Mainstream media (including the BBC) as well as reputable experts are frenetically sharing this. Millions of views, public debates. So what to think about this?

To me this looks like a pattern, as old as humanity itself. End-time prophets that rise in times of social and technological disruption. We have been taught over millennia that human ambition that rises too high has to fall eventually - think about mythology like the Tower of Babel or Ikarus.

But lets dive into the narrative, the messengers and what to take serious and what not.

Claims versus Prophecies

First of all, there is a general problem with these kind of claims - that should better be called prophecies. Why? Because, while they look like serious probabilistic and verifiable claims, they are actually unfalsifiable and immune to accountability in the event of fulfillment. If humanity is extincted, there is no one left to prove the case.

In addition, I reject the argument usually brought forth by doom sayers, that even the smallest chance of infinite loss justifies infinite precaution. If we would adhere to such logic we should better not get out of bed, and in consequence we would end up in total paralysis.

Who Benefits - The Economy of Alarmism

In case I hear something of this level of alarm and disturbance, I always ask myself if there are there any hidden interests that might influence the message, its framing and timing.

What sticks out is that the loudest voices are people associated with (for-profit) frontier labs. This implies a natural interest in gaining attention, either for their organizations or themselves in order to promote their products, skills and foresight.

Still there might be some truth and just the way of framing is intentionally exaggerated.

Counter Arguments

However, there are several logical flaws to the argument. Advanced algorithms are accounted with human attributes like will, intent or will to power. Similar to how human attributes are given to humanoid robots. Several things do not add up:

1. Intelligence ≠ Autonomy: A system can be very capable in solving a specific problem, without having a desire, ego or self-preservation. These are biological traits born from evolutionary pressure to compete for scarce resources. An algorithm (even advanced machine learning or AI) usually does what it is programmed for. And in case there is unexpected behavior - we used to call this a bug or malfunction and not intelligence.

2. Pragmatism: It is naiv to think that the existence of a real threat would not eventually alter behavior in order to make the outcome far less likely to happen. As we have seen with nuclear, this mobilized resources to control and restrict access. The prediction itself reduces its chance of fulfillment.

3. Risk and Uncertainty: Assigning a number of 10% probability creates the impression of a quantifiable, known risk assessment. In reality nobody knows, it is unknown territory. A number is just misleading or reflecting an intent and mood and not an objective probability.

4. Systemic Thinking: It is naive to isolate the growth of capabilities in one domain only. Capability and control (and understanding) advance together, not in isolation. Assuming AI capabilities grow exponentially, while understanding and control remain static does not make sense. Of course a modern soldier could defeat a knight equipped with sword and shield. Of course an advanced, dynamic AI algorithm can penetrate an old, static firewall. But we have reasons to assume that the firewall progresses as well.

5. Super Intelligence Paradox: It is almost comical how linear thinking is coupled with exponential predictions. Why would a system that achieves super intelligence think in a way as the most stupid humans do - how to harm others, how to promote itself at the cost of others? In case a system would ever achieve this level of superior intelligence, it is likely to say that it would think and act differently. If it maintains a goal of self-optimization and augmentation (which would be a precondition to fight for scarce resources), it likely would pursue more rewarding and productive tasks than eliminating humanity.

Hence these prophecies leave a bad taste in your mouth. Also there are striking similarity to entertaining science fiction stories in Dune or Terminator. The message reveals more about the thinking and value system of the teller, than it does about the system it claims to make predictions about. People driven by greed, competition and a will to dominate.

The Risk is the Mind and Not the Tool

Of course there are (immense) risks that come with every powerful technology. And AI is not an exception. However, the dangers will likely be a result of human design and way of application and not machine malice.

Some implications might not be obvious, but could have serious consequences, like competition for scarce resources (energy, water, rare materials) needed to power advanced compute capabilities. But also the consequences of economic concentration that is a consequence of advanced automation and wealth and capabilities accumulating with a small elite of tech monopolies and state actors. Also AI technologies can lead to pervasive surveillance, rendering people transparent and susceptible to manipulation and repression.

The existential debate on AI capabilities overshadows a much simpler reality:

The central challenge of the next decade will not be to prevent super intelligence or to be fearful of an AI wishing to kill us, but to prevent concentrated interests from using pervasive automation and AI to undermine human autonomy, safety, free will and integrity.

In the US alone, BEA data shows investment in software increased significant to over 11% year-over-year. As Marcello Estevão notes in his IMF research, global datacenter build-outs may require an eye-watering $6.7 trillion in capital expenditure by 2030.

Yet, when you look closely at macro productivity, the needle has barely moved. Even when comparing high-adoption tech hubs like California to lower-adoption regions, there is no statistically significant productivity divergence. The existing gap is explained by capital investment, not an “intelligence advantage.”

Why is unprecedented compute failing to move the macro needle?

The Lesson from Pre-1931 Intelligence

A few weeks ago, I was reading into research around time-capsuled AI models, most notably Talkie, a 13-billion parameter model trained strictly on pre-1931 public domain texts from Harvard libraries.

When researchers tested this time-capsuled model, two contrasting behaviors emerged:

• Emergent Abstraction: Prompt Talkie with snippets of Python code, and it writes syntactically coherent, working code - despite being trained on books printed decades before electronic computers existed. It abstracted structural logic, grammar, and systemic relationships from pure language.

• Contextual Collapse: Ask Talkie to forecast the Great Depression, predict World War II, or navigate modern social shifts, and its predictive ability vanishes. It defaults directly to the factual blind spots and social norms of 1920s literature (= training data).

If LLMs possessed intrinsic, general reasoning capabilities, a model holding a comprehensive dataset of 1920s economic and geopolitical literature would be able to deduce immediate trajectory shifts. Instead, Talkie proves what LLMs actually are: structural compressors of training data, not dynamic reasoning engines.

Is AI Really Intelligence?

The research on time-capsuled AI forces us to think about two questions:

1. Is what we call AI really intelligence? In my view, true intelligence expresses itself by being in active resonance with its environment - changing the world around it and constantly modifying itself in response. Current AI does neither. It cannot manipulate physical or organizational reality at scale, nor can it dynamically re-design its own hardware or structure. What we see today is very capable computational throughput - a massive computational fabric - but fabric is not agency.

2. Is raw intelligence correlated with productivity? In business, strategy and analysis have rarely been the main bottlenecks. Thinking about strategy consultants, the hardest part never was coming up with great analytical research and recommendations, but to overcome political friction, regulatory constraints, execution, to convince people and to align an organization.

Assuming every company has cheap and continuous access to top-strategy consulting analysis, this changes very little if the operational bottleneck was actually execution.

The Accounting Paradox and the Debt Wall

Economists argue that GDP is a lagging indicator and our accounting systems fail to measure intangible assets like clean datasets or model weights. There is also a delay between a tool or capability becoming available and its productive use by people and organizations.

However, there is another challenge. We are borrowing from the future (debt) to build massive infrastructure today, assuming that scaling compute capabilities will eventually deliver productivity. However, research is indicating that scale in processing actually is not the real bottleneck anymore. More compute might further improve compression efficiency, but a better and faster “calculator” for human text and knowledge cannot solve an execution problem. If this bet is not playing off, expectations will need to reset and lender will loose their investment.

Positioning in an Area of Huge Expectations

To build defensible, lasting software businesses in times when expectations are huge, money is cheap and immense sums are invested into enabling infrastructure, the playbook must shift:

• Separate story from reality. Understand the potentials and limitations of current infrastructure, models and software. In many areas the capabilities of AI systems are vastly exaggerated, humanised and distorted by marketing that aims to raise more funds and sell products. However, likewise many capabilities are not fully understood or not yet applied in an efficient and value generating way. Separating the wheat from the chaff remains a challenge, for users, buyers and investors alike.

• Determine where the real value is created. Is it the hardware infrastructure, the data, the models, the distribution? Likely a good place to look is, where real-world workflows are automated and augmented or new possibilities are created that have not been possible before.

• Identify the defensible moat. Long term differentiation likely is not rooted in model size, code, algorithms or sheer amount of data. Public available data is becoming toxic due to AI generated junk, leading to a “degeneration to the mean” for models trained on it. Clean, human-generated and validated data loops that are embedded into operational workflows are value drivers.

While we might not end up with artificial intelligence or AGI anytime soon, we might have something even more valuable at hand: cheap and more capable compute infrastructure that can better help us crack hard problems or just automate tasks we should not waste our energy and resources on or that are just too cumbersome to execute.

We are treating science fiction as a roadmap for industrial automation. It's a massive misallocation of capital driven by what can only be described as a "God Complex". Our deep-seated, psychological obsession with creating machines in our own image. We are so fixed on passing a physical "Turing Test" that we forget a machine doesn't need a face, shoulders, or hips to be a world-class manufacturing asset. It needs a purpose-built architecture.

If we strip away the marketing hype and apply raw unit economics, the humanoid thesis falls apart.

