Key takeaways
- Abundance moves the bottleneck. When code becomes easy to generate, understanding and maintaining it become more valuable. When products become easy to launch, distribution becomes the constraint.
- AI’s consequences are becoming investable categories. Agent evaluation, observability, governance, provenance, security and cost control are not peripheral features; they are emerging infrastructure layers.
- Look for the new scarcity. The strongest opportunities may sit on the other side of AI adoption: trust after infinite content, control after autonomous agents and reliability after intelligence becomes widely available.
Every major technology shift removes old constraints and creates new ones. The internet made information abundant, so discovery became scarce. Social media made publishing abundant, so attention became scarce. Cloud computing made infrastructure available on demand, increasing the value of security, orchestration and cost control.
AI is now producing its own version of this cycle. For the past few years, much of the startup conversation has focused on what AI makes possible. But as an early-stage investor, I’m increasingly interested in a slightly different question: What did AI make abundant? and what became scarce as a result? That question may point to some of the most interesting startup opportunities ahead.
Code became abundant. Engineering judgement became scarce.
Software development used to be constrained by the number of engineers available to write code. AI coding tools have loosened that constraint dramatically: a small team, or even one person, can now move from an idea to a working product in days. But producing code and building dependable software are not the same thing.
As the cost of generation falls, the bottleneck moves from writing to judging. Does the code behave as intended? Is it secure? How does it fit the existing architecture? Who will understand it well enough to change it six months from now? Teams are accumulating machine-generated code that nobody fully owns, while impressive demos can conceal fragile dependencies, duplicated logic and technical debt. In Stack Overflow’s 2025 developer survey, 45% of developers said debugging AI-generated code was time-consuming, while 61% wanted to understand their code fully.
This creates opportunities well beyond another coding copilot: tools that map unfamiliar codebases, recover the intent behind the code, validate tests, identify vulnerabilities, enforce architectural standards and continuously refactor technical debt. The valuable product may not be the one that generates the most code, but the one that gives teams the confidence to operate it. Our portfolio company Pandorian.ai does exactly that.
This echoes the argument I made recently in “Tokenmaxxing Was the Wrong Metric”: tokens are an input cost, not a measure of value. Lines of code are becoming much the same. The outcome that matters is reliable software.
Agents became abundant. Control became scarce.
We’re moving from AI that answers questions to AI that takes actions. That is real progress, but it creates a different class of problem: How do you test a non-deterministic agent? What is it allowed to do? How do you reconstruct its decisions when something goes wrong? Who is liable if it sends the wrong payment, deletes a customer record or gives regulated advice?
The more autonomy we give agents, the more valuable the control layer becomes. Important companies will be built around evaluation, observability, permissions, simulation, audit trails and possibly insurance. But not because agents have failed, but because they are beginning to matter.
One of the key areas of opportunity here has to do with the non-deterministic nature of LLMs – run the same prompt/agent 100 times with similar input, and you might get very different outcomes.
Content became abundant. Trust became scarce.
AI has reduced the marginal cost of producing text, images, music and video to almost zero. It has not increased the amount of attention available, and it has weakened many of the signals we once used to judge what was real, original or worth consuming.
This creates several concrete opportunities. News organisations, marketplaces and insurers need to verify the origin and editing history of images and video. Banks and communication platforms need to detect synthetic voices and impersonation. Creators and publishers need infrastructure to control, license and get paid when their work is used by AI systems. Brands need to identify fabricated content and coordinated narrative attacks before they cause financial or reputational damage.
Detection alone may become an arms race as generation models improve. The more durable businesses may sit deeper in the workflow: establishing provenance at the moment of creation, verifying identity during an interaction, managing rights and payments, and building reputation systems around trusted sources.
The important question will not simply be whether a piece of content was generated by AI. It will be: Who made it? Can they prove it? Do they own it? And may it be used here?
Tokens became abundant. Efficiency became scarce.
The industry has spent years optimising for bigger models, longer context windows and more tokens. But customers don’t want tokens; they want a task completed correctly, quickly and at a reasonable cost. A model that can answer almost anything but behaves unpredictably remains difficult to use in a high-stakes workflow.
