped every agent-stack layer except persistent reasoning memory — the single remaining gap. Honcho is the user/conversation/instance-level state pillar alongside W&B (run-level) and OpenPipe (training-level). Compounds inference revenue on Meta/OpenAI/Anthropic-scale contracts.
• Contact: Lukas Biewald (ex-W&B founder, now inside — biggest AI-fluency upgrade) and Kyle Corbitt (ex-OpenPipe, RL lead); CoreWeave Ventures for a de-risking strategic investment.
• Status: Cold but strongest validator of the agentic thesis. Tactical ask: product conversation with Biewald/Corbitt + Ventures + a 90-day design partner on Sandboxes + Inference.
• Note: Time-sensitive — open “Principal Engineer, Managed Databases” role signals a DB-as-a-service tier forming; engage now to position Honcho as the memory tier before architecture freezes. Unlike the other four, CoreWeave is Crusoe’s peer/competitor, not a customer.

Tier B — pursue deliberately, stagger behind Tier A

Anyscale

• Positioning: Commercial platform around Ray (distributed compute); the agent runtime, not memory. “Ray runs faster on Anyscale.”
• Fit: Very high. First-party agentic surface (Agent Skills, MCP, A2A, agentic tuning, Ray Serve) but zero memory product — data is explicitly BYO. Honcho deploys as a Ray Serve service and exposes itself as an MCP server; pitch “you’re the agent runtime, we’re the agent memory.”
• Contact: Robert Nishihara (co-founder, owns product, accessible); warm via Ion Stoica / Berkeley RISELab; Melkote (CEO) for enterprise co-sell later.
• Note: Most academically credentialed team in the set; named agent customers (Physical Intelligence, Cursor, Character.ai). Unknowns: revenue, whether they’re quietly building memory.

Modal

• Positioning: Developer-first serverless AI cloud (Python-defined infra); strongest agent-runtime story via Sandboxes. “AI infrastructure that developers love.” ~$300M ARR by Apr 2026 (~2.5x in four months).
• Fit: Highest of the Tier-B set. ~10M daily Sandbox launches, agents pre-marketed (Ramp, Applied Compute), no memory primitive (only Volumes/Buckets). Fits as a per-agent sidecar memory service; ship a Honcho memory decorator alongside Volumes/Sandboxes.
• Contact: Erik Bernhardsson (CEO, creator of Annoy — knows retrieval natively, reachable). Won’t need a memory 101. Warm via General Catalyst / Lux.
• Note: Flat anti-marketplace org still building its exec layer. Unknown whether they’ve considered building memory natively.

Groq

• Positioning: Inference API on custom LPU silicon — “fast, low-cost inference that doesn’t flake.” 100k+ devs.
• Fit: Strongest agentic surface (Compound productized agent system, remote MCP, deterministic inference as a reliability angle) but zero memory layer — exactly Honcho-shaped. Honcho deploys elsewhere and calls GroqCloud for reasoning. Add a memory.* tool to Compound’s tool set.
• Contact: Simon Edwards (new CEO, ex-CFO — mandate unclear); Hatice Ozen (DevRel, public face of Compound).
• Note (critical): Dec 2025 NVIDIA acquired Groq’s LPU IP + founders for ~$20B; GroqCloud survives independently but in flux. Timing-sensitive — diligence Edwards’ mandate before sinking joint eng hours; risk of wind-down / DGX Cloud merge.

Cerebras

• Positioning: Wafer-scale silicon (WSE-3) inference — “the world’s fastest AI inference,” 20x faster than GPU clouds. Public since May 2026 IPO ($5.55Braised),$510M 2025 revenue, OpenAI $20B+ deal.
• Fit: Medium — best of the silicon-vertical set. First to publicly frame fast inference as the agent unlock (Andrew Ng quote); launch partners include LangChain, LlamaIndex, AgentOps, Weaviate. No owned memory/agent runtime; thin dev platform, silicon-locked. Be the memory partner they lack, listed beside LangChain; sits above Weaviate (passive retrieval) turning it into agent state.
• Contact: Alan Chhabra (EVP Worldwide Partners); Feldman (CEO) and Sean Lie (CTO) both agent-fluent.
• Note: Risk of post-IPO BD slowdown (window H2 2026 / 2027); customer skew is big enterprise/labs vs Honcho’s startup ICP.

