CLOUDFLARE FORGE

At the center of Forge's economics sits a proprietary AI model optimized for code generation and Cloudflare-native deployment patterns. This model handles the majority of Standard Mode inference at a cost of $0.003 per credit versus $0.010 to $0.050 for equivalent calls through third-party APIs.

A Standard Mode credit maps to approximately 2,000 input tokens and 4,000 output tokens of model inference — roughly equivalent to one code generation prompt with full project context. At $0.003 per credit, the proprietary model operates at approximately $0.50 per million input tokens and $0.50 per million output tokens — comparable to DeepSeek V3.2 API pricing and Gemini Flash, but running on owned infrastructure with zero per-call API overhead.

The Forge Model is not intended to be a frontier model. It needs to be excellent at one thing: generating production-ready code that runs on Cloudflare infrastructure. A Mixture-of-Experts architecture — following the approach validated by Cursor's Composer 2 and DeepSeek V3 — allows the model to achieve high quality on code-specific tasks while maintaining efficient inference. Training begins from an open-source base, likely Qwen-Coder or DeepSeek-Coder, with continued pretraining on code corpora and reinforcement learning fine-tuning using deployment success signals in sandboxed Cloudflare environments.

The model investment is $15.5 million across all three phases: $0.5 million for fine-tuning Qwen2.5-Coder-32B on Cloudflare- specific patterns using Workers AI LoRA adapters in Phase 1, $10 million for dedicated training runs on GPU compute via the Replicate acquisition in Phase 2, and $5 million for production optimization and v2 training in Phase 3. For context, DeepSeek V3 — a 671-billion-parameter MoE model that achieved frontier performance — was trained for a reported $5.6 million in GPU compute. A competitive coding-specialized model is realistically achievable for $5 to $15 million in raw compute. The $15.5 million budget covers fine-tuning, training, evaluation infrastructure, and production serving optimization. By Year 5, the model handles 75 percent of Standard Mode queries and could save an estimated $168 million per year in API costs if utilization targets are met. The modeled cost gap between third-party APIs ($0.010 per credit) and the proprietary model ($0.003 per credit) makes the $15.5 million model investment potentially the highest-leverage component of the Forge budget — but this depends on training success, model quality, and GPU utilization rates.

The 10 to 50x price spread between efficient models (Flash, DeepSeek) and frontier models (Opus, GPT-5.4) is what makes the Standard/Power Mode split economically rational. Forge routes routine code generation to the cheap end and complex reasoning to the expensive end — with the user controlling the toggle and seeing the cost before every prompt.

BYOK: Bring Your Own Key. Available at Level 4 and above, BYOK allows users to provide their own Anthropic, OpenAI, or Google API keys. BYOK interactions consume zero Forge credits. Forge adds value through prompt caching via AI Gateway, which can significantly reduce API costs for repeated requests (savings proportional to cache hit rate). Cursor restricted BYOK access in late 2025, gating it by plan and feature. Forge takes the opposite approach — embracing BYOK as a feature, not a threat. The open-source BYOK ecosystem — Cline with 5 million-plus VS Code installs, Kilo Code, and JetBrains AI Assistant — proves that a significant developer segment prefers controlling their own model routing.

MCP integration and Headless Mode. Forge operates as both an MCP client and an MCP server. Cloudflare already has 13+ product-specific MCP servers plus the unified API server. The innovative MCP Code Mode technique reduces the apparent API surface to just two tools — deploy and query — which enables all 2,500 Cloudflare API endpoints to be accessible through a streamlined interface. Headless Mode is the primary driver of the MCP deployment capture revenue stream — projected at $175 million by Year 5. When a Cursor developer finishes building an application and runs a deploy command through the MCP server, Forge provisions the entire infrastructure automatically. The developer never opens Forge. Near-zero customer acquisition cost.

Background Agents are long-running AI agents that operate autonomously, triggered by schedules, Slack messages, GitHub events, PagerDuty alerts, or custom webhooks. Each agent runs on a Durable Object that hibernates when idle — costing fractions of a cent per hour versus Cursor's full Ubuntu VMs at cloud compute rates. Agents accumulate learned behavior in persistent state stored in Durable Objects' embedded SQLite, which does not transfer to other platforms.

Use cases include automated code review (comparable to Cursor's Bugbot at $40 per user per month, but running on Durable Objects at a fraction of the infrastructure cost), continuous integration with AI-powered test generation, documentation maintenance, dependency management, security patching, and customer support triage. The Workflows infrastructure provides durable execution with automatic retries, and the Agents SDK provides stateful agent runtime with tool orchestration. Containers add full Linux sandbox support for workloads requiring Python ML libraries, headless Chrome, or custom runtimes.

