The AI assistant market — business models, numbers, failures

As of 11 October 2026.

This file is the "business" part of the survey: who makes money on AI assistants, what customers pay for, how much it is all worth, who bought whom and what did not work out. I write it from a single perspective: I want to sell a local "on the desk" assistant to Polish professionals (lawyers, doctors, tax advisers, accountants) and I need to know which pricing models work, what the customer expects and where the competition has a gap.

⚠️ The numbers in this file go stale within weeks. Each one has a date and a source. (company data) = a claim by the company itself (blog, announcement, website), (press estimate) / (Sacra) = an estimate by journalists or analysts. ARR (annual recurring revenue) and run-rate are revenue from the latest period multiplied up to a year — not the same as annual revenue in a financial statement. I did not convert anything at an exchange rate.

Profiles of the top 15 assistants are in 01-PROFILES.md, technical patterns in 03-TECHNICAL-PATTERNS.md, and Polish law, prices and competition in 04-POLAND-AND-SOVEREIGNTY.md.

---

Table of contents

• The shortest answer
• 1. Market map: four kinds of assistants
• 2. Business models and pricing
• 3. Numbers: users, revenue, funding
• 4. Mergers, acquisitions and 2026 moves
• 5. Distribution and moats
• 6. Failures and lessons
• 7. 2026 trends
• 8. What this means for me
• Sources of this file

---

The shortest answer

• The biggest money is in professional assistants, not "general" ones. Programmers (Cursor, Claude Code, Cognition), lawyers (Harvey, Legora) and doctors (Abridge, OpenEvidence) pay the most and the fastest. According to Menlo Ventures, coding is 55% of departmental AI spending in companies (2025).
• A flat per-seat price breaks when an agent works for hours. The whole market is moving to credits and usage-based billing (Cursor 2025, GitHub Copilot 1 June 2026, Glean 2026), and the boldest to pay per outcome (Sierra).
• Open environments won the layer of "living" assistants. OpenClaw has a foundation, and Microsoft built Autopilot on it (25 September 2026). Nobody makes money on the code itself — money is made on cloud, models, hardware or support.
• The leaders fine-tune open-weight models (Harvey Tenet on Kimi K3, Cursor Composer 2 on Kimi K2.5, "Thomson" from Thomson Reuters). This confirms the "mid-sized open model + good retrieval + own data" strategy.
• No leader for professionals runs on the desk, offline. The market's privacy ceiling in 2026 is "hosting in the EU
  • no training". That is my niche.
• Polish price anchor: Beck-Noxtua PLN 590 net/month per user (plus Legalis), Noa Notes PLN 199 + VAT, LEX AI PLN 49–489/month. A professional already pays PLN 200–600 per month per seat in the cloud.

---

1. Market map: four kinds of assistants

Throughout the survey I divide assistants into four kinds. Each has a different business model and different winners.

Where the budgets flow — the best cross-sectional data I have is from 2025 (Menlo Ventures, 9 December 2025, (analyst estimate)):

• company spending on generative AI: USD 37 bn in 2025 (USD 11.5 bn in 2024, ×3.2);
• coding applications: USD 4.0 bn, i.e. 55% of departmental spending (up from USD 550M);
• vertical (industry) AI: USD 3.5 bn, of which healthcare about USD 1.5 bn;
• companies buy rather than build in 76% of use cases; 27% of application spending arrives through self-service (PLG).

In medicine: AI spending in healthcare USD 1.4 bn in 2025, of which "scribes" (automatic visit notes) USD 600M; shares: Nuance/Microsoft DAX about 33%, Abridge about 30%, Ambience about 13% (Menlo Ventures, 21 October 2025, (analyst estimate)).

Takeaway for me: money is made on narrow professional work that someone answers for with their signature. General chat is the giants' market; a professional buys a ready, checkable work product (a filing, a note, a summary).

---

2. Business models and pricing

Overview of models

Subscriptions: the 20 / 200 USD tiers

Almost all everyday assistants have converged on the same tiers:

Lesson: the expensive tier (USD 200–500) exists so that agents that work for hours have something to live on. Perplexity ties its revenue growth from about USD 450M to about USD 750M (Mar → Aug 2026) precisely to the USD 200 tier and the Computer agent (PYMNTS, Yahoo/The Information, (press estimate)).