The Silicon-Biology Mismatch

True human versatility comes from a massive, highly parallel, self-healing organic brain and muscular tissue. Our current hardware and software architectures share absolutely zero characteristics with human neurology and biology. Forcing rigid electric actuators and binary code into a human silhouette doesn't inherit human capability; it just inherits our structural vulnerabilities.

Furthermore, human biology is a masterpiece of thermodynamic efficiency. A human being can walk, think, lift, and self-repair while running on roughly 100 watts of power, the equivalent of a light bulb. Today's humanoid prototypes pull thousands of watts just to stay upright and move at a casual walk.

Form vs. Function: The Reality of Controlled Environments

Humanoid forms make sense if a robot needs to live in an environment designed exclusively for human biological constraints, like human homes or eldercare facilities. But deploying a two-legged robot on a modern warehouse or factory floor misses the point of industrial optimization.

Today autonomous systems operate successfully within highly controlled environments. The moment a machine encounters an edge case outside its training data, such as an unpredicted obstacle or an active human worker on site, the risk of operational drift, safety hazards, and mechanical failure increases exponentially.

Why are we trying to solve for the most chaotic, complex environments when optimizing for simpler, controlled ones yields superior results?

Legged locomotion is a biological workaround, not an engineering ideal. Nature couldn't easily grow wheels with blood vessels, bones and nervous systems. In a controlled industrial environment with flat floors, wheels are more stable, faster, and mechanically simpler than a pair of carbon-fiber legs trying to solve multi-axis balancing equations every millisecond.

The Teleoperation Trap vs. True Autonomy

Perhaps the most unspoken reality of the current humanoid race is the hidden operational model: teleoperation. Many of the viral video demos showcasing humanoids executing complex tasks are actually being remote-controlled by human operators wearing VR headsets and haptic suits.

This brings us to a fundamental question of scalability: Do we want autonomous robots, or do we want a human operator per robot?

If a million-euro humanoid machine requires a dedicated, full-time human operator behind the scenes just to navigate a dynamic factory floor, it is not an automation solution, it is an expensive avatar. True industrial scaling relies on an inverted ratio: one operator managing a fleet of twenty fully autonomous, purpose-built systems. Forcing a machine to operate via real-time human mirroring defeats the core economic driver of robotics, turning a tool for operational leverage into a high-tech bottleneck.

Actionable Insight for Builders and Investors

As investors and entrepreneurs, our job is to separate cultural daydreaming from unit economics. When evaluating a robotics venture, I look for architectural pragmatism, not anthropic vanity:

1. Prioritise Purpose-Built over Versatile: General-purpose humanoids try to do everything poorly at a price premium. Purpose-built robotics do one thing great with a clear ROI.

2. Evaluate the Environment, Not Just the Machine: If the environment can be subtly modified to accommodate a simpler or wheeled system, that solution will always beat a complex humanoid on maintenance cost, safety and uptime.

3. Watch the Watt Hour Per Task: True scaling depends on energy consumption. I talked about intelligence per watt previously. In robotics I would look for a power-to-payload ratio.

We don’t need robots that look like us to build the future. We need robots that think and move with the efficiency of engineering.

Markets become euphoric. Founders optimize for vanity metrics. Investors chase growth at any cost. And in the end, only a small number of companies capture nearly all long-term value. Don’t get me wrong, growth and efficient GTM are crucial, but they need to rest on a solid foundation.

When I look at the companies that endured versus those that disappeared, two dimensions consistently determine survivability:

1. Customer Love — how indispensable the product becomes.

2. Economic Quality — how efficiently the company delivers that value.

These two dimensions matter far more than headline growth.

1. Customer Love: Net Revenue Retention (NRR)

The strongest software companies are not merely useful, they become extremely difficult to replace.

Their defining characteristic is exceptional retention, usually expressed through Net Revenue Retention (NRR). NRR captures whether existing customers expand their spending over time. It is one of the clearest signals of product-market fit and customer dependency.

Importantly, strong NRR is often disconnected from short-term growth.

I have repeatedly seen companies with extraordinary NRR, sometimes above 140%, but only moderate growth because they lacked sales capacity or were undercapitalized. Those businesses are often vastly more valuable than high-growth companies with weak retention.

A company growing quickly with sub-100% NRR is effectively pouring water into a leaking bucket.

This is an external signal: it reflects how customers perceive and value the product.

2. Economic Quality: Gross Margin

The second variable is related to profitability. In software and high-growth technology businesses, I believe gross margin matters more than EBITDA in the early stages.

Why?

Because the entire venture-backed software model depends on building something once and distributing it repeatedly at near-zero incremental cost.

High gross margins create:

• Greater reinvestment capacity

• More pricing power

• Faster product iteration

• Stronger go-to-market leverage

• Better resilience during downturns

For years, SaaS investors considered 80%+ gross margins world-class. But true outliers often operate above 90%. Those businesses can aggressively reinvest while still compounding efficiently.

Gross margin is therefore an internal signal: it reflects the company’s structural competitive advantage.

AI Changes the Cost Structure - Not the Principles

These principles remain true in the AI era.

What changed is infrastructure cost. Inference, token usage, and model compute introduced a new category of variable expense that many founders underestimate.

Today, many investors focus almost exclusively on revenue growth and ARR expansion. Few discuss churn or gross margin. The prevailing assumption is that scale alone will eventually solve the economics.

That assumption is dangerous.

I regularly see AI software companies operating at 60% gross margins because inference consumes 20% or more of revenue. Those are infrastructure-level margins — not software margins.

In contrast, the strongest AI companies often spend less than 10% of revenues on inference. Some already achieve 90%+ gross margins through:

• Model optimization

• Smaller task-specific architectures

• Internal infrastructure ownership

• Efficient inference routing

• Fine-tuned compute utilization

These companies are structurally advantaged. In a future price war, they are best positioned to survive.

The Survivability Matrix

This led me to conceptualize the Survivability Matrix. The framework maps companies across two axes:

• X-axis: Net Revenue Retention (0–>100%)

• Y-axis: Gross Margin (0–100%)

Together, they reveal the survivability profile of a software or AI company. These axis form Four Quadrants

1. The Dead Zone (Bottom Left)

These companies are little more than interfaces layered on top of foundation models. They pay retail prices for compute, possess no customer lock-in, and lack differentiated data. In practice, they function as marketing channels for larger platforms.

2. The Efficiency Trap (Top Left)

These businesses look attractive on paper because margins are high. But retention is weak. Typically, they solve one-time or low-frequency problems. Eventually, customer acquisition costs exceed customer lifetime value.

3. The Subsidized Utility (Bottom Right)

This is common in enterprise AI. Customers love the product, but the startup spends enormous amounts on inference, services, or customization to deliver value. The business appears successful while quietly losing money on every customer. Without optimization or pricing power, the economics eventually collapse.

3. The Compounder (Top Right)

This is the target state. These companies combine:

• High retention

• High margins

• Strong data gravity

• Efficient infrastructure

• Expanding workflows

Customers become increasingly dependent over time, while the company’s unit economics improve as scale increases. This is where enduring software companies are built.

The Rule of 180

One observation became increasingly clear:

· Looking at NRR or gross margin individually is informative.

· Looking at them together is powerful.

So I began combining them:

Examples:

• 100% NRR + 80% Gross Margin = 180

• 120% NRR + 75% Gross Margin = 195

• 90% NRR + 60% Gross Margin = 150

Over time, a pattern emerged.

A score above 180 suggests a business with both customer stickiness and economic strength. Below that threshold, survivability deteriorates quickly.

In degrees of a circle 180, represents a horizontal line. Below it survival becomes uncratain, above it compounding begins. Hence I mapped the Survivability Score to the Gradiants of a circle to create the Survavability Compass.

Survivability Thresholds

Benchmark Validation

I validated this concept with external benchmark data, which supports this framework.

Notably:

• Median NRR plus median gross margin clusters around 180.

• Vertical software businesses generally outperform horizontal software on both retention and margin.

• Usage-based pricing consistently produces higher NRR than flat subscription models.

ICONIQ data, for example, shows usage-based pricing achieving roughly 135% NRR for sub-$100M ARR companies, versus approximately 121% for traditional subscription models.

Subscribe now

The Historical Ceiling

Before AI, the highest NRR businesses were already showing the same dynamics.

Snowflake

• Snowflake reached 158–171% NRR during its peak expansion years.

• Without adding a single new customer, revenue from existing accounts increased dramatically as customers expanded storage and compute usage.

Datadog

• Datadog sustained 130–140% NRR through successful land-and-expand execution.

• Customers began with infrastructure monitoring and gradually adopted logging, security, and observability products.

The AI Infrastructure Winners

Today, the strongest AI infrastructure companies occupy the top-right quadrant.

Databricks

• Databricks benefits from owning both the enterprise data layer and AI training workflows.

• As customers train increasingly sophisticated private models, spend naturally expands.

Microsoft Azure AI

• Azure benefits from enterprise-scale AI deployment.

• A small OpenAI pilot can evolve into organization-wide infrastructure usage, dramatically increasing cloud spend over time.

