The New Software Landscape

We are living through what Marc Andreessen, the inventor of the Netscape browser and co-founder of a16z, recently called "the biggest technological revolution in my lifetime, bigger than the internet." This document is designed to help you make sense of that claim: what it means, what has actually changed, what parallels the internet era offers, where those parallels break down, and where the real opportunities for value creation lie.

The short version: AI is not simply a better search engine or a smarter chatbot. It is a structural shift in what software can do and who can build it.

The economics of the SaaS world that dominated the last two decades are being rewritten in real time. For a product thinker with deep domain knowledge and the ambition to build, this is one of the most significant windows of opportunity in a generation.

To understand where value is created and captured in the AI era, you need to understand the new stack: the layers of technology that sit on top of each other, from raw hardware to end-user products.

The defining difference between the internet-driven era we're exiting and the AI era is the shift from connecting information to enabling action.

The internet was fundamentally a medium of access and communication. It allowed people to find information faster, communicate across distances, and transact without physical presence. The human was always in the loop, reading, deciding, clicking, buying.

AI, and specifically agentic AI, removes the human from some of that loop for an expanding class of tasks. An AI agent does not just retrieve information about contract terms, it reads the contract, identifies issues, drafts a redline, and sends it for review. It does not just show you a dashboard of project risk, it emails the subcontractor, updates the schedule, and flags the budget variance.

John Chambers, former CEO of Cisco, who navigated both the internet boom and bust, puts it directly: AI is moving at five times the speed of the internet boom, and will produce three times the outcomes of the internet age. In the internet era, a startup would spend two years building a product before taking it to market. Today, AI startups ship in weeks and reach product-market fit in one to two quarters.

The advantage goes not to those who wait for certainty but to those who ship with conviction early. A startup that waits two years to validate its idea in AI may find the window has already shifted to a different shape of problem entirely.

The scale of investment into AI is unlike anything in the history of technology. The capital is real; the question is whether the revenue can sustain it.

This is not distributed evenly. Capital is pooling at the top, foundation model leaders command the largest checks. But the application layer is catching up: more than half of enterprise AI spend in 2025 went to AI applications, not infrastructure.

One of the key arguments against the "bubble" framing is that, unlike the internet boom, AI companies are generating genuine revenue at unprecedented speed.

AI-native companies trade at materially higher multiples than their SaaS peers.

Traditional SaaS companies now cluster in the 2.5–7× EV/Revenue range, priced as proven, stable businesses. AI startups raise at 40% higher valuations than peers at Series A (Tomasz Tunguz, Theory Ventures). The premium for being AI-native, at least in early-stage pricing, is now structural rather than narrative.

But the market is getting more discriminating. Investors are paying for demonstrated impact on revenue, margin, or unit economics, not for AI as a story. The "AI inside" trade is closing; the "AI does the work" trade is opening. Founders who can show a step-change in unit economics, cost per ticket, cost per invoice, cost per lead, are getting funded at multiples that look more like venture pricing than public-market pricing.

The SaaS business model was built on a simple, elegant idea: charge per user, per month, forever. One human, one license. Revenue scaled with headcount. The entire go-to-market motion of the last 25 years was built on this logic.

Agentic AI is dismantling this logic. A single AI agent can now perform the tasks that previously required dozens of people. Why pay for 100 seats in a marketing department when three autonomous agents can execute the same volume of work through a single API?

Salesforce is attempting to pivot with its "Agentforce" platform, charging $0.10 per autonomous "action" via Flex Credits. Adobe faces an existential challenge as its creative moat, built over decades around specialized tools, is being bridged by generative AI from all sides.

The industry is in an uncomfortable transition. The old model is broken; the new model is not yet standard. Three pricing paradigms are competing.

Much of what took an 18-month development cycle and a six-figure engineering team to build in 2020 can now be prototyped in days.

The cost to train a high-performance AI model has collapsed, from roughly $100 million for OpenAI's flagship model to the $5 million claimed by DeepSeek, to just $30 for Berkeley's TinyZero reproduction in 2025. Whatever you believe about the precise figures, the direction is unambiguous: intelligence is rapidly approaching zero marginal cost.

The practical implication: software itself is becoming a commodity. The moat is no longer the code. It is the data behind the code, the trust it takes to deploy it in a specific context, and the distribution to reach the right customer.

SaaS unbundled enterprise software into specialized point solutions, a separate tool for CRM, HR, finance, analytics, and project management. The result was software sprawl: the average enterprise uses hundreds of SaaS tools, each with its own login, data silo, and integration headache.

