When one of Silicon Valley’s most influential venture capital firms articulates a new investment thesis, the technology industry pays attention. Andreessen Horowitz — better known as a16z — has been quietly advancing what partner Anjney Midha calls the “Theory of Well,” a framework for understanding where lasting value will accrue in the artificial intelligence stack. The thesis has significant implications not only for startup founders seeking funding but for the broader architecture of the AI industry itself.
The concept, as reported by AOL, centers on the idea that the most defensible and valuable positions in AI will belong to companies that control critical chokepoints in the technology supply chain — the “wells” from which all other applications must draw. It is an analogy drawn from natural resource economics: just as oil wells represent the foundational source of energy value, certain layers of the AI stack will serve as indispensable wellsprings of capability.
What the “Theory of Well” Actually Means for AI Investment
Midha, who serves as a general partner at a16z focused on infrastructure and AI investments, has argued that the companies positioned at these chokepoints — whether in foundational models, data infrastructure, or compute orchestration — will capture outsized returns. The thesis pushes back against the prevailing narrative that application-layer companies, those building consumer-facing or enterprise AI products, will necessarily be the biggest winners of the current technology cycle.
This is a notable departure from some of the conventional wisdom that has circulated through Sand Hill Road in recent quarters. Many investors have been placing aggressive bets on AI applications, reasoning that the model layer is becoming commoditized as open-source alternatives proliferate and the cost of inference continues to decline. The Theory of Well suggests the opposite: that infrastructure-layer companies will maintain pricing power and strategic importance precisely because they sit at the base of the value chain.
The Infrastructure vs. Application Debate Intensifies
The debate between infrastructure and application investing in AI has been one of the defining tensions of the current venture cycle. On one side, firms like Sequoia Capital have publicly questioned whether the massive capital expenditures flowing into AI infrastructure — estimated at over $100 billion in 2024 alone — will generate sufficient returns. Sequoia partner David Cahn wrote a widely circulated analysis last year arguing that the AI industry faces a $600 billion revenue gap between what infrastructure companies need to earn and what AI application companies are actually generating.
On the other side, a16z’s thesis suggests that this framing misses the structural dynamics at play. According to the firm’s view, the infrastructure layer is not simply a cost center waiting to be disrupted by cheaper alternatives. Instead, the companies that control the most critical resources — whether that means GPU clusters, proprietary training data, or the orchestration layers that connect models to real-world applications — will function as toll collectors on the entire AI economy. As AOL reported, Midha has drawn explicit parallels to historical technology cycles where infrastructure providers ultimately captured more value than the application companies built on top of them.
Historical Precedents: From Cloud Computing to Mobile
The historical analogy is instructive. During the cloud computing revolution, Amazon Web Services emerged as the dominant infrastructure provider and now generates more operating profit than Amazon’s retail business. Similarly, during the mobile era, companies like Qualcomm and ARM Holdings captured enormous value by controlling essential components of the smartphone stack, even as individual app developers struggled to build sustainable businesses.
The Theory of Well suggests that AI may follow a similar pattern. Nvidia’s dominance in GPU hardware is perhaps the most obvious current example — the company’s market capitalization has surged past $3 trillion on the strength of its position as the primary supplier of AI training and inference chips. But the thesis extends beyond hardware. Companies like CoreWeave, which provides specialized cloud computing for AI workloads, and Scale AI, which supplies the training data and evaluation infrastructure that model developers depend on, may also occupy “well” positions in the emerging AI stack.
a16z’s Portfolio Reflects the Thesis in Action
Andreessen Horowitz has been putting significant capital behind this conviction. The firm raised a $4.25 billion fund in 2024 that included substantial allocations for AI infrastructure investments. Its portfolio includes companies across the AI stack, from model developers like Mistral AI to infrastructure plays like Anyscale, which provides the distributed computing framework Ray that has become widely adopted for training and serving large language models.
The firm has also made significant investments in companies operating at what might be called the “middleware” layer of AI — the connective tissue between foundational models and end-user applications. This includes investments in vector database companies, model monitoring platforms, and AI safety infrastructure. Each of these categories represents a potential chokepoint where a dominant player could extract significant value from the broader market.
Critics Question Whether the “Well” Will Hold
Not everyone is convinced. Some prominent investors and technologists argue that the AI infrastructure market is moving too quickly for any single company to maintain a durable chokepoint advantage. The rapid pace of model development, the emergence of increasingly capable open-source models from Meta’s Llama family and others, and the ongoing efforts by hyperscalers like Google, Microsoft, and Amazon to build their own custom AI chips all suggest that today’s infrastructure bottlenecks may be tomorrow’s commodities.
There is also the question of whether the analogy to natural resource extraction holds up under scrutiny. Oil wells derive their value from the physical scarcity of petroleum deposits. AI infrastructure, by contrast, is built on technologies that can, in principle, be replicated and improved upon by well-funded competitors. The barriers to entry in AI infrastructure are high — building a competitive GPU cluster requires billions of dollars in capital — but they are not insurmountable, particularly for the technology giants that already possess both the capital and the technical expertise to do so.
The Broader Implications for Startup Strategy
For founders, the Theory of Well carries practical implications for how they position their companies and pitch to investors. Startups that can credibly claim to occupy a chokepoint position — controlling a scarce resource, a critical data asset, or an essential piece of infrastructure that other companies depend on — may find a more receptive audience on Sand Hill Road than those building applications that sit further up the stack.
This dynamic is already visible in the fundraising market. Infrastructure-focused AI startups have commanded some of the highest valuations of the current cycle. CoreWeave raised at a valuation exceeding $35 billion in early 2025, while Databricks closed a $10 billion round at a $62 billion valuation. These numbers reflect investor conviction that infrastructure positions will prove more durable than application-layer businesses, many of which face the constant threat of being replicated by the foundation model providers themselves.
Where the Smart Money Is Heading Next
The Theory of Well also has implications for how large technology companies think about their own strategic positioning. Microsoft’s multi-billion-dollar investment in OpenAI, Google’s aggressive push into custom TPU chips, and Amazon’s development of its Trainium and Inferentia processors can all be understood as attempts to secure “well” positions in the AI stack. These companies recognize that controlling the infrastructure layer provides both economic advantages and strategic leverage over the application developers that depend on their platforms.
As the AI industry matures and the initial wave of hype gives way to a more sober assessment of where value actually accrues, frameworks like the Theory of Well are likely to become increasingly influential. The thesis does not guarantee that every infrastructure investment will succeed — the history of technology is littered with infrastructure companies that were outmaneuvered by more agile competitors or disrupted by architectural shifts they failed to anticipate. But it does provide a coherent framework for thinking about where the most durable competitive advantages in AI are likely to emerge.
For now, the venture capital industry is watching closely to see whether a16z’s conviction in the infrastructure layer proves prescient or premature. The answer will depend not just on the performance of individual portfolio companies but on the broader structural evolution of the AI industry over the next several years. If Midha and his colleagues at a16z are right, the companies that control the wells of AI will be among the most valuable enterprises of the coming decade. If they are wrong, the Theory of Well may join a long list of elegant investment theses that failed to survive contact with reality.