Forge auto-generates configuration files that train every AI coding tool to produce Cloudflare-native code. The .cursorrules file configures Cursor to generate Workers-compatible code, use D1 for databases, and deploy via Wrangler. The CLAUDE.md file pre-configures Claude Code with Cloudflare deployment context — when a Claude Code user runs wrangler deploy, their application goes live on Workers, D1, and R2 automatically. The AGENTS.md file configures agent-based tools with Forge-specific deployment patterns.

Over time, an organization's codebase becomes idiomatically Cloudflare. These files create a subtle but powerful form of switching cost — every AI tool the developer uses is tuned to produce code optimized for Cloudflare infrastructure. Migrating away means retraining not just the developer's habits but every AI assistant's context.

Every major AI coding platform has faced user backlash over pricing. Forge's Predictable Pricing Guarantee addresses this with five mechanisms: a hard monthly spending cap defaulting to twice the plan price; a real-time credit usage dashboard; threshold notifications at 50%, 80%, and 100% of the spending cap; a cost preview before each AI generation; and optional daily credit pacing. No user will ever be charged beyond their spending cap without explicit opt-in.

Forge's visual-code sync separates it from both IDE-only tools like Cursor and visual-only tools like Squarespace. The architecture leverages Cloudflare's Browser Rendering API for full Chrome DevTools Protocol access. The sync engine operates in three stages: source-to-visual rendering via tree-sitter AST parsing, visual-to-source propagation when users modify elements on the canvas, and source-to-visual propagation when users edit code. The pipeline runs in under 100 milliseconds per change.

Forge's five-mode, five-level architecture supports a taxonomy of 18 use case categories, each representing a distinct type of software that Forge can generate, deploy, and run on Cloudflare infrastructure. Each category represents a market where Forge offers a structurally advantaged alternative to existing solutions.

Category 1: Static websites. Personal portfolios, landing pages, documentation sites, blogs. Deployed on Cloudflare Pages with R2 for assets. Competes with Squarespace ($16 to $65 per month), Wix ($17 to $159 per month), and Framer ($10 to $35 per month). Forge advantage: AI generation plus zero-egress hosting at dramatically lower cost.

Category 2: Dynamic web applications. SaaS dashboards, project management tools, CRMs, internal tools. Workers for backend logic, D1 for databases, R2 for file storage. Competes with Retool ($10 to $50 per user per month), Bubble ($29 to $549 per month), and custom development. Forge advantage: full-stack generation with integrated database and hosting.

Category 3: E-commerce storefronts. Online stores with product catalogs, shopping carts, and payment processing. Workers for API logic, D1 for product and order data, R2 for product images, Stripe integration via MCP. Competes with Shopify ($39 to $2,000 per month) and Squarespace Commerce ($23 to $65 per month). Forge advantage: zero-commission infrastructure versus Shopify's 0.5 to 2 percent transaction fees on non-Shopify Payments.

Category 4: API backends. REST and GraphQL APIs for mobile apps, third-party integrations, and microservices. Workers for endpoint logic, D1 for data, Queues for async processing. Competes with AWS API Gateway ($3.50 per million requests) and Supabase ($25 to $599 per month). Forge advantage: AI-generated API code with built-in rate limiting, authentication, and DDoS protection.

Category 5: Real-time applications. Chat systems, collaboration tools, live dashboards. Durable Objects for real-time state with WebSocket support, Workers for routing. Forge advantage: Durable Objects provide strongly consistent, single-threaded state management that eliminates distributed systems complexity.

Category 6: AI-powered applications. Chatbots, content generators, recommendation engines, semantic search. Workers AI for inference, Vectorize for embeddings, AI Gateway for model routing, D1 for conversation history. Forge advantage: inference runs on the same network as the application, eliminating latency and egress costs between the AI and the backend.

Category 7: Content management systems. Blogs, news sites, knowledge bases. Workers for server-side rendering, D1 for content storage, R2 for media, KV for caching. The Astro acquisition in January 2026 provides the industry's best content-focused web framework natively. Forge advantage: AI-generated CMS with Astro's optimized rendering, deployed globally with zero-configuration CDN.

Category 8: Forms and surveys. Contact forms, feedback collection, registration systems. Workers for form processing, Turnstile for bot protection, Email Routing for notifications. Competes with Typeform ($25 to $83 per month) and JotForm ($34 to $99 per month). Forge advantage: enterprise-grade bot protection via Turnstile included at no extra cost.