Featherless

• Positioning: Serverless flat-rate hosting for 30,000+ open models, spun from RWKV — “neutral infrastructure layer for open AI.” Series A $20M (AMD Ventures, Airbus Ventures), <50 people.
• Fit: High — highest strategic alignment in the set despite smallest size. “Agent Standard/Pro” tiers bundle sandbox + persistent storage; hosts OpenClaw, NemoClaw, and Hermes Agent (Nous) one-click; publishing memory-architecture research. Has storage but no reasoning-driven queryable memory — Honcho’s exact job. Add Honcho to the one-click launcher as the memory layer for the other agents.
• Contact: Eugene Cheah (CEO, RWKV co-lead — reachable, cares about memory, reads technical pitches); warmest intro via Nous ecosystem overlap.
• Note: Flat-rate pricing means lead with upsell/moat, not “more token revenue.” Their own RWKV memory research could become competitive rather than complementary.

RunPod

• Positioning: Developer-first “AI Developer Cloud” — Pods, Serverless, Flash, Hub; 750K+ devs, $120M+ARRononly$22M raised (very capital-efficient).
• Fit: Most agent-vocabulary-fluent IaaS player, but explicitly stays in the infra layer — /agents page, Flash lets agents self-provision GPU, and their own articles tell customers to BYO vector DB / agent memory. No owned memory primitive; roadmap is horizontal. Ship Honcho as a one-line flash deploy + Hub entry; route reasoning to Public Endpoints; joint runpod/honcho skills bundle for Cursor/Claude Code/Cline.
• Contact: Pardeep Singh (CTO, respects OSS-on-existing-primitives); the open “Director, Cloud Marketplace and AI Infrastructure Partnerships” hire. Warm via Daniel Docter (Dell Technologies Capital).
• Note: a16z speedrun + OpenAI Parameter Golf channels feed exactly Honcho’s earliest-stage agent-founder ICP; about to build a partner marketplace (early-partner status valuable).

Tier C — track, open door if easy

Lambda

• Positioning: “The Superintelligence Cloud” — raw bare-metal/on-demand GPU at industrial scale. $1.5B+SeriesE(Nov2025,$5.9B post), NVIDIA-backed, H2 2026 IPO target.
• Fit: Wound down its managed Inference API and pushed customers back to raw GPU — a clear signal they don’t want the application layer. Storage-thin (no managed Postgres/vector DB), so Honcho would be BYO-storage-on-their-compute. The pull-back is actually an asset (likelier to partner than build).
• Note: Telco-style CEO (Combes); route outreach to Paul Zhao (Head of Product) / Robert Brooks IV (CCO). Named Microsoft as a multi-billion customer.

Replicate

• Positioning: “Run AI with an API” — 50,000+ model catalog, Cog custom-model deployment; model- and dev-experience-centric.
• Fit: The target moved — acquired by Cloudflare (Nov 2025), folded into Workers AI. The conversation now belongs to Cloudflare’s AI Cloud org, which has the strongest data stack of the set (R2, Vectorize, D1, Durable Objects, Hyperdrive) and names Durable Objects as the agent-state primitive — but no reasoning-based memory layer. Build-vs-buy risk; speed matters.
• Note: Right node Rita Kozlov (Cloudflare VP Product); Ben Firshman as internal champion. Pitch: Durable Objects is state, not memory — offer Honcho as the reasoning layer above their data stack.

