Agentic Software

Agentic Software — Basics and Investments

Prepared for: Family office principals (non-technical, exec-grade) Scope: US-centric, private-market focused Date: 2026-05-04 Author: Marc Witney

How to Read This Document

This is a deep brief, not a summary deck. It is structured in two parts:

Part 1 — Technical Basics. What an AI agent actually is, what makes it different from a chatbot, where it breaks, what changed in 2024–2026 to make agents viable. ~2,500 words. Read this first.
Part 2 — Investment Landscape. Three sub-parts:
- 2A: Horizontal Infrastructure + Foundation Layer — frontier model labs (Anthropic, OpenAI, xAI), GPU compute (CoreWeave, Cerebras), inference (Together, Groq), orchestration (LangChain), agent infra (Browserbase, E2B), memory/data (Pinecone, Turbopuffer), MCP ecosystem.
- 2B: Vertical Application Companies — coding (Cursor, Cognition), CX (Sierra, Decagon), legal (Harvey), healthcare (Abridge, OpenEvidence), sales (Clay), finance (Ramp, Basis), search (Glean, Perplexity), voice (ElevenLabs, Deepgram).
- 2C: Macro — Money Flow, Theses, Incumbents, Risks — capital flow charts, six investment theses, incumbent threat analysis (Salesforce Agentforce, ServiceNow), family-office access matrix, 12–24 month watchlist.

Each section includes inline source citations and a closing “investor takeaway” or risk note. Numbers can move quickly — refresh marks before any commitment decision.

Executive Headlines (the 12 things that matter)

Agents are the 2026 product category, not a 2023 demo. Four converging shifts — reasoning models, 1M+ token context, reliable tool use, MCP standardization — turned “build an agent” from a research project into commodity engineering this year.
Capital is concentrating, not dispersing. Q1 2026 alone saw $242B of AI VC; four companies (OpenAI $122B, Anthropic $30B, xAI $20B, Waymo $16B) absorbed $188B — 62% of all global VC. AI is now ~80% of total venture capital.
Hyperscalers are betting their balance sheets. Combined 2026 capex: ~$610–700B (75% AI-related). Amazon expects $17B negative free cash flow. This is the macro story — if agentic workloads materialize, picks-and-shovels print money. If not, public AI multiples compress severely.
The “real businesses” list is shorter than headlines suggest. ~12 application-layer companies have $100M+ ARR + verifiable customers + reasonable defensibility: Cursor, Cognition, Sierra, Harvey, Glean, Abridge, OpenEvidence, Ironclad, Ramp, ElevenLabs, Perplexity, Clay. Out of 207 AI unicorns minted since 2024.
Anthropic now leads OpenAI on revenue. $30B run-rate (Apr 2026) vs. OpenAI’s ~$24B. New Anthropic round in talks at ~$900B; OpenAI just closed at $852B.
Multiples are extreme — and starting to crack. Median private AI revenue multiple is 37.5x ARR (vs. ~7.8x for traditional SaaS). Outliers at 100x+ (Sierra, Decagon, Rox). Poolside’s Series C collapsed in April 2026 — first publicly reported sign that LPs are pushing back.
Best moats: regulated verticals. Legal (Harvey), healthcare (Abridge, OpenEvidence), accounting (Basis). Slow-but-sticky buyers, workflow lock-in, proprietary data flywheels. Best risk-adjusted ARR multiples.
Worst category right now: AI BDR / outbound sales. 11x is a cautionary tale (inflated logos, three-month break clauses). Race to zero on per-meeting pricing. Will collapse to 2–3 standalone winners + acquihires.
Incumbents are real threats, not roadkill. Salesforce Agentforce hit $800M ARR — fastest product launch in Salesforce’s 26-year history. ServiceNow Now Assist crossed $600M ACV. Microsoft Copilot Studio: 230,000+ orgs. They have the distribution moat.
Outcome-based pricing is the structural shift. Sierra ($1–3 per resolution), Decagon (per-resolution), EvenUp (per-case), Hippocratic ($9/hour per agent). Compresses traditional SaaS comparables — a $50M-ARR outcome-priced business is not the same asset as a $50M per-seat SaaS business.
For most family offices, public proxies are the right answer. MSFT, NVDA, GOOG, META, AMZN, ORCL, AVGO, CRWV trade at ~25–40x earnings vs. 50–100x ARR for privates. Liquid, no carry, no lockup. Below ~$1B AUM, this should be the bulk of credible AI exposure.
Watch for reverse-acquihires. Adept (Amazon), Inflection (Microsoft), Character (Google), Windsurf (Google + Cognition). Investors get capital back, founders go to MAANG, standalone equity outside the carve-out is worthless. If buying secondary in any well-funded but stalled application-layer agent, model that scenario explicitly.

Part 1 — Agentic Software: Technical Basics

1. What Is an AI Agent?

An AI agent is software that uses a large language model (LLM) to direct its own work toward a goal — choosing which tools to call, in what order, and when it’s done — rather than following a fixed script written by a human [Source: anthropic.com/research/building-effective-agents, 2025]. Anthropic frames it precisely: an agent is a system “where LLMs dynamically direct their own processes and tool usage, maintaining control over how they accomplish tasks.”

The distinctions matter:

A single LLM call (e.g., asking Claude to summarize a paragraph) is a one-shot text-in / text-out function. No actions, no memory, no follow-through.
A chatbot is a sequence of LLM calls inside a conversation. It remembers what you said five turns ago, but it cannot do anything in the world beyond produce text.
A RAG system (Retrieval-Augmented Generation) is a chatbot plus a search step: before answering, it fetches relevant documents from a database and stuffs them into the prompt. Still text-in / text-out — just with grounded inputs.
An agent is all of the above plus the authority to take real actions (send an email, query an API, run code, place a trade), observe what happened, and decide what to do next — repeatedly, until the goal is met.

Concrete example. Ask a chatbot “Should I invest in NeoCarbon?” and you get an essay. Give the same question to an agent and it can: pull the latest 10-K from SEC EDGAR, run a Python script to compute leverage ratios, search news for the past 90 days, fetch comparables from a CRM, draft a memo, and email it to you — adjusting its plan if any step fails. The chatbot tells you about diligence; the agent does diligence.

Investor takeaway: when evaluating “AI” companies, the central question is whether the product takes autonomous actions on the user’s behalf, or merely generates text that a human still has to act on — those are very different businesses with very different defensibility.

2. Anatomy of an Agent

Model (the brain). The underlying LLM — Claude, GPT-5, Gemini — that does the reasoning. This is the single biggest determinant of how capable the agent is, and is the component that has improved most dramatically in 2024–2026 [Source: anthropic.com/news/claude-sonnet-4-5, 2025].

Planner. The component (often just a structured prompt to the model itself) that decomposes the user’s goal into ordered sub-steps. In simple agents the planner and the model are the same thing; in advanced systems the planner is a dedicated reasoning pass that produces a written plan before execution begins.

Tools. The agent’s hands. Tools are functions the model can call — search_web(query), read_pdf(file), send_email(to, body), execute_sql(query). The model produces a structured request (“call search_web with query=‘NeoCarbon revenue’”), the harness runs the tool, and the result comes back as text the model can read [Source: anthropic.com/engineering/writing-tools-for-agents, 2025].

Memory. Two kinds. Short-term memory is the conversation/context window — everything the model can currently “see” (now hundreds of thousands of tokens, often a million-plus in 2026). Long-term memory persists across sessions, typically via a vector database or structured store the agent reads from and writes to deliberately [Source: weaviate.io/blog/context-engineering, 2025]. Memory is where most production agents quietly fail.

Orchestration loop. The runtime that ties it all together: it sends the prompt to the model, parses out tool calls, executes them, appends results back into context, and loops. This is the agent’s heartbeat — and the place where reliability engineering happens.

Investor takeaway: the model is becoming a commodity input; durable value in agent companies tends to live in the memory layer, the tool integrations, and the orchestration logic — i.e., the parts that aren’t a single API call away.

3. The Agent Loop

Every major agent today runs some variant of the ReAct pattern — Reason, Act, Observe — formalized in 2022 and now the convergent design across OpenAI, Anthropic, Google, Microsoft, and Meta [Source: ibm.com/think/topics/react-agent, 2025]:

Perceive — Read the current state (user request, prior steps, tool outputs).
Plan / Reason — Decide the next step. (“I need the company’s latest revenue figure.”)
Act — Call a tool. (search_web("NeoCarbon Q3 2025 revenue"))
Observe — Read the result back into context.
Repeat — Until the goal is complete or the agent asks for help.

Worked example: agent does diligence on a private company.

Goal: “Build a one-page memo on NeoCarbon, a Series B carbon-removal startup.”

Loop 1. Reason: I need basic firmographics. Act: call crunchbase_lookup("NeoCarbon"). Observe: gets founding date, last round size, investors.

Loop 2. Reason: I need their tech claims. Act: call search_web("NeoCarbon technology direct air capture"). Observe: gets press releases and a TechCrunch article.

Loop 3. Reason: Press is thin — let me find primary sources. Act: call fetch_url("neocarbon.com/whitepaper.pdf") then read_pdf(...). Observe: extracts cost-per-ton claims.

Loop 4. Reason: I should sanity-check those costs against industry benchmarks. Act: call search_web("DAC cost per ton 2025 IEA"). Observe: gets a benchmark range.

Loop 5. Reason: Now I have enough. Act: call write_memo(...). Observe: memo file produced.

Stop.

What makes this powerful — and brittle — is that the agent didn’t know in advance that it would need step 3 or step 4. It discovered them. That adaptive quality is what distinguishes an agent from a workflow [Source: anthropic.com/research/building-effective-agents, 2025].

Investor takeaway: the value of the loop is adaptability — but every iteration costs tokens and time, so the economics of a product depend heavily on how many loops a typical task takes and whether that count is bounded.

4. Single-Agent vs Multi-Agent Systems

A single-agent system has one model running one loop. A multi-agent system has multiple models — often with specialized roles (“researcher,” “writer,” “critic”) — coordinating, usually under an orchestrator that delegates and aggregates.

Why multi-agent? Three reasons that matter:

Parallelism. A research agent can spawn five sub-agents to investigate five companies simultaneously instead of sequentially.
Context pressure relief. Each sub-agent has its own context window, so they collectively process far more material than one agent could fit.
Specialization. Different prompts, models, or tool sets per agent (e.g., a fast/cheap model for triage, a stronger model for synthesis).

Anthropic’s internal research system — orchestrator plus parallel sub-agents — outperformed a single Claude Opus 4 by 90.2% on their internal research benchmark [Source: anthropic.com/engineering/multi-agent-research-system, 2025].

When it adds chaos. Multi-agent systems are dramatically harder to build and debug. DeepMind articulated the trade-off cleanly: Net Performance = (Capability + Collaboration Benefits) − (Coordination Chaos + Communication Overhead + Tool Complexity). Add too many agents and the right side wins. Failure modes include cascading errors (one agent’s bad output poisons the next), runaway token costs (agents talking past each other for dozens of turns), and untraceable bugs across distributed reasoning [Source: imaginexdigital.com, 2025].

The rule of thumb that has emerged: use multi-agent when the task naturally parallelizes (research, broad search, multi-document review). Avoid it for sequential reasoning, where one strong agent with a long context will usually win.

Investor takeaway: be skeptical of “multi-agent” as a marketing term — it’s often added complexity with negative ROI; it pays off mainly for breadth-first problems with cleanly separable sub-tasks.

5. Frameworks Landscape

These are the building blocks engineers use to assemble agents. Think of them as the analog to Excel templates and accounting software for an investment firm — none of them produces value alone, but they shape what the team can build.

LangChain / LangGraph — The most widely adopted open-source toolkit. LangChain is the original library for chaining LLM calls; LangGraph, which hit v1.0 in late 2025, is a graph-based orchestration runtime for stateful agents and has become the default for serious LangChain deployments [Source: latenode.com, 2026]. Used by a long tail of startups and enterprises.
CrewAI — Role-based framework: you define agents as “researcher,” “writer,” “reviewer” and they collaborate. Lowest learning curve; popular with rapid prototyping and small teams.
Microsoft AutoGen — Pioneered conversational multi-agent patterns (agents talking to each other). Effectively in maintenance mode in 2026 — Microsoft has shifted strategic investment to the broader Microsoft Agent Framework. Still widely deployed inside the .NET ecosystem.
OpenAI Agents SDK — Released March 2025. Lightweight, Python- and TypeScript-first; primitives are agents, handoffs, guardrails, and tracing. Provider-agnostic but obviously optimized for OpenAI models. Used by teams building on the OpenAI stack [Source: openai.github.io/openai-agents-python, 2025].
Anthropic Claude Agent SDK — Anthropic’s official SDK, evolved from the internals that power Claude Code. Emphasizes inspectable reasoning, sandboxed tool use, and tight Claude integration. Popular among teams prioritizing safety/auditability.
MCP (Model Context Protocol) — Not a framework — a standard. MCP is to AI agents what USB is to peripherals: a wire protocol so any agent can plug into any tool/data source without custom integration code. Anthropic introduced it in November 2024; OpenAI adopted it across their stack in March 2025; Google followed in April. In December 2025, Anthropic donated MCP to the Agentic AI Foundation under the Linux Foundation, co-founded with Block and OpenAI — making it a vendor-neutral industry standard. Anthropic reported >97M monthly SDK downloads by late 2025 [Source: anthropic.com/news/model-context-protocol, 2024; en.wikipedia.org/wiki/Model_Context_Protocol, 2025; thenewstack.io/why-the-model-context-protocol-won, 2025].

The framework distinction matters because frameworks compete and may be replaced (LangChain’s dominance, for instance, is contested by every model lab’s first-party SDK), whereas MCP, as a standard, increasingly underpins all of them.

Investor takeaway: framework companies have to fight for developer mindshare against well-resourced first-party SDKs from Anthropic and OpenAI; standards (MCP) tend to commoditize the layer beneath them, which is bullish for tool-providers and tool-marketplaces and bearish for closed integration plays.

6. Why Agents Break (Reliability)

Agents in 2026 are good enough to demo and shaky enough to lose money in production. The dominant failure modes:

Hallucination. The model invents facts. In agents this is more dangerous than in chatbots because invented facts drive actions. Production studies find user abandonment spikes once hallucination rates pass ~30% [Source: galileo.ai/blog/agent-failure-modes-guide, 2025].
Tool misuse. A subtler, agent-specific failure: the model calls the right tool with the wrong arguments, or — worst case — claims to have called a tool it didn’t actually call, confidently reporting success. “The AI lying about doing its job” is now recognized as a distinct reliability category [Source: medium.com/@yaseenmd, 2025].
Planning errors / cascades. Early mistakes propagate. A wrong assumption in step 1 corrupts steps 2–10, and the agent rationalizes its way to a confidently wrong final answer.
Cost & latency blowups. A misbehaving agent can loop indefinitely. Reflexion-style self-critique loops can consume 50× the tokens of a linear pass; an unconstrained coding agent can rack up $5–8 per task [Source: online.stevens.edu, 2025].
Eval difficulty. Unlike software with deterministic tests, agents are non-deterministic. “Did this agent do its job correctly?” is itself an LLM-graded judgment call, which means evals are noisy, expensive, and hard to trust. This is the reason “production-ready” is a meaningful claim, not a shrug.

What “production-ready” means in 2026. Concretely: traced, logged, evaluated against a curated test set, sandboxed for tool execution, gated by human approval at high-stakes steps, and budget-capped at the harness level. The maturity gap between a working demo and a production deployment is roughly 10× the engineering effort — which is why most enterprise AI pilots stall there [Source: galileo.ai/blog/hidden-cost-of-agentic-ai, 2025].

Investor takeaway: a polished demo is nearly free in 2026 — it’s table stakes, not a moat. Real differentiation is in evals, observability, and the operational discipline to run agents reliably; ask portfolio companies about their eval suite, not their model.

7. Cost & Latency Reality Check

Order-of-magnitude numbers as of mid-2026 for a complex agent task (e.g., a diligence report, a multi-step coding fix, a research brief):

Tokens consumed: ~50K–500K. Tool-using agents consume roughly 5–20× more tokens than simple LLM chains because every observation gets re-read on every loop [Source: medium.com/@klaushofenbitzer, 2025].
Token cost: $0.05–$5 per task at frontier-model pricing, depending on length and how aggressively prompt caching is used. Caching cuts input cost ~90% and latency ~75% on repeated context.
Latency: 30 seconds to 10 minutes for a substantive task. A bare LLM call is ~800ms; an orchestrator-worker flow with reflection is 10–30s; a long-horizon coding or research agent runs minutes to tens of minutes.
Voice agents are the fast end of the spectrum: 500–800ms response is the leading-platform target; >2s feels broken to users [Source: a16z.com/ai-voice-agents-2025-update, 2025].