Category 9: Scheduling and booking. Appointment booking, reservation systems, event registration. Forge advantage: AI generates the complete booking system — calendar UI, availability engine, confirmation emails, payment processing — in a single conversation.

Category 10: Data dashboards. Business intelligence dashboards, reporting portals, data visualization tools. Forge advantage: dashboards deployed on the same edge network as the data sources, eliminating query latency.

Category 11: Marketplaces. Two-sided platforms connecting buyers and sellers. Stripe Connect integration via MCP. Competes with Sharetribe ($79 to $599 per month) and custom marketplace development ($50,000 to $500,000). Forge advantage: AI generates the complete marketplace including search, payments, reviews, and dispute resolution.

Category 12: Social platforms. Community forums, social networks, user-generated content platforms. Durable Objects provide the real-time infrastructure that makes social features responsive without custom WebSocket server management.

Category 13: Mobile app backends. API backends for iOS and Android, push notification infrastructure, offline sync. Forge advantage: zero-egress API responses via Workers at the edge, reducing mobile app latency globally.

Category 14: IoT dashboards. Data collection from hardware devices, sensor visualization, alert systems. Workers for data ingestion at the edge, D1 for time-series storage, Durable Objects for device state management.

Category 15: Developer tools and documentation. API documentation sites, developer portals, SDK documentation. The Astro framework provides content-optimized static generation deployed on the fastest CDN in the market.

Category 16: AI agents and automation. Custom AI agents for customer support triage, content moderation, data entry automation. Agents SDK for orchestration, Durable Objects for state, Workers AI for inference. This is both a use case and a product line through the AI Agent Marketplace.

Category 17: MCP servers. Custom integration endpoints that expose business-specific APIs as MCP tools for any AI coding assistant. Forge enables non-developers to create MCP servers for their company's internal tools. This is both a use case and a product line through the MCP Server Marketplace.

Category 18: Background automation. Always-on monitoring, automated code review, scheduled data processing, periodic report generation. Workflows for durable execution, Agents SDK for orchestration, Durable Objects for state persistence with hibernation. This is both a use case and a product line through the Background Agent Infrastructure offering.

The common thread: Forge can generate, deploy, and run applications in every category on infrastructure Cloudflare already owns, at margins that dedicated competitors — each of whom rents their infrastructure — cannot match.

Workers have no general limit on requests per second. D1 supports up to 50,000 databases per account on Workers Paid, each up to 10 GB, with Hyperdrive providing a connection-pooling fallback to external PostgreSQL for applications that outgrow D1 limits. R2 offers unlimited storage with zero egress. Durable Objects handle approximately 1,000 requests per second per object with unlimited objects. Approximately 85 percent of all web application types are fully buildable on Cloudflare's infrastructure, with the remaining 15 percent limited by three specific gaps: frontier AI inference requiring custom GPU hosting, real-time physics simulation at game-engine scale, and regulated financial systems requiring co-located exchange connectivity.

Forge's projected 79 percent gross margin in Year 5 stands out against competitors because it owns its infrastructure. Cursor reportedly runs at 55 to 65 percent gross margin, constrained by near-100 percent revenue spend on inference costs from Anthropic and OpenAI. Cursor's current business model depends on third-party inference, and any path to infrastructure margin would require significant strategic repositioning — either partnering with a hyperscaler for revenue-share terms or acquiring an infrastructure company at a valuation that would substantially dilute current shareholders. Its recent proprietary Composer models reduce inference costs but do not change this structural constraint. Replit operates at approximately 36 percent gross margin, paying Google Cloud retail prices for compute, storage, and bandwidth. GCP dependency means Replit's margins compress as usage grows. Lovable achieves 60 to 70 percent gross margin through a lighter-weight product surface, but Supabase dependency means that as Lovable users build more complex applications, a growing share of the economics flows to Supabase (and ultimately to AWS). Bolt.new operates at an estimated 50 to 60 percent gross margin, constrained by aggressive AI token consumption and WebContainers overhead.

Forge achieves structurally higher margins because its three largest cost categories — compute, storage, and bandwidth — run on infrastructure that is already amortized by Cloudflare's core security and performance business. Competitors cannot replicate this structure without building their own global network.

Forge Model fails entirely (15 percent probability). Revenue is unchanged — the model does not generate revenue directly. The impact is margin compression from 79 percent to 74 percent, as all Standard Mode inference runs through third-party APIs at $0.010 to $0.015 per credit instead of $0.003. Year 5 gross profit drops from $908 million to $851 million. The platform is still profitable. Payback extends by 12 months. The write-off is $15.5 million in model investment.