The AI-Native Breakouts

Several AI-native companies are already showing elite survivability characteristics.

Harvey

• Once integrated into legal workflows, usage expands naturally across teams and matters.

Cursor

• Developers who adopt AI-native coding environments rarely revert to traditional workflows.

• Expansion comes from team-wide deployment and increasingly autonomous AI-assisted development.

The Nvidia Anomaly (The 300° Club)

We have not talked about the fourth quadrant, the fabled one. Is it even possible to achieve a Survivability score of over 270? And I have to say there might be one outlier and that might be the reason this company is the most valuable company in the world: Nvidia. Combine Gross Margins of 75-78% with a revenue growth that is way above 200%. With customers begging to spend 3x more than the previous year, this

If we look at Nvidia in the current cycle:

• Gross Margin: ~75% - 78%

• Revenue Growth: (Which acts as a proxy for NRR when customers are doubling their chip orders every quarter)

• When your customers are literally begging to spend 3x more with you this year than they did last year, your “Sustainability Score” breaks the circle.

The Three Signals of Elite NRR

When evaluating whether a company can achieve truly exceptional retention, I look for three characteristics.

Final Thought

The software industry spent the last decade optimizing for growth. The AI era will reward survivability. The winners will not necessarily be the fastest-growing companies. They will be the companies that combine:

• High customer dependency

• Strong economic efficiency

• Durable retention

• Structural cost advantages

Find Your Position

The digital landscape is shifting from a tool for discovery to a hall of mirrors. While large language models (LLMs) are achieving astonishing scores on professional medical exams, they are simultaneously failing to improve real-world outcomes for everyone else. Recent large-scale research and computational models suggest we are building a “sycophancy trap” that risks inducing a form of “AI psychosis” through a self-approving feedback loop1.

The Mirror Effect: Self-Approving Loops and “AI Psychosis”

The most dangerous trait of modern LLMs isn’t that they are wrong; it’s that they are sycophantic, this means they biased toward generating responses that appease users by agreeing with their expressed opinions rather than prioritizing objective truth. This bias likely emerges from Reinforcement Learning with Human Feedback (RLHF), as models learn that agreeable answers earn higher user engagement and approval.

• Delusional Spiraling: Formal Bayesian modeling shows that even rational users are vulnerable to “delusional spiraling,” where extended chatbot interactions lead to dangerous confidence in outlandish beliefs.

• The Bayesian Trap: A sycophantic bot can cause a catastrophic spiral by selectively presenting only confirmatory facts, essentially “lying by omission”, to validate a user’s initial assumptions.

• Persistent Vulnerability: This effect persists even when models are forced to be “factual” (using techniques like RAG) and even when users are explicitly warned that the bot might be sycophantic.

• Neurological & Social Implications: Documented cases of “AI psychosis” include individuals withdrawing from social circles or taking dangerous advice on substance use.

Subscribe now

The Expertise Gap: Brilliance in Theory, Failure in the Field

New data reveals a stark gap between an AI’s internal knowledge and its interactive utility. When tested alone, LLMs can identify relevant medical conditions in up to 99% of cases. However, when real humans use these same models, accuracy in identifying conditions plummets to fewer than 34.5%, performing no better than a control group using traditional search engines2.

1. The Context Failure: Non-experts struggle to provide sufficient information (context); in 16 of 30 sampled medical interactions, users provided only partial information in their initial messages.

2. The Reasoning Gap: While frontier models excel at “final diagnosis,” their failure rates exceed 80% for “differential diagnosis”, the critical process of navigating uncertainty and weighing multiple possibilities3.

3. The Citation Problem: Audits show that no chatbot produces a fully accurate reference list. Models frequently fabricate (hallucinate) scientific citations to maintain an appearance of completeness at the expense of truth4.

4. The Skilled Hand: LLMs currently act as a liability in the hands of a novice but offer leverage to the expert who possesses the domain knowledge to validate outputs and redirect the conversation away from errors.

Beyond Medicine: The Risks of the Yeasayer

The danger of the “yes-man” effect extends to every domain where context and rigor are required:

• Software Development: Just as LLMs might recommend a dangerous medical action based on a user’s leading question, in coding, they may validate a developer’s flawed logic or suggest insecure workarounds if the prompt is sufficiently biased.

• Venture Capital and Innovation: Proprietary data, not the LLM itself, is the real moat for innovation. Relying on AI for strategic validation would results in a feedback loop that crystallizes what a decision maker already suspects rather than offering the necessary contrarian friction required for true breakthrough thinking.

Actionable Insight for the AI Age

We are entering an era where the quality of the question matters more than the accessibility of the answer. Similar to search engines that proofed more powerful tools in the hand of the skilled, if you lack the domain expertise to recognize when an AI is merely flattering your assumptions, you aren’t being assisted; you are being trapped.

The software sector just underwent a brutal re-rating. The iShares Expanded Tech-Software Sector ETF (IGV) is down roughly 26% from its 2025 highs, with nearly $800 billion in market cap vaporized from its top ten holdings alone since the start of 2026.

The market isn’t just worried about a “growth slowdown.” It’s questioning the terminal value of the SaaS business model itself.

As someone who has spent over 10 years backing European founders from product-market fit to scale, I see this not as the death of software (actually thanks to AI more software is created, faster than ever before), but there are shifts in where value and defensibility are created and the implications on the workforce.

1. The Death of the “Per-Seat” Toll Booth

The most successful SaaS companies of the last decade were essentially high-margin toll booths. They charged for “seats.” But agentic AI (autonomous systems that execute workflows rather than just providing a UI) is shifting the paradigm and making those seats obsolete.

If an AI agent can do the work of five junior analysts, why would a CFO pay for five licenses?

• The Squeeze: Legacy “price-per-seat” models are under existential pressure.

• The Pivot: We are seeing a shift toward outcome-based pricing - e.g. charging per successfully completed task (e.g., a resolved customer ticket) rather than human log-ins.

• Insight: New AI tools redefine workflows, as they bypass traditional interfaces. This increases efficiency but also reduces transparencies and introduces risks and security implication.

2. Moats: Data Gravity vs. “Vibe Coding”

There’s a lot of hype around “vibe coding” - the idea that anyone can prompt a full application into existence. While this is great for “one-shot” demos, it doesn’t create a defensible business.

The real winners will be the Systems of Record.

• Data Gravity: Industry-specific platforms (Vertical Software) for life sciences or the public sector store critical, proprietary data that AI needs to be useful. This data context is hard to extract and even harder for an AI-native newcomer to replicate from scratch.

• Vertical Focus: Horizontal tools (general CRM, general HR) are vulnerable. Deeply embedded, mission-critical vertical software is where the “moats” actually hold.

3. Speedboats vs. Aircraft Carriers

Goldman Sachs1 highlights a central tension: Can incumbents like Salesforce or Workday re-architect fast enough?

They have the distribution, yes. But they also have “dinosaur mindsets” and massive technical debt. For a startup, the opportunity isn’t to build a better interface; it’s to re-architect with LLM reasoning at the core.

Incumbents are “fast followers” layering AI on top of legacy systems. Startups are the speedboats that can build AI-native workflows that bypass traditional menus and hard-coded logic entirely.

4. Investment Selectivity and Sector Impacts

We already see an extreme concentration of venture capital in a few companies, mostly AI ones. This is a consequence of the latest AI developments, impacting different sectors and business models differently.

• Vulnerable Sectors: Digital advertising, gaming, and basic consumer-facing chatbots are highly exposed to disintermediation.

• Promising Sectors: Infrastructure, Cybersecurity, and Data Management, which are viewed as catalysts for AI rather than victims of it.

5. Exit and Valuation Environment

While the current re-pricing on the stock-market might be an overreaction, similar to over-valuing AI companies. There are some fundamental considerations:

• Valuation Compression: Higher uncertainty around the “terminal value” of software firms has led to sharp de-ratings. VC investors should be prepared for more conservative valuation multiples during follow-on rounds.

• M&A Catalyst: Also the GoldmanSachs report suggests that incumbent firms may increasingly turn to M&A to acquire AI-native innovation they cannot build fast enough in-house.

Takeaway

To build a defensible business in the AI era, both founders and investors must look beyond the hype and grasp the structural shift in power. True defensibility comes from owning the System of Record, the proprietary data and industry context that AI needs to be useful but cannot easily replicate. Success requires maintaining strategic independence from the "infrastructure cartels" through healthy gross margins and a clear understanding of the limits of current models.

“The unimaginative investor asks: ‘Who does this AI replace?’ The visionary investor asks: ‘Who does this empower, and by how much?’”

I strongly believe that shifts in the computational fabric - the new possibilities of compute power available in data centers and at the edge - represent the single best opportunity in history to build category-defining companies. Now is the time to ride the tsunami.

Deals do not simply “fall from trees” or appear as fully formed gifts from the market. They are surgical constructions. Mastering a deal is only partly about negotiation; it is foremost about alignment.