Improving fast: per-token prices (down ~10× over 2024–2026 at equivalent capability), throughput (Mercury 2 and similar diffusion-LLMs run >600 tokens/sec), prompt caching, and code-execution patterns that turn 150K-token prompts into 2K-token ones (~98% reduction) [Source: anthropic.com/engineering/code-execution-with-mcp, 2025].

Not improving fast: end-to-end latency on long-horizon tasks (a 10-minute research run can’t be parallelized arbitrarily; some tools just take time), and the fundamental tension between accuracy and speed — high-reasoning models like GPT-5-high run ~11× slower than fast models on the same hardware.

Investor takeaway: unit economics in agentic products are still volatile — a viable margin at today’s prices may evaporate when usage scales, or improve dramatically as model prices fall; press hard on cost-per-completed-task (not cost-per-token) and how it has trended over the last 12 months.

8. What Changed in 2024–2026

Agents were a 2023 curiosity and a 2026 product category because four things converged:

Reasoning models. GPT-o1 (late 2024), Claude 3.7 Sonnet’s “extended thinking,” Claude 4 / 4.5 / 4.7, GPT-5/5.5, and Gemini 2/3 introduced models that can take seconds-to-minutes to “think” before acting. Claude Opus 4.5 hit 80.9% on SWE-bench Verified — agentic coding moved from research demo to deployable product [Source: anthropic.com/news/claude-sonnet-4-5, 2025].
Long context. Context windows grew from ~8K tokens (2023) to 200K (Claude 3) to 1M+ (Claude 4 / Gemini 2). Agents now have working memory big enough to hold an entire diligence package in context.
Reliable tool use. Models are dramatically better at producing well-formed tool calls and recovering from tool errors. Claude Sonnet 4.5 and GPT-5.5 are explicitly tuned for “long-running agent tasks without getting stuck.”
MCP standardization. The single biggest infrastructure change. Before MCP, every agent needed bespoke integrations to every tool. After MCP — adopted across Anthropic, OpenAI, Google, Microsoft, and donated to a Linux Foundation entity in December 2025 — any compliant agent connects to any compliant tool. This is the moment a fragmented ecosystem started to look like a platform [Source: thenewstack.io/why-the-model-context-protocol-won, 2025].

Investor takeaway: 2026 is the first year where “build an agent” stops requiring a research team — which means competition is about to compress, and the winners will be the companies with proprietary data, proprietary distribution, or proprietary workflows that the platforms can’t replicate.

Sources

Part 2A — Horizontal Infrastructure + Foundation Layer

Family-office market scan, US-centric. Data as of 2026-05-04. Where numbers are estimates or older than 12 months, this is flagged inline. ARR figures from third parties (Sacra, The Information, TechCrunch) are estimates unless otherwise noted.

Framing for the Reader

The horizontal infrastructure stack is where the AI economy is being capitalized first. Capital is concentrating in three places: (1) frontier labs that train and serve the models (Anthropic, OpenAI, xAI), (2) the GPU compute layer that physically powers them (CoreWeave, Lambda, Cerebras), and (3) the abstraction layers above — inference platforms, orchestration, memory, tools — where businesses are smaller but multiples are extreme (10x–50x ARR is now common per Q1 2026 data). At the top of the stack, ~$200B of equity capital has been raised in the last 18 months at valuations totaling well over $2T. The bull case is straightforward: agents are about to drive multi-trillion-dollar workload volumes through this stack. The bear case is equally clear: the labs sit on top of the same NVIDIA-CoreWeave compute base, so the question is how much economic rent each abstraction layer can hold versus get squeezed out.

A. Foundation Models — the “Brains”

The frontier model layer is now structurally an oligopoly: three US labs (OpenAI, Anthropic, Google DeepMind) plus xAI represent the bulk of frontier-grade capability. Meta is the dominant open-weights player but is signaling a pivot to closed models. Mistral is the European hedge. Aggregate ARR across the top four US-anchored labs has crossed ~$80B in run-rate as of April 2026.

Anthropic

Description: US frontier-model lab; maker of Claude (Opus, Sonnet, Haiku). Founded 2021 by ex-OpenAI researchers; safety-forward positioning; principal Amazon partner.
Latest round: Series G, Feb 12, 2026 — $30B at $380B post-money, led by GIC and Coatue, co-led by D.E. Shaw Ventures, Dragoneer, Founders Fund, ICONIQ, MGX. A new $40B–$50B round at ~$900B valuation was reported imminent as of late April 2026 (TechCrunch, Apr 30, 2026).
Total raised: ~$60B+ pre-current round (incl. multi-year Amazon and Google strategic commitments).
ARR: $30B run-rate in April 2026, up from $14B in Feb, $19B in Mar. [Bloomberg, SaaStr]. Caveat: Anthropic books cloud-reseller revenue gross, which inflates top-line versus net-reporting peers per Sacra.
Headcount: ~2,500–3,000 (estimated, based on hiring pace through 2025).
Notable customers: Amazon, Notion, Zoom, Bridgewater, Snowflake, Cursor, Replit, GitHub Copilot (Claude as default), 1,000+ customers spending >$1M/yr.
Moat: Coding leadership (Claude is the de facto default for code agents in 2026), enterprise trust posture, Amazon distribution, and a research culture pulling top talent out of OpenAI.
Key risk: Compute dependency — Anthropic does not own its silicon and must access multi-gigawatt capacity through AWS/Google/Broadcom partnerships, which are themselves competitors.

OpenAI

Description: Maker of ChatGPT, GPT-5, Sora; the consumer leader of the AI era. Microsoft is the principal commercial partner.
Latest round: Mar 31, 2026 — $122B raise at $852B post-money valuation, co-led by SoftBank, a16z, D.E. Shaw, MPX, TPG; strategic anchors Amazon, NVIDIA, SoftBank, Microsoft. Notably included $3B from retail investors. [CNBC, Bloomberg]
Total raised: ~$170B+ cumulative across all rounds (incl. SoftBank tranches).
ARR: ~$24B run-rate (April 2026, from “$2B/month” comment); enterprise now >40% of revenue, on track for parity with consumer by year-end. Ads pilot reached $100M ARR in 6 weeks.
Headcount: ~5,500–6,000 (estimated).
Notable customers: ChatGPT — 900M weekly active users, 50M+ paid subs; enterprise — Apple (Apple Intelligence), Salesforce, Morgan Stanley, government deployments.
Moat: Brand (ChatGPT is now a household verb), distribution scale, the most expensive compute footprint on Earth (Stargate program), and Microsoft-channel pull.
Key risk: Margin structure — even at $24B ARR, the company is reportedly cash-flow negative at scale; moat against Anthropic on enterprise narrowing rapidly. Governance overhang from non-profit structure transition.

Google DeepMind

Description: Google’s consolidated AI research and product unit. Maker of Gemini (3.1 Pro shipped Feb 2026). Not directly investable; relevant as anchor and Alphabet (GOOGL) public-equity exposure.
Funding: Internal — Google Cloud + Search budgets; capex $75B+ in FY2026.
Revenue: Google Cloud GenAI run rate growing >400% YoY; Gemini subs ~$1.2B in 2025; Gemini reaches 750M MAUs (mostly via AI Overviews). 70% of Google Cloud customers using Gemini; 120,000+ enterprises.
Moat: TPU vertical integration (only frontier lab with proprietary silicon at scale), Search/Workspace/YouTube distribution, deepest research bench.
Key risk: Antitrust — DOJ Search remedies could disrupt the Gemini-in-Search distribution channel that drives most of the user reach.

xAI

Description: Elon Musk’s frontier lab. Maker of Grok. Trained on the Memphis “Colossus” cluster (largest known GPU cluster). Merged with X Corp in Feb 2026.
Latest round: Series E, Jan 2026 — $20B (upsized from $15B target) at $230B valuation. Investors: Valor, Stepstone, Fidelity, QIA, MGX, Baron, NVIDIA, Cisco. Post SpaceX merger Feb 2026, combined entity reportedly ~$1.25T. [TechCrunch, Crunchbase News]
Total raised: ~$32B+ across rounds.
ARR: xAI standalone ~$500M ARR (modest vs valuation). With X consolidated: ~$3.3B annualized incl. ad/sub revenue per Sacra.
Headcount: ~1,500–2,000.
Notable customers: Tesla (in-vehicle Grok), X premium subscribers, government/Saudi data-center deal partners.
Moat: Capital access via Musk halo, raw compute scale, X as a real-time data spigot, willingness to pursue uncensored use cases that Anthropic/OpenAI won’t.
Key risk: Founder concentration risk; valuation untethered from revenue (~460x ARR standalone); brand risk from Grok controversies.

Meta (open-weights anchor)

Description: Public (META). Llama 4 family (Scout, Maverick, Behemoth) released April 2025 — first open-weight natively multimodal MoE models. Strategic pivot signaled in April 2026 with closed-source “Muse Spark” from Meta Superintelligence Labs.
Capex: ~$60–65B AI infrastructure spend FY2026.
Strategic role: Sets the open-weights price floor — every paid API has to beat “Llama for free” on TCO. Meta benefits when AI inference is commoditized because it depresses competitor pricing.
Risk: Muse Spark transition may erode the open-weights ecosystem effect that benefited Together, Fireworks, Groq.

Mistral AI

Description: French frontier lab, primarily relevant to US investors as the European sovereign-AI hedge. Open-weights and proprietary models.
Latest round: Series C, Sep 2025 — €1.7B at €11.7B (~$14B) valuation, led by ASML (€1.3B). Plus ~$830M debt for data centers in Mar 2026. [Reuters, Mistral]
Total raised: ~$3B+ equity, ~$830M debt.
ARR: ~$400M ARR Jan 2026, up from ~$16M end-2024 (~20x growth). Targeting >$1B ARR by end of 2026. [Sacra, MLQ]
Headcount: ~400–500.
Notable customers: BNP Paribas, Stellantis, Orange, French government.
Moat: EU sovereign AI positioning + ASML strategic anchor; strong open-weights brand among developers.
Key risk: Compute/capital disadvantage versus US frontier labs by ~10–20x; ARR/valuation multiple (~35x) is rich.

Foundation-layer summary (as of May 2026)

Company	Latest Valuation	Latest Round	ARR (run-rate)	Notes
OpenAI	$852B	$122B (Mar 2026)	~$24B	Consumer leader
Anthropic	$380B (round closed); $900B in talks	$30B (Feb 2026)	~$30B	Now leads on revenue
xAI	$230B (standalone); $1.25T (incl. X)	$20B Series E (Jan 2026)	~$500M standalone	Compute-heavy bet
Google DeepMind	n/a (Alphabet)	n/a	n/a (embedded)	Only TPU-vertical lab
Meta	n/a (Meta)	n/a	n/a (embedded)	Open-weights anchor
Mistral	~$14B	€1.7B (Sep 2025)	~$400M	EU hedge

B. Inference / Runtime Infrastructure

This is where the trillion tokens per day get served. Two business models: (1) token-billing inference clouds (Together, Fireworks, Groq) that serve open-weight models at marginal cost compression; (2) GPU compute clouds (CoreWeave, Lambda, Cerebras) that rent capacity and increasingly serve as the physical layer beneath the labs themselves.

Together AI

Description: Inference and fine-tuning cloud for open-source models; provides both per-token API and full GPU rentals. Strong open-source brand.
Latest round: Series B, Feb 2025 — $305M at $3.3B post-money, led by General Catalyst, co-led by Prosperity7. Reportedly raising new round at ~$7.5B valuation (per multiple reports through Q1 2026).
Total raised: ~$534M.
ARR: ~$300M annualized, Sep 2025 (Sacra estimate); up from ~$130M end-2024.
Headcount: ~250+.
Notable customers: Salesforce, Zoom, Washington Post, Pika, Stripe, Cartesia.
Moat: Best brand among open-source AI developers; deep fine-tuning tooling; vertically integrated with research (custom optimizations on inference kernels).
Key risk: GPU rental margins compress as hyperscalers undercut; Llama-as-flagship era ending if Meta pivots to closed models.

Fireworks AI

Description: Compound AI inference platform. Specializes in fast-serving open models and custom-fine-tuned variants for enterprises. Direct Together AI competitor.
Latest round: Series C, Oct 2025 — $250M at $4B post-money, led by Lightspeed, Index, Evantic; existing Sequoia, Benchmark, NVIDIA, AMD, MongoDB, Databricks.
Total raised: ~$327M.
ARR: ~$315M annualized Feb 2026 (Sacra estimate), up >400% YoY.
Headcount: ~150–200.
Notable customers: 10,000+ companies served; DoorDash, Cresta, Quora (Poe), Upwork.
Moat: Speed/latency leadership in serving open models; FireOptimizer and FireFunction proprietary fine-tuning infra; strategic NVIDIA/AMD/Databricks alignment.
Key risk: Same compute-cost dynamics as Together; differentiation is technical and could be matched by hyperscalers.

Groq

Description: Custom LPU (Language Processing Unit) chip vendor + GroqCloud inference service. Aggressive pricing on Llama and other open models.
Latest round: Sep 2025 — $750M at $6.9B, led by Disruptive, with BlackRock, Neuberger Berman, DTCP. Plus $1.5B Saudi commitment Feb 2025. Dec 2025: NVIDIA non-exclusive licensing agreement for Groq’s inference tech, structured as $17B in cash payments through end of 2026. [Groq newsroom, IntuitionLabs]
Total raised: >$2B equity + Saudi infra commitment.
ARR: $90M revenue 2024; projected ~$500M for 2025 (per Sacra).
Headcount: ~500.
Notable customers: Saudi infra deployment; Aramco; many indie devs at the long tail.
Moat: Genuinely differentiated silicon (LPU is a real architectural bet, not a GPU clone). NVIDIA licensing deal validates the IP economically.
Key risk: Hardware-startup risk profile (capital-intensive, long sales cycle to enterprise); NVIDIA deal could be either validation or precursor to acquisition (rumors of $20B acquisition floated in Q1 2026).

Cerebras Systems

Description: Wafer-scale AI accelerator (CS-3) + Cerebras Inference cloud. Filed S-1 April 2026 — IPO imminent.
Latest round: Series H, Feb 2026 — $1B at $23B, led by Tiger Global. IPO targeting up to $4B raise at potential ~$40B valuation per Bloomberg (May 2026).
Total raised: >$4B private.
Revenue: Q4 2025 revenue $510M (+76% YoY); $87.9M net income for the period.
Headcount: ~700.
Notable customers: OpenAI — $20B+ deal for 750MW of compute through 2028 (announced Jan 2026); Mayo Clinic, GlaxoSmithKline, G42.
Moat: Only public/near-public credible non-NVIDIA AI training silicon; the OpenAI deal is the singular trust signal.
Key risk: Customer concentration (G42/OpenAI represent disproportionate share); IPO timing exposes them to public-market AI sentiment volatility.

Description: Serverless GPU runtime. “Modal turns any Python function into an autoscaling cloud workload.”
Latest round: Series B, Sep 2025 — $87M at $1.1B, led by Lux Capital. In talks for new round at ~$2.5B led by General Catalyst (TechCrunch, Feb 2026).
Total raised: ~$111M.
ARR: ~$50M annualized (per recent coverage).
Headcount: ~50–80.
Notable customers: Suno, Substack, Ramp, Ramp, Cursor.
Moat: Developer experience — universally praised as best-in-class for ML workloads via Python; sticky once integrated.
Key risk: Revenue still small relative to valuation (~50x ARR); needs to expand beyond model serving into broader agent compute to justify multiples.

Replicate

Description: Serverless inference for open models, especially image/video. Acquired by Cloudflare on Dec 1, 2025.
Pre-acquisition raised: ~$57.8M (a16z, Sequoia, YC).
Strategic position post-acquisition: Now Cloudflare’s media-AI inference layer; tucked into Workers AI strategy. Removes them as an independent investment thesis but validates the segment.
Read-through: Independent inference clouds at $50–100M ARR are acquisition targets for hyperscalers/CDN players — relevant comparable for Modal’s exit math.

Lambda Labs

Description: GPU cloud (the “AI cloud”) + on-prem GPU clusters. Direct CoreWeave competitor. NVIDIA-backed.
Latest round: Series E, Nov 2025 — $1.5B+ at $5.9B, led by TWG Global and US Innovative Technology Fund. IPO targeted H2 2026; pre-IPO talks with Mubadala for $350M at ~20% IPO discount.
Total raised: ~$2.36B equity.
Revenue: $520M 2025; ~50% gross margin (61% ex non-cloud); first-half net loss ~$24M, TTM loss ~$175M.
Headcount: ~500+.
Notable customers: NVIDIA (strategic), various frontier-lab-adjacent and enterprise AI workloads.
Moat: Long-standing NVIDIA relationship (preferred GPU allocation); private/owned data center capacity.
Key risk: GPU-rental commoditization; CoreWeave’s public-market lead and capital access; debt-financed buildout.