MCP capture underperforms 50 percent (25 percent probability). MCP deployment capture drops from $175 million to $88 million in Year 5. This is the most likely downside scenario because MCP adoption among Cursor and Claude Code users is uncertain. But every $1 of lost MCP capture is replaced by $0.40 to $0.60 of direct Forge revenue, because users who do not deploy through MCP may still adopt Forge for end-to-end development.

Market growth decelerates to 20 percent (20 percent probability). Total addressable market in Year 5 drops from $44 billion to approximately $30 billion. Forge's revenue falls to $1.1 billion. This scenario is unlikely given current growth rates — every indicator shows acceleration, not deceleration. But even at $1.1 billion, Forge operates profitably.

Cursor builds deployment infrastructure (10 percent probability). Cursor enters Forge's deployment moat. At $2 billion-plus in revenue, Cursor could theoretically invest $200 to $500 million. But building a global edge network is a 5- to 10-year project. More likely: Cursor partners with a cloud provider, which sacrifices the margin advantage that Forge gets from owned infrastructure. Forge's revenue impact: minus $150 to $250 million in Year 5, primarily from reduced MCP capture.

Aggressive free-tier abuse (15 percent probability). Bad actors use the free tier for crypto mining, spam, or bot deployment. No revenue impact, but infrastructure cost increases by $58 million annually. Mitigation: Turnstile bot detection, IP reputation scoring, and the same abuse detection systems that protect Cloudflare's existing free tier. Cloudflare has operated a generous free tier since 2010. The company's abuse detection is industry-leading.

Forge's most powerful economic mechanism is the self-funding infrastructure loop. Every application built on Forge — regardless of whether it was built with a free account or an Enterprise subscription — runs on Cloudflare infrastructure that generates consumption-based revenue. As the application grows, infrastructure revenue grows with it, at zero additional customer acquisition cost.

Target dollar-based net revenue retention of 125 percent at steady state, with 135 percent achievable under strong expansion scenarios — a range bracketing top-tier SaaS benchmarks (Snowflake at scale: 120-130 percent). This expansion is driven primarily by the mechanism described above. Users do not need to upgrade their subscription tier for Forge's revenue per user to grow — their applications simply need to succeed. As an application gains users, it consumes more compute, more storage, more bandwidth. Each of these is billed consumption-based on Cloudflare infrastructure. A 10-point swing in DBNRR moves Year 5 revenue by approximately the same percentage, so this assumption materially drives the headline revenue forecast.

What Forge specifically learns from Cursor: the .cursorrules file format for project-specific AI customization, the power of codebase-aware context management using semantic indexing and @ references, the value of tab completion with next-edit prediction rather than just cursor-position completion, and the importance of shipping at extreme velocity — Cursor released Composer 2, Bugbot Autofix, Automations, and Cursor 3.0 within months of each other.

What Forge exploits: Cursor's lack of deployment infrastructure, its pricing trust damage, its API dependency creating margin pressure, and its inability to offer visual editing for non-developers. Cursor serves Levels 3 through 5 brilliantly and has no answer for Levels 1 through 2. A Cursor user who finishes building an application faces a fragmented deployment process involving Vercel or AWS or Railway or Render or Fly.io — each requiring separate billing, authentication, and infrastructure configuration. Forge captures this moment through MCP integration. When a Cursor user types "deploy this," Forge provisions Workers, D1, R2, and DNS automatically. The developer never opens Forge. This is the deployment capture strategy: Forge makes Cursor better by making deployment invisible, and in doing so captures the infrastructure revenue from every Cursor user's application.

Shopify's $11.56 billion in fiscal 2025 revenue is built not on being the best website builder but on being the commerce infrastructure layer. Merchant solutions — Shopify Payments (processing 64 percent of GMV), Shopify Capital ($4.2 billion originated in 2025), and Shopify Fulfillment — account for 73 percent of total revenue. Forge's model mirrors the architecture directionally: the IDE is the storefront, the infrastructure is where margin concentrates.

The analogy has structural limits that should be acknowledged. Shopify retains merchants through transaction-based revenue (a 1–2 percent take rate on GMV) and financial products that create self-reinforcing dependency. Forge has no equivalent transaction layer. Shopify uses proprietary Liquid templating that makes migration technically prohibitive; Forge generates standard React and TypeScript that deploys anywhere. Forge's lock-in is infrastructure gravity — deployed applications, databases, stored objects, custom domains — not code portability barriers. That distinction matters: infrastructure gravity compounds over time as more apps deploy, but it is fundamentally weaker than Shopify's transaction-based retention in the early years before a critical mass of deployed applications accumulates.

