“In my lifetime, I’ve seen two truly revolutionary things: the first was the graphical user interface… the second is ChatGPT,” Bill Gates said not long ago. Consider the weight of that: the man who ushered in the PC era is putting a quirky AI chatbot in the same bucket as Windows. At the same time, OpenAI – the startup behind ChatGPT – hit an eye-watering $160 billion valuation barely a year after launching its product, a market cap Microsoft took decades to reach. No wonder early-stage investors in tech feel a familiar tingle: this could be the next big platform shift.
But before we all start high-fiving about finding “the next Internet,” let’s spill some coffee and ask the awkward question: if this technology is so world-changing, why do so few people actually use it every day? Yes, ChatGPT’s adoption broke records – somewhere between one-third and two-thirds of people worldwide have tried it or at least heard of it. You don’t get that kind of buzz without substance. It’s incredibly easy to try (just a website, no new hardware needed), so millions had their “wow” moment in record time. And yet: most folks used it once, said “Wow, that’s neat,” and never came back. The vast majority of us haven’t woven AI into our daily lives. In big companies, it’s the same story – nearly every Fortune 500 has an AI pilot project (even if it’s just the CIO fiddling with ChatGPT), but only a few have anything in production for their core business. There’s excitement in the air, demos in boardrooms, and slide decks galore.
So here we are: on one hand, the world is convinced AI is huge; on the other, we’re still hunting for product-market fit. Tech headlines oscillate between “AI will change everything” and “Are we sure we want to spend half a trillion dollars here with no clear ROI?”. It’s a classic case of hype outpacing adoption – a pattern we’ve seen before (remember the early days of the internet, or mobile, or honestly even electricity?). Every 10–15 years, tech goes through a paradigm shift – mainframes, PCs, the web, smartphones, and now generative AI – and everyone scrambles to figure out the new rules of the game. Benedict Evans, who’s been around for a few of these shifts, calls this moment “AI Eats the World,” riffing on that old Marc Andreessen line about software (we’ve been saying this phrase for years). The title sounds grandiose, but it captures the feeling that maybe, just maybe, this is one of those times when the ground is shaking beneath us. Still, as Evans is quick to point out, beyond the broad agreement that something big is happening, almost “all the questions are wide open”. In other words: we know this is a big deal – we just haven’t figured out exactly how.
Big Bets, No Moats (Yet): CapEx and Commoditization
Let’s talk about the supply side of this revolution – the nuts and bolts (or rather, chips and data centers). The current AI boom has a peculiar feature: the winners of this gold rush might be the guys selling shovels and pickaxes. By that I mean NVIDIA and other chip makers, cloud platforms, and all the infrastructure providers enabling these massive AI models. There’s an arms race underway to build ever-bigger, ever-smarter AI models, and it’s translating into mind-boggling capital expenditure.
Consider this: Meta’s CEO Mark Zuckerberg casually noted that the next version of their open-source model Llama would need 10× more compute than the last. How do you even grok that? Llama 3.1 already used about 16,000 NVIDIA GPUs – roughly half a billion dollars worth of chips – just for training. Today’s cutting-edge model clusters are pushing 100,000 GPUs at a time. And it’s not just Meta: across the board, the big tech platforms are pouring money into AI development. Together, the top four tech giants will spend over $200 billion on capex this year (yes, with a “b”), which is about $100B more than last year. Microsoft, historically a software company with fat margins, is now spending over a quarter of its revenue on infrastructure – dwarfing even old-school telecoms in relative investment. Satya Nadella went from selling Windows CDs to literally building datacenters so people can use AI in the cloud. It’s as if the software world remembered it runs on silicon after all, and everyone’s racing to build the biggest engine before the next guy does.
Why this frenzy? Partly it’s FOMO at a corporate scale. There’s a saying in Silicon Valley: “I don’t want to live in a world where somebody else is making it better faster than we are.” That quip was actually floating around Google and Meta as they justified these enormous bets. In plain terms: no one wants to be the company that held back on AI investment and got leapfrogged. Another reason is that, so far, size has equaled power in AI – we made recent breakthroughs largely by making models ridiculously big (and feeding them ridiculous amounts of data). It’s brute-force progress. The snag is, we don’t actually know how far brute force can take us. Will making models 10× bigger keep giving commensurate returns? Maybe it fizzles out; maybe it unleashes crazy new capabilities. Even the experts are debating like sports fans on a Friday night – and as Evans wryly notes, all you’ll conclude after watching them is “they don’t know”. We lack a theory of AI scaling; we’re basically throwing money at the problem to see how far it goes.
