Explore why AI feels like the internet in 1997 through Benedict Evans' groundbreaking analysis of agentic coding, foundation models, and the future of software.
AI in 1997: How Agentic Coding is Reshaping the Future
Key Insights
- Agentic coding has transformed from marginally useful to genuinely revolutionary, marking a fundamental shift in AI's practical applications
- Foundation models will likely become commoditized infrastructure rather than premium products, with value moving higher up the technology stack
- The AI infrastructure investment of $700 billion annually faces physical capital constraints similar to mobile networks in 2009-2010
- Only coding has achieved clear product-market fit today, while other applications remain unclear and highly speculative
- AI adoption mirrors the internet in 1997-98: exciting but undefined, with most transformative use cases yet to be discovered
Understanding the AI Moment: Why 1997 is the Perfect Analogy
When Benedict Evans describes artificial intelligence as resembling the internet in 1997, he's capturing something profound about technological inflection points. We're not yet at a stage where AI's full potential is understood or clearly defined. Just as the early internet was thrilling but its ultimate applications remained mysterious, today's AI landscape is characterized by enormous promise alongside substantial uncertainty.
The comparison works because both moments share critical characteristics. In 1997, the internet existed, was spreading, and clearly represented something important. Yet asking someone then to describe exactly how it would transform business, entertainment, communication, and nearly every aspect of human life would have yielded mostly confused speculation. Similarly, we know AI is revolutionary, we see it working in narrow domains, and we're investing unprecedented capital into it. But the contours of that revolution remain surprisingly fuzzy.
This uncertainty isn't a failure of vision or analysis. Rather, it's inherent to truly foundational platform shifts. The people living through these transitions, even exceptionally smart analysts, cannot predict the specific winners, losers, and use cases that will emerge. What Evans emphasizes is that while this ambiguity is uncomfortable, it's also a natural and necessary phase of technological evolution. The anxiety around not knowing what will happen next is actually a sign you're examining a genuinely transformative moment.
The agentic coding breakthrough exemplifies this perfectly. A few years ago, applying language models to software development seemed like an obvious experiment but not necessarily a game-changer. Yet it turned out to be the first killer application where AI achieved undeniable product-market fit. Developers actively sought out these tools because they materially accelerated their work. This wasn't a marginal improvement; it was transformational. The fact that this breakthrough surprised even experienced observers demonstrates how unpredictable technological shifts can be, even when you're watching them closely.
The Rise of Agentic Coding: Why Software Development is AI's Breakthrough Application
Over the past year, the AI industry experienced a dramatic convergence around a single application: agentic coding. This wasn't the inevitable outcome of some master plan. Rather, it emerged through a combination of technological capability, market demand, and iterative refinement.
OpenAI's strategic evolution illustrates this convergence. The company initially pursued an "everything all at once" approach, attempting to develop applications across numerous domains simultaneously. This scattergun strategy reflected genuine uncertainty about where AI would prove most valuable. However, as the year progressed and coding applications demonstrated clear product-market fit, resources and focus increasingly concentrated on this domain. Anthropic, with a smaller capital base, deliberately focused on coding and achieved comparable success, suggesting this wasn't accidental but rather a recognition of where the technology genuinely exceled.
What makes agentic coding different from other AI applications is the clarity of value delivered. When a developer uses AI-powered coding assistance, the benefits are immediate and measurable: less time spent on routine tasks, fewer bugs introduced, faster development cycles, and the ability to tackle problems previously constrained by time limitations. This isn't about marginal improvements. Some organizations report 2x to 3x productivity increases in specific coding domains. For a developer struggling with a particular problem, agentic coding can mean the difference between spending three hours debugging or three minutes.
This contrasts sharply with other applications. Ask ChatGPT to summarize a meeting? It works but might miss context. Use it for basic content writing? Functional but often requires significant revision. Apply it to coding? It frequently produces usable, functional code that developers can immediately integrate. The difference is crucial: coding has clear input specifications (the code you want to write), clear success metrics (does it run? does it do what you need?), and clear workflows (write, test, refine). These characteristics make coding ideal for AI amplification.