Credits: the unit of "agent work"

A credit combines model tokens, virtual machine time and paid APIs into a single number (Manus, credit rules).

Good practices from credit price lists: Lindy describes credits by kinds of work and stops the agent when the pool runs out, instead of issuing a surprise bill; Fellou shows a usage estimate before a task is run; Manus gives daily "refreshing" credits that build a habit (Lindy, Fellou, Manus). Rabbit went the opposite way: "rabbit intern" at USD 99.99/month for 30 tasks, advertised as an alternative to "confusing credit systems" (Rabbit, 13 June 2025).

From seats to usage

• Cursor (Jun 2025): on 16 June 2025 Pro changed from 500 requests to USD 20 of credit at API prices + an "unlimited" Auto mode. Users did not know that "unlimited" applied only to Auto, and received unexpected bills; on 4 July 2025 Cursor apologised and refunded the money (Cursor).
• Claude Code: weekly limits announced on 28 July 2025, in force from 28 August 2025 — against 24/7 background work and reselling of accounts; according to Anthropic they affected <5% of subscribers (TechCrunch).
• GitHub Copilot: from 1 June 2026 AI credits (1 credit = USD 0.01) instead of "premium requests"; Pro USD 10 with credits worth 15 + 5 USD, Max USD 100 with credits worth 200 + 100 USD (GitHub docs, pricing).
• Glean: from pure seats to a flat fee for active users + model usage (28 May 2026, TechCrunch).
• Notion: AI in the higher seat tiers, and "Custom Agents" at USD 10 per 1,000 credits (Notion).

Why: Gartner predicts that the inference cost per agentic workflow will grow more than 5× by 2028, even though tokens are getting cheaper (17 August 2026, Gartner, (analyst forecast)). Agents simply use far more tokens per task.

Pay per outcome

• Sierra (since 10 December 2024): you pay for resolved conversations, saved cancellations, upsells; "if a conversation is not resolved, in most cases there is no charge" (Sierra). The 2026 slogan: "you pay for results, not tokens" (Sierra Horizon).
• Intercom Fin: USD 0.99 per outcome (Intercom, (company data)).
• Parloa: per minute of voice or per chat interaction, "per outcome where it makes sense" (Parloa).
• Ambience Healthcare: contracts with a bonus tied to the hospital's financial results (Apr 2026, Ambience, (company data)).

Sierra itself explains its growth by picking narrow "jobs to be done" with a defined measure of success (Sierra, 6 February 2026).

Free for the professional, someone else pays

OpenEvidence gives doctors a medical search engine for free and earns from pharma and medical-device ads (CPM USD 70–150). Revenue: about USD 7.9M (Dec 2024) → 50M (Jun 2025) → about USD 300M annualised (Jul–Sep 2026), gross margin about 90% (Sacra, (analyst estimate)). The moat is licensed content (NEJM, JAMA, NCCN), not the model (Wikipedia). Doximity GPT works the same way — free for doctors, financed by pharma marketing (Wikipedia).

Open source + cloud, foundation, BYOK

• OpenClaw sells nothing. Since 8 July 2026 it has been run by the OpenClaw Foundation (a US 501(c)(3), 12 people on staff); donors and partners include OpenAI, University of Michigan (the largest donor), NVIDIA, Microsoft, Red Hat, Tencent, Atlassian, Vercel, Cloudflare, GitHub; amounts were not disclosed (foundation, (foundation data)). OpenClaw Enterprise (29 September 2026) "will always be free" (OpenClaw).
• Hermes Agent is free (MIT); Nous Research earns on Nous Portal (models and credits), Hermes Business (shared pool, per-member limits) and Hermes Enterprise (on the customer's infrastructure) (Nous Portal).
• browser-use: an MIT library, earning comes from Browser Use Cloud — browser hours, proxies, token resale (browser-use).
• BYOK (bring your own key — your own key to the model): Manus Flex (29 September 2026) — the customer pays for the model directly, Manus still charges credits for machines and tools (Manus Flex); Kortix charges only for the sandbox (Kortix); OpenHands offers models "at cost, with no markup" (OpenHands).
• Exotica: Agent Zero finances development with a cryptocurrency token A0T (Agent Zero), and nanobot with affiliate links to model providers (nanobot).