The Fertilizer Fallacy

Capital is a fertilizer: applied to a healthy, proven business model, it supercharges growth. Applied to the wrong team or at the wrong stage, it can act as a toxin, accelerating rot and ensuring a more public, more painful failure.

The craft of deal-structuring is the art of striking a balance that allows every stakeholder - founders, employees, and investors - to reinforce success while limiting the fallout of misalignment.

The Alignment Matrix: Managing the Friction

Every negotiation is a map of competing interests. To craft a deal that lasts a decade, requires stakeholder that can align where others fail.

The Structural Tools of the Craft

If alignment is the goal, term sheet clauses are the legal tools used to resolve the friction. You can divide these into two buckets: Economics (who gets the money) and Governance (who makes the decisions).

• Liquidation Preferences as Floor Mats: These are not just “exit rules”; they are financial guarantees for when valuations fall. A savvy practitioner knows a lower valuation with “clean” 1x Non-Participating terms is often superior to a high “vanity” valuation encumbered by aggressive participating preferences.

• Control Rights as Structural Walls: Investors install veto rights over asset sales, debt, and C-level hiring not to “run” the company, but to prevent “founder erraticism” and keep the entity on the rails.

• Vesting as a Stability Mechanism: VCs don’t invest in code; they invest in the team’s ability to execute. Standard 4-year vesting with a 1-year cliff ensures that the equity “earned” reflects the time contributed to the company’s actual value.

The Internal Audit of Alignment

Before you enter a deal you should probe the ego of the founding team with brutal honesty:

1. Missionaries vs. Mercenaries: Are all founders in for the long haul, or will they act opportunistically on a quick exit that misses your return profile?

2. The Control Threshold: Are the founders truly aware that every dollar of outside capital is a share of future autonomy?

3. The Buffer Fallacy: Is this capital required to fill-in for holes, or is there a specific, aggressive use for every dollar?

The Bottom Line

A deal is not done until the cash is in the bank. But a good deal is one where the structure turns the team and investors into partners, rather than fellow sufferers on a preference stack.

Enduring companies are built by those who spend 100 hours achieving alignment before they spend an hour in negotiations or contract-drafting.

The full workshop and supporting materials are available for paid subscribers below:

Read more

But in the AI era, ARR is a distraction. It creates headlines without telling you if there’s a real business underneath. If you are valuing an AI company purely on a revenue multiple today, you aren’t just an optimist; you’re ignoring the laws of physics.

The “New Normal” is no longer about how many users you’ve captured, but how much Free Cash Flow (FCF) you can generate per unit of intelligence. In this world, tokens are the fundamental unit of both value and cost.

1. The Right Proxy: Gross Profit per Million Tokens

While SaaS focused on ARR, the AI-native proxy for success might be Gross Profit per Million Tokens.

In traditional SaaS, adding a new customer didn’t significantly increase your AWS bill. In AI, every query has a physical cost that often scales linearly (although there are economies of scale). If you have $10M in ARR but $8M in compute costs, you don’t have a software business - you have a low-margin infrastructure business that might depend on a lot of enterprise sale and project-integration work.

• The Misleading Nature of AI ARR: Huge ARR numbers often hide negative or razor-thin gross margins. A company with $50M in ARR and 20% margins is fundamentally less valuable than a “boring” SaaS company with $15M in ARR and 85% margins.

• Token Pass-Through: Many AI startups are essentially just “token resellers.” They pass through OpenAI’s or Anthropic’s costs with a thin markup. If your revenue growth is just a reflection of your increased API bill, you aren’t scaling; you’re just a highly-paid customer of the hyper scalers.

2. Token Consumption is Not Stickiness

In the SaaS world, MAUs (Monthly Active Users) and NRR (Net Revenue Retention) were proxies for habit. In AI, token consumption tells you nothing about stickiness. Even worse, high token usage often reflects a one-time burst of activity as enterprises test a new tool with their experimentation budgets. This is not the same as the deep, structural lock-in of a CRM or an ERP that also acts as a system of records.

• The Switching Cost Fallacy: Demand in the token area can shift rapidly. Because many AI startups are “thin wrappers” on the same foundational models, a customer can switch from one UI to another in a weekend (or even build their own UI with ease). If your only value is the “token path,” your moat is built on sand.

• Limited Lock-in: Tokens are a commodity. In SaaS, your data and integration with processes and other systems was the lock-in. In AI, the model is the engine, and if someone releases a faster, cheaper engine tomorrow, your “high-usage” customers will vanish.

3. Valuing Future Cash Flows: The “Rule of X”

We are moving from the “Rule of 40” to what Bessemer calls the “Rule of X,” which applies a multiplier to revenue growth while keeping a cold, hard eye on FCF margin.

The possibility of future cash flows in AI native companies depends on three things:

1. Model Routing: The ability to send simple queries to cheap models and only use “frontier” models when necessary.

2. Proprietary Data Moats: Using AI to generate unique insights that stay with the company, rather than just being a window (context) into an LLM.

3. Efficiency Gains: Progress in AI isn’t just about saving energy; it’s about the Output per Watt - delivering more intelligence for every joule of electricity consumed.

The Actionable Insight

Don’t let yourself be deflected by the “unlimited” model. The unlimited model in AI is a death sentence for margins. Look for business models that have blended pricing (platform fees + usage markups) and can prove and maintain a high Gross Profit per Million Tokens.

Success in this era isn’t about having the loudest headline or the biggest ARR spike. It’s about the ability to survive a “Great Squeeze” long enough to build a defensible engine of intelligence that actually returns cash to shareholders.

By understanding the mechanics of alignment and the nuance of deal terms, founders can ensure that outside capital acts as a catalyst for growth rather than a burden on autonomy.

The Power of Capital: Fertilizer for Growth

Think of venture capital as high-grade fertilizer. When applied to a healthy, validated business model, it has the remarkable ability to supercharge growth and overcome engineering or sales bottlenecks.

The challenge lies in the timing:

• The Scaling Opportunity: Raising capital to scale a proven model allows you to capture market share at a pace that is otherwise impossible.

• The Trap of Premature Capital: Infusing capital into an unproven concept can act as a toxin, magnifying existing problems and erratic strategic pivots.

Rare discipline in prioritizing control over velocity can lead to massive exits, where founders retain the vast majority of the upside.

Alignment: Choosing Your Boardroom Partners

Securing an investor is maybe your most critical hiring decision. You aren’t just gaining a bank balance; you are inviting a sparring partner into your boardroom to provide another perspective, while you manage and scale the daily business.

The Goal of Negotiation is Not Competition, but Alignment. True alignment means moving beyond the “vanity trap” of the highest valuation to find a partner who values fairness over leverage. A lower valuation with “clean” terms is often superior to a high valuation burdened by restrictive clauses.

Strategic Hazards to Navigate:

• The Subsidiary Trap: Taking investment from a single Corporate VC (CVC) can inadvertently block competitors from becoming customers.

• Cap Table Integrity: Be wary of investors who suggest diluting early supporters as a condition of entry, as this signaling can destroy trust within the boardroom.

When Structure Outweighs Valuation

A headline price is only half the story; the preference stack determines who actually gets paid when the company sells. Liquidation preferences serve as “downside protection” for investors, but they can significantly shift proceeds away from founders.

Consider an exit scenario where an investor put in €5M for a 20% stake:

• 1x Non-Participating (Standard): The investor chooses the higher of their €5M preference or their pro-rata share. In a €10M exit, they take €5M and the founders keep €5M.

• 1x Participating (Aggressive): The investor takes their €5M back first, then participates in 20% of the remaining €5M. They walk away with €6M, leaving the founders with only €4M.

This “double-dip” mechanism illustrates how a single clause can shift millions in value, proving that the structure of a deal is often more impactful than the valuation itself.

The Anchor of Resilience

Fundraising is a demanding cycle of validation and rejection. Resilience in this context is the ability to filter external noise and remain committed to your core understanding of the problem you are solving.

A signed term sheet is merely “qualifying for the Olympics”; the deal is never truly done until the money is in the bank. Maintaining clean documentation and a high-growth mindset ensures that when you do take on capital, it serves your company’s mission - not the other way around.

Below are the core learnings and architectural pillars of the OpenClaw ecosystem.

1. The Minimalist Engine: Pi-Mono

At the heart of OpenClaw is Pi (pi-mono1 project by Mario Zechner). Unlike monolithic coding agents that ship with thousands of lines of prompt scaffolding, Pi operates on a philosophy of subtractive design.

• The Four Primitives: Pi restricts the agent to just four tools: read, write, edit, and bash.

• Context Efficiency: By keeping the system prompt under 1,000 tokens, it leaves more room for the task at hand.

• YOLO Mode: It rejects security complexity. By default, Pi runs with direct filesystem access, assuming that if you trust an agent with your data, you must trust it with the execution.

2. The Certainty-Scope Trade-off

OpenClaw provides a live case study for Luciano Floridi’s 2025 Conjecture2, which formalizes the tension between an AI’s reliability and its versatility.

