Discover the venture capital playbook for identifying breakthrough AI companies. Learn valuation strategies, risk assessment, and why token economics matter ...
How to Pick AI Winners: A16Z's Complete Investment Framework for the AI Era
Executive Summary
The artificial intelligence market has fundamentally transformed the venture capital landscape in ways that resemble—but differ significantly from—previous technology cycles. Anthropic and OpenAI are generating more monthly revenue than the combined earnings of Meta, Google, or Microsoft, despite less than 5% market penetration in most enterprise sectors. The top 1% startup valuations have skyrocketed from $10 billion (2020-2024) to $32 billion by early 2025, representing a tenfold increase in just 24 months. This unprecedented growth trajectory raises critical questions about how venture capitalists should evaluate AI companies, where value will ultimately accumulate in the ecosystem, and what distinguishes sustainable winners from hype-driven casualties. Understanding these dynamics has become essential not just for investors, but for entrepreneurs, corporate strategists, and anyone seeking to participate in what may be the most significant economic shift of our generation.
Key Insights
- Revenue Growth Outpaces All Precedent: Leading AI model companies are outearning entire technology giants with minimal market penetration, suggesting enormous runway for growth
- Valuation Explosions Are Data-Driven: The 10x increase in top-1% exit valuations in 24 months reflects genuine usage adoption, not speculation, with clear evidence in enterprise AI deployment
- Token Economics Drive Everything: The marginal cost of AI inference tokens, competitive dynamics between model providers, and open-source viability will determine which companies capture value
- Speed Advantages Are Temporary: Historical precedent shows that first-movers rarely capture the majority of long-term value; execution quality and defensibility matter far more than timing
- Portfolio Philosophy Requires High Risk Tolerance: Venture success in AI demands accepting single-digit loss ratios will eventually normalize to 60% as market matures, necessitating disciplined risk-taking
Why November 2022 Changed Everything: Understanding the AI Inflection Point
Before November 2022, artificial intelligence in the enterprise existed largely as theoretical promise. Corporate executives discussed AI through the lens of existing technologies—cloud computing infrastructure, software-as-a-service productivity tools, and incremental automation. The narrative centered on gradual adoption and uncertain ROI. Consumer-facing AI companies were evaluated as commodities, measured purely by user acquisition and pricing metrics, indistinguishable from any other consumer software business.
The launch of ChatGPT fundamentally shattered this framework. Within months, the market gained undeniable proof that large language models could perform meaningful economic work at scale. Enterprises weren't speculating about AI's potential anymore; they were deploying it, measuring productivity improvements, and allocating budgets accordingly. This wasn't hype—it was measurable value creation happening in real time across multiple sectors.
What makes this inflection genuinely different from previous technology cycles is the speed and scale of adoption. When examining the Fortune 500 and S&P 500 companies that collectively generate approximately $2 trillion in annual profit, even conservative estimates suggest that AI solutions could ultimately capture 10% or more of this economic value. Given that Anthropic and OpenAI are already generating hundreds of billions in annual revenue run rates with minimal market penetration, the math becomes staggering. By the end of this year alone, these two companies could reach a combined $200 billion revenue run rate when accounting for their expanding customer bases.
However, this spectacular growth creates a critical evaluation challenge. Previous technology waves—cloud computing, mobile, social media—followed relatively predictable adoption curves. The AI wave is moving so rapidly that traditional startup forecasting methods become unreliable. A company that appears positioned for dominance today could become obsolete within 18 months if foundational technologies shift unexpectedly. This represents perhaps the most significant shift in venture capital risk dynamics in decades.
The Market Structure Problem: Where Will Value Actually Accumulate?
The most challenging question facing every AI investor today isn't whether the technology will transform the economy—the evidence strongly suggests it will. Instead, the critical question is: which companies will capture the disproportionate value this transformation creates?