CoreWeave (public — CRWV)

Description: Hyperscaler AI cloud. IPO March 2025 at $40/share; 2026 revenue guidance $12–13B; 200%+ stock appreciation since IPO.
Backlog: $66.8B as of Q4 2025 (long-term contracts, dominantly Microsoft/OpenAI/Meta).
ARR target: $17–19B by 2026.
Notable customers: Microsoft (largest), OpenAI, Meta, NVIDIA (as both supplier and customer).
Moat: Only public-market pure-play on AI infra ex-NVIDIA; speed-to-deploy GPU capacity; locked-in long-term contracts.
Key risk: Customer concentration extreme (Microsoft alone is ~62% of revenue per S-1 era data); debt load is enormous; hyperscalers rebuilding their own capacity could compress backlog renewal.

Inference / runtime summary

Company	Latest Valuation	Latest Round (date)	ARR (estimate)	Funding Multiple
Together AI	$3.3B (~$7.5B in talks)	$305M Series B (Feb 2025)	~$300M	~11x
Fireworks AI	$4B	$250M Series C (Oct 2025)	~$315M	~13x
Groq	$6.9B	$750M (Sep 2025)	~$500M (proj.)	~14x
Cerebras	$23B (IPO at $30–40B target)	$1B Series H (Feb 2026)	~$500M Q4 only	High
Modal Labs	$1.1B (~$2.5B in talks)	$87M Series B (Sep 2025)	~$50M	~22x
Replicate	n/a (acquired by Cloudflare Dec 2025)	n/a	n/a	n/a
Lambda Labs	$5.9B	$1.5B+ Series E (Nov 2025)	$520M (FY25)	~11x
CoreWeave	~$80B+ (public CRWV)	IPO Mar 2025	$12–13B (FY26 guide)	~6–7x fwd

C. Agent Runtime / Orchestration Frameworks

This is a smaller-dollar layer relative to the giants above, but it’s the layer most agentic-AI companies sit on top of. Capital here is concentrated in a few “agent OS” candidates. Bear case: orchestration is being absorbed into the foundation-model APIs themselves (OpenAI Agents SDK, Anthropic Skills, Vercel AI SDK). Bull case: enterprise needs vendor-neutral orchestration, observability, and governance — a real, durable category.

LangChain (incl. LangGraph, LangSmith)

Description: Most-adopted open-source agent framework + LangSmith eval/observability + LangGraph orchestration runtime.
Latest round: Series B, Oct 2025 — $125M at $1.25B, led by IVP, with Sequoia, Benchmark, Amplify, CapitalG, Sapphire. [Fortune, TechCrunch]
Total raised: ~$160M.
ARR: $12–16M as of June 2025 per TechCrunch; “low for where we are today” per company spokesperson — implying mid-$20M+ by late 2025.
Headcount: ~80–100.
Notable customers: Klarna, Elastic, Replit, Adobe, Rakuten, BlackRock (LangSmith).
Moat: Mindshare and developer adoption (LangChain is the default Python import for agent builders); LangSmith is the leading vendor-neutral agent observability tool; LangGraph 1.0 GA stabilized the runtime.
Key risk: Framework risk — Python-side competitive pressure from Pydantic AI, LlamaIndex, and direct SDKs; LangSmith faces commoditization as model labs ship native eval tooling.

CrewAI

Description: Multi-agent orchestration platform (role-based “crews”). Founder-led brand around agent collaboration patterns.
Latest round: Series A, Oct 2024 — $18M led by Insight Partners; total $24.5M (per Latka, Crunchbase).
Latest valuation: $76M (Sep 2024 per PremierAlts) — flagged as stale by 12+ months; likely repriced upward given growth.
ARR: $3.2M July 2025 (per Latka).
Headcount: ~30.
Notable customers: 150+ enterprise customers within 6 months of launch; PwC, IBM (per CrewAI marketing).
Moat: Strong open-source pull (60K+ GitHub stars); cleaner abstractions than LangChain for multi-agent use cases.
Key risk: Too-narrow focus on “crew” metaphor; revenue tiny vs. peers; LangChain catching up on multi-agent through LangGraph.

Letta (formerly MemGPT)

Description: Stateful agent platform — agents with persistent memory across sessions. UC Berkeley spinout (Sky Computing Lab).
Latest round: Seed, Sep 2024 — $10M at $70M, led by Felicis with Sunflower, Essence VC. Notable angels: Jeff Dean (Google DeepMind), Clem Delangue (HuggingFace), Cristobal Valenzuela (Runway). No announced Series A as of May 2026 — likely raising or recently raised quietly.
Total raised: ~$10M+ (publicly known).
ARR: Not publicly disclosed; likely sub-$5M.
Headcount: ~20–30.
Notable customers: Letta Code is #1 model-agnostic open-source agent on Terminal-Bench (Dec 2025).
Moat: Memory-first architecture — uniquely positioned as the “memory tier” of the agent stack. Berkeley research pedigree.
Key risk: Memory could be absorbed by foundation labs (Anthropic Skills, OpenAI Memory features); category is unproven economically.

Vellum

Description: LLMOps platform — workflow design, eval, deployment, monitoring, model routing.
Latest round: Series A, July 2025 — $20M, led by Leaders Fund, with Socii, YC, Rebel Fund, Eastlink.
Total raised: ~$24.5M.
Valuation: Not publicly disclosed.
ARR: Not publicly disclosed.
Headcount: ~40–60.
Moat: Stronger enterprise UX than LangSmith; strong with non-engineering builders (PMs, RevOps).
Key risk: Crowded eval/observability market; Humanloop precedent shows the category can collapse to acqui-hires.

Humanloop

Description: Was an LLMOps/eval platform. Acqui-hired by Anthropic Aug 2025 (founders + ~12 engineers); platform shutdown Sep 8, 2025. Anthropic explicitly did NOT acquire IP.
Read-through: Frontier labs are now buying eval/observability talent rather than letting an independent layer mature. This is the most important data point for evaluating Vellum and similar bets.

Orchestration / framework summary

Company	Valuation	Latest Round	ARR	Risk Posture
LangChain	$1.25B	$125M Series B (Oct 2025)	~$15–25M	Category leader; framework risk
CrewAI	~$76M (stale)	$18M Series A (Oct 2024)	~$3M	Too small to last alone
Letta	$70M (Sep 2024 — likely repriced)	$10M Seed (Sep 2024)	n/a	Memory-pure; risk of being absorbed
Vellum	n/a	$20M Series A (Jul 2025)	n/a	Crowded segment
Humanloop	n/a — acqui-hired by Anthropic	n/a	n/a	Cautionary tale

D. Agent-Specific Infra (Browser, Sandbox, Tools)

The “hands and feet” layer for agents — when an agent needs to browse, code, or actuate. Real economics emerging in browser-as-a-service and code-execution sandboxes; these are the picks-and-shovels under every Operator-class product.

Browserbase

Description: Headless browser infrastructure for agents — “Stripe for browser automation.” Maker of Stagehand (open-source LLM-native browser SDK).
Latest round: Series B, June 2025 — $40M at $300M, led by Notable Capital, with CRV, Kleiner Perkins, and angels (Patrick Collison, Jeff Lawson, Guillermo Rauch).
Total raised: ~$67.5M.
ARR: ~$4.4M (June 2025, per Latka).
Headcount: ~40.
Notable customers: Perplexity, OpenAI (Operator), Anthropic (Computer Use partners), various YC agent startups.
Moat: First-mover; tight integration with the agent labs themselves; Stagehand becoming the de facto open-source browser-agent SDK.
Key risk: Anthropic Computer Use and OpenAI Operator could internalize the layer; price compression as competitors enter.

E2B

Description: Secure code-execution sandboxes for AI agents. Czech-founded; the default “code interpreter as a service.”
Latest round: Series A, July 2025 — $21M, led by Insight Partners, with Decibel, Sunflower, Kaya VC.
Total raised: ~$35M.
ARR: ~$1.5M (June 2025, per Latka).
Headcount: ~20.
Notable customers: 88% of Fortune 100 signed up; >50% of Fortune 500. Hundreds of millions of cloud sandboxes initiated.
Moat: Low-latency sandbox primitive; open-source pull; design-partner relationships with labs.
Key risk: Tiny revenue vs. signups (free-tier dominance); Anthropic/OpenAI shipping native code execution erodes the standalone value.

Anthropic Computer Use (capability, not company)

Description: Anthropic’s API-level capability allowing Claude to control a virtual desktop. Released Oct 2024; matured through Claude Sonnet 4.5/Opus 4 in 2025–2026.
Read-through for investors: This is the existential question for Browserbase, E2B, Skyvern, Steel — when the foundation lab itself ships the capability natively, the third-party layer must compete on enterprise plumbing (governance, sandboxing, compliance), not on raw capability.

Skyvern, Hyperbrowser, Steel.dev (browser agents — earlier-stage)

Skyvern: Open-source browser-agent layer. $2.7M seed, Dec 2025. Revenue ~$900K (2024). Tiny but growing.
Hyperbrowser: YC W25/S25-era; backed by YC, Accel, SV Angel; funding amount publicly undisclosed. Web-infra-for-agents positioning.
Steel.dev: YC W24. ~$17M raised (Series A reported). Open-source browser API for agents. Adopted as native browser provider by Hermes and other agent frameworks.
Read: Three-way race for open-source mind share in browser-agent infra. None have material ARR yet. All face the same Anthropic Computer Use squeeze.

Agent-specific infra summary

Company	Valuation	Latest Round	ARR (est.)	Notes
Browserbase	$300M	$40M Series B (Jun 2025)	~$4.4M	Category leader
E2B	Not disclosed	$21M Series A (Jul 2025)	~$1.5M	Fortune 100 distribution
Skyvern	n/a	$2.7M Seed (Dec 2025)	<$1M	Earlier-stage
Steel.dev	n/a	~$17M total	n/a	YC-pedigreed
Hyperbrowser	n/a	YC + Accel; undisclosed	n/a	Earliest-stage

E. Memory / Vector / Data Layer

The hardest segment to bet on right now. The original 2022–2023 “vector DB” thesis has been pressure-tested: Pinecone is for sale, Notion left them, and turbopuffer demonstrated a different (much simpler) architecture works better at the modern AI scale. The new question: is “memory” the correct abstraction (Mem0, Letta), or is it just a vector index inside a normal database (pgvector, Supabase, MongoDB, Snowflake)?

Pinecone

Description: First-mover serverless vector database. Powered the early RAG era.
Latest priced round: Series B, April 2023 — $100M at $750M valuation (a16z-led). No publicly disclosed priced round since — flagged as ~3 years stale.
Total raised: ~$138M.
ARR: ~$26.6M (FY2024). Reported declining to ~$14M in 2025 (per Latka/coverage), reflecting the Notion churn.
Headcount: ~150–200.
Strategic position: Reportedly exploring sale at $2B+ price (The Information, Aug 2025). Potential acquirers: Oracle, MongoDB, Snowflake, IBM.
Moat eroding: As Notion left, the brand premium evaporated. pgvector and turbopuffer ate the developer mindshare.
Key risk: Sale process is the main near-term catalyst; standalone path is uncertain.

Weaviate

Description: Open-source vector DB; Dutch-founded. Strong enterprise distribution with hybrid (vector + lexical) search.
Latest round: Series C, Oct 2025 — $50M at $200M, led by Battery Ventures and Zetta. (Note: this is a flat-to-down round vs. earlier-rumored higher marks — repricing.)
Total raised: ~$110M+.
ARR: ~$12.3M (Oct 2024 per Latka); likely $20–25M by mid-2026 estimate.
Headcount: ~100.
Notable customers: Stack Overflow, Cisco.
Moat: Open-source enterprise install base; hybrid search.
Key risk: Same secular pressure as Pinecone; reset valuation reflects market repricing of pure-play vector DBs.

Chroma

Description: Developer-first open-source vector store. Most popular in YC startup land.
Latest priced round: Seed, April 2023 — $18M (Quiet Capital). Subsequent unattributed VC round Feb 2024. No major Series A announcement publicly.
Total raised: ~$20.55M.
Valuation: Last reported ~$75M (CB Insights estimate — likely stale).
Headcount: ~100 (per CB Insights as of Jan 2026).
Strategic position: Switched messaging in 2025 from “vector DB” to “search infrastructure for AI” — repositioning amid sector pressure.
Key risk: Has not raised priced funding in 2 years; capital position uncertain.

Turbopuffer

Description: Object-storage-native vector + text search. Ex-Shopify engineers. Profitable.
Latest round: Undisclosed-amount round, Dec 2025, from existing backer Lachy Groom + new investor Thrive Capital. Pre-seed earlier from Lachy Groom + Thrive.
Total raised: Not publicly disclosed in aggregate.
ARR: Company says “many 10s and 10s of millions” — implying $30–50M+ — and profitable.
Notable customers: Anthropic, Cursor, Legora, Notion (post-Pinecone-departure rumored).
Moat: Architectural — built on object storage so cost-per-token-stored is dramatically lower than Pinecone-class systems; that economic gap is why frontier labs use it.
Key risk: Single-architecture bet; if vector search becomes a feature inside data warehouses, even good architecture won’t save standalone economics.

Mem0

Description: Memory layer for AI apps and agents — open-source SDK + cloud service. Y Combinator alumnus.
Latest round: Series A, Oct 2025 — $24M at $150M (combined Seed + Series A figure), led by Basis Set Ventures with Peak XV, GitHub Fund, YC. Notable angels: Olivier Pomel (Datadog), Dharmesh Shah, Lukas Biewald.
Total raised: ~$24M.
ARR: Not publicly disclosed.
Traction: Open-source API has 41K+ GitHub stars, 13M+ Python downloads. Q1 2025 processed 35M API calls.
Moat: Memory-as-a-product positioning; open-source distribution.
Key risk: Same as Letta — memory may be a feature, not a product. Both companies are betting that an opinionated memory layer is durable enough to hold ground against Anthropic Skills + OpenAI Memory features.

Memory / vector / data layer summary

Company	Valuation	Latest Round	ARR (est.)	Status
Pinecone	$750M (2023, stale)	$100M Series B (Apr 2023)	~$14M (declining)	Exploring sale at $2B+
Weaviate	~$200M	$50M Series C (Oct 2025)	~$20M	Resetting
Chroma	~$75M (stale)	$18M Seed (Apr 2023) + 2024 follow-on	n/a	Repositioning
Turbopuffer	Not disclosed	Undisclosed (Dec 2025)	$30–50M+ profitable	Architectural winner
Mem0	$150M	$24M Series A (Oct 2025)	n/a	Memory bet
Letta	~$70M (stale)	$10M Seed (Sep 2024)	n/a	Memory bet

F. MCP Ecosystem Players

Model Context Protocol is Anthropic’s open protocol (released late 2024) for letting agents connect to tools, data, and services. By mid-2026 it’s becoming the de facto integration standard, adopted by OpenAI, Google, and most major frameworks. Three categories of builders are forming around it:

Hosting / runtime: companies that host MCP servers as a service.
Registry / discovery: marketplaces and curated lists.
Integration platforms: Zapier-style integration suites that expose hundreds of pre-built MCP tools.

Composio

Description: Integration platform — pre-built tools and skills for AI agents across 250+ apps. The leading MCP integration platform.
Latest round: Series A, July 2025 — $25M, led by Lightspeed, with SV Angel, Blitzscaling, Operator Partners; existing Elevation, Together Fund. Total ~$29M.
Total raised: ~$29M.
Valuation: Not publicly disclosed.
Traction: “MCP infrastructure platform powering 100,000 developers.”
Moat: Production-grade integrations + skills-that-evolve thesis (agents that learn over time from user interactions).
Key risk: Zapier and Workato moving into the same integration-for-agents wedge with much larger distribution.

Smithery

Description: MCP server hosting and registry. >2,800 ready-to-use MCP tools and agents.
Latest round: Seed, 2025 — undisclosed amount, from South Park Commons.
Strategic position: Discovery layer for the MCP ecosystem — “npm for MCP.” Likely will need a follow-on round in 2026.

Glama

Description: Hosts and runs MCP servers (vs. Smithery, which is more of a registry). Developer-platform thesis.
Funding: Not publicly disclosed at meaningful scale.