However, Evans emphasizes that this breakthrough created a capacity crisis. Demand for AI coding capabilities has exploded while infrastructure capacity has not kept pace. This supply-demand imbalance has created wild price fluctuations and availability challenges, reminiscent of early mobile data pricing. In 2009-2010, mobile networks struggled with similar problems: people discovered they could stream video on their phones, capacity collapsed under demand, and pricing became chaotic. Eventually, networks upgraded infrastructure, pricing stabilized, and consumption normalized.
We're experiencing something similar with AI today. Some developers and companies built clusters of expensive hardware and run large models constantly throughout the day. Meanwhile, most other users find the technology somewhat useful but not essential—perhaps using it once weekly for specific tasks. This massive gap between enthusiasts and the broader market reflects the imbalance between supply and demand. As infrastructure scales and pricing stabilizes, we should expect broader adoption and the emergence of new use cases.
The temporal aspect matters greatly. Agentic coding went from "interesting experiment" to "genuinely transformational" in approximately six months. This speed reflects how quickly AI capabilities can shift perception and adoption. Yet nobody in the industry could have predicted this timeline or even that coding would be the breakthrough application. A year ago, many expected the first killer applications would be in content generation, customer service, or business intelligence. Instead, the breakthrough came through a different channel entirely.
Foundation Models and Commoditization: Where Will the Value Actually Reside?
One of Evans' most important and controversial theses is that foundation models themselves will become commoditized—essentially treated as low-margin infrastructure rather than premium products commanding significant pricing power. This represents a crucial challenge to assumptions that have driven massive venture and corporate investment in model development.
The reasoning builds on several interrelated observations. First, there's the question of sustainable differentiation. Why would one foundation model be fundamentally superior to another in ways that justify premium pricing? Unlike products with clear network effects—Instagram's value increases as more friends join, YouTube's recommendation engine improves with more data and engagement—foundation models lack obvious leverage points. Yes, different models have different strengths. Some might be faster, others more accurate, some better at specific tasks. But can any single company sustain a position where their model is so obviously superior that customers willingly pay significantly more, year after year, in perpetuity?
History suggests the answer is no. Consider the personal computer market: eventually, Intel chips became commodities despite their technological sophistication. Telecommunications networks are monumentally complex infrastructure, yet ISPs compete on brutal price margins. Cloud infrastructure—AWS, Google Cloud, Azure—operates on thin margins despite representing extraordinary technical achievements. The pattern repeats: highly sophisticated infrastructure eventually becomes commoditized, with competition driving prices toward marginal cost.
Second, the chatbot interface itself appears to be a limited, version-one user experience rather than the final form of AI interaction. ChatGPT and Claude are useful for many tasks, but they're simplistic. For specific use cases, you need proper tooling. You need the correct data configured appropriately. You need the right user interface designed for that specific task. You need proper guardrails, security, and domain-specific knowledge. More fundamentally, the people who are excellent at using a tool often aren't the same people who design it. The best graphic designers aren't necessarily the best at designing design software. The best tax preparers aren't necessarily the best at designing tax software. These require different skill sets.
This insight has profound implications. Foundation models will need to be integrated into purpose-built applications tailored to specific industries, workflows, and user populations. A law firm won't simply ask a chatbot to handle legal work; it will use specialized legal software that incorporates AI internally. A financial analyst won't query a foundation model directly; they'll use financial analysis tools powered by AI. A manufacturing company won't deploy a generic chatbot; it will use industry-specific software that addresses their particular processes and challenges.
Evans illustrates this with the spreadsheet analogy. Excel, Google Sheets, and similar tools are extraordinarily powerful and flexible. You can accomplish remarkable things with spreadsheets. Many companies literally run their operations on Excel files. Yet some needs inevitably exceed what spreadsheets can handle. At that point, you need a database, a custom application, or specialized software. The same will apply to chatbots and foundation models. They're powerful and flexible, but most professional and business use cases will require something more specialized built on top of them.