Venture capital professionals typically rely on recognizable patterns when evaluating technology shifts. In cloud computing, investors identified leaders like AWS relatively early and tracked their dominance. In social media, Facebook's competitive moat became apparent as network effects accumulated. But AI presents something fundamentally different: a value chain that's still being written, with multiple potential architectures, and where competitive positions could shift in unexpected ways.
Consider the historical precedent that should give every AI investor pause. Google wasn't the first search engine; it was arguably the fifth or sixth mainstream option when it launched. Facebook wasn't the first social network; it was entering a market with established competitors like Friendster and MySpace. Yet both companies captured the overwhelming majority of value in their respective categories. Conversely, many first-movers that appeared dominant proved unable to defend their positions.
An analysis of Forbes's annual AI 50 list reveals a sobering reality: approximately 40% of companies that appeared on the list in one year had disappeared from the rankings the following year. These companies didn't merely shrink or plateau; they were replaced by newer entrants or consolidated into other ventures. This suggests AI company half-lives are extraordinarily short, and market leadership positions that seem secure can evaporate with startling speed.
This market structure uncertainty manifests across multiple dimensions. First, there's the question of how many world-class model companies the market will ultimately support. If the frontier narrows to two or three dominant labs—generating what's called "peak-frontier intelligence"—token prices will remain elevated, concentrating economic value among model providers. Conversely, if five or more labs maintain competitive parity, token prices will decline, distributing value more broadly across application builders.
Second, there's the open-source question. Will open-source models become genuinely competitive with proprietary frontier models, or will they remain perpetually one generation behind? This isn't merely academic—it determines whether startups can build sophisticated AI products without dependence on expensive proprietary tokens. Companies like Anthropic and OpenAI actively resist having their models distilled into smaller, cheaper versions, even though the distillation cost represents only 2% of the original training investment. If distillation becomes economically inevitable, it could dramatically accelerate open-source viability.
Third, there's the crucial question of algorithmic efficiency. The human brain learns and processes information with stunning efficiency, requiring far fewer computational tokens to achieve intelligence. If researchers achieve a genuine algorithmic breakthrough—a paradigm shift in how we approach model training—that dramatically reduces token consumption, the entire economic structure of AI companies could transform. Currently, cutting-edge AI applications are extremely token-intensive, but this may not be inevitable.
Finally, there's the critical issue of model consolidation. Will token costs continue declining by 10x or more year-over-year as the current trajectory suggests? Or will this eventually plateau as the technology matures? Right now, token prices are dropping significantly annually, yet the total spending on AI inference is still growing because demand is growing even faster. This creates the paradoxical situation where cost per unit falls, yet total expenditure rises—exactly the opposite of what would trigger a mature market price war.
These structural questions don't have obvious answers, and they fundamentally determine where value accumulates. A venture capitalist investing in an AI application company making those decisions risks betting on a market structure that might shift. An investor backing a model company must navigate intense competition and potential commoditization. Neither position is obviously superior; both require sophisticated decision-making under genuine uncertainty.
Redefining Early-Stage Success: Why Loss Ratios Are Misleading
Traditional venture capital wisdom suggests that successful early-stage investing requires minimal loss ratios. If you're hitting 90% home runs, the thinking goes, you're selecting incredible companies and managing risk effectively. This conventional wisdom is dangerously wrong, and understanding why is critical to AI investing.
When we examine the broader venture capital industry, a troubling pattern emerges. During the 2021 emerging manager peak, seed-stage loss ratios collapsed into the single digits—essentially zero losses. This sounds ideal, but it reveals a fundamental problem: these funds weren't taking adequate risk. By definition, venture capital is supposed to fund companies that most traditional investors consider too risky to back. If your portfolio experiences no losses, you're probably funding companies that were going to succeed anyway, with or without your capital.
