Making code is cheap now. But figuring out what to build? Still hard. That's the opening insight from a16z partner Anish Acharya's "Notes on AI Apps in 2026" — AI has 10x'd our ability to make things, but our ability to think hasn't caught up.
What is this?
This is an essay published in January 2026 by Anish Acharya, a consumer investing partner at a16z (Andreessen Horowitz). He identifies five structural shifts in the maturing AI app ecosystem, and the core message is this — cheap code is just the beginning; the real transformation happens at the "what to build" stage.
The timing is perfect. In early 2026, SaaS stocks plummeted 30%, sparking "SaaSpocalypse" panic. But a16z pushed back, arguing that "AI is the best thing that ever happened to the software industry." Their companion piece, "Good News: AI Will Eat Application Software," analyzed how classic software moats (network effects, brand, process power) remain valid in the AI era.
What makes this essay special is that it reveals the actual direction of a16z's investments. In the same month, 15 partners published their individual investment theses in Big Ideas 2026 — Acharya's note is the definitive summary of the app layer thesis.
What changes?
1. The era of thinking tools is here
Every tool we use today is a "making tool" — IDEs for code, Figma for design, spreadsheets for models. But we barely have modern tools for exploration — tools that help us think. Acharya sees a near future where as coding agents improve in accuracy and time horizon, the hard problem shifts from "how to build it" to "what to build".
Imagine a PM who wakes up every morning to review 2-3 features an AI dreamt up, executed, and A/B tested overnight. But models still can't decide what to build next — ideas are bland, derivative, and lack the spark of great product thinking.
2. Every team becomes a software team
Companies have "power functions" (engineering, product, performance marketing) and "service functions" (legal, finance, HR). Coding agents collapse this distinction. Legal, procurement, finance — they all need to reach for a software toolbox before relying on processes and humans.
| Traditional | 2026 Software-First | |
|---|---|---|
| Legal contract review | External law firm + manual review | Domain AI like Harvey for auto-review |
| Marketing campaigns | Agency → 2-4 week lead time | Build tools with coding agents |
| Financial analysis | Manual spreadsheet work | AI-native analysis pipelines |
| Product prioritization | Quarterly roadmap meetings | AI suggests features daily → instant validation |
3. AI apps ≠ AI models — they compound differently
With ChatGPT at 900 million weekly active users and Claude Code hitting a $1B ARR in six months, 2026 is massive. But Acharya sees apps and models diverging further. AI apps combine cutting-edge model orchestration, domain-specific UI, and a vast feature surface that's now very cheap to build.
Even in coding — central to model progress and lab focus — startups generated over $1 billion in new revenue in 2025 alone. The labs and big tech are "jagged" in their capabilities: formidable in their focus areas but constrained by complex commitments and hard prioritization problems.
4. Consumers discover "the rest" of AI
Eugenia Kuyda (Replika founder) has been the best thinker on how the command-line UI has blocked everyday consumers from AI's best capabilities. This is changing — Wabi exposed code generation to consumers, the Images tab in ChatGPT/Grok democratized image gen, and Apps Directories are opening MCPs and prompt plugins to regular people.
No one tells you that you are living in the good old days until they are gone, so consider this your notice.
— Anish Acharya, a16z
Getting started: the essentials
- Audit your "service functions"
Legal, finance, HR, procurement — if they still rely on manual processes, start with domain-specific AI tools (Harvey, Hebbia) or coding agents (Claude Code, Codex). - Adopt "thinking tools"
Go beyond coding tools to exploration tools. Cursor's agent mode and Antigravity's "agent-first" design represent this direction. - Reset product prioritization
If AI can prototype and test features overnight, shift from quarterly roadmaps to weekly experiment cycles. "Every feature that can be built will be built." - Build a multi-model strategy
Don't lock in to a single model. According to a16z, OpenAI, Gemini, and Anthropic command 89% of enterprise wallet share, but multi-model orchestration is the core advantage for AI apps. - Have fun
Acharya's closing advice: "This product cycle is less centralized, more software-led, and simply more damn fun for technologists than any in recent memory."
Key takeaway for CEOs
Acharya's three-part advice for incumbent CEOs: (1) Collapse all customer-facing roles (sales, support, collections) into a single AI function. (2) Get non-technical functions embracing models — that's how you get broad operating leverage. (3) Demand more ambitious products and more ambitious prices.