Forge's MCP server infrastructure allows developers to deploy across AI coding tools. The MCP Server Marketplace is Forge's equivalent of the Shopify App Store — a third-party ecosystem that generates commission revenue and deepens platform stickiness.

Though not direct competitors for the AI coding market, traditional website builders define the market for Level 1 through 2 Forge users. Their combined revenue exceeds $5 billion — a market Forge partially addresses with AI-generated, portable code on zero-egress infrastructure.

Wix reported $1.99 billion in revenue for fiscal 2025, up 13.2 percent, with 250 million-plus registered users and $573 million in free cash flow. Wix adapted aggressively to AI disruption, launching Wix Harmony (traditional editing combined with vibe-coding) and acquiring Base44, a no-code app platform that reached $100 million in ARR roughly one year after founding. Forge advantage: Wix's AI generates content within its proprietary system; Forge generates standard React and TypeScript that runs anywhere.

Squarespace generated approximately $1.19 billion in revenue before being privatized by Permira at $7.2 billion in October 2024. Privatization gives Squarespace freedom to invest aggressively without quarterly earnings pressure. Forge competes by offering AI generation that produces real, portable code rather than a proprietary template system.

Webflow generated approximately $213 million in revenue in 2024, growing 66 percent year-over-year, with a $4 billion valuation. Its MCP Server now enables AI agents in tools like Cursor to interact directly with Webflow Designer — a signal that even traditional builders recognize MCP as the integration standard. Forge competes at Levels 3 through 4 with bidirectional visual-code sync plus full backend capabilities Webflow lacks.

Framer produces the best design-quality AI-generated websites in the market, starting at just $10 per month. Framer's weakness is depth: no e-commerce, limited CMS, and no backend capability. Forge competes at Level 1 through 2 on design quality while offering the full-stack depth that Framer cannot match.

June 2025: Cursor switches from fixed request counts to credit-based billing. Users report surprise charges of $200 to $350 per week. CEO issues a public apology and announces mass refunds.

July 2025: Windsurf acquisition by OpenAI falls through after exclusivity expired.^[8] Existing users face uncertainty.

October 2025: Bolt.new users report that the Attempt Fix button consumes 3 to 8 million tokens on authentication bugs — costing $15 to $40 for a single error fix attempt.

November 2025: Replit user documents $30-per-hour costs during an Agent session. The thread goes viral on Hacker News.

July 11–14, 2025: OpenAI's exclusivity on Windsurf expires. Google acqui-hires the leadership team into DeepMind for a reported $2.4 billion the same day. Cognition acquires remaining assets by July 14.

January 2026: Firebase Studio shows signs of declining investment. Feature updates slow. Community channels report reduced support responsiveness.

March 2026: Google officially announces the sunset of Firebase Studio on March 19, 2026, with full shutdown scheduled for March 22, 2027. Developers given a one-year migration window.

April 2026: Claude Code introduces pricing changes that block third-party frameworks from using Max subscription limits.

Phase 1 ($35M): Engineering: $12M (~60 engineers, 9 months at ~$27K/month fully loaded). Model fine-tuning: $0.5M (Qwen2.5-Coder LoRA on Workers AI). Infrastructure and tooling: $5M (dev environments, CI/CD, internal Workers/D1/R2 at cost). Go-to-market: $8M (developer relations, launch marketing, enterprise pilot program). Contingency: $9.5M.

Phase 2 ($42.5M): Engineering: $16.5M (~90 engineers). Proprietary model training: $10M (GPU compute via Replicate acquisition). Infrastructure scale: $4M. Go-to-market: $8M (beta marketing, enterprise sales ramp, launch events). Contingency: $4M.

Phase 3 ($30M): Engineering: $12M (~110 engineers). Model optimization and v2 training: $5M. Infrastructure and ops: $3M. Go-to-market: $8M (GA launch, enterprise expansion, international). Contingency: $2M.

Forge's partnership strategy centers on the MCP Alliance — a deliberate ecosystem play that positions Cloudflare as the infrastructure layer for every AI coding tool, not just Forge.

Anthropic partnership: Forge generates CLAUDE.md configuration files for every project. Claude Code users deploying through Cloudflare's MCP server create a natural pipeline from Anthropic's most engaged developers to Cloudflare's infrastructure. The relationship is symbiotic: Anthropic wants Claude Code users deploying real applications (it validates the product), and Cloudflare captures the infrastructure revenue from those deployments.