In fact, the most sophisticated venture firms deliberately structure their portfolios expecting 60% loss ratios at the early stage. This isn't failure; it's evidence of appropriate risk-taking. These funds invest in high-uncertainty situations where most companies will fail, but the occasional success will generate such enormous returns that the portfolio still delivers exceptional results. The famous venture capital power law means that 90% of returns come from 10% of investments. If you're trying to minimize losses, you're probably missing those 10%.
In AI, we're currently seeing loss ratios in the single digits again—approximately 4-6% across many high-profile early-stage AI funds. This is statistically unsustainable and will inevitably correct. As the AI market matures and more companies enter this space, failure rates will normalize to historical venture averages. When that happens, fund managers who expected 95% success rates will face portfolio disasters. Conversely, managers who anticipated 60% loss ratios and built their firm's economics around that reality will experience exactly what they planned for.
This philosophy drives fundamental differences in how venture firms evaluate AI companies. A16Z approaches early-stage investing by identifying promising major sectors with strong tailwinds—where the underlying technology is genuinely important—and then backing exceptional founders within those sectors. If the sector succeeds and you backed the leader, you've hit a home run. If the sector doesn't pan out but you backed a genuinely talented entrepreneur, you're perfectly comfortable with that outcome as part of the portfolio. The scenario you actively avoid is when a sector thrives but you backed the wrong company.
This framework produces several concrete consequences for how AI companies get evaluated. First, it means betting on founder quality and resilience rather than attempting to predict exact market winners. Exceptional founders can pivot as market dynamics change; mediocre founders struggle even in favorable circumstances. Second, it means accepting that most AI companies you back will fail, and that's the expected outcome. Third, it means building portfolio diversity so that even when most companies fail, the few winners generate sufficient returns to justify the entire fund's investment strategy.
The current single-digit loss ratios in AI are creating dangerous feedback loops. When venture firms see these exceptional success rates, they become overconfident and increase their conviction in their own investment thesis. This leads to higher valuations for unproven companies, which creates unsustainable expectations for future performance. When reality eventually catches up and loss ratios normalize, the portfolio destruction can be severe. The safest approach is actually to expect high failure rates, price companies accordingly, and be pleasantly surprised when loss ratios outperform expectations.
Token Economics: The Hidden Determinant of AI Company Value
Every serious AI investor eventually arrives at the same critical insight: token economics determine everything. This might seem like an oversimplification, but it's actually the most important framework for understanding where value will accumulate in the AI ecosystem over the next five years.
"Tokens" refer to the discrete units of computation and language that AI models process. When you query ChatGPT or Claude, you consume input tokens for your prompt and output tokens for the generated response. Model companies charge for these tokens—typically per million tokens processed—and this represents their primary revenue stream. The cost to the model company of providing each token depends on the computational infrastructure required, the sophistication of the model, and the competitive dynamics with other model providers.
Today's token economics are dominated by what venture capitalists call "peak-frontier intelligence." The most sophisticated models—GPT-4, Claude 3.5, and similar frontier models from leading labs—cost significantly more to run than commodity models. There's a massive willingness to pay premium prices for frontier intelligence, creating what economists call "inelastic demand." If a founder needs cutting-edge AI to build their application, they'll pay the going rate for frontier tokens because the alternative—using inferior models—doesn't work.
This inelasticity has profound implications for venture investing. It means that application companies building on top of frontier models face cost pressures that are genuinely difficult to overcome. If your business model depends on calling frontier models millions of times per second, your per-unit economics might not support commercial viability. Conversely, if you can build a compelling product using commodity models—models that are cheaper and more readily available—your unit economics become dramatically better.
This creates a critical strategic question that every AI application company must answer: Are you betting on frontier models remaining the only viable option, or are you preparing for a future where commodity models become sufficient for most applications? If you're betting on frontier models staying expensive and necessary, you're optimistic that token prices remain elevated, that only a few labs control frontier models, and that competition doesn't drive token prices down. If you're preparing for commodity models, you're anticipating either rapid algorithmic progress (making cheaper models more capable), successful open-source development, or aggressive price competition between model providers.