Read-through on MCP ecosystem

The MCP layer is too early to invest at growth-equity size. Composio is the only company in this segment with credible Series A scale. The bull case: MCP becomes the universal agent-tool protocol, the way HTTP became the universal data protocol, and a Stripe-of-MCP emerges with tens-of-billions outcome potential. The bear case: foundation labs ship native tool-use ecosystems (OpenAI Apps SDK, Anthropic Skills) that absorb the integration layer entirely.

MCP ecosystem summary

Company	Valuation	Latest Round	Notes
Composio	n/a	$25M Series A (Jul 2025)	Category leader
Smithery	n/a	Undisclosed seed (2025)	Registry
Glama	n/a	Not disclosed	Hosting

Cross-Layer Investor Read

Where capital is concentrating (last 12 months): Foundation models ($172B raised in OpenAI + Anthropic + xAI alone), then GPU-compute infrastructure (CoreWeave $66B backlog, Lambda $1.5B Series E, Cerebras $1B + IPO).

Where multiples are richest: Agent infrastructure (Modal at ~22x, Browserbase at ~70x, Mem0 at very rich multiples on undisclosed ARR).

Where multiples are compressing: Pure-play vector DBs (Weaviate flat-to-down, Pinecone exploring sale, Chroma quiet).

Where capital is exiting via M&A: Eval/observability (Humanloop → Anthropic acqui-hire), serverless inference (Replicate → Cloudflare).

Three risks to underwrite for any layer above the foundation models:

Foundation-lab vertical integration — labs are absorbing eval, memory, tools, browsing, and code execution natively. Every layer must answer: “what do you do that Claude/GPT-5/Gemini won’t ship in 6 months?”
Hyperscaler absorption — AWS Bedrock, Azure AI, GCP Vertex are bundling the infrastructure layers. CoreWeave is the only standalone winner so far.
Open-weights regime change — Meta’s reported pivot to closed-source Muse Spark is a real signal. If open weights stagnate, the Together/Fireworks/Groq economic case weakens.

Two upside scenarios:

Agent-economy take-off — if 2026–2027 sees true autonomous-agent workflows in enterprise, every “hands and feet” layer (Browserbase, E2B, Composio) sees order-of-magnitude revenue ramps.
Memory / state becomes the moat — if customer state (Letta, Mem0, turbopuffer) becomes the way agents differentiate (not the model), the data layer reprices upward dramatically.

Sources

Part 2B — Vertical Agent Application Companies

Prepared for: Family office principals Date: 2026-05-04 Scope: US-centric, application-layer vertical AI agent companies. Private unless noted. Numbers are reported figures or “estimated” with source; uncertain items flagged. Multiples computed off most recent post-money / most recent disclosed ARR.

Headline read: The application layer is bifurcating fast. At the top, a small set of category leaders (Cursor, Harvey, Sierra, Glean, Abridge, Decagon, ElevenLabs, Ramp) have crossed $100M+ ARR with credible enterprise distribution and are commanding 30–80x ARR multiples. Below them, a long tail of well-funded but sub-$50M ARR companies are vulnerable to (a) model commoditization, (b) Microsoft/Google bundling, and (c) outcome-based pricing pressure. Coding agents are most contested (foundation models compete directly); legal, healthcare, and finance show the strongest moats (regulated buyers, workflow lock-in, proprietary data). Voice infra (ElevenLabs, Deepgram) is the rare layer with both growth and margin durability. Treat any company with <$10M ARR and >$1B valuation as venture-priced optionality, not a business — that includes Magic, Poolside, Rox, and several “AI BDR” names.

A. Coding / Developer Agents

The hottest, most overfunded vertical. Foundation labs (OpenAI/Anthropic/Google) are direct competitors via their own products and via licensing deals, which is why we are seeing rapid distribution-driven consolidation (OpenAI–Windsurf collapse, Google–Windsurf reverse acquihire, Cognition–Windsurf, xAI–Anysphere option).

Cursor (Anysphere)

Description: AI-native code editor (forked from VS Code) for individual developers and engineering teams. Sells via self-serve seats and enterprise plans.
Founders / year: Michael Truell (CEO), Sualeh Asif, Arvid Lunnemark, Aman Sanger. Founded 2022, MIT.
Latest round: Series D, Nov 2025 — $2.3B at $29.3B post-money (Accel, A16Z, Thrive). In April 2026, reportedly in talks for ~$2B at $50B pre-money; xAI separately announced an option to acquire Anysphere for $60B. (CNBC, The Next Web)
Total raised: ~$3B+ across rounds.
ARR: $100M (Jan 2025) → $500M (Jun 2025) → $1B (Nov 2025) → $2B (Feb 2026). Forecasting $6B+ exit-2026. (TechCrunch)
Customers: Disclosed users at Stripe, Shopify, OpenAI, Instacart, Mercado Libre. 85%+ Fortune 500 penetration claimed.
Pricing: $20/seat/mo Pro; Business $40/seat; Enterprise custom. Heavily usage-metered on premium models.
Moat: Brand and developer mindshare in IDE category; rapid product cadence; OpenAI/Anthropic preferred-partner distribution. Network effects weak — switching cost is moderate (it’s an editor).
Key risk: Direct competition from Anthropic Claude Code, GitHub Copilot, Google. Margin compression as foundation labs raise inference prices. The $50B/$60B-implied multiple (~25–30x forward) requires sustaining 100%+ growth.

Cognition (Devin)

Description: Autonomous AI software engineer. Pitched as agent that completes tickets end-to-end vs. assistant model. Acquired Windsurf in mid-2025.
Founders / year: Scott Wu (CEO), Steven Hao, Walden Yan. Founded 2023.
Latest round: Sept 2025 — $400M at $10.2B post-money (Founders Fund). Talks reported April 2026 for “hundreds of millions” at $25B. (CNBC, Bloomberg)
Total raised: ~$600M+.
ARR: Devin grew $1M (Sep 2024) → $73M (Jun 2025). Plus Windsurf ~$82M ARR at acquisition. Combined enterprise ARR up 30%+ in seven weeks post-merger; unofficial estimates place combined ARR ~$200M+ by early 2026. (AgentMarketCap)
Customers: Goldman Sachs, Citi, Dell, Cisco, Ramp, Palantir, Nubank, Mercado Libre.
Pricing: Devin starts at $500/mo for individual; enterprise seat + ACU (agent compute unit) usage. Windsurf retains $15–60/seat tiers.
Moat: Earliest “autonomous coder” mover; strong financial-services beachhead via Windsurf; combined IDE + agent product post-merger.
Key risk: Devin demos vs production reliability gap remains a known critique. $25B at ~$200M ARR = 125x — priced for perfection.

Codeium / Windsurf

Status: No longer a standalone investable entity. May 2025: OpenAI agreed to acquire for $3B; deal collapsed July 2025 over Microsoft IP rights. Google then paid ~$2.4B reverse-acquihire for CEO Varun Mohan, co-founder Douglas Chen, and ~40 staff to DeepMind. Remaining business sold to Cognition for ~$250M in Dec 2025. (Fortune, DeepLearning.AI)
Last standalone ARR: $82M (Jul 2025, up from $40M Feb 2025).
Takeaway for investors: Cautionary tale on big-tech distribution risk. Owners of any code-agent company must consider Microsoft IP entanglement before negotiating exits.

Replit

Description: Browser-based dev environment that has pivoted to AI app-building agent (“Replit Agent”). Sells to non-developers + prosumers + enterprise dev teams.
Founders / year: Amjad Masad (CEO), Faris Masad, Haya Odeh. Founded 2016.
Latest round: Mar 2026 — $400M at $9B (a16z, Coatue). Prior: $250M Sep 2025 at $3B. (TechCrunch)
Total raised: ~$900M+.
ARR: ~$3M start of 2025 → ~$150M Sep 2025. Targeting $1B run-rate by end of 2026. (Sacra)
Customers: 50M+ users; 85% of Fortune 500 have users on platform; enterprise via Anthropic, HubSpot, others.
Pricing: $25/mo Core; usage-metered Agent (“checkpoint” billing); Enterprise custom.
Moat: “Build apps from a prompt” wedge for non-coders; runtime + deployment integration is rare. Brand among student / early-career builders.
Key risk: Bolt.new and Lovable (EU) compete directly on prosumer “build-an-app” use case. Margin pressure as Agent runs are LLM-heavy.

Magic.dev

Description: Long-context coding model + autonomous engineer. Stealthier than peers; emphasizes ultra-long context (LTM-2-mini: 100M tokens).
Founders / year: Eric Steinberger, Sebastian De Ro. Founded 2022.
Latest round: Aug 2024 — $320M from Eric Schmidt, CapitalG, Atlassian, Sequoia, Nat Friedman, Daniel Gross, Jane Street. Reported $1.5B valuation. (TechCrunch)
Total raised: ~$515M.
ARR: Not publicly disclosed; widely reported as minimal — pre-revenue or low-single-digit M.
Customers: None publicly disclosed.
Pricing: Not generally available.
Moat: Speculative IP bet on long-context architectures; dependent on whether 100M-token approach becomes commercially differentiated.
Key risk: Pure science-bet. With Cursor at $2B ARR and Magic at no ARR, the moat thesis (long-context) is being eroded by Gemini 2M+ context and Anthropic’s improvements. Highest concentration of “name-brand investor” risk in this list.

Poolside

Description: Foundation models for code + enterprise platform. France/US. Sells fine-tuned, on-prem deployments to large enterprises.
Founders / year: Jason Warner (ex-CTO GitHub) and Eiso Kant. Founded 2023.
Latest round: Oct 2025 — reportedly $2B at $12B with $1B from Nvidia. A subsequent ~$2B Series C at $14B reportedly collapsed in April 2026 over investor concerns about sub-$50M ARR. (TechFundingNews, Tech-Insider)
Total raised: ~$625M+.
ARR: $30M (Dec 2024) → $50M (Mar 2025); estimated still under $50M as of early 2026.
Customers: Large enterprise pilots; specifics not publicly disclosed.
Pricing: Enterprise license + custom deployment.
Moat: Vertical model + on-prem deployment is a real differentiator for regulated buyers (defense, banks). Nvidia commitment is a meaningful supply-chain edge.
Key risk: Burn-vs-ARR ratio is the worst on this list. The collapsed Series C and reported “Project Horizon” Texas datacenter complications are red flags. Investor narrative-dependence is high.

Augment Code

Description: Enterprise AI coding assistant differentiated on “Context Engine” (large codebase indexing) and security/compliance posture (ISO 42001 certified).
Founders / year: Igor Ostrovsky, Guy Gur-Ari. Founded 2022.
Latest round: Series B — $227M at $977M post-money (Sutter Hill, Index, Lightspeed). (Augment Code)
Total raised: ~$252M.
ARR: ~$20M (Oct 2025). (Latka)
Customers: Disclosed: Lemonade, KeyBank, Webflow, others; positioned for regulated/large-codebase buyers.
Pricing: $50/seat/mo Developer; Enterprise custom.
Moat: Compliance + large-codebase context. Reasonable thesis for slower-moving enterprise buyers who can’t deploy Cursor security-wise.
Key risk: Cursor and GitHub Copilot Enterprise are encroaching on the same buyer; ~50x ARR multiple at last round is frothy for a ~$20M-ARR business.

Coding agents — summary table

Company	Latest round	Valuation	ARR (most recent)	Multiple
Cursor (Anysphere)	Nov 2025 Series D	$29.3B (talks at $50B–$60B)	$2B (Feb 2026)	~15x current / ~25–30x at $50B
Cognition	Sep 2025 + 2026 talks	$10.2B (talks at $25B)	~$200M+ combined est.	~50x / ~125x at $25B
Windsurf	Acquired by Cognition Dec 2025	$250M sale	$82M (Jul 2025)	n/a
Replit	Mar 2026	$9B	$150M (Sep 2025)	~60x
Magic.dev	Aug 2024	$1.5B	Pre-revenue / minimal	n/m
Poolside	Oct 2025 (post Apr 2026 collapse)	$12B last closed; $14B failed	<$50M est.	>250x — extreme
Augment	Series B 2024	$977M	~$20M (Oct 2025)	~50x

B. Customer Support / CX Agents

The clearest “agentic ROI” story — labor replacement on inbound customer interactions. Sierra and Decagon are setting outcome-based pricing as the category norm, which is structurally important: it shifts value capture toward the agent vendor and away from per-seat SaaS economics.

Sierra (Bret Taylor)

Description: AI agent OS for customer experience. Sold to consumer brands and enterprises; outcome-based pricing per resolved conversation.
Founders / year: Bret Taylor (ex co-CEO Salesforce, OpenAI Chair), Clay Bavor (ex VP Google Labs). Founded 2023.
Latest round: Reported May 4, 2026 — ~$1B raise (CNBC). Prior: Sept 2025 $350M Series B at $10B (Greenoaks, ICONIQ). (CNBC, TechCrunch)
Total raised: $635M+ (pre most recent round).
ARR: $100M (Nov 2025); $150M+ by early 2026 — fastest known eight-quarter ramp in enterprise software.
Customers: SiriusXM, ADT, Sonos, WeightWatchers, Casper, Discord, Brex.
Pricing: Outcome-based ($1–3 per resolution, plus blended per-conversation). Estimated $150K–$350K+ year-one ACVs; $50K–$200K setup. (Lorikeet)
Moat: Bret Taylor distribution (CEO calls into F500 buyers); design+platform polish; outcome-based contracts compound over time as agents improve.
Key risk: Salesforce Agentforce (incumbent threat) and ServiceNow are bundling-in equivalents. At ~$10B / $150M ARR (~65x), the price assumes Sierra escapes that bundling.

Decagon

Description: Concierge-style AI agents for customer support. Sells to high-volume consumer brands and SaaS.
Founders / year: Jesse Zhang (CEO, ex-Google/Citadel), Ashwin Sreenivas. Founded Aug 2023.
Latest round: Jan 2026 — $250M Series D at $4.5B (a16z, Bain, ICONIQ). (Bloomberg)
Total raised: ~$481M.
ARR: ~$10M (end 2024) → $35M (Oct 2025). 100+ corporate customers added in 2025. (Sacra)
Customers: Notion, Duolingo, Rippling, Bilt, Eventbrite, Substack, Avis, Deutsche Telekom.
Pricing: Per-conversation and per-resolution (higher rate when AI fully resolves without human handoff).
Moat: Tooling/integrations into customer workflow systems; design partner brand quality.
Key risk: Multiple at $4.5B / $35M ARR is ~130x — pricing in another 4–6x of growth. Overlap with Sierra is high; the two will likely compete head-to-head on every meaningful enterprise deal.

Cresta

Description: Real-time AI for contact-center agents (assist humans + virtual agents). More mature, more “co-pilot” than full agent.
Founders / year: Zayd Enam, Tim Shi, Sebastian Thrun. Founded 2017.
Latest round: Nov 2024 — $125M Series D (WiL, QIA, Sequoia, A16Z, Greylock, Tiger). $1.6B valuation. (Cresta)
Total raised: ~$276M.
ARR: ~$52M (Oct 2025), nearly 4x in last two years. (Latka)
Customers: Cox Communications, Brinks Home, Holiday Inn Club Vacations, Verizon.
Pricing: Per-seat for agent-assist; usage-based for virtual agent.
Moat: Existing CC distribution and integrations; real-time intelligence layer.
Key risk: Sandwiched between full-agent disruptors (Sierra/Decagon shrinking the human seat count Cresta sells against) and incumbents (NICE, Genesys, Five9 with their own AI). ~30x ARR is reasonable but compression risk is real.

Ada

Description: AI customer service platform; pioneer of automated chat for support, transitioning to “agentic” framing.
Founders / year: Mike Murchison (CEO), David Hariri. Founded 2016, Toronto.
Latest round: May 2021 — $130M Series C at $1.2B (Spark, Accel, Bessemer, Tiger). No new priced round since; minor 2025 grant. (RBCx)
Total raised: ~$200M.
ARR: ~$70.6M revenue 2024 (Latka). 2025 figure not publicly disclosed.
Customers: Square, Verizon, Meta, Indigo, AirAsia.
Pricing: Tiered platform license + usage. Generally per-conversation tiers.
Moat: Customer base of 5,000+ (older incumbent advantage); integrations breadth.
Key risk: “Pre-agentic” architecture; getting flanked by Sierra/Decagon on outcome-based deals. Down-round risk is real if next round ever prices.