The next phase of AI may therefore be less about maximising intelligence and more about applying it efficiently. Smaller models, better routing, constrained systems, domain-specific evaluation and products that know when not to use AI could prove more useful than simply throwing more inference at every problem.
Products became abundant. Customer attention became scarce.
AI has made it dramatically easier to launch a credible product, but it has not created more hours in the day or more willingness to try another app. RevenueCat found that monthly subscription-app launches increased roughly sevenfold between January 2022 and January 2026. Yet apps launched before 2020 still generate 69% of subscription revenue, while those launched since 2025 account for only 3%. AI apps may attract curiosity and monetise well initially, but they also churn 30% faster.
Gaming makes the same imbalance visible. More than 20,000 games were released on Steam in 2025, compared with approximately 14,000 just two years earlier. Player attention has not expanded at anything close to the same rate, and the most-played games remain dominated by established titles and long-running communities. Vibe coding can produce more games, but it cannot manufacture players.
This shifts the opportunity from creation to demand. Startups can help consumers discover products that are genuinely relevant, connect developers with creators and communities, build cross-promotion networks for independent apps, or improve onboarding, retention and live operations. But the strongest products may not sell distribution as a separate tool. They will build it into the experience through collaboration, multiplayer, user-generated content, marketplaces, integrations or other loops in which usage naturally attracts more users.
This is why I’m sceptical when founders present a prototype as evidence of a moat. In the age of AI coding, a prototype is increasingly the minimum. The harder questions are how the product will reach users, why they will return and whether every new user makes the product easier, or harder, to distribute. AI democratised product creation. It did not democratise demand.
Automation became abundant. Accountability became scarce.
Employees are no longer simply using AI applications. They are assembling agents and unofficial workflows that connect models to email, customer records, internal documents, finance systems and other company tools. Microsoft reports that 29% of employees have already used unsanctioned AI agents, while fewer than half of organisations have introduced specific security controls for generative AI.
This is more consequential than traditional shadow IT. An unapproved SaaS product stores data; an AI agent may read that data, make a decision and take an action. It can inherit an employee’s access without inheriting their judgement, accountability or understanding of company policy.
That creates a concrete new infrastructure layer. Enterprises need to discover every agent operating across the organisation, give it a verifiable identity and owner, restrict it to the minimum required permissions, inspect the data it accesses and intervene before it takes a high-risk action. They also need a complete audit trail showing which model, prompt, tools and information produced each decision.
Governance cannot be limited to policies and compliance dashboards. It must operate at runtime, between the agent and the systems it is attempting to control. Every agent will eventually need an identity, an owner, defined permissions, an audit trail, a budget and a kill switch.
Look for the new scarcity
The first wave of AI startups was largely about capability: what can the models do? The next wave will increasingly be about consequences: what becomes difficult, dangerous or valuable once everyone can use them?
This is not a negative view of AI. Quite the opposite. New problems appear when a technology becomes successful enough to change behaviour. As investors, we naturally look at what a new technology enables, but abundance alone rarely determines where value accrues. Often, the better question is what becomes scarce on the other side.
AI made code abundant, but not understanding. Content abundant, but not trust. Tokens abundant, but not efficiency. Products abundant, but not distribution. Automation abundant, but not control.
The companies that solve those new scarcities may prove more durable than the ones that created the abundance in the first place.
At Remagine Ventures, we love to meet Israeli founders very early in their journey. If you’re addressing one of the new scarcities with your startup, we’d love to hear from you.
He is a former General Partner at Google Ventures (GV) in Europe, former head of Google for Entrepreneurs in Europe, and founding head of Campus London, Google's first startup hub. Eze writes on Israeli tech, venture capital, artificial intelligence, and founder strategy.
He is also the founder of Techbikers, a nonprofit that brings together the startup ecosystem on cycling challenges in support of Room to Read.
- The Next AI Opportunities Are the Problems AI Created - July 27, 2026
- Weekly Firgun Newsletter – July 24 2026 - July 24, 2026
- Vertical AI Market Map 2026: Israeli Startups and Funding Rounds - July 23, 2026