The second scenario seems more likely from an economic perspective. Throughout technology history, as markets mature, unit costs tend to decline dramatically. Cloud computing costs are orders of magnitude cheaper than they were 15 years ago. Storage is virtually free compared to the cost in 1995. Bandwidth follows the same pattern. There's no reason to expect AI tokens to be different. Some level of commoditization appears inevitable, though the timeline is uncertain.
However, there's an important counterargument that makes this prediction less certain. Token demand is growing faster than token prices are declining. Even though per-token costs might drop by 10x year-over-year, total spending on tokens is still increasing because demand for token-based applications is exploding. This creates the paradoxical situation where prices fall but total market expenditure on tokens grows, suggesting the market hasn't yet reached commodity pricing equilibrium.
What this means for venture investors is that token economics create a fundamental optionality question. The ventures that will thrive are those that can function across a range of token costs—businesses whose unit economics remain viable whether tokens become 10x cheaper or remain expensive. Conversely, companies whose entire business model depends on tokens staying expensive, or conversely, on tokens becoming incredibly cheap, are taking on unnecessary risk relative to more economically flexible approaches.
For founding teams, this suggests a critical strategic imperative: benchmark your business model against multiple token cost scenarios. Run your unit economics assuming tokens drop 10x. Also run them assuming tokens stay expensive and frontiers remain concentrated. Build your product and business model to succeed across both scenarios. This dramatically increases the probability that your company survives the inevitable changes in AI's market structure.
Building Companies for a Supply-Constrained World: Infrastructure Constraints That Won't Disappear Soon
Most discussions of AI focus on model capabilities, application potential, or market size. But venture investors increasingly understand that the constraint limiting AI growth isn't demand—it's supply. Specifically, there's a severe shortage of data center capacity, memory, computing hardware, and electrical power necessary to train, deploy, and run AI models at scale.
This supply constraint manifests across the entire infrastructure stack. You cannot currently book data center capacity at scale until late 2028 or early 2029—that's a three to four-year wait just to get physical infrastructure online. This is genuinely remarkable when you consider that data centers can theoretically be built in 18-24 months. The delay suggests that we're a full year behind the expected schedule for data center buildout in the United States, and we're catching up slowly.
Beyond data centers, the supply constraints cascade through the entire hardware ecosystem. TSMC, the critical semiconductor manufacturer that produces the specialized AI chips required for training and inference, is deliberately showing restraint in ramping production. Part of this reflects a balanced approach to supply and demand; part of it reflects genuine manufacturing constraints around the components themselves. Other specialized hardware components that are essential for data centers—power supplies, cooling systems, networking equipment—face similar bottleneck dynamics.
These constraints have profound implications for how startups should think about their competitive positioning. In a supply-constrained world, getting access to compute capacity is itself a competitive advantage. This is why major companies with relationships with chip manufacturers and data center operators can access resources that startups cannot. It's why infrastructure partnerships matter enormously. A startup that can access compute capacity at reasonable terms has a genuine advantage over competitors who face the same constrained supply.
What's particularly notable is how irrational resistance to new data centers has become. The best data center operators genuinely offer communities significant benefits: funding for nature preserves, investment in high-speed internet infrastructure for local schools, job creation, and substantial tax revenue. Yet communities continue to resist data center development based on water consumption concerns that often don't hold up under scrutiny. A single data center consumes less water than several residential lawns in the surrounding area, yet the political resistance remains strong.
This dynamic matters to venture investors because it affects the timeline for when supply constraints might ease. If communities continue blocking data center development, supply constraints could persist for longer than current projections suggest. Alternatively, if regulatory approaches shift and data centers become more socially accepted, capacity might come online faster. Either way, the supply constraint dynamics remain one of the most underrated factors in AI company evaluation.