The crucial question then becomes: if foundation models become commoditized infrastructure, can model providers still capture significant value? Evans argues the answer is likely no—or at least not through the models themselves. The value will shift further up the stack, to the applications, software, and services built on top of the models. This resembles what happened with telecommunications networks or cloud infrastructure: the infrastructure providers eventually competed on price and margins, while the value accrued to companies building services on top of that infrastructure.
Consider what happened with mobile networks. Companies spent hundreds of billions building sophisticated global infrastructure. That infrastructure transformed how we communicate, work, and live. Yet telecommunications companies operate on modest margins and capture a fraction of the value their infrastructure enables. Why? Because network capacity became a commodity that competing carriers provided, and competition drove prices toward cost. Meanwhile, companies like Google, Meta, Apple, and Spotify built services on top of that infrastructure and captured most of the profit.
The same dynamic likely applies to AI. Several companies (probably three to six major players) will provide foundation models. They'll spend hundreds of billions annually on compute, infrastructure, and development. These models will be powerful and essential. But because there will be multiple competing providers, because models become increasingly efficient each year, and because there are few genuine network effects or lock-in mechanisms, competition will eventually commoditize pricing. The real value will accrue to companies building specialized applications, domain-specific solutions, and integrated services using these models as components.
This creates a challenging situation for current model providers and investors. It's not that foundation models won't be important. They absolutely will be. But the economics might not sustain the current investment levels or pricing power indefinitely. Eventually, pricing and profitability will stabilize at lower levels as competition increases, models become more efficient, and consumer surplus captures most of the benefit.
The Infrastructure Investment Puzzle: Can Trillions in CAPEX Be Justified?
The scale of capital investment flowing into AI infrastructure is staggering. Major technology companies—Microsoft, Google, Meta, Amazon, Apple—are collectively investing roughly $700 billion annually in capital expenditures for data centers, compute infrastructure, and model development. To contextualize this figure, consider that the entire telecommunications industry, historically one of the most capital-intensive sectors, invests about $300 billion annually. The global oil and gas sector invests between $700 billion and $1 trillion annually.
This creates an apparent puzzle. Justifying $700 billion in annual investment requires generating commensurate returns. Yet the actual business models generating revenue from these investments remain surprisingly unclear. OpenAI's $20 monthly subscription for ChatGPT Plus serves thousands of users, but this revenue stream doesn't justify billion-dollar infrastructure investments. Enterprise AI services are still emerging. The broad commercial case remains underdeveloped.
Several factors drive this investment despite the uncertain returns. The first is competitive necessity. If AI represents the future of computing—and the stakes seem that high—then falling behind risks catastrophic competitive disadvantage. Microsoft in the 2000s, IBM in the 1990s, and Meta in the 2010s all faced periods where they risked missing transformative shifts. Companies want to avoid that fate. This creates a "fear of missing out" dynamic where rational economic analysis yields to strategic necessity. Every major technology company feels compelled to invest heavily or risk irrelevance.
The second factor is the genuine difficulty of predicting returns on transformative technology investments at early stages. In the late 1990s, companies couldn't precisely calculate the ROI of internet infrastructure investment. The value was real—the internet clearly mattered—but the specific numbers were elusive. Surveys from the Federal Reserve, Deloitte, and other institutions suggest that enterprise productivity gains from AI, while real, remain difficult to measure financially. An organization might observe that customer service improved, analytics capabilities expanded, or employee productivity increased. These benefits are real and valuable, but translating them into exact financial metrics is genuinely hard.
Evans notes that this situation resembles the 1990s internet investment wave. Companies knew the internet would be important but couldn't precisely quantify returns. They invested based on conviction about the technology's importance rather than crystal-clear financial projections. Many miscalculated. But enough got it right that massive value was eventually created, albeit often in unexpected ways.
However, there are genuine physical constraints. These companies cannot simply double their annual infrastructure investment to $1.4 trillion. Capital availability, borrowing capacity, and shareholder tolerance for massive CAPEX limits exist. At some point, returns must justify continued investment. If AI infrastructure investments consistently fail to generate adequate returns, capital will eventually retreat.