Cursor partnership: Forge operates as a deployment destination for Cursor through MCP integration. Cursor benefits because seamless deployment improves their product experience. Forge benefits because it captures infrastructure revenue from Cursor's millions of monthly active users without competing for IDE market share.

Open-source ecosystem: Forge provides first-class support for Cline, Aider, OpenHands, and other open-source AI coding tools through MCP servers and configuration file generation. The open-source community's preference for BYOK aligns with Forge's BYOK strategy — these users generate zero AI credit revenue but full infrastructure revenue when their applications deploy on Cloudflare.

Forge requires specific executive and technical talent:

VP/GM of Forge: A product leader with experience shipping developer platforms at scale. Ideal profile: former VP Product at a developer tools company (GitHub, Vercel, Netlify, Supabase) with both PLG and enterprise GTM experience.

Head of AI/ML: A technical leader capable of directing the proprietary model program from research through production. Must have experience training and deploying large language models for code generation. The Replicate acquisition (November 2025) provides a nucleus of ML infrastructure talent.

Head of Developer Relations: Forge's free-tier and MCP capture strategy depends on developer community trust. This role bridges engineering and marketing — building credibility with the developer audience that drives bottom-up adoption.

Enterprise Sales Lead: Phase 3 enterprise sales require someone who has sold developer infrastructure to Fortune 500 companies. The target: Cloudflare's existing 4,298 large customers paying over $100,000 annually.

The broader hiring strategy prioritizes engineers with experience on VS Code, CodeMirror, tree-sitter AST parsing, and language server protocols for the editor team; experience with Durable Objects, Workers, and D1 for the platform team; and experience with model training pipelines for the AI team. Cloudflare's existing engineering organization of 1,500-plus provides an internal talent pipeline.

Cloudflare already trades at a significant premium to infrastructure peers — approximately 30x trailing revenue versus 20x for CrowdStrike, 12.5x for Snowflake, and 10x for both Datadog and MongoDB. AI coding companies command even higher multiples: Replit at approximately 60x ARR and Cursor at 15x (with reports of a $50–60 billion round that would imply 25–30x). The market pays extraordinary premiums for companies that combine infrastructure stickiness with AI-driven growth acceleration — precisely what Forge would deliver.

Templates serve beginner and intermediate users with one-time purchases ($5 to $50). AI agents serve professional developers with subscription-based recurring revenue ($5 to $100 per month). MCP servers serve the entire AI coding ecosystem — including Cursor, Claude Code, and Copilot users who may never open Forge.

By separating the three marketplaces, Forge allows each to develop its own velocity, its own community, and its own economics. The MCP Server Marketplace creates the most durable lock-in: a three-sided network effect between server developers, external tool users, and Forge users. Forge positions its marketplaces as the "App Store" of AI-built software — curated, trusted, and worth paying for.

Each marketplace enforces automated quality gates before items go live. Templates must pass a build verification (no errors, loads in under 3 seconds on simulated 3G), accessibility audit (WCAG 2.1 AA minimum), and security scan (no hardcoded secrets, no known vulnerabilities in dependencies). AI agents must pass a sandbox execution test (completes primary task in under 60 seconds), a resource consumption audit (no runaway compute), and a data privacy review (no unauthorized data exfiltration). MCP servers must pass an API specification validation, a latency benchmark (p99 under 2 seconds), and an OAuth flow verification.

Submissions that pass automated gates enter a 48-hour review queue staffed by a curation team that evaluates utility, differentiation, and documentation quality. Rejection rate is expected to be 30 to 40 percent initially, declining to 15 to 20 percent as creators learn the requirements. High rejection rates are a feature, not a bug — they signal quality and differentiate Forge's marketplaces from undifferentiated app stores.

General availability launches through three channels simultaneously. Product Hunt and developer community launch — a coordinated 24-hour launch with live demos, template showcases, and a free-tier promotion. Live demo strategy — a real-time coding demonstration where Forge builds, deploys, and runs a complete application in under 10 minutes. The demo deploys to a live URL on Cloudflare's network, demonstrating the full build-to-deploy pipeline that no competitor can match live on stage. Enterprise 90-day trial — targetingCloudflare's 4,298 customers spending over $100,000 annually with a free 90-day enterprise trial that includes dedicated onboarding, migration assistance, and a named account manager. The 90-day window is long enough for organizations to build real applications and accumulate switching costs.