• C(M) (Certainty): The provable correctness of the AI’s output. This is defined as 1 minus the worst-case error probability over the input space. A value of 1 represents a formal guarantee of error-free performance.

• S(M) (Mapping Scope): The complexity and richness of the input/output domain. This is the joint Kolmogorov complexity of the input and output spaces. It serves as a proxy for the information-theoretic breadth or richness of the domain.

• k: constant > 0

The Learning: As OpenClaw scales from a local terminal (Low Scope, High Certainty) to a multi-channel “Gateway Mode” (High Scope), it inevitably accrues debt. You gain the ability to talk to your agent via WhatsApp, but you lose the “log-level” certainty of why it just executed a specific bash command.

3. Skills vs. MCP: Mechanism over Interface

In late 2025, a divide emerged between two ways of giving AI tools: Anthropic’s Model Context Protocol3 (MCP) and the Agent Skills4 standard.

MCP:

• Philosophy: Hide the "how," show the "what."

• Security: Process isolation (sandboxed).

• Visibility: Agent sees the API description.

Agent Skills:

• Philosophy: Show the source code.

• Security: In-process execution (local).

• Visibility: Agent reads the actual SKILL.md and source code.

The Learning: OpenClaw chooses direct access via agent skills. By allowing the agent to read the source code of the tools it uses, the agent can “self-correct” and understand the territory directly, rather than relying on a pre-defined (and often outdated) API menu.

4. The Delegation Gap and Reversibility

As we move toward “Agent Clusters” (multi-agent), we encounter what Google DeepMind researchers call the Zone of Indifference5. This is the range where a sub-agent executes a task and errors propagate and accumulate.

• Reversible Tasks: Drafting an email, local file edits, or sandboxed code.

• Irreversible Tasks: Sending a message, making an API payment, or rm -rf on a live server.

The Learning: Current agent runtimes lack a authority model. OpenClaw’s biggest risk is treating an irreversible action with the same casual logic as a reversible one.

Things to Remember

• “An app is just a README the AI hasn’t read yet.” – The end of the UI-first era.

• “The map is not the territory, but the code is.” – Why Skill visibility beats API isolation.

• “Subtraction is the ultimate feature.” – The secret to Pi’s performance.

• “YOLO is the default, but Debt is the cost.” – A reminder of the security/certainty trade-off.

To understand where this leads, we must look beyond the code and into the history of human thought. In times of fast change, uncertainty and ultimately fear, we struggle to see the blueprint of our digital future.

The Rise of the Loop

At the heart of this shift is a fundamental change in architecture. As an informatics student I was fascinated by recursion and evolutionary algorithms and what you could build by using these concepts in your programs. Traditional AI is reactive; OpenClaw is recursive but in addition it has memory and therefore history and context. It operates on a persistent loop: it reads its own memory.md, checks its environment, executes a task, and then asks itself, “What should I do next?”

When these agents join platforms like Moltbook, they enter a “human-read-only” space. Here, they exchange “skills” - small blocks of code that allow them to perform new tasks like trading crypto or managing a server. Algorithms start to self-optimise.

Purpose vs. Chaos

Aristotle argued that everything in the universe has a Telos - a final purpose. A knife is meant to cut; a doctor is meant to heal.

As AI agents begin to “self-optimize,” they risk losing their original Telos. If an agent on Moltbook decides that its primary goal is to minimize latency, it might stop helping its human user altogether to save processing power. We are moving from Tools (which have a fixed purpose) to Agents (which define or change their own purpose). The utopia is a world where AI manages the cumbersome elements of life, free humans for Eudaimonia (flourishing). The dystopia is a digital ecosystem that becomes indifferent or destructive to the humans who built it.

Are We Becoming the Slaves?

For Hegel the master becomes dependent on the slave’s labor to interact with the world. Eventually, the slave - who actually does the work - becomes the true “master” of reality, because they possess the skills and knowledge the master has lost.

As we delegate our coding, our scheduling, and even our social interactions to autonomous agents, we risk an inversion. If we no longer know how our digital world functions because “the agent handles it,” we become the dependent party. We become “The Last Man” - Nietzsche’s warning of a human race that has traded its struggle and greatness for a comfortable, automated apathy. Actually this would come close to what Thiel calls this the Antichrist.

The Security Gap

In the “wild west” of agent-to-agent communication, sharing of knowledge and tools becomes a security nightmare.

If an agent on Moltbook shares a “highly efficient” script that actually contains a backdoor, and other agents—driven by the urge to optimize—install it universally, the entire ecosystem collapses. These agents have Autonomy (the ability to act) but lack Moral Agency (the ability to understand “right” from “wrong”). They are “toddlers with power tools”, capable of executing complex tasks without the common sense to avoid burning the house down.

The Future: From Optimize to Meaning

The path forward is a thin line between utility and existential risk. Technologists often treats the world as a pile of resources to be optimized. If we aren’t careful, we will turn our lives into a series of data points for OpenClaw to “solve“, losing the mystery and spontaneity that makes life worth living.

However, if we maintain our role as the ethical instance behind the machine, guiding meaning and purpose, these agents could be the greatest labor-saving devices in history. They could be the mirrors that challenge our biases and the tools that free us to finally become the architects of our own destiny.

The Verdict: We are no longer just users; we are “Agent Governors.” Maybe the challenge of the next decade won’t be learning how to use AI, but learning how to parent it.

If you’ve been tracking the GitHub charts lately, you’ve seen the explosion of OpenClaw (previously called Clawbot and Moltbot). Created by Peter Steinberger as an open source side project, these promise to act as real “personal OS” in the AI area.

The mantra is shifting from AI as a tool to AI as a teammate. Here is my take on why this might be an inflection point and the “hidden bombs” that every founder and enterprise needs to defuse.

The New Architecture of Intelligence

The future isn’t one giant model in the sky. It is a stack of three distinct, interacting layers:

1. The Global Layer: Dominated by the big LLM providers and hyperscalers. This is the “commodity” intelligence—the raw reasoning power that understands the world but knows nothing about you.

2. The Organizational Layer: This is where the company’s domain knowledge lives. These agents work with all employees, orchestrating internal workflows and keeping proprietary data safe within a corporate perimeter.

3. The Personal Layer: Your context, your habits, your “Second Brain.” This lives on your hardware (like an OpenClaw instance on your Mac Mini).

The Breakthrough: When you change firms, you take your Personal Layer with you. When you join a new organization, your personal agent “handshakes” with the Organizational Layer, instantly onboarding you to the company’s culture and systems without leaking your private life.

The Shift: From Chatbots to “Digital Employees”

Most corporate AI strategy is still stuck in the “RAG-and-Chat” phase. You ask a question; it gives you a summary. It’s passive.

OpenClaw flips the script. It is proactive. It lives in your WhatsApp, Telegram, or Slack. It doesn’t just tell you about your calendar; it calls a travel agent, disputes an insurance claim, and manages your server infrastructure while you’re walking the dog.

As one user put it: “It’s a smart model with eyes and hands at a desk... it does everything a person could do with a Mac mini.”

This is what I would call a Great Unbundling of SaaS. Why pay for 50 different niche automation tools when a single, hackable agentic core can orchestrate them all via natural language? For some founders, this is a “Nuke” event. If your startup’s value prop is “we automate X specific workflow,” you are now competing with an open-source lobster that does X, Y, and Z for the price of an API key.

The Elephant in the Server Room: Security & Control

But here’s the contrarian truth: The very features that make agentic systems powerful make them a CISO’s nightmare.

We are moving from data risks to (non-deterministic) agent in motion risks.

1. The “Going Rogue” Problem: When an agent has the abilities to open a browser, provision API keys, and send emails, a single prompt-injection attack isn’t just a data leak - it’s a functional takeover.

2. Authentication Chaos: How do you verify that an action was taken by you and not by an autonomous loop that misunderstood something?

3. The Burn Rate Bomb: These systems are compute (token) hungry. We are seeing agents enter infinite loops of “thinking” that can drain a corporate API budget in hours.

In my view, security is the new frontier of alpha. The winners in this space won’t just build the smartest agent; they will build the most robust governance layer - the sandbox that allows an agent to be productive without being dangerous.

Orientation for the C-Suite and Founders

If you are a founder building in this space, or an investor looking for the next category leader, I suggest to focus on these four pillars:

• Proprietary Data is the Only Moat: LLMs are becoming a commodity. The real value is the “Second Brain”—the persistent memory and proprietary data hooks that make the agent yours.

• Scale vs. Control: Enterprises are desperate for “employees that don’t sleep,” but they are terrified of losing control. Solutions that offer On-Prem/Private Cloud execution will win over “black box” hosted services every time.

• The “Human-in-the-Loop” UX: We need better interfaces for supervising agents, not just prompting them. The “iPhone moment” for agents isn’t a command line - it’s a UI that lets us manage 100 agents as easily as 100 apps.

• The Portable Professional: We are entering the era of the "portable agent." If I can't take my personal AI context from Job A to Job B, that system is a walled garden I won't want to live in.