The question isn't whether to invest—the competitive pressure ensures substantial investment will continue—but rather how much. What level of infrastructure investment is sustainable given actual and projected returns? This represents one of the central economic questions facing the industry. Currently, the answer remains somewhat speculative, driven more by conviction about AI's future importance than by clear financial models.
The Unknown Unknowns: Physical Limits and Breakthrough Moments
One way AI differs from previous technological shifts is the uncertainty surrounding physical and performance limits. With mobile networks, the internet, and personal computers, the physical constraints were relatively clear. In 1995, you knew telecommunications companies wouldn't deploy global broadband overnight. You knew consumers wouldn't suddenly purchase $3,000 computers en masse. The limiting factors were understood, even if the timeline was fuzzy.
AI presents a different challenge. We don't know the physical limits of model capability, cost reduction, or efficiency improvements. A breakthrough in architecture, training methodology, or hardware could suddenly shift capabilities dramatically. A new discovery about model training could reduce compute requirements by 50% or 90%. Conversely, we might be nearing fundamental limits where scaling becomes increasingly difficult. The truth is we simply don't know.
This uncertainty matters enormously for planning and prediction. If model efficiency improves 100x over the next three years, current economics would be completely transformed. Companies currently unable to afford large model deployments could suddenly access them cheaply. New applications would become economically viable. Conversely, if improvement slows significantly, current trajectories must be revised downward.
This is why Evans emphasizes that breakthrough applications in new domains remain likely but unpredictable. Coding achieved clear product-market fit. What's next? Law? Banking? Healthcare? Manufacturing? Some domain will likely experience its own breakthrough moment where AI proves genuinely transformational. But predicting which domain or when that breakthrough occurs is genuinely difficult. It requires the unusual combination of:
- Technology maturity sufficient for real-world application
- A problem significant enough to drive adoption
- Recognition of that problem by decision-makers
- Economic incentives sufficient to drive migration
Often, the people experiencing problems don't even recognize them as problems. Middle managers in finance might struggle with cash flow prediction, not realizing they face a solvable problem. Manufacturing companies might have inefficient processes they view as inherent to their industry. In these cases, even when technology becomes capable of solving the problem, adoption happens slowly because decision-makers don't initially recognize the opportunity.
Evans illustrates this with consulting firm experience. Strategy consultants like McKinsey, Bain, and BCG provide value partly by identifying problems organizations didn't know they had. This is difficult, specialized work requiring deep industry knowledge, personal relationships, and permission to investigate across organizational silos. Even when AI becomes capable of similar analysis, adoption may depend on whether someone recognizes and articulates the opportunity. This can take years.
Beyond Software: The Enterprise Transformation Challenge
While agentic coding addresses the software development industry directly, the broader challenge is applying AI throughout the economy. As Evans notes, making the software industry more productive might be worth $1 trillion. But that's just one industry. Applying AI effectively across law, finance, manufacturing, healthcare, retail, energy, and dozens of other sectors represents a vastly larger opportunity—potentially orders of magnitude larger.
However, applying AI in these domains is qualitatively different from applying it to software development. Software development has clear workflows, explicit requirements, and success metrics. Enterprise operations are messier. Workflows are often implicit rather than documented. Processes evolved through decades of organizational history, incorporating lessons, mistakes, and accumulated knowledge. People don't follow documented procedures; they follow practical workarounds developed through experience.
This is where consultants provide value. Strategy firms can come into an organization, investigate how it actually operates (as opposed to how it's supposed to operate), identify inefficiencies, and recommend changes. This work requires domain expertise, because inefficient processes often reflect rational business decisions most people outside the industry wouldn't understand.
Evans describes this as a "hybrid question"—half AI, half industry-specific. The AI component is powerful: analyzing data, identifying patterns, making recommendations. But the industry-specific component is equally important: understanding why processes work the way they do, recognizing which changes would be valued, and determining whether recommendations align with industry economics and competitive dynamics.
This reality has driven partnerships between AI companies and consulting firms. OpenAI, Anthropic, Google, and others are partnering with Accenture, Deloitte, McKinsey, and other service providers. Why? Because these consultants understand specific industries deeply. They understand what problems organizations face, what solutions might work, and how to navigate the organizational dynamics of implementing change. AI provides the analytical power, but applying that power effectively requires domain expertise that AI alone doesn't possess.