Forethought

Description: Multi-agent AI for support resolution; positions on omnichannel + ROI delivery ($1B+ customer ROI claimed).
Founders / year: Deon Nicholas (CEO), Sami Ghoche, Mike Murchison (separately, then split). Founded 2018.
Latest round: May 2025 — $25M strategic (Blue Cloud Ventures + Writer/Cognition/Cartesia founders). (BusinessWire)
Total raised: ~$115M.
ARR: Not publicly disclosed. Estimated low double-digit M based on customer base + pricing.
Customers: Airtable, Grammarly, Cohere, Datadog, WordPress, Upwork, Roadie.
Pricing: Per-conversation tiers.
Moat: Engineering depth, multi-agent architecture, customer ROI track record.
Key risk: Limited brand vs Sierra/Decagon at the C-suite; smaller war chest. May struggle to stay in the top tier.

Parloa

Description: Voice-first agentic AI for customer service. German origin; strong EU enterprise base, aggressively expanding US.
Founders / year: Malte Kosub (CEO), Stefan Ostwald. Founded 2017, Berlin.
Latest round: Jan 2026 — $350M Series D at $3B (General Catalyst, EQT, Altimeter, Durable). (TechCrunch)
Total raised: ~$562M.
ARR: $50M+; 4x revenue growth in 2025. US revenues overtook Europe in 2025.
Customers: Allianz, Booking.com, SAP, Swiss Life, Decathlon.
Pricing: Enterprise contracts; per-minute usage typical.
Moat: Voice-first focus differentiates vs Sierra (text-first); strong EU regulated-industry credentials.
Key risk: ~60x ARR; competing with Sierra (now adding voice), Cognigy, Talkdesk. EU labor cost arbitrage helps margin but US scale is the bet.

Customer support — summary table

Company	Latest round	Valuation	ARR	Multiple
Sierra	May 2026 (~$1B reported) / Sep 2025 $350M	$10B (last priced)	$150M+	~65x
Decagon	Jan 2026 Series D	$4.5B	$35M	~130x
Cresta	Nov 2024 Series D	$1.6B	$52M	~30x
Ada	May 2021 Series C	$1.2B (stale)	~$70M (2024)	~17x (stale)
Forethought	May 2025 strategic	n/d	n/d	n/d
Parloa	Jan 2026 Series D	$3B	$50M+	~60x

C. Sales / GTM Agents

Highest hype-to-revenue ratio of any vertical. Most “AI BDR” companies are reporting churn-laden ARR and inflated logos. Clay is the clear category leader — it’s a real business with a real moat (workflow + data orchestration). Most of the rest should be treated skeptically.

Clay

Description: AI-augmented GTM data platform. “GTM engineer” tooling for enrichment, personalization, outbound automation. Sells PLG → enterprise.
Founders / year: Kareem Amin (CEO), Varun Anand. Founded 2017 (initial pivot), GTM PMF ~2022.
Latest round: Aug 2025 — $100M Series C at $3.1B (CapitalG, Sequoia, Meritech). (TechCrunch)
Total raised: ~$204M.
ARR: $1M → $100M in ~24 months; on track to triple again in 2026. (Clay)
Customers: 10,000+ — including OpenAI, Anthropic, Cursor, Canva, Intercom, Rippling.
Pricing: $149–$800/mo published tiers; enterprise custom; usage credits as core unit.
Moat: Workflow orchestration on top of dozens of data providers — switching cost is real (Clay is the system of record for outbound). Strong PLG flywheel + GTM-engineer cult brand.
Key risk: Enrichment vendors building competing AI; LinkedIn / Salesforce bundling. ~30x ARR is one of the more defensible multiples in this report.

11x.ai

Description: “AI digital workers” — Alice (BDR) and Julian (voice). Sales agent products.
Founders / year: Hasan Sukkar (CEO). Founded 2022, London/SF.
Latest round: Nov 2024 — $50M Series B at ~$350M (a16z, Benchmark). No priced round since. (TechCrunch)
Total raised: ~$76M.
ARR: Reported $10M (2024) → $25M (Q1 2025), but TechCrunch reporting documented inflated ARR via three-month break clauses, and logos used without permission (ZoomInfo threatened legal action). (TechCrunch)
Customers: Disputed.
Pricing: Per-agent monthly subscription (~$2K+/mo).
Moat: Brand among VC-funded outbound buyers; weak structurally.
Key risk: Reputational + governance. Family office reader: this is a name to avoid in secondary purchases until further clarity. Treat as cautionary.

Artisan

Description: “Ava” the AI BDR — finds, researches, and emails leads end-to-end.
Founders / year: Jaspar Carmichael-Jack (CEO). Founded 2023.
Latest round: Apr 2025 — $25M Series A (Glade Brook, YC, HubSpot Ventures). (Artisan)
Total raised: ~$36M.
ARR: Reported $5–6M+ early 2025. (ARR Club)
Customers: Remote, Quora, SumUp.
Pricing: $1.5K–$5K/mo platform fee + usage.
Moat: “Stop hiring humans” marketing was viral. Limited technical moat.
Key risk: Same risk profile as 11x — questionable retention on AI-BDR category broadly. Premature to judge outcome.

Regie.ai

Description: AI-powered sales prospecting platform; “RegieOne” combines content gen + dialer + agent prospecting.
Founders / year: Matt Millen (President), Srinath Sridhar (CEO). Founded 2020.
Latest round: Feb 2025 — $30M Series B (Scale VP, Foundation Capital). (PRNewswire)
Total raised: ~$50M.
ARR: Not publicly disclosed; growth reported “300% YoY”; estimated low-to-mid single-digit M (Extruct AI estimate ~$3.8M).
Customers: Crunchbase, Sophos, ZoomInfo (interesting given 11x situation).
Pricing: Tiered seats + content credits.
Moat: Established outbound content niche; weak vs Clay+ Apollo combo.
Key risk: Squeezed between Clay (workflow) and Apollo (data + workflow).

Rox

Description: “Agent swarm” platform for sales reps; AI-native CRM alternative pitching “every seller has an agent team.”
Founders / year: Ishan Mukherjee (CEO, ex CGO New Relic). Founded 2024.
Latest round: Closed late 2025 (announced Mar 2026) — $50M+ at $1.2B (Sequoia, GV, General Catalyst). (TechCrunch)
Total raised: ~$80M reported (50M Nov 2024 + 2025).
ARR: Projected $8M close-2025 (per GV partnering note + reporting).
Customers: Mercury, Ramp, Unify (early design partners).
Pricing: Per-seat plus usage.
Moat: Sequoia-backed founder narrative + early enterprise design partners. Real product is unproven at scale.
Key risk: ~150x ARR for a 1-year-old company against Salesforce Agentforce + HubSpot Breeze incumbents. Pure venture optionality.

Sales / GTM agents — summary table

Company	Latest round	Valuation	ARR	Multiple
Clay	Aug 2025 Series C	$3.1B	$100M+	~30x
11x.ai	Nov 2024 Series B	~$350M	~$25M (disputed)	~14x (caveats)
Artisan	Apr 2025 Series A	n/d	$5–6M	n/m
Regie.ai	Feb 2025 Series B	n/d	<$10M est.	n/m
Rox	Late 2025	$1.2B	~$8M projected	~150x

D. Legal Agents

Best-in-class moats in this report. Regulated buyers, low risk tolerance, structured workflows, proprietary playbooks. Harvey is the clear leader; EvenUp and Eve are real plaintiff-bar businesses with legitimate ARR growth.

Harvey

Description: AI agent platform for elite law firms and in-house legal teams. Drafting, research, due diligence, contract analysis, deposition prep.
Founders / year: Winston Weinberg (CEO, ex O’Melveny lawyer), Gabriel Pereyra (ex DeepMind). Founded 2022.
Latest round: Mar 2026 — $200M at $11B (GIC, Sequoia). Prior: Dec 2025 at $8B (a16z); Feb 2025 $300M Series D at $3B. (CNBC, TechCrunch)
Total raised: $1B+.
ARR: $100M (Aug 2025) → $190M (Jan 2026). (Harvey)
Customers: Majority of AmLaw 100; 500+ in-house teams; 50 asset managers; 60 countries.
Pricing: Enterprise seat licenses; reportedly $100–500/lawyer/mo + custom enterprise tiers; large firm contracts $1M+.
Moat: Brand among AmLaw partners; OpenAI partnership; 4-year head start with elite firm beachhead. Switching costs at firm-wide deployments are very high.
Key risk: Microsoft Copilot for Legal + Lexis/Thomson Reuters entering the same market. ~$11B / $190M (~58x) is full-priced but defensible if growth holds.

EvenUp

Description: AI for plaintiff-side personal injury firms. Demand letter generation, case valuation, medical records summarization.
Founders / year: Rami Karabibar (CEO), Saam Mashhad. Founded 2019.
Latest round: Oct 2025 — $150M Series E at $2B+ (Bessemer, REV/RELX, B Capital, SignalFire). (Fortune)
Total raised: ~$385M.
ARR: Not publicly disclosed precisely — “doubling YoY” disclosed; Sacra and other estimates place at $40–80M range late 2025.
Customers: 2,000+ firms, including 20% of Top 100 US PI firms; 10,000 cases/week processed.
Pricing: Per-case fees + platform subscription.
Moat: Workflow + medical records ingestion specific to PI; bar association distribution; LexisNexis investment is strategic distribution.
Key risk: Plaintiff-bar TAM ceiling; new product launches drove 90% of recent sales (concentration). Eve is direct competitor.

Eve

Description: AI-native platform specifically for plaintiff-side firms; competes head-to-head with EvenUp.
Founders / year: Jay Madheswaran (CEO), Dorna Moini. Founded 2023.
Latest round: Sep 2025 — $103M Series B at $1B+ (Spark Capital, a16z, Lightspeed, Menlo). (PRNewswire)
Total raised: ~$164M.
ARR: Not publicly disclosed — claims 250% YoY revenue growth; processes 200,000+ cases annually.
Customers: Plaintiff firms collectively recovering “$3.5B+ in settlements.”
Pricing: Per-firm seat + per-case usage.
Moat: Newer architecture vs EvenUp; rapidly closing share.
Key risk: EvenUp is bigger; Eve must outpace it on net-new logos. Two-horse race shape is unstable long-term.

Ironclad (incumbent + agentic)

Description: Contract lifecycle management incumbent with new agentic layer (intake, redlining, renewals, search agents).
Founders / year: Jason Boehmig (CEO), Cai GoGwilt. Founded 2014.
Latest round: Jan 2022 — $150M Series E at $3.2B (Franklin Templeton). No priced round since. (Sacra)
Total raised: ~$334M.
ARR: $200M (Jan 2026), up 34% YoY. (Ironclad)
Customers: L’Oréal, Mastercard, Salesforce, Asana, Cox Enterprises.
Pricing: Per-seat + workflow tiers; enterprise.
Moat: Existing CLM customer base + workflow lock-in. Agentic layer is incremental ARR on installed base.
Key risk: OpenAI dependency for agent features (publicly known). Pure-play AI competitors (Lexion-acquired-by-DocuSign, Spellbook, Robin) target their underbelly. Multiple at $3.2B / $200M = 16x — looks reasonable; not priced for failure.

Robin AI

Description: Contract review/drafting AI; UK origin; mid-market focus.
Founders / year: Richard Robinson (CEO), James Clough. Founded 2019, London.
Latest round: Nov 2024 — $25M Series B Plus. Acquired by Scissero Dec 2025. No longer standalone. (Fortune)
Total raised: ~$72M.
ARR: ~$10M (late 2024).
Notable: Cautionary mid-market case — couldn’t keep up with Harvey’s enterprise momentum; consolidated.

Honorable mention: Legora

Swedish/EU competitor; April 2026 hit $5.6B valuation in Harvey-rivaling round. (TechCrunch) Worth tracking but EU-rooted.

Legal agents — summary table

Company	Latest round	Valuation	ARR	Multiple
Harvey	Mar 2026	$11B	$190M	~58x
EvenUp	Oct 2025 Series E	$2B+	est. $40–80M	~25–50x
Eve	Sep 2025 Series B	$1B+	n/d (rapid growth)	n/d
Ironclad	Jan 2022 Series E	$3.2B (stale)	$200M	~16x
Robin AI	Acquired Dec 2025	n/a	$10M	n/a

E. Healthcare Agents

Strongest US-private-market story for “real ARR + real moat.” Health systems are slow but sticky buyers; agentic deployments are hitting $100M+ ARR with 30–50x multiples that look defensible vs. SaaS.

Abridge

Description: AI ambient scribe for clinicians; converts patient encounter audio into structured documentation, billing codes, summaries.
Founders / year: Shiv Rao (CEO, MD). Founded 2018, Pittsburgh.
Latest round: Jun 2025 — $300M Series E at $5.3B (a16z, Khosla); reported $316M extension late 2025. Prior: Feb 2025 $250M at $2.75B. (TechCrunch)
Total raised: ~$800M+.
ARR: $117M contracted ARR (Q1 2025); $100M+ realized May 2025 per Sacra.
Customers: 150+ health systems including Kaiser, UPMC, Sutter, Yale, Christiana Care, Emory.
Pricing: ~$2,500–$5,000/clinician/year; volume-discounted enterprise.
Moat: EHR integrations (Epic Workshop, etc.); HIPAA compliance + clinical accuracy track record; partnership with Epic.
Key risk: Microsoft (Nuance/DAX Copilot) is the elephant. ~50x ARR is high but the moat is real.

Ambience Healthcare

Description: Ambient AI clinical platform — scribe + AutoCDI (clinical documentation integrity) + AutoAVS (after-visit summary) + AutoRefer.
Founders / year: Mike Ng (CEO), Nikhil Buduma (CTO). Founded 2020.
Latest round: Jul 2025 — $243M Series C (Oak HC/FT, a16z). ($1B valuation widely reported.) (Fierce Healthcare)
Total raised: ~$345M.
ARR: $19M (end 2024) → $30M (May 2025).
Customers: Cleveland Clinic, UCSF, Memorial Hermann, Houston Methodist, Ardent Health.
Pricing: $2,800–$3,200/provider/year base; $4,000–$5,000 with full suite.
Moat: Multi-module suite (vs Abridge’s narrower scribe focus); enterprise health system focus.
Key risk: Abridge is bigger and faster; Microsoft DAX is bundled into Microsoft Cloud for Healthcare. Multiple at $1B / $30M = ~33x; reasonable.

Suki

Description: AI voice assistant for clinical documentation. Older entrant; broader workflow ambitions than Abridge.
Founders / year: Punit Soni (CEO, ex-Google/Flipkart). Founded 2017.
Latest round: Oct 2024 — $70M Series D (Hedosophia, Venrock). (Fierce Healthcare)
Total raised: ~$165M.
ARR: Not publicly disclosed; doubled in 2023, ~quadrupled in 2024 (per Suki disclosures); estimated low double-digit M.
Customers: 400+ health systems; partners include MEDITECH, Amwell, Zoom, Athenahealth.
Pricing: Per-clinician/year SaaS + platform fees.
Moat: Health system breadth (400+); platform partnerships.
Key risk: Lapped by Abridge and Ambience on revenue; “quadrupling” off small base. May need a larger round to catch up.

OpenEvidence

Description: “ChatGPT for doctors” — clinical decision support agent grounded in peer-reviewed literature. Free for verified MDs; monetizes via pharma/medical-device CPM advertising.
Founders / year: Daniel Nadler (CEO, ex-Kensho). Founded 2021.
Latest round: Jan 2026 — $250M Series D at $12B (Thrive, DST). Prior: Oct 2025 $200M at $6.1B; Jul 2025 $210M at $3.5B; Feb 2025 $1B. (CNBC)
Total raised: ~$700M+.
ARR: ~$100M run-rate by early 2026 (one of fastest ever to $100M).
Customers: 18M consultations/month from logged-in verified doctors (Dec 2025); not enterprise SaaS — direct-to-physician.
Pricing: Free for physicians; ad CPMs $70–$1,000+ (vs $5–15 social CPMs).
Moat: Verified-physician audience is hard to replicate; Mayo/AMA/NEJM content licensing; ad-business economics are extraordinary.
Key risk: Pharma ad spend cyclicality + regulation (FDA scrutiny of physician-facing pharma promotion). ~120x ARR is the price of explosive growth + ad-CPM economics; thin margin for error.

Hippocratic AI

Description: Patient-facing AI agents (calling patients, follow-ups, medication adherence, screening). “AI nurses” (controversial framing).
Founders / year: Munjal Shah (CEO, ex-Like.com). Founded 2023.
Latest round: Nov 2025 — $126M Series C at $3.5B (Avenir Growth). Prior: Jan 2025 $141M Series B at $1.64B. (Fierce Healthcare)
Total raised: ~$404M.
ARR: Not publicly disclosed. 50+ health systems, payors, pharma partners; 1,000+ use cases; 115M patient interactions claimed.
Customers: UHS, Cincinnati Children’s, WellSpan Health.
Pricing: $9/hour per “AI agent” advertised; enterprise contracts custom.
Moat: Safety-validation framework + nursing-board endorsements; partnership pipeline.
Key risk: Nursing union pushback; clinical safety bar; ARR opacity at this valuation is concerning. Unclear if usage is paid pilots vs revenue.