That uncertainty hasn’t slowed the capex tsunami – if anything, it has fueled it. Interestingly, there’s no consensus “moat” in AI right now. Unlike past platform shifts, where one company might build an ecosystem that gives lasting advantage, in AI the core tech is leaking out into the open. An internal Google memo famously admitted “we have no moat… and neither does OpenAI.” In other words, the raw models are becoming commodities – everyone will soon have access to similar brains. Open-source models (like Meta’s Llama) are freely available or cheap, and they’re catching up to the best proprietary models at a breathtaking pace. In fact, we’ve seen a rapid convergence of model capabilities and costs: today you can get a model that’s, say, 90% as good as the best-in-class for maybe 5% of the cost. Meta is open-sourcing models essentially for free to bust others’ business plans, and Apple is training models to run locally on your iPhone – treating AI as just another built-in feature. The implication is wild: having a great model won’t be enough, because everyone and their cousin will have one. As Evans puts it, the “moat” might simply be $$ – capital itself. If it takes billions to train a frontier model, only those with deep pockets can play that game… at least until open-source and innovation find cheaper paths (which they are, quickly).
So we have a paradoxical landscape: Huge money is being spent to create something that will likely be cheap and ubiquitous. Nvidia’s revenue chart is shooting up like it’s 1999, and history teaches us that such spikes often precede a bubble correction. Are we over-investing in AI infrastructure ahead of actual demand? Possibly – and that’s a key risk for VCs to weigh. It’s reminiscent of telecoms in the late 90s building fiber everywhere for an internet that hadn’t arrived yet. The internet did arrive (and changed the world), but not without wiping out a lot of early investment. AI could follow a similar pattern: transformative in the long run, but punishing to those who bet on the wrong layer and/or timeframe.
What does this mean for me?
For an early-stage VC, betting on a startup just because they boast a superhuman model is a fool’s errand. If Google and OpenAI have no inherent model moat, a five-person startup likely doesn’t either. Instead, ask the founders what proprietary advantage they have – is it a unique dataset? A novel algorithm that dramatically cuts costs? A specific customer integration others can’t easily replicate? If their answer is “we have a slightly better fine trained LLM,” be skeptical. Also, watch the capital requirements. Ambition is great, but if a pitch sounds like “just give us $100M for GPUs and trust us to out-research OpenAI,” that’s probably not your deal. Encourage plans that leverage the commoditization of models – e.g. startups using open models or efficient techniques to do more with less, rather than those requiring heaps of cash to keep up. The infrastructure is becoming a commodity; the application of it is where new defensible businesses can emerge.
The “Infinite Interns” & the Hard Question of Now What?
So if the supply side is overclocked, what about demand? What are people actually doing with AI? Here’s where things get both exciting and perplexing. Generative AI is like having an army of eager interns at your disposal – but many people have no idea what to delegate to them. Evans uses this brilliant metaphor: modern AI “gives you infinite interns.” Think of all the back-office, low-level cognitive tasks you’d happily hand off to a bright 10-year-old if you could – scanning hours of call transcripts for angry customers, summarizing a week’s worth of news, drafting boilerplate emails. You used to need a human (or ten) for that, so most of it simply didn’t get done at scale. Now you can automate that entire class of tasks that just require a “mammal brain,” not deep expertise. You can have AI listen to every support call and flag which customers sound frustrated – no senior analyst needed, not even the 10-year-old you’d have bribed with ice cream. That’s a game-changer!
The catch is figuring out which tasks to give the interns, and how to integrate those interns into the workflow. It turns out, this is non-trivial. Let’s rewind to a classic analogy Evans draws: the spreadsheet in 1979. When the first spreadsheet software (VisiCalc) appeared, it was mind-blowing for accountants – it automated a task that was basically their whole job (recalculating rows of numbers) and turned weeks of work into minutes. Accountants who embraced it could finish a month’s work in a day and spend the rest of the month golfing. No joke – it was that dramatic. If you were an accountant back then, you had to have this software. But show that same spreadsheet to a lawyer in 1979, and they’d shrug: “That’s neat, maybe my accountant finds it useful, but I don’t have that problem.” In fact, 90% of people had no use for a spreadsheet in their daily work at first. They recognized it was clever, but it wasn’t relevant to their jobs.