From a startup perspective, three strategic insights emerge from this supply-constrained environment. First, think carefully about whether your business model requires enormous compute to scale, or whether you can build a compelling offering with constrained resources. Second, consider partnerships or preferential access to compute capacity as a critical moat. Third, be realistic about timeline expectations—products that require massive infrastructure to be viable probably won't be viable for longer than you might expect.
The Defense Question: Why First-Mover Advantage Might Not Matter in AI
One of the most persistent myths in technology investing is that the first mover to address a new market captures disproportionate advantages. The first browser, the first email client, the first search engine—conventional wisdom suggests these companies should dominate their categories. Reality has been far more complex. Few first-movers actually capture the majority of long-term value. Google wasn't the first search engine; it was better. Facebook wasn't the first social network; it was more compelling. Amazon wasn't the first online retailer; it was more operationally excellent.
Yet in AI, we're seeing venture firms express enormous conviction that being first—getting out of the gate immediately with an AI application—represents a critical advantage. This conviction isn't entirely unfounded, but it deserves scrutiny. The concern isn't whether being early matters; it's whether being early provides lasting defensibility.
Several factors suggest that first-mover advantages in AI might be particularly ephemeral. The underlying technology is changing extraordinarily rapidly. A product built on GPT-3 might become obsolete when GPT-4 arrives with dramatically different capabilities. A startup that optimized its product for a specific model's architectural quirks might find that optimization worthless when the next generation model changes those dynamics. This technological churn makes it difficult to build defensible positions based on product lock-in—the traditional first-mover advantage.
Furthermore, the switching costs for AI products remain relatively low. Unlike enterprise software systems that become deeply embedded in organizational workflows, most AI applications today are relatively easy to replace with alternatives. If a customer is using an AI application built by startup X, but startup Y releases a materially better product, switching costs are minimal. This means that product quality and continuous improvement matter far more than temporal first-mover status.
However, there are exceptions where defensibility might accumulate more readily. If you build a product that generates massive network effects—where the product becomes more valuable as more people use it—you might establish defensible dominance. If you build a product with genuine switching costs—where users have invested substantial data or configuration into your system—defensibility improves. If you build relationships with major companies that create organizational inertia around your offering, you might establish durability.
But many AI applications don't naturally fall into these categories. They're good products, not essential platforms. And in a market where the technology is changing so rapidly that entire categories of products can become obsolete in 18 months, good products built by first-movers might not significantly outperform products built by faster followers with superior execution.
This has concrete implications for how venture investors should think about AI company valuation. When evaluating an early-stage AI startup, the question isn't primarily "did they enter the market first?" It's whether they're building defensible advantages that will still matter when the underlying technology transforms. Companies that are disciplined about product-market fit, that continuously evolve as the technology evolves, and that build genuine switching costs or network effects, will likely outperform companies that simply moved first.
The Platform Dynamics: What Bill Gates Knew About Value Distribution
There's a famous Bill Gates observation that provides remarkable clarity on how value distributes across technology platforms: "If you're a platform, the value of the companies that are built on top of you needs to exceed the value of the platform itself."
This principle explains why Microsoft became more valuable than the sum of its dependent software ecosystem, why Apple's platform commands premium pricing, and why Google's advertising platform enabled the creation of countless highly profitable company. The insight is that if you control the platform—the foundational layer—you can capture enormous value, but only if you leave sufficient economic rents for companies built on top to thrive.
In the AI context, this raises a profound question about whether the model companies (Anthropic, OpenAI, Anthropic) will capture the majority of AI's economic value, or whether the real value creation will occur in the application layer—companies built on top of AI models.
Current venture investing patterns suggest deep uncertainty about this question. Six months ago, many venture investors were convinced that model companies would capture nearly all value, relegating application companies to commodity status. The narrative shifted again when model companies themselves started building applications—OpenAI developing premium consumer products, Anthropic building enterprise offerings. This suggested that the model companies recognized that the highest-value opportunities might actually be in applications, not in providing generic model access.