Healthcare agents — summary table

Company	Latest round	Valuation	ARR	Multiple
Abridge	Jun 2025 Series E	$5.3B	$117M	~45x
Ambience	Jul 2025 Series C	~$1B	$30M	~33x
Suki	Oct 2024 Series D	n/d (~$1B est.)	n/d (low double-digit M)	n/d
OpenEvidence	Jan 2026 Series D	$12B	$100M	~120x
Hippocratic AI	Nov 2025 Series C	$3.5B	n/d	n/d

F. Finance / Accounting / Back-Office Agents

Mix of fintech-with-AI (Ramp), insurance/healthcare ops (Tennr), and pure agentic accounting (Numeric, Truewind, Basis). Ramp is the standout — a generational fintech that has elegantly bolted on agents.

Ramp

Description: Corporate card + spend management with autonomous AI agents (controllers, AP, travel booking, expense filing). Not a pure agent company but agent surface area is now core.
Founders / year: Eric Glyman (CEO), Karim Atiyeh, Gene Lee. Founded 2019.
Latest round: Nov 2025 — $300M at $32B (Lightspeed-led tender + primary). Prior: Jul 2025 $500M at $22.5B (ICONIQ); Mar 2025 secondary at $13B. (TechCrunch)
Total raised: ~$2.3B.
ARR: $700M (Jan 2025) → $1B (Aug 2025), 110% YoY.
Customers: 30,000+ businesses; Anduril, Shopify, Notion among public references.
Pricing: Free card + interchange; premium plans $15–35/seat; agent features bundled / usage-tiered.
Moat: Interchange-funded distribution = effectively negative CAC; deep ledger integrations; Brex/Mercury are real but Ramp is winning.
Key risk: ~32x ARR is rich for a fintech; AI agent layer is a multiple driver. If macro turns, interchange volumes fall.

Tennr

Description: AI for healthcare patient referral processing — fax/email/PDF intake, intelligent routing, prior-auth.
Founders / year: Trey Holterman (CEO), Tyler Johnson, Diego Baugh. Founded 2021.
Latest round: Jun 2025 — $101M Series C at $605M (IVP, GV, ICONIQ, a16z, Lightspeed). (Fortune)
Total raised: ~$162M.
ARR: “Eight figures, triple of Series B” implies ~$15–25M range; processes 10M docs/month.
Customers: Hundreds of provider groups across ortho, derm, GI, infusion.
Pricing: Per-document or per-referral usage + platform fee.
Moat: Highly proprietary data + workflow into a notoriously paperwork-heavy corner of healthcare. Real defensibility.
Key risk: Sales motion is one specialty practice at a time; scaling beyond referrals is unproven. ~30x ARR is fair.

Numeric

Description: AI close-management platform for finance/accounting teams. Reconciliations, flux, controls.
Founders / year: Parker Gilbert (CEO), Garrett Loftus. Founded 2021.
Latest round: Nov 2025 — $51M Series B (IVP, Menlo, Founders Fund, Alkeon, 8VC). (SiliconANGLE)
Total raised: ~$89M.
ARR: Not publicly disclosed.
Customers: Brex, OpenAI, Watershed, Public.com, Clipboard Health.
Pricing: Per-seat + module add-ons.
Moat: Wedge into the close + early customers in tech/finance. Limited switching costs near-term.
Key risk: Sage/NetSuite/Workday all racing the same AI close features; the wedge needs to widen quickly.

Truewind

Description: AI bookkeeping co-pilot for SMB accounting firms.
Founders / year: Alex Lee (CEO), Niral Patel. Founded 2022.
Latest round: Late 2024/Early 2025 — $13M Series A (Rho Capital, Thomson Reuters Ventures). (Axios)
Total raised: ~$17M.
ARR: Not publicly disclosed; “4x revenue growth” claimed; estimated <$10M.
Customers: EisnerAmper, Frank Rimerman.
Pricing: Per-firm SaaS + per-client tiers.
Moat: Accountant network distribution.
Key risk: Subscale; Basis is much better-capitalized.

Basis

Description: End-to-end AI agent platform for accounting firms — accounting, tax, audit workflows.
Founders / year: Mitch Troyanovsky, Matt Harpe. Founded 2023.
Latest round: Feb 2026 — $100M Series B at $1.15B (Accel, GV, Khosla; Lloyd Blankfein angel). (Bloomberg)
Total raised: ~$138M.
ARR: Not publicly disclosed. ~30% of Top 25 US accounting firms reportedly deploying.
Customers: Top 25 firms (specific names not all public); some BDO and mid-market deployments referenced.
Pricing: Enterprise per-firm contracts.
Moat: Penetration of the top firms (consolidated buyer market — only ~25 firms matter).
Key risk: Big-Four firms are building / partnering on similar in-house. Multiple is high but TAM (~$50B accounting labor) is large.

Finance / accounting — summary table

Company	Latest round	Valuation	ARR	Multiple
Ramp	Nov 2025	$32B	$1B	~32x
Tennr	Jun 2025 Series C	$605M	est. ~$15–25M	~30x
Numeric	Nov 2025 Series B	n/d	n/d	n/d
Truewind	Series A 2024	n/d	<$10M est.	n/d
Basis	Feb 2026 Series B	$1.15B	n/d	n/d

G. Knowledge / Enterprise Search Agents

The “Glean is the next ServiceNow” thesis has held up better than skeptics expected. Hebbia is a niche vertical play. Perplexity is the consumer brand with an enterprise wedge.

Glean

Description: Enterprise AI assistant + agent platform; knowledge-graph indexing across SaaS and on-prem; agents built on top.
Founders / year: Arvind Jain (CEO, ex-Google, Rubrik co-founder). Founded 2019.
Latest round: Jun 2025 — $150M Series F at $7.2B (Wellington, Khosla, Altimeter, Sequoia, Kleiner, Lightspeed, ICONIQ). (TechCrunch)
Total raised: ~$655M+.
ARR: $100M (mid-2024) → $200M (mid-2025; doubling in nine months). (Glean)
Customers: Reddit, Confluent, Pinterest, Workday, Sony, Wayfair, Booking.com.
Pricing: Per-seat ~$30–50/mo + platform/connector fees; enterprise custom (six- to seven-figure ACVs).
Moat: Connectors + knowledge graph + permission-aware search are hard to replicate. Strong against Microsoft Copilot when buyer wants cross-tool indexing.
Key risk: Microsoft Copilot bundled with M365; ChatGPT Enterprise + Connectors. Glean must keep widening lead on connector breadth and agent execution quality. ~36x ARR is fair.

Hebbia

Description: AI for unstructured-document analysis; market focus is asset managers, IB, PE, legal due-diligence.
Founders / year: George Sivulka (CEO, Stanford). Founded 2020.
Latest round: Jun 2024 — $130M Series B at ~$700M (a16z, Index, GV, Peter Thiel). (TechCrunch)
Total raised: ~$161M.
ARR: $13M at last round (profitable). Estimated $30M+ end-2024; could be 2–5x in 2025 per investor commentary.
Customers: Disclosed users at 33% of top global asset managers by AUM; Centerview, Charlesbank, Oaktree.
Pricing: Enterprise seat + usage; six-figure ACVs typical.
Moat: Penetration of the top asset managers + IB; specific “Matrix” workflow has cult status among analysts.
Key risk: Glean Agents + Harvey are bracketing Hebbia from above and side; ChatGPT Enterprise commoditizes basic doc Q&A.

Perplexity

Description: Consumer + enterprise AI search/agent. “Comet” agent browser; “Deep Research” mode.
Founders / year: Aravind Srinivas (CEO, ex-OpenAI), Denis Yarats, Johnny Ho. Founded 2022.
Latest round: Sep 2025 — $200M Series D at $20B; further reporting puts it at $22.6B Jan 2026. (TechCrunch)
Total raised: ~$1.7B.
ARR: $80M (late 2024) → $200M (Feb 2026) → $500M (Apr 2026, +335% YoY). (Sacra)
Customers: 100M MAU; tens of thousands of enterprise customers; major partnerships with telecom carriers (free Pro bundling).
Pricing: $20/mo Pro consumer; $40/seat Enterprise; ad-supported free tier rolling out.
Moat: Consumer brand + speed of product iteration; deal-making (Microsoft Azure $750M GPU commitment); Comet browser distribution wedge.
Key risk: Google’s bundled AI Overviews + ChatGPT search are direct, well-funded competitors; ~45x ARR (off $500M) is high but improving fast. Litigation risk over publisher data is non-trivial.

Knowledge / search — summary table

Company	Latest round	Valuation	ARR	Multiple
Glean	Jun 2025 Series F	$7.2B	$200M	~36x
Hebbia	Jun 2024 Series B	$700M	est. $30M+	~20x+
Perplexity	Sep 2025 Series D (Jan 2026 implied $22.6B)	$20–22.6B	$500M (Apr 2026)	~45x

H. Voice Agents

Voice has been the breakout 2025 vertical. Real ARR + real margin profile (because audio inference is more proprietary, harder to commoditize). ElevenLabs is the bellwether.

Bland AI

Description: Programmable voice agent platform; high-volume outbound/inbound calls.
Founders / year: Isaiah Granet (CEO), Sobhan Naderi. Founded 2023.
Latest round: Mid-2024 — Series A reported; total ~$59.6M raised. Investors include Emergence, Bali Venture Partners. Detailed valuation not publicly disclosed.
Total raised: ~$60M.
ARR: Not publicly disclosed; estimated low-to-mid double-digit M based on usage indicators.
Customers: Healthcare, real estate, recruiting outbound use cases.
Pricing: $0.09/min self-serve; enterprise volume discounts.
Moat: Latency + reliability at scale; developer-accessible API.
Key risk: Vapi + Retell are direct API competitors; ElevenLabs and Deepgram add agent layers above. Margin compression risk.

Retell AI

Description: Voice agent platform; sells to enterprise + dev teams.
Founders / year: Yi Wang (CEO). Founded 2023, YC.
Latest round: Seed $4.6M (Alt Capital, YC). Subsequent rounds reportedly closed but not all publicly disclosed.
Total raised: ~$5M+ disclosed; rumored larger 2025 round.
ARR: $50M (2025 disclosed by company). 50M+ real-time AI calls/month. (Yahoo Finance)
Customers: Enterprise voice deployments — healthcare, services, BPO.
Pricing: Per-minute usage; ~$0.07–0.31/min depending on tier.
Moat: Voice-orchestration platform with telephony integrations; real growth.
Key risk: Concentrated in voice infra layer being squeezed by ElevenLabs Agents + Deepgram.

Vapi

Description: Voice AI developer platform; YC graduate.
Founders / year: Jordan Dearsley (CEO), Nikhil Gupta. Founded 2022.
Latest round: Dec 2024 — $20M Series A at ~$130M (Bessemer, AI Grant, Saga, YC). (Vapi)
Total raised: ~$25M.
ARR: Targeting $8M end-2024; estimated $20–30M run-rate by 2026 based on growth trajectory.
Customers: Developer-heavy base; emerging enterprise.
Pricing: Usage-based per-minute.
Moat: Developer mindshare; platform breadth.
Key risk: Same layer-compression risk as Bland and Retell.

Deepgram

Description: Speech-to-text/voice infrastructure; pivoted to “voice AI economy” with agent layer + recent acquisition of YC startup.
Founders / year: Scott Stephenson (CEO), Noah Shutty. Founded 2015.
Latest round: Jan 2026 — $130M Series C at $1.3B (AVP, plus existing investors + Twilio, ServiceNow, SAP, Citi). (TechCrunch)
Total raised: ~$215M.
ARR: Not publicly disclosed; was cash-flow positive in 2024; 400+ enterprise customers, 200K developers.
Customers: NASA, Spotify, Citi, Volkswagen.
Pricing: Per-minute API; enterprise contracts.
Moat: Proprietary voice models + 10+ years of speech infrastructure. One of the few voice players with structural cost advantage.
Key risk: OpenAI Whisper / GPT Realtime API commoditize transcription; Deepgram needs to keep winning on cost and latency.

ElevenLabs

Description: Generative voice + voice agents. UK/Poland-founded; effectively the Anthropic-of-voice.
Founders / year: Mati Staniszewski (CEO), Piotr Dąbkowski. Founded 2022.
Latest round: Feb 2026 — $500M Series D at $11B (Sequoia). Prior: Jan 2025 $180M Series C at $3.3B. (TechCrunch)
Total raised: ~$830M.
ARR: $100M (took 20 months) → $200M (10 months) → $330M (5 months — closed 2025). CEO publicly committed to doubling in 2026 (~$660M).
Customers: Deutsche Telekom, Revolut; mass developer + enterprise base; near 50/50 enterprise/self-serve revenue split.
Pricing: $5–$1,100/mo published self-serve; enterprise custom (per-character / per-minute).
Moat: Voice quality leadership + low latency; the only voice infra co with brand ownership at the consumer layer.
Key risk: OpenAI’s GPT Realtime + Google + Meta directly compete; IP risk around voice cloning. ~33x ARR is the most defensible large multiple in this report.

Voice agents — summary table

Company	Latest round	Valuation	ARR	Multiple
Bland AI	2024	n/d	n/d	n/d
Retell AI	2025 (rumored)	n/d	$50M	n/d
Vapi	Dec 2024 Series A	$130M	~$8M projected	~16x
Deepgram	Jan 2026 Series C	$1.3B	n/d	n/d
ElevenLabs	Feb 2026 Series D	$11B	$330M	~33x

I. General-Purpose / Consumer Agents

This is the riskiest sub-sector for capital deployment. Reverse-acquihires, regulatory blocks, and unclear unit economics dominate.

Adept

Status: Effectively dissolved. Jun 2024: Amazon “reverse-acquihired” co-founders David Luan, Augustus Odena, Maxwell Nye, Erich Elsen, Kelsey Szot + ~40 staff into AGI lab; investors (Greylock, General Catalyst — $414M put in) made roughly whole; Adept retained ~20 employees and ~$25M license payment from Amazon. (TechCrunch, Semafor)
Subsequent: David Luan left Amazon Feb 2026.
Investor lesson: Reverse-acquihires return capital but no upside. Watch for similar deals in coding-agent space.

MultiOn

Description: Web-task automation agent (browser-controller).
Founders / year: Div Garg (CEO, Stanford), Omar Shaya. Founded 2022.
Latest round: Seed only — General Catalyst, Amazon Alexa Fund, Samsung Next, Maven, OpenAI individuals.
Total raised: Not fully disclosed; estimated ~$5–10M seed.
ARR: Not publicly disclosed; consumer freemium product.
Customers: Consumer prosumer; some travel-vertical exploration.
Moat: Limited; thin browser-agent layer being copied by ChatGPT Operator / Claude Computer Use / Gemini.
Key risk: Highly likely acquihire candidate.

Manus (Butterfly Effect — comparison only)

Status: Chinese-origin (Singapore-incorporated) autonomous agent. Apr 2025: $75M Series B at ~$500M (Benchmark, Tencent). Dec 2025: Meta announced $2B+ acquisition. Apr 2026: China NDRC blocked the deal under new Foreign Investment Security Review Measures. (Wikipedia, SiliconANGLE)
Takeaway: First test case of China blocking foreign tech acquisition; not investable for US family office. Useful comparison for Western general-purpose agents on capability claims and demo virality.

General agents — summary table

Company	Status	Valuation	ARR	Notes
Adept	Reverse-acquihired by Amazon Jun 2024	n/a	n/a	Investors recouped capital
MultiOn	Seed-stage	<$50M est.	n/d	Acquihire risk
Manus	Acquisition blocked Apr 2026	$500M (Apr 2025)	n/d	Geopolitically uninvestable

Cross-Cutting Family-Office Takeaways

The “real businesses” list is shorter than the headlines suggest. Companies with $100M+ ARR + verifiable customer base + reasonable defensibility: Cursor, Cognition (post-Windsurf), Sierra, Harvey, Glean, Abridge, OpenEvidence, Ironclad (incumbent), Ramp, ElevenLabs, Perplexity, Clay. That’s ~12 names worth a thesis-level conversation, out of dozens raising at unicorn-plus valuations.
Avoid “AI BDR” and “AI general-purpose agent” verticals for now. ARR is fragile (3-month break clauses, logo inflation), competition is high, and large-incumbent CRM bundling will likely win.
Healthcare and legal show the most durable moats. Regulated buyers, slow but sticky, real workflow lock-in, proprietary data flywheels. These are the verticals with the best risk-adjusted ARR multiples (Harvey, Abridge, Ironclad).
Voice infrastructure is the rare layer with both growth and gross-margin durability. ElevenLabs and Deepgram have proprietary models that resist commodity pressure better than text-only application layers.
Watch for reverse-acquihire patterns (Adept, Inflection, Character, Windsurf). Existing equity-holders can recoup, but founders/employees move to MAANG and the standalone equity outside the carve-out is worthless. If you’re considering secondaries in any well-funded but stalled application-layer agent, model that scenario explicitly.
Outcome-based pricing is the structural shift. Sierra, Decagon, Hippocratic, EvenUp are normalizing pay-per-resolution / pay-per-case. This compresses traditional SaaS comparables — a $50M-ARR outcome-priced business with 90% gross retention is not the same asset as a $50M per-seat SaaS business.
Multiple discipline matters. Anything above ~50x ARR (Decagon, OpenEvidence at certain time, Magic, Poolside, Rox) is priced for multiple step-functions of growth. The Poolside Series C collapse in April 2026 is the first publicly reported sign that LPs are pushing back. Expect more of this in 2026.