Enterprise licences and content bundles

Elite legal tools do not publish prices:

• Harvey: an estimated USD 1,200 per seat per month for small deployments (minimum 20 seats) and USD 100–200 at Am Law 100 scale (Vaquill, (external estimate, unverified));
• Legora: about USD 3,000 per seat per year (minimum 10) (Layer3, (external estimate));
• Abridge: about USD 2,500 per doctor per year (Sacra, (analyst estimate));
• CoCounsel sold in a bundle with Westlaw and Practical Law (Thomson Reuters).

Tools for small firms publish prices: Clio Work USD 199 per user per month (LawSites), Blue J USD 1,498 per year "with no add-ons" (Blue J). Reference level: Microsoft 365 Copilot Business USD 18–21 per user per month (Microsoft).

---

3. Numbers: users, revenue, funding

Only numbers from my research files. "No data" means: I did not find a reliable number, not that none exists. GitHub stars from the API on 11 October 2026.

The top 15

The next ones — for comparison

How to read it:

• The fastest revenue ramps in the history of software are developer tools: Cursor from USD 100M (Jan 2025) to 2 bn ARR (Feb 2026; according to Bloomberg 3 bn, May 2026), Claude Code >2.5 bn run-rate after less than a year.
• Law is the second hottest vertical: Harvey from a USD 3 bn valuation (Feb 2025) to 15.5 bn (Sep 2026), Legora from about USD 1M to 100M ARR in 18 months (TNW).
• Open projects are measured by reach, not revenue. The numbers of "clones" and "claws" are claims — they cannot be verified, but stars and the number of security reports can.

---

4. Mergers, acquisitions and 2026 moves

Chronologically, from the end of 2025:

The Windsurf story (2025) as a cautionary tale: the acquisition by OpenAI fell through, Google paid USD 2.4 bn to hire the leaders and license the technology, and the rest of the company was signed over to Cognition's acquisition in Jul 2025 (Wikipedia). Windsurf brought Cognition over 1M users and 4,000+ organisations (secondary source) — distribution was bought.

The pattern: the big players do not fight open runtimes, they absorb them — they hire the creators (OpenAI), build products on their code (Microsoft), sell hardware for them (NVIDIA NemoClaw) or buy the teams (Cognition).

---

5. Distribution and moats

A moat is what protects a business from being copied. In 2026 a language model is not a moat — everyone has access to similar models. Here is what works.

Distribution through the OS, browser and messenger

• Gemini in Chrome and Android: Google has a browser with about 2/3 of the market, a password manager, Gmail and Calendar; the agent works where the user is already logged in. Auto browse reached Chrome on Android on 18 August 2026 (USA) (Google, 36Kr). All AI browsers together had <1% of the market (Aug 2026).
• Apple Messages: Poke was the first AI agent approved on Apple Messages for Business (4 June 2026) — "the only AI agent allowed to write natively in Apple Messages" (company data) (TechCrunch). Lindy got a permanent ban in the same channel (below).
• WhatsApp, Telegram, Signal and the rest: OpenClaw and Hermes have gateways to a dozen or so messengers at once — this is their main interface (OpenClaw docs).
• Microsoft 365 and Windows: 30M+ paid Copilot seats (FY2026) and Autopilot in Teams (Microsoft IR).
• EHR in medicine: Epic introduced its own AI Charting (Feb 2026), and hospitals are "re-evaluating the point of separate contracts" (HLTH). Whoever owns the documentation system owns the scribe customer.

Data and content

• Legal content: at LexisNexis AI products are 90% of new sales (23 July 2026, LexisNexis); Harvey has an alliance with LexisNexis, Noxtua belongs to C.H.Beck, Clio bought vLex.
• In Poland commentaries are controlled by Wolters Kluwer (LEX Expert AI) and C.H.Beck (Legalis → Beck-Noxtua) (prawo.pl). A local assistant will not beat them on content — it should complement their subscriptions by working on the client's own files (details in 04-POLAND-AND-SOVEREIGNTY.md).
• Medicine: OpenEvidence wins with NEJM, JAMA, NCCN licences, not with a model.

Habit and workflow

• The winners are those who deliver a ready, checkable work product (a filing, a note in the EHR, edits in Word), not an answer in a chat. Harvey has add-ins for Word and Outlook, Abridge writes the note straight into the EHR, Granola sits in the calendar.
• Memory per case and per user (Harvey, Hebbia, Sierra) — the longer a customer uses it, the harder it is to leave.