Today, we're seeing model companies simultaneously trying to be both platforms and application builders. OpenAI and Anthropic are licensing their models to third parties while also building proprietary products that compete with those same third parties. This creates an inherent tension: if you're too generous with platform pricing, you cannibalize your own application business. If you're too expensive, you're not leaving enough value for application builders to succeed.
The most probable outcome, based on technology history, is that genuine value creation will occur both in platforms and in applications. The model companies—assuming they maintain frontier capabilities—will generate enormous value. But the ecosystem of companies built on top of those models will generate value that dwarfs even the model companies' revenue. This is the "and" that venture professionals keep referencing: both the platforms and the applications can be extraordinarily valuable.
What this means for venture investors is that the opportunity set is broad enough to generate multidecade returns across multiple categories. You don't need to perfectly predict whether model companies or application companies will generate higher returns. Both categories can generate transformational outcomes. The imperative is to back exceptional teams building valuable products, wherever in the stack they're positioned.
For founders, this framework suggests that competitive position against the model companies themselves shouldn't be the primary organizing principle for your strategy. Instead, focus on building something genuinely valuable for customers, something that would be harder for model companies to build themselves either due to market focus, organizational differences, or other competitive factors. Some of the highest-value AI companies will be built by teams that aren't directly competing with OpenAI or Anthropic, but are instead building unique offerings in verticals or use cases where they have asymmetric advantages.
The Public Markets Inflection: What Four to Five Trillion Dollars of IPO Activity Means for Venture
Something extraordinary is about to happen to the public markets. SpaceX is expected to conduct an IPO, likely in the second half of this year. Following SpaceX, both OpenAI and Anthropic are expected to pursue public listings. These three companies alone could create four to five trillion dollars in total value on public markets—a staggering sum that dwarfs most technology companies that have gone public in the past decade.
This represents a profound shift in market dynamics. Over the past 20 years, the number of publicly traded companies in the United States has shrunk by approximately 50%. This consolidation meant that retail investors had decreasing opportunities to invest in high-growth technology companies. The mega-cap "Magnificent Seven" stocks (Apple, Microsoft, Google, Amazon, Tesla, Nvidia, Meta) have become increasingly concentrated, absorbing investor capital that might otherwise have been distributed across a broader ecosystem.
What's remarkable about the upcoming wave of AI company IPOs is that these companies represent genuinely exceptional growth rates. The Magnificent Seven companies, despite their excellence, are now growing in the 20-30% annual range. That's exceptional for companies of their scale, but it's modest compared to the 70%+ annual growth rates we're seeing from companies like Palantir. OpenAI and Anthropic, should they go public, would likely be in the hypergrowth category—companies genuinely rare in public markets.
This creates a rare opportunity for public market investors. For the past decade, if you wanted exposure to high-growth technology, you essentially had to invest in private markets or place significant capital in mega-cap stocks that were growing more slowly than historical precedent. The upcoming IPO wave reverses this dynamic. Suddenly, investors of all sophistication levels—including people with retirement funds in index accounts—will have the opportunity to own pieces of companies at the absolute frontier of technology development.
The question many institutional investors raise is whether the public markets have sufficient capacity to absorb four to five trillion dollars in new value creation. The answer appears to be yes, for several reasons. First, public market capitalization globally is measured in the hundreds of trillions of dollars, so absorbing a few trillion in new technology value is entirely manageable from a pure capacity perspective. Second, there's genuine scarcity of compelling investment opportunities in public markets today. If SpaceX, OpenAI, and Anthropic become available for public ownership, capital will flow into these opportunities readily.