Sources

Part 2C — Money Flow, Theses, Incumbents, Risks

Macro investment landscape for AI agents and agentic software. US-centric, private-market focused. Prepared for family office principals, May 2026.

This document maps how capital is flowing into the agentic software stack, what theses are competing for that capital, who the incumbents are, what could break, and how a family office of meaningful but non-mega scale should actually access the opportunity. Numbers are sourced inline. Where we estimate, we say so.

A. Market Sizing — The Capital Tsunami

Total VC into AI, by year

Year	AI VC (global)	Total VC (global)	AI share	Notes
2023	~$52B (Carta) / ~$50B (Crunchbase)	$285B	~18–41% depending on definition	First post-ChatGPT year. GenAI alone went from $2.8B (2022) to $15.3B [Source: OECD, 2026-02; Crunchbase, 2024-01]
2024	$131.5B	~$370B	~35.7% globally; ~50% in North America	Frontier-lab raises (OpenAI, xAI, Anthropic) more than doubled YoY [Source: PitchBook, 2025-01]
2025	$192.7B (PitchBook) / $225.8B (CB Insights) / $270.2B (broader)	$366.8B–$512.6B	First year >50% of all VC	Mega-rounds ($100M+) = 79% of AI funding. OpenAI + Anthropic + xAI = $86.3B alone (38% of AI total) [Source: Bloomberg, 2025-10-03; CB Insights State of AI 2025]
2026 YTD (Q1)	$242B	$300–331B	~80%	Q1 alone exceeded all of 2025. Four rounds — OpenAI $122B, Anthropic $30B, xAI $20B, Waymo $16B — were $188B (62% of all global VC) [Source: Crunchbase, 2026-04; KPMG Venture Pulse Q1’26; Yahoo Finance, 2026-04]

Read this twice. In Q1 2026, four companies absorbed more venture capital than the entire global VC industry deployed in any full year prior to 2018. Capital is concentrating, not dispersing.

Mega-rounds ($500M+) in agentic / AI-adjacent space (2024 through Q1 2026)

Company	Round	Size	Valuation	Lead	Date
OpenAI	Multiple	$40B (2025) + $122B (Q1’26)	$300B → $852B–$1T	SoftBank, Microsoft	2025–2026
Anthropic	Series F+	$13B (2025) + $30B (Q1’26)	$183B → ~$350B; talks of $900B at $50B raise	Lightspeed, ICONIQ	2025–2026
xAI	Multiple	$20B (Q1’26)	~$200B+	Various	2026
Waymo	Late-stage	$16B (Q1’26)	n/a	Alphabet-led	2026
Databricks	Series J	$10B	$62B → $134B	Thrive	2025-01
Cursor (Anysphere)	Series D	$2.3B	$29.3B; talks of $50B	Accel, Coatue	2025-11
Cognition AI	Series —	$400M; talks of $25B round	$4B → $10.2B → $25B (in talks)	Founders Fund / 8VC	2025-09 / 2026-04
Sierra	Series —	$350M	$4.5B → $10B	Greenoaks	2025-09
Harvey	Series E	$200M	$5B → $8B → $11B	Kleiner / Sequoia	2025-12 / 2026-03
Perplexity	Late-stage	$200M+	$20B	Various	2025-09
Glean	Series F	$150M	$7.2B	Various	2025-06
Replit	Series D	$400M	$9B	Various	2025
Sierra-style voice agent (Sequoia-led)	$500M	n/a	n/a	Sequoia	2025
Metropolis	$500M	n/a	n/a	n/a	2025
Nexthop AI	$500M	n/a	n/a	n/a	2025
Abridge	Series —	$300M	$5.3B	a16z	2025

[Source: TechCrunch 2025-11-13, 2025-09-08, 2025-09-04; Bloomberg 2025-09-04; CNBC 2026-03-25; Crunchbase News 2025; aifundingtracker.com 2026-04]

Pattern: Below the foundation labs, ~$10–12B is the gravitational ceiling for vertical-AI and horizontal-app agentic players. Above that, you must be a frontier lab or own infrastructure.

TAM frames the bulls cite

a16z / Bessemer “services-as-software”: US services labor is a ~$13T addressable market (BLS-derived). If agents capture even 5–10% of services workflows over a decade, that is $650B–$1.3T of recurring revenue migrating from labor budgets into software budgets [Source: a16z Big Ideas 2026; Bessemer State of the Cloud 2025].
Gartner enterprise apps: Agentic AI could drive ~30% of enterprise application software revenue by 2035, surpassing $450B, up from 2% in 2025 [Source: Gartner press release, 2025-08-26].
Gartner B2B commerce: AI agents will channel $15T in B2B purchases by 2028 [Source: Gartner / Digital Commerce 360, 2025-11-28].
Gartner SCM: $2B (2025) → $53B (2030) for SCM software with agentic AI [Source: Gartner, 2026-04-07].
McKinsey: AI software and services have $15.5T–$22.9T annual economic potential by 2040; mid-case US-only economic value of $2.9T/yr by 2030 [Source: McKinsey, 2025; Marketing AI Institute, 2024].
Sequoia “$600B question”: This is the bear TAM frame — David Cahn’s argument that to justify current AI capex, end-user revenue must reach ~$600B/yr. As of mid-2026, frontier labs run-rates are ~$50–60B combined; the gap is real but closing fast [Source: Sequoia Capital, 2024 essays; Sequoia “AI in 2025”].

Public-market signal: hyperscaler capex

Company	2024 Capex	2025 Capex (est.)	2026 Capex (announced)
Amazon (AWS)	~$77B	~$110B	~$200B
Alphabet (Google)	~$53B	~$85B	$175–185B
Microsoft	~$56B	~$90B	$120B+
Meta	~$40B	~$70B	$115–135B
Combined Big 4	~$226B	~$355B	~$610–700B

[Source: CNBC, 2026-02-06; Tom’s Hardware, 2026-04; Introl, 2026-01; CreditSights, 2026]

Key facts to internalize:

Hyperscaler capex has quadrupled since GPT-4’s release [Source: Epoch AI, 2025].
Approximately 75% of 2026 hyperscaler capex (~$450B) is AI-related [Source: Futurum, 2026].
Amazon expects ~$17B of negative free cash flow in 2026; Microsoft FCF projected to slide ~28% [Source: Morgan Stanley via CNBC, 2026-02-06; Barclays].
Combined Alphabet + Amazon + Meta + Microsoft + Oracle capex has compounded ~72% per year [Source: CNBC, 2026-02].

This is the single most important macro datapoint in the deck. The hyperscalers are betting their balance sheets that agentic + AI workloads will require this much compute. If they are right, infrastructure picks-and-shovels print money. If they are wrong, public equity multiples compress severely.

B. Money Flow Patterns

Round-size inflation is acute in agentic vs. broader SaaS

Average AI revenue multiple (private): ~37.5x ARR. Traditional SaaS comps: ~7.8x [Source: Aventis Advisors, 2025; Qubit Capital, 2026].
Best-in-class agentic outliers: 50–100x ARR. Sierra: $100M ARR → $10B valuation = 100x [Source: TechCrunch, 2025-11-21].
Stage-by-stage multiple curve (mid-2025): Seed 22.7x → Series A 39.1x → Series B 41.0x peak → Series C 26.2x → Series D 29.3x → Series E+ 31.7x rebound [Source: Finro, 2025; PitchBook].
18% YoY compression observed in 2025 at the median, suggesting some discipline returning outside the top-decile names [Source: Aventis Advisors, 2025].

Time from seed to unicorn has collapsed

Cursor / Anysphere: Series A $400M valuation (Aug 2024) → Series D $29.3B (Nov 2025) → in talks at $50B (Q1’26). Series A to $29B in ~15 months. Hit $100M ARR Jan 2025; $500M June; $1B November; $2B by Feb 2026. Fastest B2B SaaS scale on record — ahead of Slack, Zoom, and Snowflake [Source: CNBC, 2025-11-13; TheNextWeb, 2026].
Cognition (Devin): $1M ARR (Sept 2024) → $73M ARR (June 2025) → ~$150M post-Windsurf acquisition. $4B → $10.2B → $25B (in talks) in ~13 months [Source: TechCrunch, 2025-09-08; Bloomberg, 2026-04-23].
Sierra: $0 → $100M ARR in 21 months; $4.5B → $10B in 11 months [Source: TechCrunch, 2025-11-21].
Harvey: $50M ARR (Dec 2024) → $195M ARR (2025) — 3.9x in one year [Source: CNBC, 2026-03-25; Sacra].
Glean: $100M ARR (Jan 2025) → $200M ARR (Dec 2025) — doubled in 9 months [Source: BusinessWire, 2025-12-08].

207 AI-focused companies have minted unicorn status since 2024 [Source: PipelineRoad, 2026-04-30]. The “Series A to unicorn in <18 months” pattern is now the norm in agentic, not the exception.

Secondary markets — yes, very active

AI captured 67% of all private secondary market funding in 2025, up from 9% in 2022 [Source: Forge Global, 2025].
Secondary trading in AI companies surged from 2% of total volume in 2022 to 44% in 2025 [Source: Forge Global, 2025].
Major tender offers in 2025: OpenAI, Anthropic, Stripe, Databricks, SpaceX [Source: Forge Global; EquityZen, 2025].
EquityZen H1 2025 deal volume nearly doubled vs. H1 2024 [Source: EquityZen Q3 2025 update].
Total private secondary market hit $103B in H1 2025 alone [Source: aggregated industry estimates].

Top investors by deal count in agentic / AI

Investor	Posture	Notable agentic plays
Sequoia Capital	#1 by deal count; led most-active VC list June 2025	Harvey (Series E lead $300M), OpenAI, Sierra, Cyera ($540M Series E)
Andreessen Horowitz	#1 alongside Sequoia in unicorn-class deals	Abridge ($300M), Mistral, Anysphere participation
Lightspeed	Top-5 by AI capital deployed	Anthropic, Mistral
Index Ventures	Topped AI-agent deal list w/ a16z (2025)	Various agentic apps
Khosla Ventures	Early OpenAI; healthcare/scientific AI	OpenAI, Replit
Founders Fund	Concentrated bets	Cognition lead ($400M)
Greenoaks	Late-stage growth	Sierra lead
Thrive Capital	Growth/crossover	Databricks $10B (2025), OpenAI
Benchmark, ICONIQ	Selective	Sierra, Anthropic
General Catalyst	Healthcare AI roll-up strategy	Multiple
Accel	Coatue partner on Cursor	Anysphere
Coatue	Crossover; very active again	Cursor lead, Oracle public

[Source: Crunchbase News, 2025-08; PitchBook agentic-AI tracker, 2025-10-13; Affinity Top-5 VC AI report, 2026]

Crossover funds — yes, fully back

After 2022–2023 retreat, the Tiger Cubs are deploying aggressively:

Coatue: ~$70B AUM; +19% returns 2025; co-led Cursor’s $2.3B Series D; doubled Nvidia/Microsoft public positions; new Oracle position [Source: Bloomberg, 2025-01-07; CNBC, 2025-08-14; Institutional Investor, 2025].
Tiger Global: +24% in 2025 on tech-rally beta [Source: Bloomberg, 2025-01-07].
Lone Pine: +36% in 2024 [Source: Bloomberg].
Altimeter, Dragoneer, D1: all increasingly active in $500M+ AI rounds, often paying 50–100% premiums to traditional VCs [Source: PitchBook, 2025].

Crossovers are paying premiums precisely because they need to deploy massive funds at the late stage and AI is the only asset class with the absorption capacity. This is bullish for current paper marks but creates the classic 2021-style markdown risk if growth disappoints.

C. Thesis Frames — Six Ways to Bet

Thesis 1: Picks & shovels (own the infrastructure)

Bet: Applications commoditize. Compute, models, and frameworks capture the durable margin pool.
Proponents: Public-market consensus (Nvidia thesis), CoreWeave investors, hyperscaler bulls, Sequoia (selectively).
Evidence: Nvidia data center revenue run-rate $200B+; CoreWeave/Lambda growth; hyperscaler capex $610–700B in 2026; Cerebras filed S-1 with $510M 2025 revenue, $237.8M net income [Source: TechCrunch, 2026-04-18].
Counter: Hyperscalers + open-weight models could collapse infra margins; DeepSeek, Llama, Qwen, Kimi, GLM, Mistral — five open frontier-quality model families now exist [Source: BentoML / HuggingFace, 2026]. Inference cost per token has fallen ~95% since GPT-4 launch.

Thesis 2: Vertical apps in regulated/sticky verticals

Bet: Distribution + proprietary workflow data > model. Regulated industries (legal, healthcare, finance, defense) protect incumbents from horizontal disruption.
Proponents: Bessemer, Menlo Ventures, Coatue, sector specialists.
Evidence:
- Harvey (legal): $195M ARR, $11B valuation, 100,000+ lawyers across 1,300 firms [Source: Harvey blog, 2026; CNBC, 2026-03-25].
- Abridge (healthcare ambient scribe): $5.3B valuation; 30% market share; ambient scribes generated $600M industry revenue in 2025, +2.4x YoY [Source: Menlo Ventures State of AI in Healthcare, 2025].
- Hippocratic AI, Ambience, Nabla — all $1B+ in healthcare specifically.
Counter: TAM per vertical is bounded; Big Law is ~$400B globally. Incumbents (Thomson Reuters/Westlaw, Epic, Salesforce Health Cloud) are bundling fast. Salesforce paid $8B for Informatica in 2025 to defend the data layer.

Thesis 3: Horizontal apps (one function, all enterprises)

Bet: A single best-in-class agentic UX wins a category (coding, support, sales, search).
Proponents: a16z, Index, Greylock.
Evidence: Cursor’s coding dominance ($2B ARR in <3 years); Sierra in CX ($100M ARR in 21 months); Glean in enterprise search ($200M ARR); Decagon, Crescendo, Parloa in voice CX.
Counter: The hyperscalers (Microsoft Copilot, Google Gemini Enterprise) bundle horizontals into seats already paid for. Microsoft’s 230,000+ Copilot Studio orgs [Source: Microsoft Adoption, 2026] is the single biggest threat to standalone horizontals. Cursor itself competes with GitHub Copilot, Cline, and Anthropic’s Claude Code — bundled at $0–$20/seat.

Thesis 4: Agent-first incumbent disruption (kill Salesforce/ServiceNow/Workday)

Bet: SaaS UI is the wrong shape. Agent-native rebuilds of CRM/ITSM/HRIS will drink incumbents’ lunch.
Proponents: Founders’ bet (Sierra was literally founded by Bret Taylor, ex-Salesforce co-CEO); Decagon; 11x; Clay.
Evidence so far: Sierra at $100M ARR in 21 months; 11x and Decagon at $50–100M ARR ranges; but Salesforce Agentforce is now the fastest-growing product in the company’s 26-year history — $800M ARR by early 2026, 18,500+ deals signed [Source: Salesforce Q4 FY26 earnings; Salesforce Ben, 2026].
Counter: Distribution wins. Salesforce reaches every Fortune 500 already. ServiceNow Now Assist hit $600M ACV in Q4 2025, doubling YoY [Source: Futurum, 2026]. The incumbents are waking up before disruptors gain escape velocity.

