The Twenty Minute VCCognition CEO Scott Wu on Acquiring Windsurf: The Process, The Deal, The Rationale
CHAPTERS
- 0:00 – 1:50
Windsurf news breaks: why Cognition saw a “treasure chest,” not a husk
Scott explains Cognition learned about the Windsurf situation at the same time as everyone else and immediately recognized a strategic fit. He frames the opportunity as valuable product, customer book, IP, and team—especially go-to-market talent—rather than leftover scraps.
- •Cognition hears about the Windsurf/Google split on Friday
- •Perception vs reality: not “nothing left,” but meaningful assets remain
- •What Cognition lacked (GTM) matched what Windsurf had; product fit also strong
- •Core assets: product usage, customers, codebase, proprietary IP, team
- 1:50 – 6:36
From cold outreach to a 72-hour acquisition: structuring speed over diligence
Scott walks through the execution mechanics: a cold reach-out Friday evening, verbal alignment Saturday, legal/terms Sunday, signature and announcement by Monday. He compares the moment to a bank receivership—speed is essential to prevent customer and talent erosion.
- •Cold outreach Friday night; first call immediately
- •Decision principle: resolve uncertainty for customers/team ASAP
- •Compressed timeline: verbal Saturday, paperwork Sunday, signed Monday
- •Trade-off: less diligence depth, but enough confidence in business fundamentals
- •Goal: stabilize operations and reassure stakeholders quickly
- 6:36 – 7:20
Was Windsurf “left in the lurch”? Options the team weighed after the split
Harry presses on whether Windsurf leadership and staff were stranded by the prior deal dynamics. Scott says notice was limited, but leadership thoughtfully laid out paths: remain independent, raise new VC (with no investors), or find the right acquirer—fast.
- •Limited advance notice, but leadership tried to manage options transparently
- •Three paths: stay independent, raise new capital, or partner/acquire
- •Time pressure: recruiting inbound and customer anxiety accelerate fallout
- •Emphasis on doing right by the remaining team
- 7:20 – 8:24
Google’s miss and the rise of talent-first deal norms
Scott acknowledges there’s truth that valuable pieces can be left behind in modern deal structures. He argues norms are shifting: founders historically “go down with the ship,” but talent and asset carving has changed, and he finds that disappointing.
- •“Valuable pieces get left behind” in certain acquisitions
- •Deal structures are evolving toward talent/IP carve-outs
- •Founder obligation norms shifting; Scott views it as disappointing
- •Broader question: will this become standard M&A playbook?
- 8:24 – 10:40
Why the AI talent war is ‘reasonable’: massive shift, few key drivers
Scott argues the fierce competition for AI talent is rational given AI’s magnitude—even with capabilities frozen today. He estimates the number of people materially shaping AI’s trajectory is surprisingly small, likely in the hundreds to low thousands.
- •AI as the biggest technology shift—even if progress stopped today
- •Productization and distribution matter even without new breakthroughs
- •Only a small cohort meaningfully steers AI progress
- •Scott’s estimate: at least ~100, far fewer than 10,000 (likely well below)
- 10:40 – 14:25
Can apps compete with model giants? Differentiation, dependency, and collaboration
The discussion turns to the application layer’s viability amid dependence on foundation labs like Anthropic. Scott’s view: value accrues wherever real differentiation exists, and foundation labs and app-layer products solve different problems that benefit from collaboration.
- •Dependence risk: the Windsurf/Anthropic cutoff example and vendor leverage
- •Value accrues where differentiation is strongest, not a single layer by default
- •Cognition focuses on workflows: humans+AI producing code in real environments
- •Why foundation labs collaborate: different focus areas; both sides needed
- 14:25 – 16:56
Will model progress plateau? RL as the engine and why AI coding already changed work
Scott predicts continued capability gains, citing reinforcement learning (RL) as the biggest recent breakthrough and a scalable technique. He adds a practical claim: in software engineering, AI is already a no-brainer—engineers are slower without it.
- •Confidence drivers: known techniques still have room to scale
- •RL framing: define environments and success/failure, then train to behavior
- •Even with zero future progress, AI coding would still be transformative
- •AI coding adoption: already makes engineers materially faster
- 16:56 – 19:44
50% AI-written code vs real productivity: measuring speed, not lines
Responding to claims that 50% of net-new code is AI-generated, Scott argues the better metric is effective productivity—how much an engineer accomplishes per hour. He estimates today’s aggregate uplift at ~1.5–2x, with potential to reach ~10x in three years, and expects far more total code to be written (Jevons paradox).