ChatGPT and generative AI today are in a similar spot. Nearly everyone who tries it says, “Wow, that’s very clever.” But a huge chunk follow that with, “…but I don’t do that all day.” The use cases aren’t obvious for the majority – yet. It’s telling that the early successes of generative AI are in fields where the fit is natural: software developers are using Cursor or similar tools to get 20-30% productivity boosts in coding, content marketers are letting AI help draft and brainstorm copy (no more staring at a blank page), and customer support teams are experimenting with AI-generated replies and issue categorization. In these domains, AI feels like a turbocharger on an existing workflow – it slots in with relative ease. Developers don’t mind an AI suggesting code because it’s like a superpower in their IDE; support agents appreciate a draft answer they can tweak rather than writing from scratch.
But for many other professions, it’s not obvious what to do with an AI intern. A doctor or a project manager or a sales exec might try ChatGPT and get a cute answer, but it doesn’t immediately change how they do their job. In some areas (legal, medical) there’s even a high barrier: if the AI makes stuff up or errs, the cost is serious. A lawyer can’t happily let an AI draft a contract and not worry about hidden mistakes – the risk is too high without robust verification. So adoption is uneven: surveys show maybe 20% of workers in the U.S. claim to be using generative AI regularly now, but those numbers skew to certain fields. And even those using it will admit they’re still experimenting.
This leaves us in a strange “in-between” phase: everyone knows AI is powerful, but many people don’t yet know what to do with it. We technologists are basically handing a magical toolbox to users and saying “Here, we built this – now you figure out how it fits into your life.” That’s backwards. Steve Jobs famously said “It’s not the customer’s job to know what they want”, and here we are kind of ignoring that wisdom. Historically, successful new tech comes with obvious use cases and user-friendly packaging. Early smartphones were phones + iPod + internet in your pocket – easy to grasp. Early PCs did spreadsheets and word processing – clear value if you did accounting or typing. AI in its current form is more like a raw super-ingredient waiting for recipes. It’s powerful, but in the kitchen most people are a bit unsure how to cook with it.
For venture investors, this smells like opportunity. The world “gets it” – there’s a huge latent demand to make AI useful – and whoever figures out the compelling use cases will unlock that demand. It’s also a caution: many startups will peddle AI snake oil or cool demos with no real workflow integration. As an investor, we have to be especially sharp-eyed now.
What does this mean for investors?
When meeting AI founders, listen for the specific problem they’re solving. If a pitch is all “AI AI AI” with vague talk of disruption, and no concrete pain point being alleviated, treat that as a red flag. The best pitches sound like “X is a tedious/important task for [target customer]. We use AI to do Y, which saves [time/money] or enables [new capability].” In other words, they focus on the solution, not the shiny tech. As Benedict Evans notes, customers pay for solutions, not technologies. The story of Everlaw (a legal tech company) is instructive here – they succeeded not by touting their AI, but by solving a real problem for lawyers (making discovery faster and easier). Also, gauge whether the founders expect users to figure out the magic. Are they basically throwing GPT-4 in an app and saying “the user will prompt it to do whatever they want”? That’s not good enough. Look for teams that deeply understand their user’s workflow and have slotted AI in to eliminate a step or ten from the usual process. AI isn’t a product by itself – as one highlight from Evans’ talk put it, ChatGPT is more of a cool demo than a fully-fledged product. Successful startups will bridge that gap between cool demo and daily must-have. Don’t be fooled by novelty; demand utility.
From Models to Workflows: Where Is the Moat?
Because raw AI models are becoming commodities, the value in the AI stack is shifting above and below the models – to data, product, and distribution. Think of the AI model as an engine. If everyone can rent the same engine (be it GPT-4 or Llama-2 or whatever comes next), what makes one solution win over another? It’s everything around the engine: the fuel (unique data), the chassis (user interface and workflow integration), the drivetrain (integration into other systems), and the driver (execution, strategy, partnerships). Let’s unpack that with insights from Evans’ perspective:
First, owning a “frontier model” doesn’t guarantee you a business. Sure, OpenAI has GPT-4, the most famous kid on the block. But if tomorrow an open-source model can match GPT-4 at a fraction of the cost (which is not far-fetched), what stops everyone from using that cheaper brain? Network effects are minimal here – one company’s model doesn’t get inherently better because more people use someone else’s model (unlike, say, Facebook, where more users does make the product better for everyone). As Evans bluntly puts it, having the best model isn’t a durable advantage if it doesn’t also get cheaper with scale or stickier with network effects. This is why OpenAI, Anthropic, and others are scrambling to improve efficiency and lock in partnerships. It’s also why so many startups that tried to be “the OpenAI of X” by training their own models are pivoting – they realize they’re burning cash to reinvent a wheel that giant companies or open communities are also building.