Agentic AI is reversing this; a rebundling is underway. Bain's 2025 Technology Report describes the emerging enterprise stack as three layers: systems of record (the source-of-truth data layer), agent operating systems (the orchestration layer), and outcome interfaces (what the human sees and controls). A single agentic system can orchestrate workflows across multiple tools without the human ever touching a dashboard. IDC envisions a future where users instruct an AI agent to "approve last week's expense reports" or "generate next quarter's sales forecast", and the agent orchestrates workflows across HR, finance, and CRM systems behind the scenes, with no direct interaction with any individual SaaS product.

The internet boom of the late 1990s rhymes with the AI boom in important ways. Understanding the parallel helps you avoid repeating history's mistakes while recognizing the genuine opportunity.

During the internet boom, enormous capital flowed into infrastructure, laying fiber, building data centers, and manufacturing networking equipment. Most of those companies failed or saw their stock price fall 80–90% from their peak. Cisco, which briefly became the most valuable company in the world at $550 billion in March 2000, is still trading well below that peak today.

The same dynamic is emerging in AI. Goldman Sachs warned in 2024 that companies worldwide are poised to spend $1 trillion on AI infrastructure with "little to show for it." Sequoia partner David Cahn identified a "$500 billion revenue hole", the gap between what AI infrastructure spending implies in revenue expectations and what is actually being generated.

This does not mean the infrastructure is worthless, Global Crossing's fiber became a gift to the cloud era, it means the value migrates up the stack, to the companies that use that infrastructure to build things people actually pay for.

The internet gave us platform wars, Microsoft vs. Netscape, Google vs. Yahoo, Facebook vs. MySpace. In each case, the winner was the company that controlled the most valuable layer of the stack, and worked to shut other companies out (the internet spawned a lot of anti-competitive activity, and anti-trust regulation followed).

The AI era has its own early platform war at the model layer: OpenAI, Anthropic, Google DeepMind, and Meta are competing to become the default intelligence substrate. The dynamics are similar, massive subsidized access to attract developers, followed by eventual monetization. But the stakes are higher and the switching costs are potentially greater, because switching your underlying model is not quite like switching your search engine.

In the internet era, the greatest value was created not by those who built the infrastructure but by those who built the applications: Google, Amazon, Facebook, Salesforce, and hundreds of vertical SaaS companies. The infrastructure companies were important, but the application builders captured more sustained value.

The same is likely true in the AI era. The foundation models are impressive and important. But the $37 billion in enterprise AI spending is going increasingly to applications, to the products built on top of the models, not to the models themselves.

The parallels are useful, and they have limits. The next section covers the five places the AI era is materially different from the internet era. Pattern-matching from the 1999 playbook past those five places will mislead you.

Together these are the places where "the AI era is just the internet again, faster" stops being a useful analogy and starts being a misleading one.

1. Speed. The internet era unfolded over a decade. The AI era is unfolding in years. A startup that waits two years to validate its idea in AI may find the window has already shifted, to a different problem shape entirely.
2. Barriers to entry. The internet required physical infrastructure, servers, bandwidth, datacenters, that created real barriers to entry. AI infrastructure is virtualized and being commoditized at speed. This means it is easier to build, and that your competitive moat needs to come from somewhere deeper than in prior paradigm shifts.
3. The human displacement question. The internet created new categories of jobs while disrupting old ones, it automated some work while expanding the workforce that could do new things. The AI era poses a more direct challenge: agents do not just assist workers; they replace workflows. The economic, political, and ethical implications differ in scale and speed.
4. Data as moat. In the internet era, network effects and data were emerging concepts. In the AI era, proprietary data is the primary moat. A model trained on generic internet data is a commodity; a model fine-tuned on 20 years of your industry's specific projects, contracts, and outcomes is a defensible asset.
5. The regulatory environment. The internet grew up largely unregulated, especially in the U.S. The AI era is arriving in a world actively grappling with how to regulate it. The EU is moving quickly with the AI Act (adopted May 2024), perhaps far too quickly given it does not largely originate this technology. The U.S. has retreated from thoughtful anti-competitive regulation and is torn between federal and state efforts. California is emerging as the most mature regulatory district in the nation.

Marc Andreessen (a16z).

Andreessen argues that AI will be fundamentally deflationary, intelligence costs will "drop like a stone" as competition intensifies, similar to how the cost of computing, storage, and bandwidth all fell to near-zero over the internet era.

Martin Casado (a16z). While acknowledging that "private valuations are crazy," Casado professes no concern about an AI bubble. "This stuff is magic. The users are real. The demand is real. The GPU usage is real." a16z has committed a $3 billion fund specifically to AI infrastructure bets.