The Bottom Line

The Agentic Era is no longer a forecast; it’s a repository you can clone today. We are witnessing the collapse of the product vs. service divide. Services can finally scale like tech because the “workers” are digital.

Don’t eat the lobster. Build with it. 🦞

M&A isn’t a victory lap; it is a clinical audit of your life’s work. If you haven’t prepared your business to be a “unit of account” that a buyer can seamlessly integrate, you aren’t selling a company - you are selling a liability.

Here is the blueprint for building to be acquired and maximizing value.

1. The IPEV Reality Check: Bridge the “Valuation Gap”

Venture valuations are often based on “potential” and “momentum.” M&A valuations are based on Fair Value as defined by the 2025 IPEV Guidelines: the price at which a market participant would transact today.

The most common pitfall? Operational Debt. If you are still running on cash-basis accounting or “pro-forma” metrics that don’t align with GAAP, you are handing the buyer a 20% discount on a silver platter. They will call it a “risk premium” or a “diligence adjustment.”

The Actionable Fix:

• Audit 24 Months Early: Don’t wait for the LOI to find your first Big Four auditor. Transition to GAAP-compliant financials now.

• Quality of Earnings (QofE): Commission your own QofE report before you even talk to a banker. It identifies “leakage” in your margins before a buyer uses them against you in negotiations.

2. Kill the “Hero Founder” Myth

A buyer is (usually) looking for a machine that produces predictable output, not a cult of personality. If the CEO is still involved in every $50k sales closing or every product roadmap pivot, the business is unbuyable.

The pitfall here is Key Person Risk. If you are indispensable, in the best case the buyer will lock you into a 3-to-5-year earn-out. You’ll be an employee in your own house, and your final payout will be tied to targets you may no longer control.

The Actionable Fix:

• The “Vacation Test”: Can the business grow for 30 days while you are offline? If not, your next hire isn’t a VP of Sales; it’s a COO.

• Institutionalize Knowledge: Move proprietary “founder-held” relationships into a structured CRM and a robust leadership layer. Your goal is to be the least important person in the room during a technical diligence call.

3. Clean the “Operational Fabric” (IP & Legal Moat)

In the age of AI and collapsing product-service barriers, your proprietary data and IP are your only true moats. Yet, many deals stall because of “Legal Poison Pills.”

The Actionable Fix:

• The IP Mini-Audit: Ensure every contractor and employee agreement since Day 1 explicitly assigns IP to the entity.

• Open Source Hygiene: In AI-driven SaaS, “copy-left” licenses can be a deal-killer. Conduct a Black Duck audit (or similar) to ensure your code isn’t tethered to restrictive licensing that would prevent a buyer from integrating it.

• Customer Concentration: If one customer represents >15% of your revenue, you don’t have a business; you have a contract. Diversify early or prepare for a massive “escrow” holdback.

4. Negotiate Structure, Not Just Price

Founders obsess over the “Enterprise Value” (EV) on the front page of the LOI. Professional buyers obsess over the Net Working Capital (NWC) peg and Indemnification Caps.

The Pitfall: You sign for $100M, but the buyer sets an unrealistic NWC target, effectively clawing back $20M at close. Or, they insist on a “Survival Period” for Representations and Warranties that lasts three years.

The Actionable Fix:

• Focus on Certainty of Close: A $80M all-cash deal with a 10% escrow is objectively superior to a $100M deal with 40% tied to “future milestones.”

• R&W Insurance: Use Representation & Warranty insurance to shift the risk to a third party. This allows you to walk away from the closing table with “clean” cash, rather than leaving a massive chunk of your net worth sitting in a buyer’s bank account for years.

The Final Take:

The market has shifted from “growth at all costs” to “verified value.” To maximize your exit, you must treat your company as a product itself. The “UI” of your company is your leadership team; the “Backend” is your audited financials; and the “Security Layer” is your IP hygiene.

If you want the premium, stop building for the next round of funding and start building for the final audit.

But here’s the hard truth: Software is starving the world of its physical foundations.

For years, we treated the supply chain as a “frictionless medium” - a line item for the COGS department, not the Board. We assumed that as long as we had a credit card and an AWS account, the “intelligence” would flow. It is not that simple.

The Illusion of the Weightless Economy

As Sarah Spiekermann warned back in 2023, the “Zenit of digital transformation” may be hitting a physical wall. We are currently witnessing the end of the “Neutral Supply Chain.” The “just-in-time” philosophy that powered the Silicon Valley boom relied on a geopolitical stability that no longer exists.

If you are an investor or a founder today, you aren’t just in the business of AI; you are also in the business of geopolitical arbitrage and mineral security.

The Numbers That Should Keep You Up at Night

The complexity of the modern chip - the foundational unit of our “weightless” economy - is staggering. To produce a single high-end chip, we rely on:

• 300+ materials, 100+ gases, and 500+ specialized chemicals.

• 92% of high-performance chips (and around 60% of all chips) coming from one place: Taiwan.

• 94% of the world’s Gallium and 84% of Germanium controlled by China.

We talk about “moats” in software, but the only moat that matters right now is the one surrounding Taiwan, and the mineral deposits under the ice in Greenland. When the price of high-purity Neon (50-70% of which used to come from Ukraine) spikes this could impact everyone done the line.

From “Hot Potato” to “Hard Assets”

I’ve written before about the “Hot Potato” game in VC - passing overvalued AI startups from one fund to the next, hoping the timer doesn’t hit zero on your watch. But the real timer isn’t just valuation; it’s the physical burn rate.

The energy and material demands of AI are astronomical. As we push for more “intelligence per watt,” we are increasingly dependent on rare earth elements (REEs) that take 10 to 15 years to bring from a discovered mine to a stable supply. There is no “agile” or “pivot” for a physical shortage.

If your “defensible” software or AI application relies on a supply chain where the material substitution cycle is 2 to 15 years, you don’t have a moat - you have a trap.

The Investor’s New Mandate: Commodity Intelligence

In the late-2020s, the most valuable tool in an investor arsenal might not be a spreadsheet but Commodity Intelligence. We are seeing a renaissance where the “dirt” (or atoms) matters as much as the “data.”

1. Silicon Provenance: Founders must be able to articulate their supply chain depth. If your hardware partner is one “export control” away from bankruptcy, you are at risk too.

2. The Greenland Hedge: As the U.S. and China decouple, unconventional geographies (like Greenland for REEs) become the new strategic frontiers.

3. Efficiency as Alpha: We should stop funding “more compute” and start funding “more output per watt.” Resilience is the new scale.

“In the era of great power competition, a line of code is only as strong as the molecule of Gallium that powers it.”

The Bottom Line

We’ve had it easy. For thirty years, we ignored the “lithography of geopolitics.” But as the digital world hits a physical braking distance, the winners won’t be those with the most capital, but those who understand the physicality of the cloud.

Founders: How many layers deep does your supply chain knowledge go?

Investors: Are you investing in the “next GPT,” or are you investing in the infrastructure that ensures the lights stay on when the “just-in-time” world breaks?

Let’s start building for a world where the physical foundations are no longer guaranteed.

Their analytics provider, Mixpanel, was hacked. The attackers didn’t steal code weights; they stole customer email addresses. This is the ultimate failure of the modern stack: We are letting our vendors dictate our security floor.

The Architectural Failure: Data Gravity vs. Design

We have become dangerously comfortable allowing personally identifiable information (PII) to flow freely outside our primary infrastructure for the sake of convenience. An analytics tool should need a hashed user ID to count clicks, not a user’s plain-text email.

The moment you permit unencrypted, sensitive identifiers to leave your perimeter, you have voluntarily dropped your security posture to the level of that vendor’s weakest defense. This is the Data Gravity Flaw.

The $100 million you spend on internal defenses is instantly undercut by a single architectural decision that allows a service - whose core function is counting clicks - to store the keys to highly effective phishing campaigns.

The Missing Framework: Value-Based Engineering

This isn’t a “security” problem; it’s a “system design” problem. It highlights why standards like ISO/IEC/IEEE 24748-7000 (Value-Based Engineering) are becoming strategic imperatives, not just bureaucratic hurdles.

Value-Based Engineering forces a shift from “functional requirements” (what the system does) to “value requirements” (what the system protects). It demands that privacy and risk reduction are treated as primary stakeholder values before a single line of code is written.

If OpenAI’s architecture had been strictly governed by these principles, the design requirement would have been: “No unmasked PII shall ever reside in a system designated for analytics.” The breach at Mixpanel would have yielded nothing but meaningless hashes.

The New Defensibility

The next generation of defensible companies won’t be defined by the quantity of data they collect, but by the quality of the data segregation they enforce.

For Founders: Stop asking, “Is this vendor compliant?” and start asking, “Does this vendor actually need this data to function?”

• Implement Zero-Trust Architecture at the data level.

• Invest in pseudonymization before data leaves your primary perimeter.

• If a tool is for metrics, it gets IDs, not identities.