There's also been debate about whether these companies should be included in major indexes like the S&P 500 or total market indexes. From a venture capitalist's perspective, index inclusion is tremendously important. It means that passive index investors—individuals with retirement funds, people with 401(k)s, everyday investors who aren't sophisticated technology investors—will automatically own pieces of these companies as part of their broad market exposure. This is genuinely positive for two reasons. First, it's democratic—regular people benefit from these extraordinary value creation opportunities, not just venture capitalists and early investors. Second, it's economically efficient—capital can flow to its highest-value uses without artificial barriers.
What this means for venture capital's future is that we're likely entering an era where the venture industry's exit opportunities are significantly enhanced. When you can take companies public that are still in genuine hypergrowth phase—where the peak opportunity still lies ahead, not behind—you create conditions where venture capital generates returns rivaling the highest-performing asset classes. This is genuinely excellent news for the ecosystem. It means that capital can cycle into new venture investments even as existing ventures achieve successful exits.
From a market composition perspective, the arrival of high-growth technology companies in public markets should drive a shift in ownership. Some capital will rotate out of lower-growth mega-cap stocks and into newly public high-growth AI companies. But the total market should expand to absorb these new opportunities without creating a crisis of confidence or market dysfunction. The market has adjusted to far more dramatic transformations in the past century. Absorbing genuine high-growth technology companies back into public markets is a relatively straightforward transition.
Consumer AI: The Untapped Market That Could Shift Digital Attention
While the venture capital discussion around AI has concentrated overwhelmingly on enterprise applications—B2B companies solving business problems—there's an equally enormous opportunity in consumer AI that remains fundamentally underdeveloped.
For the past decade, the consumer technology narrative centered on a troubling dynamic: the biggest technology companies (Google, Meta, Amazon, TikTok) have collectively captured the vast majority of user attention and time spent online. This concentration created an essentially insurmountable competitive moat. If you're a startup trying to build a consumer application and you're competing against Facebook, which has billions of users and decades of engagement optimization, the odds are heavily stacked against you. The network effects are so strong, and the switching costs for users are so high, that new consumer companies struggle to achieve meaningful scale.
AI changes this dynamic in potentially transformative ways. Consider how users interact with consumer technology today. Your time spent on social media generates value for the platform through advertising. Your time spent on search generates value for the search company. The dynamics of engagement and monetization haven't fundamentally changed in 20+ years. But AI introduces the possibility of genuinely personalized, adaptive experiences that might shift how users allocate their attention and time.
What if AI applications could provide genuinely superior experiences compared to incumbent platforms? What if a startup could build an AI-powered interface that's more helpful, more personalized, and more aligned with user interests than the experience provided by massive platforms with misaligned incentives? In a world where the principal incentive is user engagement rather than advertising revenue, entirely different product designs become possible.
This isn't hypothetical. We're already seeing early evidence of this shift. Consumer applications powered by AI are capturing meaningful usage. Young people are spending time interacting with AI in new ways that don't neatly fit into existing categories. The dynamics of how users spend time and allocate attention appear to be in genuine flux, which creates extraordinary opportunity for new consumer applications.
What makes this even more important is that consumer businesses, when they work, scale to enormous valuations. If you can build a consumer application that genuinely captures meaningful user attention and time, and you can build a business model around that attention, the value creation can be staggering. This is potentially where some of the highest-value AI companies could be created—not in B2B automation or enterprise efficiency, but in genuinely reimagining how people spend time and interact with technology.
For venture investors, this means that the consumer opportunity in AI is genuinely underfunded relative to its potential importance. There's been a relative flood of capital into B2B AI applications, but consumer AI remains undercapitalized. This creates a rare opportunity: areas where the venture industry hasn't yet recognized the true scope of value creation potential.
From a founder perspective, this suggests that consumer AI might be one of the higher-variance, higher-potential-return opportunities available. Building consumer applications is difficult and faces serious competitive challenges from incumbent platforms. But the potential outcomes—companies that capture genuine shifts in how people spend time and attention—could be enormous.
