Industry expert explains why AI is transformative like mobile/internet but not AGI, revealing investment trends, job market impacts, and real opportunities a...
AI's Future Unpacked: Why the 1997 Internet Comparison Matters Most
Key Takeaways
- AI is as significant as the internet or mobile revolution, but comparing it to an industrial revolution lacks grounding—the 1997 internet analogy better captures where we are now
- Job automation fears are overblown: New technologies historically create new jobs while automating old ones; attempting to quantify specific percentages of automated work is "bullshit"
- Foundation models likely face margin compression: As commodity infrastructure providers, AI labs may see pricing power erode, with value accruing to the application layer instead
- Professional services investment boom reveals the real opportunity: AI labs acquiring consultancies signals where actual value creation happens—in implementation and strategy, not just model superiority
- Distribution matters more than raw capability: When products become commodities (like browsers or cloud computing), distribution channels determine winners, not technological superiority
Where AI Actually Stands: The 1997 Internet Moment
We're living through one of technology's most transformative moments, yet we're treating it like we understand the destination. The truth? We don't—and that's actually the healthiest perspective you can adopt.
The speaker's controversial yet grounded take: AI is as significant as the internet or smartphones, but not necessarily more so. While some in the tech industry believe AI's impact will rival an industrial revolution, comparing AI's transformative potential to that of the internet or smartphones is a more rational benchmark. This distinction matters because it anchors expectations in historical precedent rather than speculative superlatives.
The 1997 internet comparison is particularly instructive. It was an exciting time—the infrastructure was clearly revolutionary. Yet most internet applications hadn't been built yet, and we couldn't predict which ones would matter. We couldn't have foreseen Google, YouTube, or Netflix. We certainly wouldn't have predicted that a "search company with a weird logo" would dominate the industry, nor would anyone have anticipated that a "has-been PC company from Krypton" (Apple) would reshape technology entirely.
Today's AI landscape mirrors this uncertainty perfectly. Survey data shows adoption remains highly fragmented. Among 13-18-year-olds, only 15-20% are daily active users, another 20% use it weekly, and 60% don't use it at all. This "jagged frontier"—where AI works brilliantly in some contexts and fails spectacularly in others—defines our current phase. We simply don't know which applications will become essential and which will fade into obscurity.
The current phase of AI development is characterized by remarkable capability in specific domains paired with genuine uncertainty about broader implications. Much like accounting professionals first encountering VisiCalc spreadsheets in the late 1970s, some professionals (particularly software developers) are experiencing paradigm-shifting moments. They see AI assistance in code generation (Claude Code, for instance) as transformative—a true "before and after" moment. Meanwhile, lawyers, accountants, and consultants view the same technology with puzzlement: "Yes, this is clever, but how exactly do I use this in my practice?"
This divergence in perception reveals something critical: the impact of any technology depends entirely on whether it addresses the actual hard part of your job. For accountants calculating interest rates, VisiCalc was revolutionary because it eliminated the tedious, manual recalculation work. For lawyers, it offered no immediate value. The word processor—arriving shortly afterward—proved far more useful to legal professionals.
The same principle applies today. AI can write competent code, generate reasonable PowerPoint presentations, and handle specific analytical tasks. But it often produces mediocre versions of these outputs. When you hire McKinsey or Bain, you're not paying $100,000 for a PowerPoint deck; you're paying for deep organizational understanding, strategic insight, political navigation, customer discovery, and problem-solving expertise. The presentation is merely the visible output. This distinction between task and job proves absolutely critical for understanding where AI will create or destroy value.
The Job Automation Myth: History Repeats, But Not How You Think
The "doomer" sentiment about job losses pervades social media, yet it fundamentally misunderstands how technological change actually works. Dismissing these fears entirely would be naive, but accepting them uncritically would be equally foolish.
Historically, new technologies automate away existing jobs while simultaneously creating new ones. This process isn't magical; it's a structural economic feature driven by several mechanisms. When something becomes cheaper to produce, price elasticity typically kicks in—you don't necessarily do the same thing for less money; instead, you do more of it at the same cost, or even more for higher prices due to increased ROI.