Own models (fine-tuned open-weight)

This is the most important market news for a local assistant: the leaders themselves show that a mid-sized open model, fine-tuned on domain tasks and plugged into good retrieval, is enough for many tasks — and is many times cheaper.

Trust and compliance

• All the leaders link quotes to sources (CoCounsel, Lexis with Shepard's, Abridge "Linked Evidence", LEX Expert AI).
• The same set of certifications everywhere: SOC 2 Type II, ISO 27001, increasingly ISO 42001; in Germany BSI C5.
• Security enforced outside the model: Harvey has an MCP Policy Engine blocking the "lethal trifecta" (access to data + untrusted content + an exit channel) (Harvey).
• In medicine in the EU the moat is medical device certification and public procurement, not privacy alone: Tandem has MDR class IIa (21 May 2026), Heidi won the NHS Midlands framework for 70,000 clinicians (TNW, NHS).

Network effects

• OpenClaw: ClawHub (a "skills" marketplace), 391.7k stars, hardware partners — an ecosystem attracts an ecosystem. But the same skills marketplace is also an attack surface (section 6).
• Poke: shared "recipes" (automations in plain text); the creator gets USD 0.10–1.00 per acquired user (TechCrunch).

---

6. Failures and lessons

I learn more from them than from successes — each one says what not to do in a product for a professional.

Common denominator: failures come from a promise bigger than the product (11x, Rabbit, Martin), bad defaults (OpenClaw, Granola, Replit) or dependence on someone else's platform (Lindy, Atlas, Mariner). For a professional who answers with their signature, each of these three is disqualifying.

---

7. 2026 trends

Each trend with evidence. Technical details — in 03-TECHNICAL-PATTERNS.md.

1. Agents in messengers

• Poke in Apple Messages (4 June 2026), OpenClaw and Hermes with gateways to WhatsApp, Telegram, Signal, Teams.
• Microsoft Autopilot in Teams (25 September 2026); Lindy moved from SMS to a "teammate" in Slack (31 August 2026, Lindy).
• What people actually delegate: "work around work", not life management. In Claude Cowork 33.4% is business operations, 16.4% content (Anthropic, Jul 2026); the morning brief is the universal entry point (ChatGPT Pulse, Google CC, Lindy).

2. Background and cloud agents

• ChatGPT dots (29 September 2026): always-on agents with their own computer in the cloud (TechCrunch).
• Manus 2.0 (28 September 2026): a persistent "Cloud Computer" and event-triggered automations (Manus).
• Gemini Spark: up to 15 tasks in parallel (Google).
• Cognition calls the Poke purchase a bet on "always-on agents in the cloud" (Cognition).
• In parallel the dedicated box is coming back: Cloudflare wrote about a "wave of Mac mini buying" for OpenClaw (Jan 2026), and Ghost sells a local AI computer for USD 3,499 (5 October 2026) (Cloudflare, TechCrunch).

3. Fine-tuning open-weight models

Harvey Tenet, Cursor Composer 2, Thomson Reuters "Thomson", Noxtua Athena, Canva (section 5). In addition, open GUI models: Qwen3.8-27B (Apache 2.0) claims 84.3% on OSWorld-Verified (Aug 2026, Hugging Face, (company data)). The model layer is becoming a commodity.

4. Sovereign stacks in the EU

• Money: Mistral EUR 3 bn (Sep 2026), Cohere + Aleph Alpha (Apr 2026), Noxtua EUR 100M+ (Sep 2026), Tandem and Lovable with the EU Scaleup Europe Fund.
• Policy: the technological sovereignty package (3 June 2026) and the draft CADA with four levels of sovereignty — from L1 (data in the EU) to L4 (full control of the supply chain) (European Commission). CADA is still a draft (Oct 2026).
• The limit: an "EU region" at an American provider is L1, not sovereignty; Microsoft's EU Data Boundary does not cover Anthropic models (Microsoft).
• At the same time some agentic features launch without the EEA: Gemini Spark "globally except the EEA, Nigeria, Switzerland and the United Kingdom", dots in the Pro tier also without the EEA (Google, Emergent).

5. Pay per outcome and per usage

Evidence in section 2 (Sierra, Intercom Fin, Parloa, Ambience; Copilot, Cursor, Glean, Notion) and Gartner's forecast (cost per agentic workflow ×5 by 2028). The cost risk falls on the customer or the vendor — unless the model is local.