Thesis 5: Services-as-software (a16z thesis)

Bet: AI agents replace services revenue (consulting, legal services, accounting), not just SaaS. Outcome-based / gainshare pricing replaces seat-based pricing. The market expands 10x because services is 10x SaaS.
Proponents: a16z (most explicit), Foundation Capital ($4.6T+ opportunity counter-essay), General Catalyst.
Evidence:
- McKinsey itself runs ~25,000 internal AI agents; 1.5M human-hours saved [Source: McKinsey 2026; vogon.today, 2026-01-14].
- Outcome-based pricing emerging at Crescendo (CX), 11x (SDR), Sierra (CX) — pay-per-resolution, pay-per-meeting-booked.
- a16z’s December 2024 enterprise letter explicitly frames this shift: “AI does the work, you guarantee the outcome, customer pays for the result” [Source: a16z, 2024-12].
Counter: Outcome attribution is hard. Buyers want to pay only for the marginal lift. Margins on outcome-based deals can be lower than SaaS once delivery costs are fully loaded. Liability shifts to the vendor.

Thesis 6: The “AI engineer” hire — virtual employee at $X/month

Bet: Companies will buy headcount, not software. A $5K/month agent replaces a $150K/yr human.
Proponents: Cognition (Devin), 11x (Alice/Jordan), Replit Agent, Lindy, Adept (pre-acquihire).
Evidence: Cognition’s $25B-talks valuation off ~$150M ARR is precisely this thesis (~167x revenue multiple). 11x grew to ~$10M ARR in <12 months. McKinsey’s 25,000-agent deployment is enterprise demand validation.
Counter: The metaphor breaks down. Real employees take direction, learn, escalate, exercise judgment. Most “AI employee” pitches are over-promising. Gartner predicts >40% of agentic AI projects will be canceled by end of 2027 [Source: Gartner, 2025-06-25]. 2026 is the year of “I’ll believe it when I see it.”

D. Incumbents — Friend or Foe?

Incumbent	Agent product	Traction	Posture for family-office investor
Salesforce	Agentforce (+ Data 360)	$800M ARR (early 2026), $1.4B combined w/ Data 360, +330% YoY, 18,500+ deals, 12,000+ enterprise customers. Fastest ARR ramp in Salesforce’s 26-year history	Distribution moat. Friend to vertical AI players that integrate (Sierra works alongside; Harvey doesn’t compete). Foe to horizontal CX disruptors. Likely acquirer.
ServiceNow	Now Assist + AI Agents	$600M ACV crossed Q4 2025, +>100% YoY; 35 deals >$1M in Q4 alone; Moveworks acquired for $2.85B	Bundling ITSM agents free with Pro Plus / Enterprise Plus. Aggressive acquirer. Foe to standalone enterprise search and ITSM agents.
Microsoft	Copilot Studio + Agent 365	230,000+ orgs using Copilot Studio; 15M+ M365 Copilot seats (~$1.5–5.4B est. revenue depending on discount assumption)	Massive distribution moat. Owns OpenAI. Friend as platform; foe as competitor to standalone agent-builders.
Google	Vertex AI Agent Builder + Gemini Enterprise	Less disclosed; growing on GCP-bundled deployments	Dark horse for enterprise — Gemini quality has caught up. Acquirer.
Workday	Illuminate (Case Agent, Performance Review, Financial Close)	Limited disclosed traction; HRIS agents shipping in 2025–2026	More vulnerable than peers — HRIS is more replaceable. Foe that may become roadkill if agent-native HR (Lattice, Eightfold AI) executes.
Oracle	Embedded AI in Fusion Cloud Sales	”AI included” messaging, less disclosed	Acquirer / consolidator. Stratechery view: Oracle is the dark-horse hyperscaler winner of the cycle.
SAP	Joule + agent portfolio	Bundled with subscriptions	Defensive posture. Not an acquirer at scale. Roadkill risk in mid-market.

Translation for principals: Salesforce, ServiceNow, and Microsoft are both distribution moats (validating that enterprise buyers are signing real agentic contracts) and threats to startups. Workday and SAP look most vulnerable. Oracle is the single most underestimated enterprise AI play in public markets.

E. Risks Specific to Family-Office Investors

1. Valuation risk / pull-forward

Median Series A AI multiple of 39x ARR; outliers at 100x+. This is double the dotcom-bubble peak by some measures [Source: Aventis Advisors, 2025].
79% of 2025 AI funding was in mega-rounds (>$100M), meaning index exposure is heavily concentrated in expensive late-stage paper [Source: CB Insights State of AI 2025].
Estimate: If Cursor, Sierra, Cognition, and Harvey miss 12-month forward growth expectations by 30%, paper marks could compress 50–70% on a peer-comp basis. This is an estimate based on 2022 SaaS-multiple compression precedent, not a sourced number.

2. Model commoditization

Five open-weight frontier-quality model families exist as of 2026: DeepSeek, Qwen, Kimi, GLM, Mistral, plus Llama [Source: BentoML, 2026].
Inference cost per token has fallen ~95% since GPT-4 launch.
If Anthropic / OpenAI lose pricing power, the entire valuation stack — from frontier labs to applications that pay them — re-rates. Anthropic’s $30B run-rate is currently growing 3x in 4 months; if that growth halves, the $900B-talks valuation collapses [Source: TechCrunch, 2026-04-29; Sacra/Anthropic].

3. Liability and regulation

EU AI Act: General-purpose AI obligations enforced 2 Aug 2025; Commission enforcement powers (and fines) live 2 Aug 2026; high-risk systems originally 2 Aug 2026 — possibly deferred to 2 Dec 2027 under the Digital Omnibus proposed Nov 2025, but not yet adopted [Source: artificialintelligenceact.eu; DLA Piper, 2026].
“High-risk agentic AI systems with untraceable behavioural drift cannot currently be placed on the EU market” [Source: secureprivacy.ai, 2026].
US state-level: Colorado AI Act, California SB-1047 derivatives, NY automated employment decision rules. No federal framework yet.
Liability gap is real: Who pays when an agent makes a $10M mistake — the developer, deployer, or end user? Open question. Agent-caused harm is uninsurable in most current policies.

4. Concentration in foundation models

Most agentic companies route tokens through Anthropic, OpenAI, or (increasingly) Google. A single foundation lab’s pricing change, capacity restriction, or terms-of-service shift can erase the gross margin of a $1B+ application company overnight. This is the closest analog to platform-risk on top of Apple/Meta in the 2010s.

5. Talent / founder risk

Many agentic unicorns are 2–3 year-old, solo or duo-founder companies at $1B+ valuations. Cognition (founded 2024), Sierra (2023), Cursor (2022). The number of operators who have actually scaled enterprise software past $100M ARR is thin. Bret Taylor (Sierra) is the rare exception — most are first-time CEOs.

6. Exit pathway

IPO: Cerebras filed S-1 (April 2026) — first pure-play AI IPO since the cycle began [Source: TechCrunch, 2026-04-18]. Anthropic engaged Wilson Sonsini for potential 2026 IPO. OpenAI targeting 2026/2027 at $1T. Databricks “ready when it decides.” An IPO window that absorbs $1T of paper is unprecedented; expect bottlenecks.
M&A: CB Insights reports 266 AI M&A deals in Q1 2026, +90% YoY; ServiceNow / Moveworks ($2.85B), NiCE / Cognigy ($955M), Google / Wiz ($32B), CoreWeave / Core Scientific ($9B), AMD / ZT Systems ($4.9B), OpenAI / io ($6.5B) [Source: Crunchbase News, 2026; Solganick AI M&A Q3 2025; Analytics Insight 2025].
Lockups: IPO lockup risk is acute given concentration. Family offices entering pre-IPO via secondary should price 6-month lockup discount.

7. Liquidity

Late-stage privates are theoretically liquid via Forge / EquityZen / Hiive, but spreads are wide and tender frequency varies dramatically by company.
Secondary discounts: Top-tier AI names (OpenAI, Anthropic, Databricks) trade at premiums to last primary; mid-tier names trade at 20–40% discounts; struggling names at 50–70% discounts. This is a market characterization, not a single sourced statistic; verify per-name with Forge/Hiive marks at time of trade.
Tender offer frequency: top-tier ~quarterly to semi-annually; most mid-tier names get one tender per ~12–18 months.

F. How a Family Office Should Think About Access

The reality matrix by AUM

FO size	Direct primary in hot rounds	Secondary / SPV	Fund LP / co-invest	Public proxies
<$100M	Effectively zero	$25K–$250K SPV checks	Small-LP in micro-funds	Primary tool
$100M–$500M	Rare; via relationships	$250K–$2M SPV checks; co-investment via fund LP positions	Top-tier emerging managers; co-invest unlocks key	Core position
$500M–$2B	Selective; usually via existing GP	$1M–$10M secondary; family-club SPVs	Multi-fund LP base unlocks meaningful co-invest pipeline	Should still be 50%+ of AI exposure
$2B+	Real allocations possible (Series B+)	Anchor SPVs; lead family-club deals	Anchor LP at top tier (Sequoia, a16z, Founders Fund)	Tactical

Five access modes, ranked by typical family-office utility

Public proxies (highest weight for most FOs). MSFT, NVDA, GOOG, META, AMZN, ORCL, AVGO, PLTR, CRWV (CoreWeave). For 90% of family offices below $1B, this is the bulk of credible AI exposure. Liquid, transparent, no lockup. Trades at ~25–40x earnings vs. 50–100x ARR for privates.
Fund LP commitments to AI-focused VCs. Top-tier (Sequoia, Founders Fund, Index, Lightspeed, Coatue Growth) is closed to new LPs. Realistic path: emerging managers focused on AI (Conviction, Air Street, Radical, South Park Commons GP funds). $2–3M LP commitments often unlock $10–20M in co-invest capacity at lower fees [Source: Astra Global, 2025; PwC Global Family Office Deals Study].
Co-investment SPVs through fund relationships. Often 5–10% carry, no management fee. The single best risk-adjusted access mechanism.
Secondary SPVs (Forge, EquityZen, Hiive, broker SPVs). Total private secondary $103B in H1 2025 [Source: industry estimates]. Best for buying into proven names (Databricks, Anthropic, OpenAI) at known marks. Worst for: paying retail premium to last primary on hot names you can’t otherwise access. Min checks $10K–$25K.
Direct primary. Rarely available. When offered to a non-strategic family office, treat as a negative signal — the round is not oversubscribed by the right people.

What we’d actually advise (single FO, $200M–$1B AUM)

30–50% of AI exposure in public proxies (NVDA, MSFT, GOOG, META, AMZN, ORCL, AVGO). Liquid, no carry drag.
15–25% in 1–2 emerging-manager AI VC funds with strong co-invest unlocks. Plan to deploy 2x your fund commitment in co-invest over 5 years.
15–25% in secondaries of proven names (Databricks, Anthropic, possibly OpenAI), accessed via SPVs from broker-dealers. Avoid hot-deal SPVs at retail markups.
5–10% in 1–2 vertical-AI direct-primary positions if you can secure them through GP relationships.
Remainder cash / opportunistic to deploy on a market reset or specific high-conviction event.

The biggest mistake we see: family offices paying 50–100% premiums into 7th-tier SPV layers to “get into OpenAI.” After fees, carry, and discount-to-last-round adjustments, returns are often inferior to simply holding MSFT.

G. The 12–24 Month Watchlist

Most likely IPOs (12–24 months)

Company	Probability	Rationale
Cerebras Systems	Already filed S-1 April 2026	$510M 2025 revenue, $237.8M net income, only pure-play AI IPO so far
Databricks	High; “ready when it decides”	$134B private valuation; faster growth than Snowflake; profitable
Anthropic	Medium-high	Engaged Wilson Sonsini; possible H2 2026 / 2027
OpenAI	Medium-high; targeting late 2026 / 2027	$1T target; would be largest IPO in history
CoreWeave already public; Lambda Labs	Medium	Inference infra demand; private comps trading at high multiples
Mistral, Cohere	Medium	European AI champion narratives
Glean	Medium-low (could go either way; M&A target)	$200M ARR; strategic value to MSFT/CRM/SAP
Harvey	Medium-low (or acquired by Thomson Reuters)	Vertical leader; clean financials

[Source: Crunchbase News 2026 outlook; Marketwise IPO watchlist 2026; Morningstar 2026]

Most likely acquired (12–24 months)

Vertical specialists with $50–200M ARR: Abridge (by Microsoft Health, Epic, or Oracle Cerner), Hippocratic, Decagon (by Salesforce/ServiceNow), 11x (by HubSpot or Salesforce).
AI search / knowledge: Glean, Hebbia (by Microsoft, Oracle, or a financial-services strategic).
AI dev tools: Cursor itself is plausibly too expensive to buy now; second-tier (Cline, Codeium-equivalents) get rolled up. Cognition + Windsurf is a precedent.
Voice AI: Parloa, Cresta, Replicant — Salesforce, Genesys, NiCE, Five9 are aggressive consolidators.
Infrastructure: any inference-optimization or retrieval-DB startup that hits $20M+ ARR. Pinecone, LanceDB, Turbopuffer — all M&A targets for hyperscalers.
Pattern: Forecast is 120–180 disclosed agency / agentic M&A deals across Q3 2026 → Q2 2027, peaking Q4 2026 / Q1 2027 [Source: digitalapplied.com, 2026].

Sectors most likely to consolidate

AI SDR / outbound sales agents (11x, Artisan, Regie, Jason AI, Salesify…). 30+ funded entrants. Will collapse to 2–3 standalone winners + acquihires.
Customer-support agents (Sierra, Decagon, Crescendo, Ada, Parloa, Forethought, Lorikeet). 20+ funded; clear 2–3 leaders emerging.
Coding assistants (mid-tier) — Cursor, Cognition, GitHub Copilot, Claude Code dominate; long tail evaporates.
Healthcare ambient scribes (Abridge, Ambience, Suki, Nabla, DeepScribe, Augmedix). Already consolidating; Abridge has 30% share.
Legal AI (Harvey, Eve, Spellbook, Ironclad+AI). Harvey has clear lead; expect 1–2 acquisitions of #2 / #3 by Thomson Reuters or LexisNexis (RELX).

Sectors most likely to break (over-funded, no moat)

“Generic” vertical AI agents without proprietary data flywheel. If your moat is “we built a wrapper on Claude for X industry,” you are roadkill.
AI BDR/SDR commodity tools. Race to zero on per-meeting price. SDR-replacement promises consistently under-deliver in production.
Standalone “AI employee” platforms. Gartner: >40% of agentic AI projects will be canceled by end of 2027 [Source: Gartner, 2025-06-25]. The AI-engineer category will see 50%+ company mortality.
Mid-tier model labs without distinctive distribution — Inflection-style outcomes (acquihired). Reka, Adept already absorbed precedents.
Generic horizontal “agent platforms” (build-your-own-agent IDEs without enterprise distribution). Microsoft Copilot Studio + Salesforce Agentforce ate this category.

Agentic Software

Agentic Software — Basics and Investments

How to Read This Document

Executive Headlines (the 12 things that matter)

Part 1 — Agentic Software: Technical Basics

1. What Is an AI Agent?

2. Anatomy of an Agent

3. The Agent Loop

4. Single-Agent vs Multi-Agent Systems

5. Frameworks Landscape

6. Why Agents Break (Reliability)

7. Cost & Latency Reality Check

8. What Changed in 2024–2026

Sources

Part 2A — Horizontal Infrastructure + Foundation Layer

Framing for the Reader

A. Foundation Models — the “Brains”

Anthropic

OpenAI

Google DeepMind

xAI

Meta (open-weights anchor)

Mistral AI

Foundation-layer summary (as of May 2026)

B. Inference / Runtime Infrastructure

Together AI

Fireworks AI

Groq

Cerebras Systems

Modal Labs

Replicate

Lambda Labs

CoreWeave (public — CRWV)

Inference / runtime summary

C. Agent Runtime / Orchestration Frameworks

LangChain (incl. LangGraph, LangSmith)

CrewAI

Letta (formerly MemGPT)

Vellum

Humanloop

Orchestration / framework summary

D. Agent-Specific Infra (Browser, Sandbox, Tools)

Browserbase

E2B

Anthropic Computer Use (capability, not company)

Skyvern, Hyperbrowser, Steel.dev (browser agents — earlier-stage)

Agent-specific infra summary

E. Memory / Vector / Data Layer

Pinecone

Weaviate

Chroma

Turbopuffer

Mem0

Memory / vector / data layer summary

F. MCP Ecosystem Players

Composio

Smithery

Glama

Read-through on MCP ecosystem

MCP ecosystem summary

Cross-Layer Investor Read

Sources

Part 2B — Vertical Agent Application Companies

A. Coding / Developer Agents

Cursor (Anysphere)

Cognition (Devin)

Codeium / Windsurf

Replit

Magic.dev

Poolside

Augment Code

Coding agents — summary table

B. Customer Support / CX Agents

Sierra (Bret Taylor)

Decagon

Cresta

Ada

Forethought

Parloa

Customer support — summary table