Where, then, can a startup build a moat? Data is one answer. A startup solving a niche problem can accumulate proprietary data (user feedback, domain-specific documents, specialized sensor data, etc.) that, when used to fine-tune a model, gives consistently better results for that niche. That data is not easily acquired by competitors, especially not the big generalist models. We saw this in earlier tech waves: Google’s search quality advantage came in part from data (links, clicks) no one else had at that scale; Tesla’s self-driving lead comes from billions of road miles its cars have driven. In AI software, a vertical SaaS that caters to, say, medical coding could, over time, build a dataset and tuning that makes its AI far better for medical coding than any general model. That becomes a self-reinforcing moat if executed well.
Another angle: workflow integration and UI/UX. If AI becomes “just another API call” in software – which Evans suggests is likely for the steady-state of this tech – then the end-user won’t care which model is under the hood. They’ll care what the product actually does for them. Think about it: most of us don’t know what database our favorite apps use, or what programming language they’re written in. Similarly, if in 2-3 years every app has some AI features, users will gravitate to the ones that are designed in the most intuitive, helpful way. The winners might be those who figure out the UX of AI – how to make interacting with AI feel natural and valuable within a particular context. That could be as simple as a better prompt interface for a given task, or as deep as rethinking an entire workflow so that the AI does 90% of the heavy lifting and the human just supervises.
Distribution can be a moat too. If you can embed your AI into existing platforms that people already use, you have a leg up. For instance, a startup that partners deeply with, say, Salesforce to be the recommended AI tool for sales teams might beat out an arguably better AI product that has no distribution channels. We saw this in the past with enterprise software – often the “good enough but widely available” product beats the “excellent but hard to find” product. In AI, this might mean being the one that comes bundled with Office 365, or the default AI in a popular industry-specific software suite, etc. There’s a strategic ambiguity right now: will the value accrue to apps (the new AI-powered products) or to infrastructure (the big platforms or APIs)? Or maybe to those who manage to bundle AI into large offerings vs those who unbundle specific features out of incumbents? It’s still up in the air, and it may not resolve to a single winner-take-all – different layers might capture different slices of value.
Evans gives a nice framework: in new tech waves, initially incumbents try to bundle the new tech as just a feature in their existing products, while entrepreneurs unbundle things – picking off one specialized task at a time. We saw this with the SaaS explosion: one giant Oracle or SAP feature could be split into 10 startups each doing one thing better. We’re now seeing a swarm of AI startups doing exactly this – taking one slice of a workflow and using an LLM to supercharge it. “There are two ways to make money in business: bundling and unbundling,” the old Jim Barksdale quote goes, and it rings true now. For the moment, unbundling is rampant. Every week I hear a pitch for “the AI-powered solution for ___” (fill in the blank with a very specific use case). Many of these will find some traction. But I also suspect we’ll see re-bundling down the line: the big players aren’t sitting still, and if they can fold a given AI capability into their suite (or the AI itself gets advanced enough to handle multiple tasks), then yesterday’s neat standalone tool could become tomorrow’s trivial feature.
In fact, there’s an extreme scenario to consider: what if the AI itself becomes the ultimate bundler? If one day a single AI agent can handle any task you throw at it – the fabled general AI that arranges your travel, does your taxes, designs a website, all in one – that would collapse many app categories into one interface. Evans calls this the “AI maximalist” view, where “the LLM sits on top and runs everything else”. In that world, you wouldn’t need thousands of specialized apps, you’d just speak your needs and the Big AI would handle it. It sounds wild, but not totally science fiction – more like a distant cousin of where we are. On the flip side, the more pragmatic view (and the one Evans leans toward for now) is that LLMs will be “another API call” in the toolkit. They’ll certainly enable lots of new products, but they won’t replace the concept of distinct apps solving distinct problems, at least not in the foreseeable future. Not when we’re still wrestling with AI writing plausible BS and making mistakes.
