Somewhere in the quiet hours before dawn, a venture capitalist is staring at a ceiling, running numbers that don’t add up. The AI boom has minted paper fortunes, inflated valuations to levels that would make dot-com investors blush, and created a gold-rush mentality not seen in Silicon Valley since the iPhone launched. But behind the euphoria, a growing cohort of sophisticated investors is nursing a set of anxieties that no earnings call or product demo can fully soothe.
The worries are specific, structural, and in some cases existential — not just for individual companies but for the entire investment thesis underpinning generative AI.
According to The Information, investors across the venture capital and public markets spectrum are increasingly fixated on a handful of risks that threaten to undermine the sector’s staggering growth trajectory. These aren’t the concerns of Luddites or skeptics. They’re coming from the people writing the checks.
Here’s what’s eating at them.
The Revenue Mirage and the Infrastructure Mismatch
The most fundamental anxiety is deceptively simple: Are AI companies actually going to make money commensurate with the capital being poured into them? The gap between AI infrastructure spending and AI revenue generation has become a chasm that even the most optimistic financial models struggle to bridge. Nvidia’s data center revenue has exploded. Microsoft, Google, Amazon, and Meta have committed hundreds of billions to AI-related capital expenditure. But the downstream revenue — the actual money that AI applications generate for the companies deploying them — remains stubbornly modest relative to the investment.
This isn’t a theoretical problem. It’s an accounting one.
Capital expenditures by the major hyperscalers are projected to exceed $200 billion in 2025 alone, much of it directed at AI infrastructure — chips, data centers, cooling systems, power generation. The question investors keep circling back to: Who is going to pay for all of this? Enterprise adoption of AI tools is growing, but slowly and unevenly. Many companies are still running pilots. Proof-of-concept projects. Internal experiments that haven’t translated into the kind of large-scale, recurring software contracts that would justify the infrastructure build-out.
Sequoia Capital’s David Cahn made waves last year when he estimated that AI companies would need to generate $600 billion in annual revenue just to cover the cost of the infrastructure being built to support them. That number has only grown. And the revenue isn’t there yet.
Some investors point to the early internet as a parallel — massive infrastructure investment that took years to pay off but eventually created trillions in value. Others counter that the internet analogy is dangerously misleading because the economics are fundamentally different. Running AI models is expensive. Every query costs money. Unlike a web page that can be served for fractions of a cent, an AI inference requires significant compute. The marginal cost problem hasn’t been solved, and it may not be solvable with current architectures.
So the money keeps flowing in, and the returns keep not flowing out. Not yet, anyway.
There’s also the competitive dynamic that’s making investors lose sleep. OpenAI may have kicked off the generative AI era, but it’s now fighting on multiple fronts — against Google’s Gemini, Anthropic’s Claude, Meta’s open-source Llama models, and a growing roster of well-funded startups. The moat question haunts every AI investment. If the underlying models are converging in capability, and if open-source alternatives continue to close the gap with proprietary ones, where does pricing power come from?
Nowhere good, if you’re an investor.
The commoditization risk is real. Model performance benchmarks show diminishing differentiation among the top-tier foundation models. What was a clear OpenAI advantage eighteen months ago has narrowed considerably. Anthropic’s Claude 3.5 and Google’s Gemini Ultra perform competitively on most tasks. Meta’s Llama 3 models, available for free, are good enough for many enterprise applications. This means the value may not accrue to the model makers at all but to the companies that build applications on top of them — or to the incumbents who already own the customer relationships.
That’s a problem for venture investors who’ve bet billions on foundation model companies at eye-watering valuations. OpenAI’s reported valuation of $300 billion. Anthropic at $60 billion. xAI at $50 billion. These numbers assume winner-take-all or winner-take-most dynamics. But the market is looking increasingly like it will support multiple viable players, which compresses margins and limits upside.
The Regulation Overhang and the Talent Squeeze
Then there’s regulation. The European Union’s AI Act is already in effect, imposing compliance requirements that will cost companies real money and constrain how models can be deployed. In the United States, the regulatory picture is murkier but no less concerning. The Trump administration has taken a lighter-touch approach compared to what a second Biden term might have pursued, but state-level regulation is proliferating. California’s proposed AI safety bills, while vetoed by Governor Newsom in 2024, will return in some form. And the bipartisan consensus that something needs to be done about AI — even if no one agrees on what — creates persistent uncertainty.
Uncertainty is what investors hate most.
Copyright litigation adds another layer of risk. The New York Times’ lawsuit against OpenAI and Microsoft is proceeding, and its outcome could reshape the economics of training AI models. If courts determine that training on copyrighted material constitutes infringement, the cost of developing new models could increase dramatically. Licensing deals would become mandatory rather than optional. Some estimates suggest this could add billions to the cost of training next-generation models.
And then there’s the talent problem. The pool of researchers and engineers capable of doing cutting-edge work on large language models and other AI systems is remarkably small. The competition for this talent has driven compensation to absurd levels — senior AI researchers at top labs command packages worth $5 million to $10 million annually. This creates a cost structure that’s difficult to sustain, particularly for startups that don’t have the cash reserves of a Google or Microsoft.
The talent wars have also produced an unusual amount of executive instability. OpenAI’s boardroom drama in late 2023 was the most visible example, but departures and defections have become routine across the industry. Key researchers leaving to start competitors. Founding teams fracturing over strategic disagreements. This kind of organizational turbulence is normal in fast-growing industries, but it adds risk to already risky bets.
Beyond the company-specific concerns, there’s a macro worry that few investors want to discuss publicly but many acknowledge privately: What if AI just doesn’t deliver the productivity gains that justify the investment? The economic evidence so far is mixed. Some studies show significant productivity improvements from AI tools in specific contexts — coding, customer service, content generation. But economy-wide productivity statistics haven’t moved meaningfully. The gap between the micro evidence and the macro data is troubling.
It echoes Robert Solow’s famous 1987 observation about computers: “You can see the computer age everywhere but in the productivity statistics.” The productivity gains from computing eventually materialized, but it took more than a decade and required complementary investments in organizational change, worker training, and business process redesign. AI may follow the same pattern. But investors operating on three-to-five-year fund cycles don’t have a decade to wait.
There’s also the energy problem. Training and running AI models requires enormous amounts of electricity. Data center power consumption is projected to double or triple by 2030. Utilities are struggling to keep up. New natural gas plants are being built. Nuclear power is making a comeback specifically because of AI demand. But permitting and construction timelines for new power generation capacity are measured in years, not months. This creates a physical bottleneck that no amount of software optimization can fully overcome.
Some of the smartest money in the room is hedging. Tiger Global, which was among the most aggressive AI investors in 2023, has reportedly become more selective. Venture firms that previously competed to lead AI rounds are now spending more time on due diligence and pushing back on valuations. The secondary market for AI company shares has shown signs of softening, with some late-stage companies trading at discounts to their last primary round.
None of this means AI is a bubble about to burst. The technology is real. The use cases are growing. The largest technology companies in the world are betting their futures on it. But the distance between “AI will be transformative” and “this specific AI investment will generate attractive returns” is vast. And it’s in that gap where the anxiety lives.
The investors lying awake at night aren’t questioning whether AI matters. They’re questioning whether the math works — whether the valuations, the capital requirements, the competitive dynamics, and the regulatory risks all add up to returns that justify the risk. For many of them, the answer is still yes. But it’s a quieter, more conditional yes than it was a year ago.
And in venture capital, when the yeses start getting quieter, the market is usually about to change.
WebProNews is an iEntry Publication