6. AI browsers in retreat

• Atlas closed (9 Aug 2026), Mariner closed (4 May 2026), all AI browsers <1% of the market.
• Gartner (8 Dec 2025): "Cybersecurity Must Block AI Browsers for Now" — the defaults put convenience above security (The Register).
• OpenAI (22 Dec 2025): prompt injection "is unlikely to ever be fully solved" (TechCrunch).
• Akamai buys LayerX for about USD 205M (14 May 2026) — AI browser security is becoming a separate market (BankInfoSecurity).

7. Consolidation and platform resistance

• Consolidation: section 4 (Cursor, Poke, Windsurf, Aleph Alpha, vLex, Vercept, Moltbook).
• Platforms defend themselves: Amazon sued Perplexity over shopping in Comet — on 4 August 2026 the Ninth Circuit vacated the injunction, holding that the site is "visited" by the user, not by the agent's vendor (Jones Day); WeChat, Taobao and Alipay blocked the Doubao phone agent (Dec 2025) (Business Standard).
• Takeaway: APIs, MCP connectors and the user's own session instead of scraping pages.

8. The big players move into verticals

Anthropic extended "Claude for Legal" with domain plugins and connectors to Westlaw, Box and Docusign (12 May 2026, TechCrunch); Thomson Reuters is building the new CoCounsel on the Claude Agent SDK (LawSites). Verticals defend themselves with content, workflow and integrations.

9. The European gap and weak retention of consumer agents

• Europe gets new agentic features later or not at all. As of 11 October 2026 Gemini Spark is not available in the EEA, Switzerland and the United Kingdom (press); ChatGPT Health (launched 7 January 2026) and the first "dot" in the Pro plan also exclude the EEA (Wikipedia, blog); Siri AI (iOS 27, 14 September 2026) is not available on iPhones in the EU because of the DMA dispute (heise); Alexa+ is not offered in Poland (Amazon, the country is not on the list). In Poland ChatGPT Free and Go has had contextual ads since 24 August 2026. This is a gap a local product can enter.
• A background agent does not mean a habit. Meta's Muse (launched 8 September 2026, its own cloud virtual machine per user) had 5 million downloads in the US in 22 days, but only about 700k daily users at the end of Sep 2026 (CryptoBriefing, press); available only in the US and Canada (Meta). Lesson: the mere ability to work in the background does not retain the user — a concrete, repeatable benefit counts, like a visit note every day.
• Muse Glimmer (10 August 2026, Apache 2.0, about 30 bn parameters, 4-bit on a 24 GB GPU) is Meta's open agentic model; its quality in Polish is unverified (TechCrunch, VentureBeat).

---

8. What this means for me

Purpose: a local assistant on the desk for a Polish professional (law firm, medical practice, accounting office), data does not leave the room. The legal background (professional secrecy, GDPR, the AI Act, the Polish AI systems act) is in 04-POLAND-AND-SOVEREIGNTY.md — this is not legal advice.

Price anchors in Poland (checked 11 October 2026)

Takeaway: a professional already pays about PLN 200–600 per month per seat for cloud AI. Gaius-Lex is the closest Polish counterpart of "data does not leave the office", but it targets firms of 20–99 lawyers and project-based deployments. The segment of a small office of 1–10 people, a ready box for the desk — I found nobody in it (this is a search result, not proof that nobody is there).

Unit economics: local vs cloud

Cloud — cost grows with the agent's work:

• Manus: an input-to-output token ratio of about 100:1, on average about 50 tool calls per task; cached tokens cost about 1/10 of regular ones (Manus).
• Claude Opus 5.5: USD 4 per million input tokens, USD 20 per million output tokens (22 September 2026, Anthropic).
• My calculation (example): a task with 50 steps of 20k input tokens and 500 output tokens is 1M input tokens (USD 4) + 25k output (USD 0.50) = about USD 4.5 without cache, about USD 0.9 with a cache cutting input about 10× (as in Manus's example). A dozen or so such tasks a day per person is hundreds of USD per month (my calculation).
• Long frontier tasks on OSWorld 2.0: USD 15.70–26.62 per task (28 August 2026, Simular, (company data)).