Consider accounting as an example. Despite the introduction of adding machines, punch cards, mainframes, databases, ERP systems, spreadsheets, and cloud computing over the past century, the number of people employed as accountants has continuously risen. This counterintuitive reality reflects a fundamental truth: automation doesn't eliminate professions; it elevates their nature. The hard part shifts from routine calculation to strategic analysis, advising, and interpretation.
Software development demonstrates this pattern strikingly. Before integrated development environments (IDEs), libraries, and operating systems, developers had to write nearly every line of code themselves. Today, if you write an iPhone app, Apple provides device drivers, graphics drivers, the file system, and countless frameworks. You don't write any of that yourself. Yet we don't have one-tenth as many engineers; the demand for engineers has actually increased dramatically.
The question isn't whether jobs will change—they certainly will. The insidious part of automation discourse is the attempt to quantify exactly which jobs disappear and by what percentage. Saying "17% of a lawyer's work can be automated" is simply bullshit. We lack the analytical framework to break down complex professional services into discrete, automatable components. Lawyers aren't hired to write documents or research precedent; they're hired for judgment, strategy, client relationships, and institutional knowledge. You cannot cleanly separate the automatable portion from the irreducible human element.
Even more telling: the most advanced AI companies—Anthropic, OpenAI, and others—are increasing their human headcount, not decreasing it. If these companies truly believed AI rendered humans obsolete, they'd be rightsizing their operations. Instead, they're expanding staff dramatically. This action speaks louder than any public statement about AI's transformative potential.
The anxiety about entry-level job displacement deserves serious consideration, but context matters enormously. When Dario Amodei (CEO of Anthropic) discusses disappearing entry-level roles, he's drawing from AI expertise, not labor economics expertise. These are fundamentally different domains. Someone expert in machine learning isn't automatically expert in comparative advantage, labor markets, or macroeconomic cycles. This matters because predictions from authority figures often carry unwarranted weight when they venture outside their domains.
What we should actually be watching is how quickly industries adapt and whether implementation timelines match hype cycles. Enterprise software adoption typically takes 18+ months just for sales cycles alone. Large organizations in aerospace, healthcare, and similar sectors don't instantly rip out existing systems (like decades-old SAP installations) and replace them. Real-world transformation takes years, not quarters. This glacial pace provides crucial adjustment time for workers and institutions—it's not perfect, but it's far slower than venture-capital-fueled narratives suggest.
The historical pattern is clear: from the 1800s onward, when roughly 90% of the population were subsistence farmers whose main anxiety was crop failure, technology has continuously automated certain jobs while creating new opportunities. Elevator operators disappeared when buttons replaced manual lever-pulling, but elevator manufacturing, maintenance, installation, and building code development became entire industries. Railway engineers seemed inconceivable before railways existed. Software engineers didn't exist as a profession a century ago. This cycle of destruction and creation has proven resilient across two centuries of technological disruption.
Foundation Models as Commodities: Why AI Labs Lose the Value War
One of the most misunderstood dynamics in AI investment circles concerns where actual value will accumulate. The narrative often assumes foundation model companies (OpenAI, Anthropic, Google) will capture outsized returns. The data suggests otherwise—and that mismatch represents the most underexplored question in AI discourse.
Consider the telecommunications industry as an instructive parallel. The global mobile industry generates roughly $1 trillion in annual revenue while spending $200 billion yearly on capital expenditure (20% of revenue). Mobile data consumption has grown approximately 1,500-2,000 times since 2010—an exponential, staggering growth trajectory. Yet mobile carrier stock prices have been essentially flat for 25 years. Why? Because they're selling commoditized, low-margin utility infrastructure. The innovation and value creation happened above the infrastructure layer—at Apple, Google, Netflix, and countless app developers. The carriers captured only marginal rents despite providing essential physical infrastructure.
The foundation model business exhibits nearly identical characteristics. These companies spend hundreds of billions annually on compute infrastructure to train and run models. They're building undifferentiated commodity infrastructure, regardless of marginal differences in capability. Once you reach "good enough" performance for most use cases, incremental improvements matter less than cost and availability. This creates structural downward pressure on pricing power.
