Innovation Is Not Enough: Survival in the Age of AI

The extraordinary expansion of artificial intelligence is creating an unusual economic environment. OpenAI, Anthropic, Google, and other frontier AI developers are competing in a market where slowing investment may itself be more dangerous than investing too aggressively. As long as improvements in models, inference capacity, and agentic systems continue to generate meaningful productivity gains, each company has a strong incentive to secure as much computing capacity as possible. This creates something resembling a survival game.

This blog post argues that innovation is not enough, being first is not enough, and scientific excellence is not enough. In a rapidly changing technological environment, none of these advantages guarantees survival. What may matter most is an organization's ability to repeatedly reorganize itself as the nature of competition changes. This is particularly important because yesterday's strengths can become tomorrow's constraints. A company that pioneered a technology may become too attached to the business model built around the previous generation. A research organization that has achieved extraordinary scientific breakthroughs may develop incentives and institutional prestige around fundamental research that make rapid commercialization more difficult. A company with enormous financial resources may accumulate layers of management and competing projects that slow decision-making rather than accelerate it.

The paradox is that organizations may therefore need to challenge the very structures that made them successful. Kodak's early development of digital photography did not guarantee leadership in the digital era. Similarly, Google's extraordinary contributions to modern AI, including the Transformer and Nobel Prize–winning research, do not by themselves guarantee leadership in the commercial AI race. DeepMind has historically demonstrated extraordinary strength in fundamental research, producing achievements such as AlphaGo and AlphaFold. This scientific orientation has generated enormous intellectual value and should be recognized as one of the great achievements of modern AI research. Yet the competitive environment created by ChatGPT placed increasing importance on a different capability: the ability to move rapidly from research to product, from product to users, and from users back into continuous improvement.

OpenAI's achievement was not the invention of the Transformer itself. Rather, it recognized how transformer-based models could be developed into a broadly accessible, general-purpose interface and then rapidly improved through massive real-world deployment. In doing so, OpenAI demonstrated that technological leadership can depend not only on the quality of research but also on the speed with which research is converted into products, distributed to users, and refined through feedback.

However, OpenAI's early leadership in generative AI similarly does not guarantee that it will remain dominant. The organizational structure that enabled a company to succeed during one stage of technological development may become less suitable for the next. As AI moves from experimental models and chatbots toward large-scale infrastructure, reasoning systems, agents, and enterprise deployment, OpenAI may face the same fundamental challenge as its competitors: adapting its organization as rapidly as the technology itself changes.

Survival may therefore require something more difficult than innovation itself: the willingness to repeatedly reorganize a successful organization before external forces make that reorganization unavoidable. In a technological survival game, the greatest danger may not always be technological failure. It may be becoming too successful at the organizational model that worked yesterday.

P.S. The survival game among frontier AI companies raises another question: who benefits even if it remains uncertain which AI company ultimately wins? During the California Gold Rush, many durable businesses were not the miners searching for gold, but the merchants selling picks, shovels, clothing, and other necessities. A similar economic structure may be emerging in AI. OpenAI, Anthropic, Google, and other frontier developers compete intensely, while the companies supplying the infrastructure required by all of them may benefit from the expansion of the industry itself.

Nvidia and Broadcom provide particularly interesting examples. Both supply critical AI infrastructure, but their products have different economic characteristics. Nvidia primarily sells relatively general-purpose GPU computing platforms supported by the CUDA ecosystem. Broadcom has become increasingly important in custom AI accelerators, or ASICs, as well as high-performance networking.

An ASIC, or Application-Specific Integrated Circuit, is optimized for a narrower set of workloads than a general-purpose GPU. At sufficient scale, this specialization can provide substantial advantages in performance, power consumption, and cost per computation. This is why hyperscalers and frontier AI companies are increasingly interested in custom accelerators.

Specialization, however, introduces another kind of risk. If one frontier AI company fails, Nvidia GPUs may still be useful to another model developer, cloud provider, research institution, or enterprise. They can potentially be reassigned to training, inference, scientific computing, or other workloads. Nvidia GPUs therefore have some characteristics of relatively liquid computing assets.

A custom ASIC may have a different residual-value profile. Its architecture, software environment, compiler stack, and system design may have been optimized for a particular customer or workload. If that customer fails, finding another buyer capable of using the hardware efficiently may be more difficult. The same specialization that creates superior economics during normal operation can therefore reduce collateral value during liquidation.

This distinction becomes particularly important as AI infrastructure financing grows. The capital required to build the next generation of AI infrastructure has become so large that financial institutions are increasingly involved in financing computing equipment and data-center capacity. Technology suppliers may also provide contractual commitments or other forms of support that reduce the risks borne by financing partners.

Such arrangements are not necessarily signs of weakness. They may simply indicate that AI infrastructure demand has become too large to be financed entirely through conventional corporate capital budgets. Nevertheless, investors have good reasons to examine them carefully. Semiconductor companies traditionally manufacture equipment, sell it, and receive payment. Once a supplier materially supports customer financing, some of the customer's credit risk can migrate back toward the supplier.

The relevant risk therefore depends not only on the probability that a customer defaults, but also on how much value can be recovered from the underlying equipment. This may give Nvidia an important structural advantage. If an individual AI company fails, its Nvidia GPUs could potentially be redeployed elsewhere. The greater risk would be a systemic decline in the economic value of AI computation, severe industry-wide overcapacity, or technological change that makes existing GPUs obsolete much faster than expected.

Broadcom's custom ASIC exposure may have somewhat different economics. If an accelerator is deeply optimized for one customer, its secondary market may be narrower and its recovery value lower. In financial terms, specialization can simultaneously improve operating economics and reduce collateral liquidity. This does not make Broadcom's strategy unattractive; custom silicon could become enormously valuable if AI workloads continue expanding. The important question is whether financing-related exposure grows proportionately with genuine economic demand.

This brings us back to the Gold Rush analogy. Selling picks and shovels can be an excellent business when miners compete aggressively for gold. But the economics change if the shovel manufacturer must also lend miners the money to buy the shovels. At that point, the supplier's profitability depends not only on how many shovels it sells, but also on whether the miners can repay their obligations—and on how much the recovered shovels are worth if they cannot.