Companies are also building AI-native software for specific vertical markets. Rather than trying to apply general-purpose AI tools across diverse domains, they're building specialized applications for legal practices, medical offices, manufacturing plants, and other specific use cases. This approach incorporates domain knowledge directly into the software, making it more useful than generic tools.
Where Will People Actually Use AI? The Automation vs. Augmentation Question
Evans poses a fundamental question that has surprised him: when he shows ChatGPT to people and asks what they'd use it for, most struggle to articulate meaningful applications. This isn't a failure of AI capability—it's a failure of imagination or recognition. People don't often think about how they could radically change their processes using a transformative technology.
Part of this reflects the challenge of imagining novel workflows. When new technology emerges, people initially use it to do existing things faster or cheaper. First spreadsheets were used to replicate paper ledgers. Early email replicated memos. The truly important shifts come when people realize they can do things previously impossible. At that moment, new problems become solvable, new businesses become viable, and new value can be created.
Evans identifies several frameworks for thinking about this:
Price elasticity: If something becomes cheaper to do, do people do more of it? This is the Jevons Paradox—efficiency improvements often lead to increased consumption rather than decreased costs. If AI makes certain analysis cheap, organizations might not reduce spending; they might instead conduct vastly more analysis, discovering insights previously inaccessible due to cost.
Barrier removal: What things were previously impossible due to cost or complexity that now become viable? The example Evans uses is Spotify. Spotify works because streaming music for $15/month became possible. Before that, you owned CDs or paid per-song; you didn't have unlimited access to all recorded music for $15/month. That business model couldn't exist before the infrastructure existed to make it economically viable.
Computational possibility: What was computationally cost-prohibitive and thus unthinkable, but now becomes within reach? YouTube and Spotify emerged when bandwidth became cheap enough. Express trains became possible when steam engines provided sufficient power. AI might enable similar breakthroughs—not in physical transportation or communication, but in analysis, decision-making, and information synthesis.
The challenge with these frameworks is that they describe general possibilities but don't predict specific outcomes. In the late 1990s, you could articulate that the internet would destroy physical distribution economics. But that had completely different implications for newspapers (devastating) versus movie studios (minimal). The framework was correct; the specific predictions would have been wrong.
Applied to AI, this suggests:
In domains with clear workflows, standardized processes, and measurable outcomes, AI will generate obvious value. This includes coding, data analysis, specific writing tasks, technical support, and similar domains.
In domains where the "why" matters as much as the "what," where creative judgment drives value, where novel approaches are preferred over standard ones, AI will be less transformational. This includes strategy, creative arts, original research, and roles requiring judgment and opinion.
In between these extremes lies enormous uncertainty. For most work, some components are standardizable and automatable while others require judgment and innovation. The optimal approach might vary dramatically across organizations, industries, and even teams within organizations.
Market Structure and Pricing: How Will AI Infrastructure Economics Stabilize?
Evans suggests that current AI pricing dynamics represent temporary disequilibrium that will eventually stabilize, resembling what happened with mobile data. In 2009-2010, some people received bills for thousands of dollars due to unexpected data usage. Networks were congested because usage exploded while capacity remained limited. Pricing was chaotic and unpredictable.
Eventually, the market rationalized. Networks upgraded infrastructure, pricing models clarified (capped plans, throttling, metered usage), and pricing reached levels reflecting actual costs and competitive pressure. The chaos wasn't permanent; it reflected the gap between explosive growth in demand and the time required for supply to catch up.
We're experiencing similar chaos with AI today. Some users spend thousands monthly on API calls. Others spend $20 and encounter no usage limits. Pricing is unpredictable and volatile. As infrastructure scales, pricing models clarify, and supply catches up with demand, this should stabilize. The equilibrium point isn't yet clear—pricing might remain relatively expensive, might decrease substantially, or might follow some pattern we haven't anticipated.