Several technical questions will determine value distribution, but the likely scenario involves foundation models following the cloud computing trajectory rather than the Windows/iOS model. Windows achieved pricing power through ecosystem lock-in—developers standardized on Windows because customers demanded it, creating a virtuous cycle of leverage. Cloud computing never achieved this. When selecting cloud infrastructure, companies primarily care about price, reliability, and feature coverage—not which cloud provider holds dominance. Developers build for AWS, Azure, and Google Cloud interchangeably because customers don't standardize on platforms the way they did with operating systems.
Foundation models appear headed toward the cloud outcome. If the optimal user experience requires building custom applications tailored to specific use cases (rather than simply using a chatbot), then the model companies won't capture value from those applications. If models become genuinely interchangeable commodities available from multiple providers at competitive pricing, pricing power evaporates. If no single model achieves overwhelming network effects (and current evidence suggests they won't), competition will persist indefinitely, further compressing margins.
This dynamic explains the surprising and counterintuitive investment trend: OpenAI, Anthropic, and other AI labs are aggressively acquiring consultancies and forward-deployed engineering firms. This seems backward—shouldn't AI eliminate the need for consultants? The truth is more nuanced. The real value in AI transformation lies not in the model itself but in implementation, strategy, and organizational change management.
Every large organization has a fundamental problem: they don't have dozens of idle employees sitting around waiting to redesign internal workflows, analyze whether operations can be automated, figure out which processes should be reimagined, or determine where to allocate new resources. That's precisely why they hire Bain, BCG, McKinsey, Accenture, or Publicis. These consulting firms provide the human expertise and bandwidth to identify opportunities, design solutions, and manage implementation. Adding AI capabilities to this process doesn't eliminate the need for consultants; it makes consulting more valuable because the scope of possible transformation expands.
The AI labs understand something crucial: the leverage isn't in owning the model; it's in owning the relationships and implementation expertise. By acquiring or investing in consulting firms, they're positioning themselves to capture value further up the stack, where the real money is made. This represents a tacit acknowledgment that foundation models will become commoditized.
The Professional Services Boom: Where AI Value Actually Gets Created
The investment in professional services by AI labs represents one of the most revealing trends in the industry, yet it receives remarkably little analysis compared to model capability comparisons.
Consider the structure of professional services work. Consulting firms don't sit around hoping for big new projects to materialize. They're hired because organizations have strategic questions: Why is customer churn too high? Where should new stores open? What should our digital transformation strategy be? How can we redesign internal workflows? How do we use AI to reimagine our operations?
These aren't one-week projects; they're multi-month engagements requiring five to ten people to spend significant time understanding the organization, identifying opportunities, designing solutions, and managing implementation. Then implementation is a separate project entirely. Large organizations can't staff permanent teams for every possible initiative—it's economically inefficient. That's why consulting as an industry exists.
The same logic applies to forward-deployed engineers. Anthropic and OpenAI are hiring forward-deployed engineers—essentially embedded specialists who sit within client organizations to help integrate and customize AI solutions. This represents genuine recognition that the gap between "we have an AI model" and "we've successfully transformed our organization to use this AI model" is vast. It requires deep domain expertise, institutional knowledge, change management skill, and sustained problem-solving.
When you contrast this with the "AI will replace everyone" narrative, the disconnect becomes obvious. If AI were truly going to eliminate the need for human expertise and implementation effort, these companies wouldn't be making multi-billion-dollar bets on consultancies. Instead, they're recognizing that implementation, change management, strategy, and organizational redesign represent the actual value creation—the model is just a tool enabling that value creation.
This doesn't mean consultants are safe forever. What it means is that consulting transformation will mirror what happened to accounting: the nature of the work shifts, but overall demand increases. Senior consultants will spend less time on routine analysis and more time on strategy and client relationships. Junior consultants might find their traditional roles changed (analyzing spreadsheets shifts toward prompt engineering and model evaluation). But the total demand for strategic consulting likely increases as organizations try to understand how to apply AI across their operations.