The Five-Year Outlook: What the VC Industry Becomes in the Age of Trillion-Dollar AI Companies
Looking ahead to 2030, the venture capital industry itself will likely look substantially different than it does today, with the transformation driven primarily by three uncertain factors: the market structure of the model industry, the role of open-source, and the accumulation of competition for tokens.
The paramount question is whether the model industry will remain highly concentrated with one or two dominant frontier providers, or whether five, ten, or more companies will maintain genuine competitive parity. If the market concentrates, token prices will remain elevated, model companies will be extraordinarily valuable, and the economic opportunity in applications will be constrained. If the market becomes more competitive, token prices will decline, and applications will capture more value.
Second, open-source is on a trajectory to become genuinely important, but its ultimate role remains uncertain. Will open-source models achieve parity with proprietary models? Will they remain perpetually one generation behind? The answer probably depends partly on whether distillation proves economically viable, and partly on whether open-source communities can sustain the infrastructure investment required to remain competitive.
Third, there's the algorithmic efficiency question. Will researchers achieve breakthroughs that fundamentally reduce the token intensity of AI applications? Or will current approaches prove to be near the efficiency frontier? This question is genuinely uncertain, but the potential impact is enormous.
Given these uncertainties, venture capitalists are positioning themselves accordingly. The firms that will thrive are those that can maintain exposure to multiple possible outcomes: backing model companies, backing application companies, backing infrastructure providers, and backing consumer experiences. Portfolio diversification across these categories increases the probability that you'll have meaningful exposure to wherever value ultimately accumulates.
Beyond portfolio composition, the venture industry will likely become more operationally sophisticated. Early-stage AI companies face "big company problems" far earlier in their lifecycle than previous generations of startups. A company might have hundreds of millions or billions in revenue while still being genuinely early-stage from an organizational perspective. This requires venture firms to provide scaling support in areas like international expansion, complex sales operations, supplier negotiations, and cloud infrastructure management. The firms that scale their operational support platforms will win larger equity ownership percentages and generate better outcomes.
We're also likely to see continued consolidation of venture funds. The ability to back early-stage companies, then provide capital and support through growth phases, and have the ability to syndicate into later-stage rounds, becomes increasingly important when companies are growing at AI-driven speeds. Smaller venture funds that lack the capital to support companies through multiple rounds will find themselves diluted out of successful investments.
The consumer opportunity in AI represents a genuine frontier for venture capital. If consumer AI applications do begin capturing meaningful shifts in user time and attention, the venture industry that bet on this early will generate extraordinary returns. Conversely, if consumer AI fails to break through incumbent platform advantages, the venture capital deployed there will result in losses. This uncertainty is exactly the kind of scenario where venture capital thrives—genuine optionality on transformational outcomes.
Conclusion
The artificial intelligence market is simultaneously the most exciting and most uncertain frontier in venture capital. Anthropic and OpenAI are generating revenue that was unimaginable two years ago, yet we're still in the earliest stages of broader AI adoption. The companies that will define this era's winners aren't yet obvious. The market structure that will determine where value accumulates remains undecided. The role of open-source, the efficiency of algorithms, and the competitive dynamics between model providers remain open questions.
For venture investors, this uncertainty is actually ideal. Venture capital thrives when you can identify massive opportunities that most investors haven't yet recognized. AI represents multiple such opportunities: breakthrough applications that could shift consumer behavior, enterprise solutions that could fundamentally transform how companies operate, platform dynamics that could reshape the technology industry, and infrastructure requirements that will demand enormous capital deployment.
The opportunity to participate in transforming how humans live and work—this is what venture capital is ultimately designed for. The AI era represents perhaps the clearest such opportunity in decades. The investors, founders, and companies that navigate this transformation wisely, maintain flexibility as technologies change, and build genuinely defensible competitive advantages will generate returns that rival any in venture capital history. The next five years will determine the shape of technology and business for the next two decades.
Original source: The Rule for Picking AI Winners | The a16z Show
powered by osmu.app