Key variables determining the equilibrium include:
Number of significant model providers: Evans predicts three to six major foundation model providers will dominate globally. If true, this creates limited competition and possible pricing power. If ten or more viable providers emerge, competition intensifies and pricing pressure increases.
Model efficiency improvements: If models become 50% more efficient annually (requiring 50% less compute for the same capability), this dramatically affects pricing economics. Alternatively, if efficiency improvements slow, current infrastructure scaling remains necessary.
Competing business models: Google, which monetizes through advertising, might price models differently than OpenAI, which relies on direct usage payments. Meta and Microsoft bring different incentive structures. This heterogeneity might prevent pure commodity pricing.
Application layer value capture: If applications built on foundation models capture most economic value (as with mobile networks), model providers might accept thinner margins. Alternatively, if models capture more value (like operating systems), margins might remain substantial.
Evans emphasizes he's genuinely uncertain about which scenario prevails. His position is: "Here's a logical chain suggesting foundation models will become commodities. Explain why they won't be." That's not a strong prediction; it's a framework for discussion. Better arguments might convince him otherwise.
What he's confident about is that the current situation—extreme supply-demand imbalance, unpredictable pricing, constrained capacity—is temporary. Markets adjust. Prices will reach equilibrium. Whether that equilibrium involves high or low prices, few or many providers, or strong or weak model provider profitability remains uncertain. But sustained disequilibrium isn't the end state.
Why These Questions Matter: The Return on Investment Paradox
Perhaps the most under-examined question is simple: will all this AI infrastructure investment generate adequate returns? Companies are spending more on AI infrastructure than the telecommunications industry spends on networks. Yet the revenue-generating businesses built on this infrastructure remain underdeveloped.
Evans notes that this parallels earlier technology cycles. In the late 1990s, companies invested in internet infrastructure without knowing exactly what revenue would be generated. Some investments failed entirely. Others generated enormous returns, but often in unexpected ways. The internet enabled Google, Amazon, Netflix, and other companies that wouldn't have been imagined in 1995. The investment eventually proved worthwhile, but at the moment when the investment was being made, that outcome wasn't predetermined.
One mechanism that captures value without direct revenue is consumer surplus. If AI reduces the cost of analysis from days to seconds, you don't necessarily charge more for analysis. Instead, you conduct vastly more analysis, discovering insights worth far more than the analysis cost. McKinsey, Bain, and BCG might handle 5x more analyses without increasing headcount or pricing, improving their economics through efficiency. Customers benefit through better analysis without paying more. The value is real; it just doesn't show up as revenue growth.
Similarly, applying AI to back-office processes might dramatically reduce costs without generating new revenue. A company using AI for cash flow prediction might reduce finance headcount by 20% and redirect those resources to higher-value work. The return on investment exists, but it comes from cost reduction and efficiency rather than new revenue.
These mechanisms are real and important. But they create a puzzle: if AI's primary benefit is enabling companies to do existing things more cheaply, where does the revenue come from that justifies $700 billion annual infrastructure investment? Some of that investment is speculative—companies betting on future applications that don't yet exist. Some reflects competitive necessity—companies investing because competitors are investing and falling behind seems worse than investing speculatively.
But at some point, realized returns must justify continued investment. How much longer the speculative phase lasts, and how returns eventually manifest, remains unclear. This uncertainty explains why Evans remains agnostic about AI's ultimate impact on different industries and organizational structures. The technology is powerful and will matter enormously. But predicting the exact contours of that impact, even for an exceptionally thoughtful analyst, remains beyond current knowledge.
The Broader Economic Transformation: Why Margin Structure Will Change
One of Evans' more fundamental observations is that new technology typically changes margin structure substantially. When spreadsheets emerged, accounting and finance functions transformed. Workers who previously spent hours manually calculating became more productive, which could mean fewer staff or more analyses per person. Either way, the economics of financial functions changed.
AI will have similar effects across numerous domains. But the direction and magnitude of those changes vary significantly by industry and function. Consider legal services. If AI dramatically reduces the time required for legal research and document analysis, what happens? Law firms might reduce staff, maintain pricing, and improve margins. Alternatively, clients might demand lower prices, capturing the benefit. Or firms might conduct more thorough legal analysis without charging more, improving service quality without raising prices.