The critical insight: when a technology is genuinely transformative, the infrastructure itself doesn't capture most of the value. The value accrues to whoever helps organizations understand how to use that technology to solve their actual problems. Google and Apple created vastly more wealth than telecommunications carriers despite being utterly dependent on telecom infrastructure. Similarly, the AI labs might create less wealth than the software companies, consultancies, and service providers that figure out compelling applications and implementation strategies.
The Unexpected Power of Distribution: Why Google's Advantage Isn't What You Think
As foundation models become increasingly commoditized, an unexpected variable gains disproportionate importance: distribution. This principle has proven decisive across every major platform shift in technology history.
The browser wars of the 1990s-2000s illustrate this perfectly. Microsoft leveraged Windows distribution to drive Internet Explorer adoption, eventually dominating the browser market. Yet this dominance translated into remarkably little long-term value. The winning browser didn't actually matter because the value was happening above the browser layer, in applications and services. Users didn't switch browsers to get a particular website; they used whatever browser came with their operating system. Competition from Firefox, Chrome, and Safari proved that browser dominance provided no durable advantage.
Today's AI landscape shows similar distribution dynamics playing out in real-time. Google has begun pushing Gemini through all its existing distribution channels. For an average user without deep AI expertise, Gemini and Claude are roughly equivalent—unless you're using AI tools constantly all day. Meta, often overlooked by Silicon Valley insiders, represents a significant AI player simply because they have massive existing distribution and user adoption.
OpenAI spent much of last year pursuing what their team called "everything everywhere"—attempting to build a flywheel where early adoption would create inertia and switching costs before competitors (Google, Meta, Amazon) deployed their own models through their existing distribution networks. This is a distribution game, not a capability game. The company needed to get millions of users dependent on their product before established players leveraged their advantages.
Apple's delayed AI integration, showcased at WWDC 2024, demonstrates distribution power in the other direction. Apple Intelligence presented the most compelling vision of a personal AI assistant anyone has shipped or even clearly articulated: tool-using, agentic, on-device AI with no prompt injection risks, no hallucinations, and a standardized API for thousands of applications. The technological vision is genuinely sophisticated. Yet Apple hasn't shipped it yet—and neither has anyone else. This delay doesn't diminish the vision's power; it highlights how distribution matters more than shipping timeline when you have 1 billion installed devices.
The underlying dynamic: when products become commodities (which foundation models increasingly are), distribution determines winners more than raw capability. This explains why Google and Meta won't be disrupted by OpenAI despite being "late" to AI—they already have distribution that makes any adequate AI product competitive. This also explains why investors should be far more interested in companies building compelling user experiences on top of models than in the model companies themselves. The browser didn't matter; the search engines and social networks built on top of browsers mattered enormously.
The Anti-AI Sentiment: Legitimate Concerns vs. Manufactured Panic
Surveys showing AI is less popular than ice cream, protests against data center construction, and public figures being booed for mentioning AI suggest growing backlash. Understanding this sentiment requires disentangling legitimate concerns from manufactured panic and genuine misunderstandings.
The "big fuzzy mess" of anti-AI sentiment encompasses multiple distinct issues. Some are genuinely real: if a data center substantially increases local electricity costs, that's a legitimate problem. Some are exaggerated: the water usage claim sounds dramatic until you examine actual data. According to research from Lawrence Livermore Lab (2024), U.S. data center water consumption represents approximately 0.017% of total U.S. water usage. If you live in a small town with one well that suddenly gets capped for a data center, you have a legitimate grievance—but that's a planning and resource allocation problem, not a data center problem generically.
The employment impact question generates the most concern and deserves serious attention. Yet honest analysis reveals genuine uncertainty. Economic data shows employment slowdowns for 18-24-year-olds, but similar slowdowns appear across both AI-exposed and non-exposed fields, and both degree-holders and non-degree-holders. We're collecting relatively little actual data on how organizations are deploying AI and what impact it's having on hiring. Model labs won't publish meaningful usage statistics or daily active user numbers. Consultancies spend significant money surveying 20,000 people about AI usage. Academic economists attempt back-calculations from Bureau of Labor Statistics surveys. For a technology supposedly transforming the economy, we have remarkably poor data.