International expansion follows Cloudflare's existing geographic coverage. The 330-city network already provides low-latency access worldwide. Localization efforts focus on the five largest developer markets outside the United States: India, the United Kingdom, Germany, Brazil, and Japan. Forge's UI is translated into 12 languages. Pricing is localized using purchasing power parity — the Starter tier may be priced at $15 in India and $20 in Brazil, matching local market expectations while maintaining positive unit economics due to the proprietary model's low marginal cost.

Template localization focuses on region-specific use cases: Indian e-commerce storefronts with UPI and Paytm payment integrations, Japanese marketplace templates with local payment rails, German GDPR-compliant data dashboards with Impressum generators, and Brazilian marketplace templates with PIX payment support. Cloudflare's network already spans 125-plus countries, eliminating the infrastructure expansion typically required for international launches.

Forge offers a white-label program for digital agencies and hosting companies. Agencies rebrand Forge as their own platform, offering AI-powered website and application creation to their clients. Cloudflare provides the infrastructure; the agency provides the brand, customer relationships, and support. Revenue model: agencies pay a per-seat license ($99 to $199 per seat per month) plus infrastructure consumption. Cloudflare captures both the license revenue and the infrastructure revenue from every application built on the white-labeled platform. The agency pays Forge Scale pricing per seat and charges their end clients at whatever margin the market supports. This channel is capital-efficient — agencies handle customer acquisition, support, and billing for their clients, while Forge provides the technology and infrastructure. Estimated Year 3 to 5 contribution from the agency channel: $30 to $60 million in annual revenue.

Forge's billing architecture operates on three layers that maximize both transparency and revenue capture.

Layer 1: Subscription. The monthly plan fee ($0 to $299 per seat) that includes a base allotment of AI credits. This is the primary number users see and the anchor for pricing expectations. It covers platform access, storage quotas, deployment slots, and team seats.

Layer 2: Credit consumption. AI inference credits consumed beyond the monthly allotment. Standard Mode at one credit per prompt, Power Mode at three to five credits per prompt. Overages are billed at $0.08 per credit for Standard, $0.25 for Power Mode equivalents, with volume discounts through credit packs. All overages are subject to the spending cap.

Layer 3: Infrastructure consumption. Pay-as-you-go billing for Workers compute, D1 database operations, R2 storage, Durable Objects, and other Cloudflare services consumed by deployed applications. This layer is invisible during creation — it activates when applications go live and scales with application success. This is the "self-funding loop" that drives dollar-based net revenue retention into the 125-135 percent range at steady state.

The three-layer model aligns incentives: Forge succeeds when users create (Layer 1), iterate (Layer 2), and deploy successfully (Layer 3). Competitors who charge only for AI credits (Cursor) or only for subscriptions (Lovable) leave infrastructure revenue on the table.

The roadmap beyond general availability focuses on three priorities: deepening the moats through proprietary model v3/v4 and expanding marketplaces to 10,000+ items; expanding the addressable market through adjacent products like e-commerce, mobile app generation, and cloud cost management; and scaling revenue through enterprise sales team growth from 5 to 30 AEs and driving ARPU growth through infrastructure consumption.

The Forge Model v2 training run incorporates six months of production user interaction data from v1, using reinforcement learning from deployment success signals — a training methodology that no pure-play competitor currently matches because none owns both the model and the deployment infrastructure.

Adjacent market revenue is not included in the $1.15 billion base-case projection. These figures represent directional optionality, not modeled forecasts. Each category would require dedicated product investment and is contingent on Forge achieving its core AI coding platform goals first.

Year 1: Domain registration via Cloudflare Registrar, website security monitoring, SEO analysis tools. Estimated adjacent revenue: ~$8M.

Year 2: E-commerce storefronts with Stripe/Shopify integration, transactional email, lightweight CRM. Estimated adjacent revenue: ~$25M.

Years 3–5: Analytics/BI dashboards, white-label platform for agencies, mobile app generation. Estimated adjacent revenue growing from ~$65M to ~$180M, contingent on platform maturity and competitive dynamics.

Cloudflare financial data from SEC filings (10-K, 10-Q, proxy statements) and earnings press releases through Q4 2025 results (fiscal year ending December 2025). Balance sheet data current through the most recent quarterly filing.

Competitor data from company announcements, investor presentations, verified reporting by The Information, Bloomberg, TechCrunch, Fortune, and company blogs. ARR estimates cross-referenced against multiple independent sources. Valuation data from Sacra, PitchBook, and press reporting of funding rounds.