The actual outcome likely depends on market structure and competitive dynamics. In concentrated markets with few providers, margins might expand. In competitive markets, prices might decrease while margins remain stable or shrink. Across industries, we'll likely see mixed outcomes: some sectors improving margins, others compressing, depending on competitive dynamics and whether suppliers or consumers capture the efficiency gains.
This creates profound questions about future organizational structures. Software engineers initially feared AI would eliminate their jobs. But Evans notes the reaction has been more nuanced. Experienced engineers recognize that AI amplifies their effectiveness, making them more valuable rather than less. The question isn't whether engineering jobs will exist, but how they'll evolve. Junior engineers might find career paths altered, as AI handles some tasks they'd previously learn from. Alternatively, organizations might grow engineering teams faster because each engineer can accomplish more.
For other professions, the implications are less clear. Management consulting has built an enormous business on human expertise and analysis. If AI can perform similar analysis more cheaply and faster, what happens? Consulting firms might shrink as consulting services become commoditized. Alternatively, they might adapt by focusing on higher-level strategy and implementation, using AI to handle lower-level analysis. The outcome depends on organizational flexibility, market dynamics, and whether consultants adapt effectively to new tools.
The Open Questions That Will Define the Next Decade
Evans identifies several crucial questions that remain genuinely unanswered:
First, can models achieve differentiation? Will one foundation model be so superior that customers willingly pay premium prices for decades, or will multiple competent models commoditize pricing?
Second, where will product-market fit emerge next? Coding achieved it. What's next? Which domain will find that transformational application where AI proves obviously and measurably valuable?
Third, how will organizational structures adapt? If certain entry-level work becomes automated, how do professionals learn and advance? If team productivity increases, do organizations grow or shrink? The answers affect millions of career paths.
Fourth, what will new applications and markets look like? Just as YouTube and Spotify couldn't have existed before bandwidth became cheap, what new applications become viable only because AI is cheap? What new problems become solvable? What new markets emerge?
Fifth, how will value be distributed? Will model providers, application companies, service providers, or end-users capture most of the benefit? This partly reflects technological factors but also market structure, competitive dynamics, and policy choices.
Sixth, what are the actual physical limits? Can model capabilities continue improving at current rates, or are we approaching fundamental limits? How much further can efficiency improvements extend?
These questions don't have clear answers, and Evans emphasizes that not having answers isn't a failure. It's evidence we're examining a genuinely transformative moment where the outcome remains contingent and uncertain. Clarity about the future will come not from analysis but from the unfolding of events. Eventually, the market will answer these questions through competition, innovation, and experimentation. For now, acknowledging the uncertainty and examining the chains of logic driving different outcomes represents the most honest approach available.
Conclusion
Benedict Evans' analysis fundamentally reframes how we should think about AI in 2024-2025. Rather than presuming we understand AI's trajectory, the more accurate stance is recognizing we're living through a moment resembling 1997—genuinely important, transformative, but with the ultimate implications yet to be determined.
This framework is simultaneously humbling and liberating. Humbling because it acknowledges the limits of prediction, even for experienced analysts. Liberating because it validates focusing on building and experimenting rather than waiting for clarity that may not arrive until choices are made and markets respond.
The investment, infrastructure deployment, and competitive intensity we're witnessing make sense in this context. Companies are betting on AI's importance while simultaneously facing genuine uncertainty about which bets will prove right. Some investments will fail spectacularly. Others will spawn enormous value. Most will generate returns that seem obvious only in retrospect.
For twenty years, we'll accept AI capabilities as normal, just as we've accepted that computers are part of our lives. The achievements that seem magical and impossible today will become mundane. And the truly transformative applications—the ones that create entirely new markets and possibilities—will likely come from places we haven't yet imagined. That's not a failure of vision; it's simply how transformative technologies work. And that possibility, more than anything else, explains why the world is racing forward into this uncertain future.
Original source: Why AI Feels Like the Internet in 1997 | Benedict Evans on a16z
powered by osmu.app