What we do know: there's no clear consensus showing major job losses attributable to AI at this moment. This absence of evidence isn't evidence of absence—the impact might not be visible yet. But it also might not be as catastrophic as social media narratives suggest.
The legitimate concerns warrant serious attention. Deepfakes represent a genuine new challenge compared to Photoshop era. Fifteen-year-olds can't create explicit fake videos of their entire school and distribute them in an afternoon using Photoshop. They can now. This represents a real escalation, not merely a faster version of existing problems. The capacity to generate synthetic content at scale introduces novel risks around misinformation, harassment, and abuse.
The "AI slop" phenomenon—where 30-40% of new podcasts are AI-generated, where book cover artists see their work automated, where synthetic content floods platforms—represents genuine creative industry disruption. These aren't imaginary harms; people are legitimately losing work and income.
Yet parallels to social media backlash prove instructive. Social media enabled genuine goods: LGBTQ+ individuals in isolated communities could connect with others. It also enabled harms: hate groups could organize, extremists could find community, and child exploitation material could proliferate. Both outcomes were real. Society didn't solve this by rejecting social media; it's still figuring out governance structures decades later.
Similar complexity characterizes AI. Some harms are real. Some proposed harms are exaggerated. Many critiques fall into ambiguous middle ground. The productive path forward involves neither dismissing concerns as Luddite panic nor accepting doomsday narratives uncritically. Instead, it requires sober assessment of specific risks and evidence-based policy responses.
The "AI is bullshit, I'm never using it" position feels morally superior and provides psychological comfort. It's also strategically worthless. If you're interviewing at a law firm that hired 100 associates last year but will hire only 50 this year, walking into the interview declaring your principled rejection of AI is not the optimal career strategy.
What Actually Changes When Technology Shifts: The Frame.io Lesson
One of the most underexamined questions in technology history concerns timing: when did technology enable something that seemed inconceivable before it existed?
The answer often surprises: frequently, the technology existed long before anyone deployed it to solve the problem. Frame.io provides the instructive example. The company offers video collaboration technology—allowing teams to review, annotate, and discuss video files in real-time. The underlying technology could have existed many years before Frame.io's founding. The limiting factor wasn't technical capability; it was someone realizing the problem it solved within a specific professional context and devising a compelling solution. The delay wasn't in technology; it was in the hard part—understanding the job deeply enough to recognize the opportunity.
Similarly, Google Docs' impact wasn't immediate. Real-time collaborative document editing could have theoretically existed earlier. The technology matured, adoption grew, and people invented related features and workflows. Years elapsed between basic functionality and the ecosystem of related products and practices that made Docs transformative.
This temporal mismatch between capability and impact has massive implications for AI predictions. Even assuming foundation models remain genuinely transformative, the period between "model exists and works well" and "business models have fundamentally changed and employees have adapted" involves years, not months. Large organizations don't instantly restructure themselves. Professions don't overnight reimagine their hierarchies. New workflows don't materialize because technology is available; they require discovery, experimentation, and iterative refinement.
This reality offers a degree of comfort: the dislocation won't be instant. It also provides a realistic timeline for individuals and organizations to figure out adaptation strategies. The early 2000s job market for accountants didn't collapse overnight when enterprise resource planning systems became sophisticated. The accounting profession is still adapting to cloud-based systems decades later. The adjustment process is ongoing, uneven, and often uncomfortable—but it's gradual enough that most people navigate it successfully.
Jevons Paradox and the Elevator Problem: Why Automation Doesn't Simply Save Time
One of history's most counterintuitive economic principles, Jevons Paradox, explains why technological efficiency improvements don't necessarily translate to reduced work hours or smaller workforces. The name derives from 19th-century economist William Stanley Jevons, who observed that improving coal engine efficiency didn't reduce total coal consumption; instead, it increased it by making coal-powered applications more economically viable.