Nebius

• Positioning: “AI cloud designed for the agentic era” (Jensen’s words at the $2B NVIDIA partnership, Mar 2026); Yandex international spinoff (Nasdaq: NBIS).
• Fit: Strong — dossier says “pitch aggressively.” Best storage primitives of the group (Token Factory ships a pgvector-powered managed vector store, so Honcho can plausibly run on Nebius-native Postgres — a major topology advantage). Already uses “agentic” in product naming (no education needed). Gap is reasoning-based memory; they’ve solved retrieval (Tavily + pgvector) but not the cognitive layer.
• Note: Acquired/integrated Tavily (agentic search); on DGX Cloud Lepton marketplace. Token Factory monetizes per-token, so Honcho memory ops drive volume — structurally aligned.

Lepton

• Positioning: Now “NVIDIA DGX Cloud Lepton” — a global GPU compute marketplace under NVIDIA’s software interface (NIM/NeMo/NVCF). Acquired by NVIDIA (Apr 2025).
• Fit: Low-to-medium, structurally awkward — not a direct BD target. Selling “into Lepton” means negotiating with NVIDIA’s NIM/Blueprints team (multi-year enterprise motion). No storage primitive. But it passively aggregates demand from underlying neoclouds (Nebius, Together, CoreWeave, Lambda) — winning those gets Lepton-routed traffic for free.
• Note: Founder Yangqing Jia (Caffe/PyTorch/ONNX creator) now a VP at NVIDIA — keep warm as a relationship via PyTorch/Meta AI alumni, not as a pitch.

DeepInfra

• Positioning: “LOW-COST · FAST · SIMPLE · RELIABLE” inference cloud; 100+ models, owned hardware. $107MSeriesB(May2026,NVIDIA+Supermicro).Costleader($0.09/M input tokens), SOC2 / ISO 27001 / zero retention.
• Fit: Lowest of the C batch for co-sell — better as a Honcho inference backend than a partner. No agentic posture, no agent product, no memory positioning, no up-stack roadmap. Operator-fluent CEO (pitch dollars, not architecture).
• Note: Recommended posture: use them as Honcho’s reasoning-call backend (especially regulated deployments) via DeepStart; re-tier only if they move up-stack.

Tier D — low fit, honestly

Vast

• Positioning: Two-sided GPU rental marketplace (17K+ GPUs, 350+ hosts); “market operating system for the agentic economy” where agents procure their own compute.
• Fit: Low. Pure marketplace, no managed-services layer and no salesforce calling enterprise buyers — zero co-sell leverage. No state plane (per-host storage), so Honcho-on-Vast = BYO-everything. Founder Jake Cannell’s anti-centralization ideology runs against vertical integration into managed memory.
• Note: Financially under-disclosed/constrained. Narrow opening: agents-as-buyers (a Honcho agent renting its own GPUs) — speculative product integration, not co-sell.

Hyperbolic

• Positioning: “The open-access AI cloud” — decentralized GPU marketplace (~75% cheaper) + OSS inference API, with verifiable-inference/privacy as a web3-flavored differentiator. ~$20M raised.
• Fit: Low (with one twist). No managed services to attach to, no state plane (BYO database), decentralization ideology in tension with bundled SKUs. The twist: a real wedge in on-chain/crypto-native agents where memory must be auditable/verifiable — niche co-design, not flagship co-sell.
• Note: CEO Jasper Zhang (Berkeley math PhD, ex-Citadel/Ava Labs), crypto-native; HuggingFace inference-provider integration is the main distribution edge. Stay friendly/informal, don’t burn BD cycles.

SambaNova

• Positioning: Vertically integrated custom silicon (RDU chips: SN40L, SN50 shipping H2 2026); “purpose-built for agentic inference,” exited training to focus on inference.
• Fit: Low (structural lock-in). Custom-silicon lock-in gives Honcho no special leverage — it just calls their OpenAI-compatible API. Sharp vocabulary clash: their “memory” / “agentic cache” means silicon-level HBM/SRAM/DDR tiering, not agent state. No platform storage primitive.
• Note: Financial pressure (analyst est. <$50MARR,$200M+ burn); Intel acquisition talks collapsed Oct 2025, then Intel joined a $5.1B refinancing Feb 2026. Listing them as a supported inference backend is sufficient. Narrow indirect wedge: engage their sovereign-cloud operators (Argyll UK, SCX, Infercom), not SambaNova.