Local — the cost is upfront, the marginal cost almost zero (electricity and depreciation):

• My machine: 4× Radeon AI PRO R9700, about PLN 8,900 per card (Ceneo, 11 October 2026) and 300 W per card (AMD); that is about PLN 35.6k for the cards alone and up to 1.2 kW under full load (my calculation).
• Throughput: 4 agents × 13.6 tok/s = about 54 tok/s of total generation (my calculation). The task mentioned above with 25k output tokens is about 30 minutes of generation alone on a single slot — slower than the cloud, but with no token bill and no limits.
• Boxes for customers: ASUS GX10 (GB10, 128 GB) from PLN 24,499 (Ceneo), Mac mini M5 Pro from PLN 8,499 (Apple), Strix Halo 128 GB PLN 13,490–19,999 (Ceneo) — prices from 11 October 2026, comparison in 04-POLAND-AND-SOVEREIGNTY.md.

Comparison example (list prices net, no discounts, excl. VAT — illustration only):

• 3 lawyers on Beck-Noxtua: 590 × 3 × 24 months = PLN 42,480 over two years, plus Legalis. One GX10 box is PLN 24,499 plus a support subscription.
• One doctor on Noa Notes: 199 × 24 = PLN 4,776 over two years — less than the Mac mini M5 Pro alone. For a single doctor the argument has to be privacy, not price; the economics work with a shared box in a clinic (3+ doctors).

A hybrid with Claude Opus 5.5 as the planner changes the calculation only a little: one planner call with 10k input tokens and 2k output is about USD 0.08 (my calculation). But then the planner sees the data — in professional secrecy mode the cloud planner may receive only an anonymised task description or nothing at all. I describe the privacy boundary in course 1, part 10 and course 2, part 17.

Pricing model — my hypothesis to test

This is my opinion based on the facts above, not market data:

1. A flat licence per seat or per office + hardware, not credits. With a marginal cost close to zero, "no meters, no token limits" is a selling point against usage-based price lists (section 2). Beck-Noxtua already advertises "no token limit" — the customer understands it.
2. Hardware sold at cost or leased, margin on the subscription (model updates, security patches, box replacement, quality tests in Polish before every model change). The pattern: Ghost ("works even if the vendor disappears") + Hermes Business (a paid layer around a free core).
3. A subscription price below the cloud anchors per seat over 24–36 months, including hardware — then the comparison with Beck-Noxtua (PLN 590) wins on paper, and privacy is a bonus.
4. Do not replace publishers, complement them: LEX and Legalis stay with the customer; my assistant works on their files, emails, KSeF and documentation, to which cloud players have no access (or should not have).
5. A compliance package as part of the price: a mapping of § 23e of the Code of Ethics of Advocates (NRA, 12–13 June 2026), a DPIA template, a KRiBSI opinion under art. 12 of the Polish AI systems act — details in 04-POLAND-AND-SOVEREIGNTY.md.
6. A "pay per outcome" option for later: e.g. per accepted visit note or prepared filing — only once I can measure it reliably (Sierra's lesson: a narrow task, a clear measure).
7. None of the sins from section 6: no "will replace the trainee" promises, sandbox and encryption on by default, recording only with explicit consent, human consent before every irreversible action, a channel I control.

Demand signals and market risks

• Poland is at the tail of the EU: 8.7% of Polish companies use AI (GUS, 2025) against about 20% in the EU (Eurostat, 2025), but in professional services 23.2% and +8.8 p.p. in a year (GUS, Eurostat).
• The barriers in the EU are "unclear legal consequences" (52.52%) and "data protection" (48.83%) — exactly what a local box solves (Eurostat, as above).
• Risk: 77% of Polish companies without AI will not adopt it "until they have to", every fifth points to cost (PIE, 24 September 2025).
• Risk: Polish programmers are among the most sceptical — only 26% trust AI (Stack Overflow 2025, Stack Overflow). Trust is built with citations, a review step and undo.
• I lack data on the willingness to pay of Polish advocates, legal advisers and doctors — I found no Polish studies. This is the first thing to check in conversations with customers.

Ranking and profiles — README.md and 01-PROFILES.md. A survey of open harnesses on which a core can be built — ../agent-harnesses/README.md.

---

Sources of this file

Every number has a link next to it; analyst estimates (Menlo Ventures, Gartner, Sacra) are marked in the text, and the price lists I checked on 11 October 2026. The full list of sources with a description of what I took from each — SOURCES.md.