The same dynamic plays out across industries. Consider the elevator: before electric elevators, manual lever-operated elevators required human operators. Once buttons replaced levers, the task became trivial—anyone could press a button. You might expect this to eliminate elevator operating jobs entirely while reducing the labor required. In reality, the change shifted the entire industry. Building codes evolved. Elevator manufacturing became an industry. Maintenance, installation, and safety protocols developed into professions. The "automated" job didn't simply vanish; the industry transformed.
Goldman Sachs associates provide another instructive example. Before Excel, junior investment bankers worked extraordinarily long hours performing complex financial modeling and analysis manually. Excel didn't reduce the workload; bankers still work long hours. Instead, the technology enabled more complex analysis, larger models, and more sophisticated financial engineering. The same tool that should have reduced hours instead enabled more work at higher complexity.
In accounting and professional services generally, the pattern repeats. Despite a century of computing advancement—from mechanical adding machines to modern cloud platforms—the number of accountants has only increased. Why? Because automation reduced the hard part (arithmetic and rote calculation) while expanding the scope of what became possible and valuable (comprehensive financial analysis, advisory services, strategic guidance).
This principle applies directly to AI. The hard parts of most professional work aren't executing the task; they're deciding what the right task is, understanding the client's underlying needs, navigating organizational politics, translating requirements into solutions, and managing change. These aren't easily automated. They require human judgment, relationship management, and contextual understanding. A lawyer's real value isn't writing a brief (which AI can approximate); it's understanding which legal arguments actually matter for this particular client's specific situation, which courts respond to particular framings, and which solutions align with the client's broader business strategy.
The implication: automation of routine components typically expands the scope of the profession rather than shrinking it. Lawyers won't disappear; the nature of legal practice will shift toward higher-value strategy and away from routine documentation. But total demand for lawyers may actually increase because more organizations will consider legal strategy worth investing in if the routine documentation work becomes cheaper.
Platform Shifts: Why Past Winners Often Lose Unexpectedly
Mobile represented the last major platform shift, and it offers crucial lessons for predicting AI's impact. The internet was transformative—and it completely failed to disrupt roughly half the tech industry. Some companies barely adapted; some thrived; some collapsed entirely.
Google barely noticed the mobile transition. Their search-dominant business translated directly to mobile because mobile users still need search. Facebook thrived in mobile—it was arguably a better platform for social networking than desktop, with cameras, notifications, and always-on availability. Amazon's business was completely unaffected; e-commerce works on desktop and mobile equally. Yahoo Mail failed to transition—a significant company that couldn't adapt its email experience to mobile. eBay struggled—while it exists on mobile, it never dominated like on desktop.
This uneven impact matters. The shift from PC to mobile was epochal, yet half the industry didn't care. Similarly, AI might be transformative overall while barely impacting some industries and companies. It's quite possible that a major internet company gets completely disrupted by AI. It's equally possible that most incumbents adapt successfully through acquisition, partnership, or internal development. It's also possible that some areas of the economy remain stubbornly resistant to AI transformation despite theoretical potential.
The historical lesson: don't assume symmetrical disruption. Assume that some companies will miss the shift entirely (and fail), some will adapt successfully, and some will thrive specifically because they've been early in adoption. The distribution of outcomes tends toward concentrating further than most people expect—winner take most dynamics emerge in some categories while commodity competition locks in others.
This unpredictability argues for humility about AI predictions while recognizing transformation is genuine. You can make reasonable arguments about likely outcomes in specific domains. But making definitive statements about which companies will dominate or how entire industries will reorganize is largely speculation dressed up as analysis.
The Pricing Power Question: Foundation Models as Utilities
One of the most consequential questions about AI's economic impact concerns pricing power for foundation model companies. Will OpenAI, Anthropic, and similar organizations maintain the ability to charge premium prices, or will their products become commoditized utilities sold at marginal cost?
The evidence increasingly points toward margin compression. Sam Altman's famous statement—"We're going to be selling AI intelligence on a meter like water or electricity"—reflects the aspiration that AI would function as a core utility with pricing leverage. But utility industries offer very different economics than Altman suggests.