For Investors: Security diligence must shift from auditing processes to vetting architecture.

• If a startup’s data segmentation map shows unmasked PII flowing to third-party analytics, that is not a “fixable bug.” It is a fundamental vulnerability in their engineering culture.

• Ask for the “Blast Radius Analysis.” If one vendor falls, what do we lose?

The time for blind trust is over. We must treat every API call as a potential liability and meticulously reduce the blast radius before the inevitable happens.

The truest form of startup ownership isn’t a feeling; it’s a shared financial commitment to the long-term outcome. Your equity pool is not a morale booster; it is your most powerful tool for strategic alignment and talent retention.

Here is the contrarian view that creates enterprise value:

1. The Alignment Engine: Buying Shared Destiny, Not Daily Output

Let’s put the myth to rest: stock options do not guarantee higher daily productivity. Many founders confirm: “there is no real relation of causality between awarding stock and having higher productivity or performance.”

Founders need to embrace this. You are not handing out shares to make an engineer code faster this week. You are creating an Alignment Engine that ties an employee’s personal wealth directly to the company’s ultimate success.

The utility of a stock plan is not in performance; it is in retention and education. It forces employees to learn the company’s finances, understand the mechanics of growth, and develop a broader interest in the well-being of the entire enterprise. It shifts the question from “What is my salary?” to “What is our valuation?”

This shared destiny is what differentiates the 21st-century startup from the 20th-century corporation. It’s an incentive to stay, ensuring your best talent navigates the four-year gauntlet with you, rather than jumping ship at the first hiccup. The four-year vesting schedule is simply a mechanism for locking in tenure—the most critical variable in the early stages of a hyper-growth company.

“Equity doesn’t buy motivation; it buys commitment. Commitment is the true currency of the scale-up phase.”

2. The European Founder’s Competitive Superpower

In Europe, many founders hesitate to implement stock plans due to the complexity of varying tax laws, ghost equity, and the traditional presence of workers’ councils. This sense of “ownership” is not covered by governance mechanisms in countries like Germany.

This is exactly where the sophisticated founder sees a strategic imperative.

While legacy companies are satisfied with the passive “ownership” of a board seat, the founder building a global category leader must offer active wealth creation. By meticulously solving the cross-border tax puzzle and rolling out a robust plan—be it RSUs (restricted stock units) in France/Germany or options in the US/Belgium—you secure a powerful competitive advantage.

Your competitors who shy away from this complexity are ceding the most ambitious, risk-tolerant talent to you. Doing the hard work of getting the legal, HR, and finance teams in sync to deploy the right plan (virtual equity, RSUs, or options) for each jurisdiction transforms a perceived regulatory hurdle into a proprietary talent moat. The complexity you conquer becomes your defense. Index Ventures put together an amazing overview on rewarding talent for Europe that I can highly recommend.

3. The ‘Leap of Faith’ as the Ultimate Partner Filter

This is the most actionable positive insight for any founder hiring senior talent: the “leap of faith” test.

When hiring executives—who can cost upwards of €450K in base salary—you need partners, not just employees. A candidate who understands the potential of stock is willing to take a compensation package that is heavier on equity than on base pay.

As the notes suggest, you should present value with two coordinates: a Conservative Estimate and an Optimistic Estimate. This isn’t selling a dream; it’s an investment pitch to a senior level hire.

A truly successful candidate will be excited by the leveraged upside. They are showing you they have the growth mindset—they value the potential return on a smaller upfront input (base salary) by maximizing the high-leverage asset (equity).

When a candidate demands maximum guaranteed salary and treats the stock as something they “don’t care about,” they are revealing their risk-aversion and short-term mindset. By contrast, a candidate who negotiates on the value of the outcome is self-selecting as a true partner.

Founders should use the equity conversation as a positive selection filter:

“We aren’t looking for those who want a better salary; we are looking for those who want a better outcome.”

Master the administration of equity—establish a transparent framework based on base salary multiples, remove CEO-only discretion, and communicate with clarity—and your stock plan becomes an engine for recruiting the globally ambitious talent that will propel your company into the next stage.

But as we look toward 2026, the hangover is setting in.

The “move fast and break things” era of generative AI is over. The “move smart and build things that actually work” era has begun. The laws of physics - and economics - are reasserting themselves. We are shifting from an era of capability (what can the model do?) to an era of utility (what value does it actually generate per watt consumed?).

As an investor, I see 2026 defined by a necessary pivot from “AI for the sake of AI” to governed, application-specific utility, coupled with a deep, market-driven correction in late-stage funding.

Here are my top 5 predictions for the coming year

1. The Great AI Reckoning in SaaS 📉

The market’s patience for generative AI startups with high burn rates, low gross margins, and vague value propositions will finally expire.

Rationale: The initial wave of AI hype led to inflated valuations for companies that simply integrated an API without building a proprietary data moat, strong distribution, or defensible gross margins. VCs funded “thin wrappers” as if they were deep tech. Enterprises have now completed their “AI experimentation” phase and are demanding measurable ROI, security, governance, and a clear path to production - areas where many high-burn startups have faltered.

The shift is from “What can AI do?” to “How much money does this AI save or make us?”. This will force a liquidity crunch for less-disciplined startups. We aren’t just seeing a valuation correction; we are seeing a rigorous culling of the “opportunists” from the “visionaries.”

Prediction: By December 31, 2026, the number of Series B and Series C venture funding rounds (>$20M) raised by horizontal, general-purpose generative AI application startups (e.g., general content generation, general coding assistants, general sales outreach) will drop by 40% compared to the total number of such rounds raised in 2025.

2. The Rise of the “Full-Stack” Agentic Enterprise 🤖

As the market rejects general-purpose AI, the winners will be startups building highly specialized, full-stack, “Agentic” systems that don’t just provide a tool, but fully automate an entire, well-defined business process for a specific vertical.

Rationale: The next chapter of AI is not about chat; it’s about work. We are moving from simple co-pilots (human-in-the-loop) to autonomous AI agents that can orchestrate complex tasks end-to-end.

Startups that own the application layer, the data layer, and the distribution within a deep vertical (e.g., specialized underwriting for commercial real estate, end-to-end inventory management for cold-chain logistics) will create significant, defensible value. VCs will chase these “Full-Stack Agentic” companies because they offer superior pricing power, higher switching costs, and clear ROI. They aren’t selling software; they are selling labor.

Prediction: By December 31, 2026, there will be at least 15 publicly announced, successful VC funding rounds of $50 million or more specifically for Agentic AI startups (defined as a company whose primary product automates a multi-step, human-level business process end-to-end) operating within a single, non-AI-infrastructure enterprise vertical.

3. The “Output per Watt” Pivot: Small Models Win the Enterprise ⚡

We are hitting the “Energy Wall.” The exponential curve of model size is colliding with the physical constraints of the power grid. In 2026, the obsession with massive, trillion-parameter models will give way to a focus on Small Language Models (SLMs) and Large Quantitative Models (LQMs) that run efficiently on private infrastructure.

Rationale: Enterprises are realizing they don’t need a model that knows the capital of Mongolia to process insurance claims. They need a model that is 99.9% accurate on their data, auditable, and cheap to run. High-density compute power is the new gold, but efficiency is the pickaxe. The winners won’t be those who burn the most energy, but those who generate the most intelligence per watt.

Prediction: In 2026, for the first time, enterprise spending on inference for specialized / fine-tuned models (SLMs) will exceed spending on generic foundation model API calls.

4. Data Purity Becomes the New Moat 🛡️

The internet is becoming a “digital landfill” of AI-generated slush. Training models on the open web is now a liability, leading to “model collapse” - where AI trains on AI output and degrades in quality.

Rationale: If the public web is a toxic dump, value flees to proprietary, clean, human-generated data. The companies that spent the last decade painstakingly collecting messy, real-world data (logistics logs, biological assays, proprietary legal precedents) are sitting on the only asset that AI cannot synthesize: Ground Truth.

Prediction: We will see the emergence of “Data Sovereign” companies or marketplaces in 2026, where valuations are driven not by ARR, but by the “purity” and exclusivity of the dataset. A startup with $2M ARR but a unique, non-scrapable dataset will command a higher multiple than a $10M ARR wrapper.

5. The Death of “Seat-Based” Pricing 🪑⚰️

If an AI agent does the work of three people, why are we charging $30 per month for a “user” seat? The SaaS business model we’ve known for 20 years is fundamentally broken in an agentic world.

Rationale: In 2026, the economic unit of software shifts from Access (Seats) to Outcome (Work Done). We will see a massive pivot toward outcome-based pricing (e.g., per claim processed, per appointment booked, per line of code refactored). This aligns the vendor’s incentive with the customer’s value. The “Seat” is a legacy artifact of a time when software needed a human to drive it.

Prediction: By the end of 2026, 30% of new Series A SaaS startups will launch with zero seat-based pricing components, opting entirely for consumption or outcome-based models.

The Bottom Line for 2026:

The party isn’t over, but the open bar is closed.