- •Line-count metrics are misleading; importance of code varies
- •Better benchmark: time-to-output with best AI tools vs none
- •Current uplift estimate: ~1.5–2x; potential: ~10x over ~3 years
- •Expect 10x more code overall as software demand expands (Jevons paradox)
- •Many products are far from “top-tier” polish; huge room for improvement
- 19:44 – 21:24
What agents still can’t do: deep context, ownership, and shifting skill value
Scott outlines what’s missing for agents to truly ‘own’ work: deep contextual understanding across complex codebases and tooling. He predicts programming shifts toward intent expression and higher-level architecture/product decisions, making solution design the core skill.
- •Agents need to reach true ownership: operate with high-touch context
- •Future interface: express intent, manipulate product directly vs editing code
- •Most valuable skill: defining problems, designing solutions, architecting systems
- •Less about syntax; more about technical product/architecture judgment
- 21:24 – 24:36
Pricing and value capture: usage-based Devin and where AI value accrues
Harry challenges whether AI coding tools will capture enough of the value they create. Scott describes Devin as usage/hour-based, priced roughly 10x cheaper than the value of an engineer’s time, and argues the bigger mission is accelerating capability adoption over perfect monetization optimization.
- •Devin pricing: usage-based, roughly hourly
- •Target: ~10x cheaper than an engineer’s time value
- •Market size: ~30M engineers; expectation of big productivity gains
- •Debate: what fraction of created value tools can capture
- 24:36 – 29:45
Devin vs Cursor/Windsurf attention: growth reality, enterprise workflow, and IDE head start
Harry suggests Devin ‘fell out of the zeitgeist’ compared to IDE-first brands; Scott counters with internal traction: 5–10x usage growth in six months and strong enterprise/team-based workflows (Slack/Linear/GitHub PRs). He explains IDE assistants had a head start, while agents are just now becoming familiar to the mainstream.
- •External buzz vs internal numbers: Devin usage up 5–10x in 6 months
- •Primary adoption pattern: real engineering teams integrating into workflows
- •Agents vs IDEs: IDE experiences matured earlier; agents catching up fast
- •Short-term expectation: agents become as familiar as IDE copilots within 6–12 months
- 29:45 – 36:47
Market end state: ‘too early to call’ and the next human-computer interface
Asked how the developer market will shake out, Scott says no one is close to the end state. He reframes the whole category as the next generation human-computer interface—moving from code to intent—using the ‘Tony Stark/Jarvis’ analogy and projecting rapid iteration cycles.
- •Rejects simplistic segmentation (bottom-up vs enterprise) as premature
- •Core thesis: software engineering tools are evolving into a new HCI
- •End state: intent-driven interaction, generative UI, single-use software
- •Rapid cadence: many ‘levels’ of UX progress, each taking months
- 36:47 – 43:10
Quick-fire: RL’s underappreciated power, curated data, and foundation model consolidation
In quick-fire, Scott argues RL is still underappreciated: once you define the right benchmark/feedback loop, agents can be trained for many tasks. He also notes a shift from ‘more data’ to smaller curated datasets plus compute, discusses Kevin-32B, and weighs how many foundation model winners remain after consolidation.
- •RL’s promise: training to explicit success/failure criteria across verticals
- •Key bottleneck: defining the right ‘benchmark’ for real-world jobs
- •Shift: quantity of data → curated data + compute (example: Kevin-32B/kernelbench)
- •Foundation model landscape: likely consolidation to ~2–6 major players
- •Scott’s investment take: both OpenAI and Anthropic can be great at current prices
- 43:10 – 47:46
M&A details, messaging critique, and integrating IDE + agent workflows
Harry probes deal composition (cash vs stock) and critiques Cognition’s acquisition video as too scripted; Scott agrees and notes the time crunch. Scott then shares the core post-deal product goal: a unified workflow combining IDE planning/retrieval with agent execution and IDE-based review—moving seamlessly between synchronous and async work.
- •Deal split: mixture of cash and stock (no exact disclosure)
- •Brand/communication feedback: more personality and narrative shaping needed
- •Immediate operational focus: reassure customers, integrate teams
- •Product vision: tight IDE+agent loop—plan in IDE, execute via agent, review locally
- •Maintain philosophies of both Devin and Windsurf while finding the intersection
- 47:46 – 51:05
Closing: leadership under stress, personal calm, and the case for telling the story
Scott reflects on exhausting days closing the deal and integrating teams, then answers a personal question: he’s most proud of staying emotionally calm under pressure. Harry urges Scott to build a stronger public narrative—sell the ‘human enhancement’ story and proactively shape growth perceptions to avoid rumor-driven narratives.
- •Post-deal reality: minimal sleep; rapid customer communication and integration
- •Scott’s proud trait: composure in stressful, high-stakes moments
- •Marketing advice: tell customer ‘superpower’ stories; publicize growth deliberately
- •Narrative control: shaping the story reduces gossip and improves recruiting/funding
