Cognition
Usage-based + subscription SaaS: Agent Compute Units (ACUs) and per-seat plans for Devin, plus Windsurf IDE subscriptions; bottoms-up developer adoption feeding enterprise land-and-expand.
Post-money marks from priced primary rounds. Valuation more than doubled ($10.2B to ~$26B) in ~8 months across Series D to E, reflecting the ~13x ARR ramp and the July 2025 Windsurf acquisition. Series A at ~$2B is widely reported; Series D and E marks are the verified anchors.
Earnings, margins, COGS & capex
Revenue is ARR-run-rate only (no audited P&L). ARR went from ~$1M (Sep 2024) to ~$37M (May 2025) to ~$73M (Jun 2025) to ~$492M (May 2026), an ~13x single-year jump propelled by the July 2025 Windsurf acquisition (which alone added ~$82M ARR and 350+ enterprise customers) and rapid enterprise Devin adoption. Cost of revenue is dominated by LLM inference; margins are undisclosed and the primary financial unknown.
Revenue trend
Margins
inference-cost dependent; own SWE model on Cerebras aims to improve it
assumed deeply negative (growth stage)
COGS structure
Not disclosed. Cost of revenue is dominated by frontier-model inference (Devin routes to Claude, GPT-5, Gemini for longer/complex work) plus serving its own SWE-1.5/1.6 coding models via Cerebras wafer-scale inference (~950 tok/s fast tier). Owning the model narrows but does not remove third-party token cost.
Capex
Not disclosed; asset-light. Compute is consumed as opex through inference partners (notably Cerebras) rather than self-built GPU clusters.
Latest earnings
not applicable (private)
No formal guidance; company frames trajectory as continued rapid enterprise expansion of Devin.
- ARR run-rate (May 2026)
- ~$492M
- Post-money valuation
- ~$26B
- Enterprise customers (from Windsurf)
- 350+
- Code written by Devin internally
- 89%
Growth drivers
- Windsurf acquisition (Jul 2025) — +~$82M ARR, 350+ enterprise customers, a forked-VS-Code IDE surface, and the SWE model team
- Enterprise Devin land-and-expand (reported +50% MoM usage for six months)
- Own SWE-1.5/1.6 coding models served on Cerebras for speed + cost control
- Product consolidation: Windsurf folded into Devin Desktop; windsurf.com pricing redirects to devin.ai
- Dogfooding credibility — company says 89% of its own committed code is written by Devin (up from ~13% in Dec 2025)
Bull & bear
Cognition is a top-two or top-three winner in the fastest-growing software category since cloud, compounding ARR ~13x a year while vertically integrating model, agent, and IDE - a $492M run-rate that can plausibly become multi-billion, making ~$26B look cheap in hindsight.
- AI coding crossed ~$2B+ category revenue in 2026 and is still early; Cognition holds one of the strongest brands (Devin) and a real enterprise book
- Windsurf gave it distribution (a forked-VS-Code IDE + 350+ enterprise logos) that pure-agent rivals lack
- Owning SWE-1.5/1.6 on Cerebras is a genuine cost/latency edge (~950 tok/s) that can widen gross margin as own-model usage grows
- 89%-internal-Devin dogfooding is a credible proof point that the autonomous agent works at production scale
- Blue-chip investors re-upped and more than doubled the valuation in ~8 months, signaling strong insider conviction on the growth curve
At ~$26B on a ~$492M ARR run-rate (~53x), Cognition is priced for near-flawless execution while its best models are rented from the very competitors (Anthropic, OpenAI, Google) who sell the same coding agents and can bundle them at zero marginal price - a margin-and-moat squeeze that a hot funding market is masking.
- ~53x ARR run-rate leaves no room for a growth deceleration or a funding-sentiment reset
- Third-party frontier-model dependency caps gross margin and exposes pricing/availability to suppliers who are also rivals
- The model is not the moat (all tools route to Claude/GPT-5/Gemini for peak quality); differentiation compresses toward UX and price
- Microsoft/Google/Amazon can give away enterprise coding AI as an ecosystem feature; Cognition must charge for its core
- Autonomous-agent reliability is still imperfect; enterprise trust and security concerns bound fully-hands-off deployment
- Two products to integrate (Devin + Windsurf) and undisclosed burn on a category with brutal talent-cost inflation
What it is worth
Private-market comps + reverse read on the last priced round (EV/ARR-run-rate; no public DCF). Peers: Anysphere/Cursor (private, high-multiple), Microsoft/Google/Amazon (public, embedded).
~$8-12B on a growth-and-sentiment reset: supplier-competitors (Anthropic/OpenAI) win the agent layer, hyperscaler bundling compresses pricing, and the ARR multiple re-rates toward the Sep-2025 ($10.2B) level or below.
~$26B (the May 2026 Series E mark) holds if growth stays strong but decelerates from the +50% MoM burst and margins remain undisclosed/thin.
~$40B+
if ARR compounds past ~$1B at improving own-model gross margin and Cognition holds a top-two enterprise position - the multiple looks justified in hindsight.
The ~$26B post-money on a ~$492M ARR run-rate implies ~53x ARR - a top-decile AI multiple that requires sustained ~50%+ growth toward multi-billion ARR AND a path to healthy gross margin as own-model usage scales. Supported by the growth rate and category size; unsupported if either the growth or the margin (still undisclosed) disappoints. Not financial advice.