The global mobile industry demonstrates this perfectly. Carriers spend 15-20% of annual revenue on capital expenditure every year, supporting exponential data growth (1,500-2,000 times since 2010). Yet mobile company stock prices have been flat for 25 years. Why? Because when you become essential infrastructure, you become a low-margin commodity business. Customers care about price and reliability, not technology company prestige. Profits compress toward marginal cost as competition persists and customers have alternatives.
The conditions that enable pricing leverage—like Windows operating systems achieved—require either genuine differentiation or network effects that lock in customers. Windows had both: meaningful technical differentiation (initially) and ecosystem lock-in (developers targeting Windows because customers used Windows created a reinforcing cycle). Foundation models show neither characteristic. Multiple providers can build models of roughly equivalent capability. Developers and users show minimal lock-in—they'll readily switch between Claude, Gemini, and GPT depending on specific use cases and pricing.
Without network effects and facing genuine differentiation challenges, foundation models should expect indefinite competition. In competitive commodity markets, pricing power disappears. Customers evaluate options based on price, performance metrics, and availability. Margins compress toward the cost of compute plus reasonable operational overhead. The 2024 observation of "$1.5 million monthly token bills" and "50k mobile data bills in 2010" are temporary price disequilibrium states. They'll normalize as competition increases and supply becomes abundant.
This dynamic doesn't mean foundation model companies can't be profitable. It means they'll be profitable at utility-level margins rather than software company margins. This still represents billions in annual revenue, but it's substantially less valuable than the software company valuations currently assigned. Where will the value accrue? To companies solving the implementation, strategy, and application layers—exactly the reason AI labs are acquiring consultancies.
The Surprising Persistence of Spreadsheet Logic: Templates as Billion-Dollar Opportunities
One seemingly mundane observation contains surprising implications: the template. If Claude or another foundation model solves a particular job extremely well but requires customization, what emerges? Organizations create templates—standardized starting configurations for specific use cases.
This mirrors Excel's ecosystem. Excel itself is a powerful general-purpose tool. But the actual value capture happens in the million small templates, specialized tools, and industry-specific applications built on top of it. Specialized financial modeling packages, inventory management systems, and industry-specific analytics all leverage Excel's core functionality while providing domain-specific solutions.
The same pattern seems likely with foundation models. While raw Claude or GPT can handle many tasks, organizations will want specialized versions: Claude for legal analysis, Claude for financial modeling, Claude for customer service, and so on. Each of these specialized applications could plausibly become a billion-dollar business. Each solves a specific job better than the generic model because it's tuned, trained, and optimized for that particular domain.
This observation suggests value won't simply be binary (model companies winning or losing). Instead, value will distribute across the stack: some accrues to model companies for foundational capability, some to companies providing specialized templates and applications, and significant value to consultancies and services organizations helping companies implement and customize these solutions.
Conclusion
Benedict Evans' framework for understanding AI emphasizes radical humility about what we actually know. We're in the 1997 internet moment—a time of obvious transformation and profound uncertainty about how that transformation will unfold. Some professionals will experience paradigm shifts. Others will barely notice change. New jobs will emerge that currently seem inconceivable. Existing professions will adapt rather than disappear.
The productive path forward isn't choosing between techno-optimism and doomsay panic. It's neither assuming AI will solve everything nor that it represents an existential threat to employment. The evidence suggests something more nuanced: genuine transformation happening at a pace slower than venture-backed hype suggests, with uneven impact across industries and professions, creating both genuine disruption and genuine opportunity.
For individuals and organizations, this reality suggests clear strategic implications: dive deep into understanding what AI can do, experiment extensively with applications, think hard about which components of your work represent the real value (as opposed to routine task execution), and remain alert to how your industry's value chain might reorganize. Don't reject AI from moral principle; that offers comfort without practical benefit. Don't assume AI will instantly solve your problems; implementation and integration take years. Instead, engage seriously with the technology, understand its limitations as clearly as its capabilities, and figure out how it fits into your specific situation.
The future remains genuinely uncertain. That uncertainty itself is valuable information—it suggests you're not missing an obvious trend that everyone else has figured out. It suggests there's still time to learn, adapt, and position yourself for whatever emerges.
Original source: A rational conversation on where AI is actually going | Benedict Evans
powered by osmu.app