Ben Horowitz (a16z). On the firm's AI portfolio: "It is one of the best funds, like, I have ever seen.", while also cautioning it is too early to judge.

David Cahn (Sequoia). Published the most prominent bear case in the VC world, identifying a growing revenue gap, expectations of $600 billion in AI revenue against infrastructure buildout that implies even greater spend.

Gary Marcus (AI researcher). Calls OpenAI "the WeWork of AI" and is so confident current approaches cannot achieve AGI that he bet Anthropic CEO Dario Amodei $100,000 that AGI would not be achieved by the end of 2027. His argument: without AGI, the valuations of leading AI startups stop making sense.

Goldman Sachs. Companies worldwide are poised to spend $1 trillion on AI infrastructure with limited visible return. Investment risk stems from "a handful of tech stocks accounting for an uncommonly high share of market capitalization."

• Revenue concentration. Too much capital going to too few players at the model layer, creating systemic risk.
• The DeepSeek shock. A Chinese model matching frontier U.S. performance at allegedly 95% lower cost destroyed the assumption of an insurmountable U.S. moat. The directional implication, that model intelligence will commoditize, is widely accepted.
• ROI anxiety. Enterprises ran pilots in 2025 and are demanding evidence that AI spending translates to outcomes. Some are seeing sticker shock: the cost of effective intelligence often exceeds the cost of human staff today.
• Safety vs. speed. Is the industry moving too fast for safety guardrails to keep up? Largely a regulatory/policy battle, not a product one.
• Job displacement. What happens when AI agents automate not just repetitive manual work but knowledge work, the cognitive tasks that constituted white-collar employment for decades?

Not all AI businesses are equal. The ones that create durable value share specific characteristics. It helps to think in three categories.

Type 1 · AI-Enabled, AI at the edge (weakest moat). Traditional software companies that add AI features, a chatbot on top of an existing product, AI-powered search, GPT integration in a dashboard. These are table stakes by 2026, not differentiators. The moat here is distribution and brand, not the AI itself.

Type 2 · AI-Native, AI at the core (stronger moat). Products designed from the ground up around AI capabilities, where the AI is the product, not a feature. Cursor (AI code editor, valued at $29.3B in November 2025), Midjourney, ElevenLabs. These companies can price for outcomes because the AI is doing the primary work. The moat is the workflow design, the proprietary data they accumulate, and the speed of iteration.

Type 3 · AI Vertical, specialized AI at the core (strongest moat, hardest to build). Companies that combine AI-native architecture with deep domain knowledge, proprietary data, AND earned relationships inside a specific industry. This is where the most durable value creation lies for teams with domain expertise. The moat is four-layered: the AI capability, the domain data, the trust of a specific customer base, and the relationships and credibility it takes years to build in that field. Relationships matter as much as the data, often more, because they are what get the agent deployed inside the workflow in the first place.

In the SaaS era, moats came from network effects, switching costs, and scale. In the AI era, the moat hierarchy looks like this, from weakest to strongest:

Every instance of that pattern is a candidate for an AI agent. Every industry where that work happens at scale, construction, legal, healthcare, insurance, finance, is a market where the transition from human-executed cognitive work to agent-executed workflows is inevitable. The question is only who gets there first and with whose data. For a builder with domain depth, e.g. real estate or construction data intelligence, the opportunity is to take what has been built (a passive intelligence platform that showed patterns) and evolve it into an agentic system that acts on those patterns.

Fifteen signals in three buckets. Watch the last five as a sign of pace and transformation investment speed.

The AI era is real, the revenue is real, and the opportunity is genuine. But the lesson of every major technology transition, from the PC to the internet to mobile, is that the lion's share of the value does not go to those who build the rails. It goes to those who build the trains.

The rails in this era are the foundation models and the GPU infrastructure. They are impressive, important, and largely controlled by well-capitalized incumbents. The trains are the vertical applications, the products built for specific industries, using specific data, solving specific problems for specific buyers whose workflows will need to be upgraded to the current technological level in the coming decades.

For a product thinker with deep domain knowledge in an industry like real estate and construction, where the data model is complex, the workflows are paper-heavy, and the buyers are notoriously skeptical of generic technology, the timing is unusually good. The AI capability is now accessible to drive deeper than the SaaS era could. The domain knowledge required to use it well is rare. The window to establish a position is open, but will not remain open indefinitely.

Primary references behind the data, analysis, and quoted material in this brief. We have cited the freshest accessible source for each figure; where multiple sources align, we cite the most rigorous methodology.