What does this mean for investment? It means we should prepare for both possibilities without betting the farm on either. In the near term, I’m inclined to back startups that unbundle effectively – those laser-focused on a use case where an AI+human workflow can deliver 10× better results. Those companies can acquire users, build datasets, learn fast, and potentially become acquisition targets for bigger players who need to fill a gap. In the long term, if someone comes along claiming a more general AI platform, I’d press them on why they’ll win against incumbents bundling the same tech. The key question is: Are you building a feature, a product, or a platform? And whichever it is, what’s your plan when the giants come hunting in your territory?
What does this mean for me?
In diligence, try to pin down where a startup sits in this bundler/unbundler dynamic. If it’s a feature disguised as a company, be wary – could an incumbent swat it away by releasing their own version next week? (E.g., how many “AI writing assistants” got pummeled the moment Google Docs and Microsoft Word introduced built-in AI writing features?) On the other hand, if it’s a true workflow re-imagination (a full product) or a platform play, ask: what’s their moat against the eventual bundlers? Are they accumulating proprietary data or users at scale? Could they become a de-facto standard in a niche that later gives them power to expand? There’s no formulaic answer, but one thing is clear: don’t get star-struck by a demo; dig into the durable edge. Remember, ChatGPT’s core tech is available to everyone – so why will this team serve customers better or cheaper than others using the same engine? Finally, keep an eye on enabling tech. Maybe the big winners aren’t applications at all, but the picks-and-shovels that allow many apps to thrive (think of developer tools, privacy and compliance layers for AI, model optimization tools, etc.). Those can be gold mines if a general trend (like every company adding an “AI layer”) sustains. Just make sure any such infrastructure play isn’t prematurely overbuilt for a demand that might not materialize (we’ve got enough capex bubbles brewing).
A New Era… That Feels Kinda Familiar
If you feel a bit dizzy, you’re not alone. This phase of the AI revolution is a whirlwind of contradictions – extraordinary breakthroughs and exuberant funding, mixed with tentative usage and an unclear path to ubiquity. In our coffee chat here, we’ve bounced between optimism and caution, concrete examples and big-picture speculation. That’s exactly what it’s like talking to a sharp friend about AI these days – one minute you’re marveling at what’s possible, the next you’re scratching your head about what’s practical.
Perhaps the best way to wrap up is with a dose of perspective. Evans suggests that ultimately, we have two buckets to put questions in. Either “AI will work just like every other platform shift” – meaning it will create opportunities for lots of new companies, empower some incumbents, disappoint others, and gradually become a normal part of life – or the answer to our big questions is simply “We don’t know yet.” There are things we can predict by looking at history (for example, that commoditization drives prices down and spreads technology widely, or that enterprises adopt slower than the hype would suggest). And then there are things that are genuinely new – scientific unknowns about AI capabilities or emergent behaviors – where even the best-informed admit uncertainty.
For a venture investor, the job is to balance those two. Rely on the patterns to guide you in asking the right questions (Where is the moat? Who is the customer? Is this the picks or the shovels? Are we bundling or unbundling?). But also stay humble in the face of the unknowns. We might be at a stage akin to the internet circa 1995: everyone senses it’s big, lots of crappy dot-coms are being funded alongside a few Amazons and Googles that will actually last. Many people “got” the internet was the future, but few knew how it would unfold – that selling books online would be a hit, but delivering groceries online would take 20 more years to get right, for example. AI could surprise us in similar ways.
In the meantime, I find it useful to remember that while we futurists fixate on what’s coming in 2025 or 2030, most of the world is still catching up with tech from a decade ago. The cloud, SaaS, mobile apps – these are still rolling out and generating new winners in many industries. AI will layer on top of that, not replace it overnight. The truly seismic shifts (think electricity, the internet itself) took years to diffuse and didn’t follow a straight line. AI is likely to be the same: it will eat the world, but maybe one bite at a time.
So, as we finish our coffee, the takeaway for early-stage VC friends is this: Keep your excitement, but arm it with insight. Enjoy the sweeping narrative of AI’s promise (it’s indeed a thrill), yet channel your inner skeptic when evaluating the here-and-now of a pitch. The best founders will be those who can articulate a vision that’s grand and a go-to-market that’s grounded. Those are the folks who will build real value while others are spinning sci-fi tales. Invest in the “infinite interns” – the AI that actually works for someone, doing something useful – and you’ll do well. After all, AI might be a new kind of electricity, but someone still has to figure out the lightbulb. That’s where you should be looking, and that’s where the world-changing startups will emerge. Good hunting out there!