How each neocloud actually talks, so our pitch speaks their dialect without thinking about it. Companion docs: Distribution Thesis · Landscape Map · Reference Architecture · Vendor Dossiers. Built from direct quotes in the per-vendor dossiers.
Why this exists
A BD meeting goes one of two ways. Either we use the neocloud’s verbs, product names, and coded shibboleths and they recognize us as someone who did the homework — or we speak the wrong dialect and they spend the meeting mentally translating and don’t take a second one.
Master glossary
How each Tier-A / high-priority Tier-B vendor refers to the five concepts that matter most. When writing a deck for a vendor, use their word.
”Inference”
“Managed Inference” dominates among stack-up Tier-A vendors (Crusoe, Together, Baseten). “Serverless Inference” is the scale-to-zero alternative (Fireworks, Together, RunPod). Vendor-specific brand discipline: CoreWeave AI Inference (caps), Crusoe Managed Inference, Baseten Dedicated Inference (“Inference is everything”), Fireworks Serverless Inference (“From Inference to Intelligence”), Together’s four “Inference” SKUs, Groq differentiates on LPU + Compound, RunPod Public Endpoints.
”Agents”
Everyone above Tier C uses “agentic” as a first-class word — but the framing differs, and you mirror their register:
- CoreWeave — adjective modifying workloads: “Honcho is the memory layer for agentic workloads.”
- Baseten / Fireworks — a use-case category: “Honcho is the missing piece in the AI agent use case on Chains” / “the memory primitive your Agentic AI Systems blog tells customers to build.”
- Anyscale / Modal — a product surface (Agent Skills, Modal Workflows).
- Featherless — a pricing tier (Agent Standard / Agent Pro).
- Groq — a product name (Compound).
”Memory” — the most important pattern
Fireworks is the only Tier-A vendor that names memory as a first-class agent primitive in marketing — and they tell customers to build it themselves with vector stores. Every other Tier-A vendor avoids the word. This is exactly the gap Honcho fills, and the linguistic risk: we must introduce “memory” into vocabularies that don’t use it without triggering a coopetition reaction.
- Fireworks pitch: quote their own blog back — “Lin Qiao writes agents need ‘memory + planning + reasoning + tool use.’ Honcho is that memory layer, productized.”
- Together / Baseten / CoreWeave: introduce memory as a concept, not a category — start with customer pain (“Cursor and Cognition each build their own state layer in-house”) and let “memory” emerge as the name for the gap.
- SambaNova: avoid the word entirely until you’ve disambiguated from their silicon-memory (“agentic cache” = HBM/SRAM/DDR).
”Compound AI” / orchestration
Owned by Baseten (Compound AI, Chains) and Fireworks (Compound AI System f1); also Groq’s Compound product. Using it in a Baseten/Fireworks pitch signals belonging — avoid it elsewhere (sounds borrowed at Together, doesn’t map at CoreWeave).
”Storage / state / persistence”
Together has the strongest named product (Managed Storage, “zero egress fees”). CoreWeave has the most storage products but no application DB (yet — Managed Databases hire open). None of Tier A sells a managed application database today — the substrate of the Honcho-fits-the-gap pitch.
Per-company vocabulary bank (signature phrases)
Preserve capitalization on proprietary product names — lower-casing reads as casual.
Crusoe — “The AI factory company”; “vertically integrated, power → DC → cloud → managed inference.” Products: Crusoe Cloud, Managed Inference, Intelligence Foundry, AutoClusters, MemoryAlloy (inference engine). Avoid generic “GPU cloud.”
Baseten — “The Inference Cloud for the multi-model era”; “Inference is everything.” Products: Chains, Loops, Frontier Gateway, BEI. “Forward Deployed Engineers” (Palantir-coded). Avoid “AI factory” (Crusoe), “AI Native Cloud” (Together).
Fireworks — “From Inference to Intelligence”; “Fastest Inference for Generative AI.” Products: Compound AI System f1, Serverless Inference. Namecheck: Jensen Huang’s “Fireworks is the TSMC of AI Factories” (GTC 2026).
Together — “The AI Native Cloud.” Products: Sandbox, Kernel Collection, GPU Clusters, Model Shaping. Research-prestige names to reference: FlashAttention-4, ATLAS, DeepSWE, Mamba-3. Avoid “Compound AI” and “AI factory.”
CoreWeave — “The Essential Cloud for AI” / “The AI-Native Cloud” (note the hyphen — distinct from Together). Products: Mission Control, SUNK, LOTA, ARENA, CKS, Sandboxes, Serverless RL, W&B Inference, AI Object Storage. Namecheck ClusterMAX Platinum, Cline/W&B Inference.
Anyscale — “Any workload. Any cloud. Anyscale.” Products: Anyscale Runtime, Agent Skills, Ray Serve/Train/Data. Mirror: agents, agentic, MCP, A2A, RAG. Absent: memory, persistence, context — introduce carefully.
Modal — serverless function primitive. Products: Functions, Endpoints, Sandboxes, Modal Workflows, Volumes, Mounts.
Groq — speed-first (tokens/sec on LPU). Products: LPU, GroqCloud, Compound. Caveat: post-NVIDIA acquisition, vocabulary in flux — confirm before committing copy.
Cerebras — ultra-fast wafer-scale inference. Products: WSE-3, CS-3, Cerebras Inference. Namecheck Andrew Ng’s launch quote on agentic workflows.
Featherless — agent-tier hosted inference for open models. Products: Agent Standard, Agent Pro, Managed OpenClaw, Daytona. Speak the RWKV / open-model / Nous dialect.
RunPod — dev-cloud, horizontal control plane. Products: Serverless, Public Endpoints, Flash, Hub, runpod/skills. Signal-of-belonging: cite npx skills add runpod/skills.
Universal lexicon patterns
- A. “Inference” is universal; “model serving” is dead. Say “Honcho’s reasoning operations route to your inference layer,” not “model serving endpoints.”
- B. “AI Native Cloud” is the anchor phrase of the moment — Together (“The AI Native Cloud”), CoreWeave (“The AI-Native Cloud”, hyphen), Baseten (“Inference Cloud for the multi-model era”), Crusoe (“The AI factory company”). Mirror exact phrasing, hyphen included.
- C. Capitalized proprietary product names are brand discipline — preserve caps (Chains, SUNK, LOTA, Sandbox).
- D. Endorsement language is a belonging signal — drop the right name: “Loved the Cline / W&B Inference case study” at CoreWeave; “Nous on Together is the most exciting thing in the open-research stack” at Together.
- E. The words that don’t exist yet — “persistent memory / user-level state,” “cognitive layer,” “identity-level personalization,” “reasoning-driven memory,” and “Honcho” are conspicuously absent from neocloud vocabulary. Our job is to put them in the room. Suggested deck primitives: reasoning-driven memory (the category we own), stateful agents (the outcome), user representation (the primitive), memory-as-attach (the commercial concept), non-fungible state (the moat), memory in space (the cultural anchor).