Market sizing data from Gartner, IDC, Forrester, and verified industry analyses. AI coding market figures represent consensus estimates for calendar year 2026, triangulated across multiple analyst reports.

Technical capabilities from official product documentation, developer changelog entries, GitHub repositories, npm download statistics, and SWE-bench leaderboard data through April 2026. Pricing data from publicly listed prices on product websites, verified through account creation where possible.

MCP ecosystem statistics from Anthropic announcements, the Agentic AI Foundation under the Linux Foundation, npm download data, and GitHub repository counts through April 2026.

Revenue projections use a bottom-up subscriber model with tier-level pricing, conversion rates, churn assumptions, and ARPU calculations. Cost of goods sold is modeled per-credit with separate assumptions for proprietary model inference, third-party API inference, and infrastructure consumption. Operating expenses use headcount-based modeling for R&D and standard SaaS benchmarks for S&M and G&A as a percentage of revenue.

Six key financial model assumptions:

1. Cloudflare's infrastructure marginal cost is modeled at 15 to 25 percent of retail pricing for equivalent AWS/GCP services, reflecting the amortized fixed-cost structure of an existing global network.

2. The proprietary Forge Model achieves competitive code generation quality (within 5 percent of frontier models on SWE-bench) by Phase 3, enabling 75 percent of Standard Mode queries to route to the proprietary model by Year 5.

3. Free-to-paid conversion rates rise from 2.2 percent in Year 1 to 4.8 percent in Year 5 — already at the upper end of the 2–5 percent freemium SaaS norm. Cursor's exceptional 36 percent rate reflects a self-selected professional-developer audience and is not used as the benchmark; the Forge model assumes its broader, mixed-skill audience converts at typical freemium rates.

4. MCP deployment capture assumes 1.2 million applications deployed at $15 average monthly infrastructure spend by Year 5, based on Cloudflare's current pricing and typical application resource consumption profiles.

5. Dollar-based net revenue retention targeted at 125 percent at steady state, with 135 percent achievable under strong expansion — driven primarily by infrastructure consumption growth as deployed applications gain users, not by subscription tier upgrades. A 10-point swing in DBNRR moves Year 5 revenue by approximately the same percentage.

6. Market growth of 30 to 40 percent annually through Year 5 is based on consensus analyst estimates and current demonstrated growth trajectories of top-5 AI coding tools.

This is not investment advice. It is not a recommendation to buy, sell, or hold Cloudflare stock (NET). It is not a product specification or engineering plan. It is a strategic analysis that models a hypothetical product opportunity using publicly available data and the author's independent assessment of market dynamics.

The financial projections are estimates, not predictions. They are intended to frame the scale and structure of the opportunity, not to provide precise forecasts. Actual results would depend on execution quality, market conditions, competitive dynamics, and dozens of variables that cannot be modeled with certainty.

This analysis is independent. It was not commissioned by, reviewed by, or endorsed by Cloudflare, Inc. or any company mentioned in the analysis. The author has no financial relationship with any company discussed.

This analysis is provided for informational purposes only. It does not constitute investment advice. The author has no financial relationship with Cloudflare, Inc. and holds no position in NET stock. The projections are the author's estimates based on publicly available information and may differ materially from actual results. The competitive assessments represent the author's interpretation of publicly available product information and may not reflect internal capabilities or strategic plans.

Structural Gap is committed to accuracy. If you identify a factual error, outdated data point, or mischaracterization in this analysis, please contact corrections@structuralgap.com. Verified corrections will be incorporated with a dated changelog appended to the analysis. Analytical disagreements and alternative interpretations are welcome but will not result in changes unless they identify factual errors.

Corrections — April 17, 2026. Following an external editorial review:

• MCP ecosystem size revised from "tens of thousands of servers" to "more than 5,800 registered servers" (the 97 million monthly SDK downloads figure is unchanged).
• Five-year return metric relabeled from "ROI" to "MOIC" (multiple on invested capital), with the corresponding IRR of approximately 40% disclosed alongside.
• SWE-bench citation grounded in the swebench.com leaderboard and reframed as model-progression scores rather than a tool-specific benchmark.
• Cursor's 36% free-to-paid conversion benchmark contextualized against the 2–5% freemium SaaS norm; Forge's modeled conversion (rising from 2.2% in Year 1 to 4.8% in Year 5) is framed as already at the upper end of that norm.
• The companion v2 spreadsheet (nine sheets, three scenarios, MOIC/IRR reconciliation) is now downloadable from the financial model page.

The financial model's underlying numbers — investment, revenue, FCF, scenarios — are unchanged.