SWOT
Strengths
- Category-defining brand in autonomous coding agents (Devin) with credible enterprise traction
- ~13x single-year ARR growth to ~$492M run-rate
- Owns both an agent (Devin) and a distribution surface / IDE (Windsurf) after the acquisition
- Vertically integrating the model layer with own SWE-1.5/1.6 models on Cerebras for speed and cost leverage
- Deep-pocketed cap table (Founders Fund, Lux, General Catalyst, 8VC) and ~$1B+ fresh Series E cash
Weaknesses
- Financial opacity: gross/operating margins undisclosed; economics gated by third-party inference cost
- Still materially dependent on rival labs' frontier models (Anthropic, OpenAI, Google) for peak quality
- Devin's autonomous task success historically imperfect (early SWE-bench-style success rates well below human), limiting fully-hands-off use
- Two overlapping products (Devin + Windsurf) to integrate without cannibalization
- No public profitability path disclosed; burn assumed high
Opportunities
- Enterprise standardization on an agentic 'AI software engineer' as a budget line, not a seat add-on
- Expanding from assisted coding into fuller autonomous SDLC (test, deploy, maintenance)
- Own-model economics reducing COGS and third-party dependence over time
- International + regulated-industry enterprise expansion via Windsurf's on-prem/enterprise posture
Threats
- Cross-subsidized incumbents — Microsoft (GitHub Copilot), Google (Gemini/Antigravity), Amazon (Q/Kiro) bundle coding AI into larger ecosystems
- Anthropic's Claude Code and OpenAI's Codex — the same labs Cognition buys tokens from - compete directly at the agent layer
- Cursor (Anysphere) contesting the daily-driver IDE surface
- Commoditization: the model is nearly identical across tools, compressing pricing power
- Valuation reset risk if AI-coding growth or funding sentiment cools (~53x ARR run-rate)
Moats, dependencies & bottlenecks
Moats
Strong mindshare but not a structural lock-in; rivals contest the same narrative.
enterprise install base) The editor surface is stickier than a plugin; 350+ enterprise customers acquired.
own SWE-1.5/1.6 models + Cerebras inference Speed/cost edge that could deepen into a real economic moat if own-model quality holds.
repo integrations, ACU spend commits) Growing with land-and-expand but early; multi-tool experimentation is the norm.
Dependencies
Peak-quality routing depends on rivals' models (Devin defaults to SWE-1.x, escalates to Claude for hard tasks); they are also direct competitors and set token pricing.
Supplier (inference) Serves own SWE models at high throughput; concentration risk on a single wafer-scale inference partner.
Supplier (compute/hosting) Standard hyperscaler dependency for hosting and enterprise deployment.
Supplier (upstream compute) Indirect exposure via inference partners' hardware availability and cost.
Pre-profit; ~$26B valuation and growth spend assume ongoing access to venture capital.
Advantages
- First-mover brand in autonomous coding agents
- Agent + IDE + own model under one roof (rare full-stack coverage)
- Cerebras-served own models give a latency/cost lever competitors relying purely on rented tokens lack
- Strong, re-upping investor base and a large war chest
Weaknesses
- Undisclosed, likely-negative margins and burn
- Structural dependence on competitor-owned frontier models
- Rich valuation with little margin for a growth stumble
- Commoditizing model layer erodes differentiation to UX + price
Bottlenecks
- Gross margin gated by third-party inference cost until own-model usage dominates
- Autonomous-agent reliability/security ceiling for fully-hands-off enterprise deployment
- Integrating Devin and Windsurf into one coherent product without cannibalization
- Talent-cost inflation in a market where hyperscalers out-bid on AI engineers
Top signals & trends
Top signals
bull if rising · Direct read on future gross margin and supplier independence.
Watch for deceleration as the base scales past ~$500M.
Suppliers competing downstream is the core structural risk.
Tests whether the ~$26B mark holds in a cooler market.
Trends
Directly favors Devin's positioning.
Windsurf gives Cognition an owned surface; validates the acquisition.
Lowers input cost but erodes tool-level differentiation.
Pressures standalone pricing power.
A handful of names now take most paid usage; scale accrues to leaders.
Ecosystem & competitor graph
Suppliers feed the company; customers pull from it. Line thickness shows the strength of each tie (supply-chain dependency, customer earnings contribution). Hover to isolate a tie.
Wafer-scale inference serving Cognition's own SWE-1.5/1.6 models at ~950 tok/s; IPO'd on Nasdaq May 2026.
Frontier model (Claude) routed by Devin/Windsurf for hard tasks; also a competitor.
Frontier model (GPT-5) routed by the products; also a competitor.
Gemini models routed; Google Cloud infra; also a competitor.
Cloud hosting and enterprise deployment substrate.
Upstream GPU supply underpinning partners' inference capacity.
Enterprise software / engineering orgs (350+ from Windsurf) Core buyer: dev teams adopting Devin per-seat + ACU usage; named logos include Goldman Sachs, Citi, Dell, Cisco, Palantir.
Land-and-expand enterprise contracts; the growth engine.
Bottoms-up funnel via Devin plans and the Windsurf IDE.
Private (~$965B post-money, Series H, May 2026). Both a key model supplier and a direct agent-layer competitor - the sharpest structural threat.
Private. Frontier model supplier (GPT-5) and direct coding-agent competitor.
Private (~$9-29B range through 2026). Leading daily-driver AI IDE; contests Windsurf's surface directly.
20M+ lifetime users, ~90% Fortune 100 penetration; enterprise default, cross-subsidized by the GitHub/Azure ecosystem.
Bundles coding AI into Google Cloud + Gemini; deep model + distribution.
AWS-native coding agent bundled into cloud spend.
Private. Full-stack agentic build-and-deploy platform, strong with prosumer/SMB.
Enterprise code-intelligence and IDE-native assistants competing on integration.