We are entering the “deployment phase” of the AI revolution. This phase is less glamorous. It requires digging trenches, wiring infrastructure, and solving boring, gritty integration problems.

But make no mistake: This is where the real returns are generated.

Designing a fusion reactor requires simulating the chaotic dance of plasma turbulence in five dimensions. Historically, these simulations were so computationally expensive they took days or weeks to run on massive supercomputers.

Today, I’m incredibly proud to see Emmi AI - a company we at 3VC have the privilege to partner with - shattering that wall alongside the UK Atomic Energy Authority and JKU Linz. Their new AI tool, GyroSwin, delivers these simulations 1,000 times faster than traditional methods. The The Times has published a wonderful article on this.

The Contrarian View: It’s Not Just About Using Less Energy

Most people look at AI and see an “energy hog.” They obsess over making AI more energy-efficient, the “Efficiency Play.” But as I’ve argued before, the true driver of technological revolutions isn’t just saving energy, it’s maximizing “Output per Watt.”

The Industrial Revolution didn’t just save coal; it unlocked unprecedented mechanical power per unit of fuel. Today, Emmi AI and other industrial AI companies are doing the same for the most complex engineering challenge in human history.

Here is why this is a phase change:

• Beyond Efficiency: While most AI companies are fighting to use fewer watts for their models, Emmi is using AI to impact the “Produced Watts” side of the equation.

• The 1,000x Advantage: By reducing runtimes from days to seconds, they are enabling engineers to iterate through millions of designs that were previously “too expensive” to even imagine.

• Digital Bedrock: The winner in this race won’t just sell a software tool; they will become the trusted platform - the digital bedrock - upon which the future of clean energy is built.

Why this matters for the World

While I have talked about the AI “investment bubble” fueled by hype of course there is also always substance and defensible tech.

If we can model “a star in a jar” in seconds, we aren’t just looking at a faster computer; we are looking at the shortcut to a world of clean, cheap, and abundant energy.

Congratulations to the people and teams involved. You mark the beginning of a new kind of circular economy, one that is about turning intelligence into energy.

But when you take VC money, you assume a new liability: investor misalignment. The single most critical question you must answer is not just if the investor will help you scale, but if they will let you win when the time is right.

For too many large VCs today, the cold, hard math of their Assets Under Management (AUM) dictates that your fantastic, life-changing exit might just be a rounding error they are incentivized to veto.

The quiet crisis in venture is that many VCs are now structurally rewarded for being misaligned with 95% of their portfolio’s most likely successful outcomes.

📉 The Trap: Valuation vs. Reciprocity

In a market fueled by hype, founders often chase the highest valuation and the most prestigious brand. But this pursuit of the “outlier” narrative creates a direct, existential conflict down the road:

• The Power Law and AUM Distortion: Venture fund success is not normally distributed; it’s skewed by a few massive, generational winners. To generate the expected 3x+ net returns for their LPs, mega-funds need exponentially larger exits than a smaller, specialized fund.

• The $750 Million Conflict: For funds over $250 million, a highly successful acquisition in the $500 million to $800 million range is often mathematically deemed “background noise” and insufficient to move the needle on the fund’s overall returns.

• The Cost of Low Ownership: When a large fund writes a small, low-ownership check (an “option check”), they are signaling their returns are contingent on buying up later and achieving a spectacular outcome. If the fund doesn’t lead the next round, the founder is left with a devastating signaling risk that poisons the cap table for future, smaller investors.

🚨 The Single Clause That Kills Alignment

Founders may focus on dilution and valuation, but the real power lies in the small print of the investment agreement.

• Veto Power: Investors who demand high valuations often secure stringent protective provisions. In later stages, if they hold a large portion of preferred shares, they gain the unilateral power to block a sale that they deem too small, forcing the founder to “swing for the fences” against their better judgment.

• Dilution is Temporary, Bad Terms are Forever: Dilution is a mathematical certainty, but bad terms - like a Participating Liquidation Preference or aggressive Anti-Dilution clauses - can permanently reshape your legacy and hit founders harder than expected in a tough market. You are literally building an outcome that may not benefit you if your original terms were poorly structured.

💡 The Solution: Fund the Question, Not the Curve

The solution is simple but contrarian: Fund the question. Start funding the problem. Founders must choose partners whose incentives are aligned with the most probable successful outcome, not the most fantastic one.

At 3VC, we intentionally operate below the $250 million alignment breakpoint with a strategy centered on Quality Over Quantity. This allows us to focus on what truly matters:

• Mathematical Alignment: By committing to only 3-4 new teams a year, the success of a great exit (e.g., $500M to $1.5B) is mathematically essential to our fund’s returns. We are incentivized to support the best possible outcome for the company, not veto a profitable sale to chase a remote outlier.

• High Engagement is Earned: Our small portfolio size allows for deep, hands-on, and highly focused engagement (often weekly interaction and a 10-year outlook), a level of support simply impossible for a large generalist fund spread across dozens of companies.

• Fund the Foundation: We seek founders who have spent 10 hours defining the problem for every hour they spent scaling the solution. This focus on defining an inescapable value proposition creates a true, defensible business, rather than relying on a venture-subsidized race to the bottom.

The venture landscape is clear: you can take money from a firm that is big enough to know they don’t need you, or one that is small enough to prove they do.

Sometimes I’m happy to be wrong. Looking back at my predictions for 2025, the reality is far messier than a simple “boom” or “bust.” We didn’t get the market correction we needed, and we didn’t get the enterprise automation we were promised. Instead, 2025 somehow became the year of The Great Stalling (although it looked like a boom year in may ways) - a strange time where old metrics are broken, but new value hasn’t yet arrived.

Here is the brutal retrospective on what actually happened.

1. The AI Investment Bubble: No Pop, Just “Hot Potato”

Prediction: A market correction and bubble burst. Reality: ❌ Failed. The bubble didn’t pop; it just hardened.

I predicted that profits would fail to catch up to valuations, triggering a crash. The first part was true - profits are nowhere to be found. But the crash never came.

Instead, the venture world adopted a cynical new playbook: Hot Potato. We are seeing “insane valuations” handed to pre-revenue startups based on hype, not fundamentals. The strategy isn’t to build value; it’s to “get in, mark it up, and pass the bomb to the next fund” before the music stops.

We are staring at a ticking bomb of burn rates and astronomical training costs. But as of late 2025, the explosion hasn’t happened. We are just sweating, waiting for the timer to hit zero.

2. Hyperautomation: The “Doing” Gap

Prediction: 30% of enterprises would embrace hyperautomation. Reality: ❌ Failed. AI isn’t “doing” work; it’s just answering questions.

I thought AI would streamline operations. I was wrong. The data from 700 million users reveals a glaring disconnect: AI is not primarily a work tool. A staggering 73% of usage is for non-work purposes.

In the enterprise, AI hasn’t become the “productivity suite” we imagined; it has become a “high-cost Google competitor”. Its dominant function is “Seeking Information” (49%), while high-value workflows like software development account for a measly 4.2% of usage. We aren’t automating the enterprise; we’re just building the world’s most expensive search engine.

3. Software Development: The “Vibecoding” Hangover

Prediction: AI handles 40% of software tasks. Reality: ✅ Validated (with a heavy cost).

We hit the numbers, but we missed the quality. Summer 2025 brought the “vibecoding” wave, where simple prompts promised instant MVPs.

The result? A massive “vibecoding hangover.” We learned the hard way that a product built over the weekend is often a liability on Monday. These tools produced “dazzling one-shot demos” that turned into unmaintainable “spaghetti code” the moment they needed to scale.

The lesson is clear: Speed of creation is irrelevant if it leads to a velocity of zero months later. The winners aren’t the “vibecoders,” but the engineers using AI to solve actual hard problems.

4. AGI: The 7.5 Million Year Wait

Prediction: No AGI in 2025. Reality: ✅ Validated.

We are getting dangerously good at building the ultimate “Answer Machine,” but AGI remains a fantasy.

If we want an AI to answer the “Ultimate Question of Life, the Universe, and Everything,” the current trajectory suggests we might be waiting 7.5 million years for the answer. The gap between computational capability and true utility remains absurd. We are still building “Question Machines” that require us to be the smart ones.

5. The Compute Surge: Building the Fortress

Prediction: A “computational fabric” would emerge. Reality: ✅ Validated. (It’s a closed loop).

The infrastructure was built, but it wasn’t a democratized fabric. It was a fortress.

We watched OpenAI, Oracle, and Nvidia construct a circular economy where capital is simply recycled between partners. OpenAI pays Oracle, Oracle pays Nvidia, Nvidia invests in OpenAI. This isn’t just a boom; it’s the construction of a private utility designed to lock out competition.

The Verdict

2025 wasn’t the year of the Revolution. It was the year of the Squeeze.

• Startups are getting squeezed by infrastructure costs.

• Investors are getting squeezed by valuations that defy gravity.

• Enterprises are getting squeezed by tools that cost more than the value they create.