The biggest risk facing the AI industry may not be that artificial intelligence fails to live up to its promise, but that it becomes too cheap to support the extraordinary sums being spent to build it.
OpenAI and Anthropic are committing enormous amounts of capital to computing power, data centers, chips and other infrastructure in a race to remain at the frontier. At the same time, Chinese developers such as DeepSeek, Alibaba’s Qwen, Zhipu AI and Tencent are narrowing the performance gap while offering models at substantially lower costs.
That combination is raising a difficult question for investors: what happens to the economics of the frontier AI business if customers can obtain sophisticated models without paying the premium required to support the industry’s enormous infrastructure buildout?
Register for the next Tekedia Mini-MBA.
Register for Tekedia AI in Business Masterclass.
Join Tekedia Capital Syndicate and co-invest in great global startups.
Scott Wilson, whose early investment in SpaceX generated a multibillion-dollar windfall for Washington University’s endowment, believes the answer could be painful for the industry’s most highly valued companies.
“These trillion-dollar-plus frontier companies are not worth the liabilities that they signed up for,” Wilson said. “There’s going to be a ton of free alternatives.”
Wilson’s argument challenges one of the central assumptions behind the AI investment boom: that spending more than competitors on infrastructure will necessarily create a durable competitive advantage. His view is that the opposite could happen if advances in open-weight and lower-cost models make AI increasingly difficult to monetize.
The leading AI companies are spending hundreds of billions of dollars on computing capacity, data centers, chips, and related infrastructure in an effort to maintain an advantage over competitors. The underlying assumption is that superior models will command enough revenue to justify that investment.
But if increasingly capable models become available at dramatically lower prices, the economics could change. Customers may become less willing to pay premium prices for proprietary systems if open-weight or lower-cost alternatives deliver comparable results.
That is the scenario Wilson believes the market is underestimating.
The China Challenge Is Changing The Economics
Wilson’s skepticism has hardened as Chinese model developers have narrowed the performance gap with US frontier laboratories. Companies including DeepSeek, Alibaba’s Qwen, Zhipu AI and Tencent have emerged as credible alternatives, while open-weight models have made it easier for businesses to deploy AI without relying entirely on the largest US providers.
Wilson said conversations with colleagues operating in China have reinforced his view that progress in open-weight models is happening rapidly.
He has also been hearing similar signals from companies in Washington University’s investment portfolio.
“Whenever we talk to our portfolio companies, especially the ones who are heavy of AI, they are all moving aggressively towards open source,” he said. “It’s like any other high-cost U.S. good that has to compete with a low-cost import, particularly from China.”
That comparison captures the heart of Wilson’s thesis. If AI models become increasingly commoditized, the economics could begin to resemble other industries in which a high-cost producer struggles to maintain pricing power against a lower-cost competitor. The issue would not necessarily be whether US laboratories can build better models, but whether the incremental improvement is large enough for customers to justify paying substantially more.
OpenRouter data provide some evidence of the shift, although the platform represents only one slice of the AI market.
DeepSeek accounted for 25.3% of text-model requests on OpenRouter, compared with 18.6% for OpenAI and 2.9% for Anthropic.
The figures do not establish overall market share or revenue, and OpenRouter users are not necessarily representative of enterprise AI customers. They nevertheless illustrate the growing willingness of developers to experiment with alternatives to the dominant US laboratories.
That creates a difficult proposition for OpenAI and Anthropic. Their enormous capital requirements are based partly on the expectation that customers will continue to value frontier performance enough to sustain premium pricing.
If the performance gap narrows faster than the cost gap, that assumption becomes harder to defend.
Khosla Sees Infrastructure As The Moat
Vinod Khosla, the billionaire venture investor and early OpenAI backer, sees the economics very differently.
He believes that Wilson is focusing too heavily on the model itself and not enough on the infrastructure required to build, train, and operate it.
“People like that are silly, and they don’t understand how this works,” Khosla said when told of Wilson’s view. “They have this notion that the model is the value.”
Khosla’s argument is that a closed-model company can eventually control more of the technology stack, allowing it to lower its underlying costs even if competitors offer models for less.
That stack includes computing infrastructure, electricity, data centers, chips, software, and inference systems.
OpenAI, for example, is working on its own inference hardware, including its Jalapeo inference chip, which Khosla says could reduce the company’s dependence on Nvidia and third-party cloud providers.
“I’m not talking price, I’m talking about cost,” he said. “From power to data center to chips, to infrastructure software to inference models, the cost of the stack is almost certainly going to be lower in closed-source models than open-source.”
This is the central counterargument to Wilson’s thesis.
An AI company may be able to charge less for its models while still maintaining attractive margins if it can lower the cost of producing each response. Owning or co-designing critical infrastructure could give frontier laboratories advantages that are not visible when models are compared simply on headline API prices.
The difference is between price and unit economics.
A cheap open-weight model can be attractive to customers, but running that model may still require expensive computing infrastructure. A vertically integrated AI company could potentially offset higher model development costs by lowering its cost per inference through specialized chips, optimized software, and dedicated data centers.
If Khosla is right, today’s infrastructure spending is not simply an expense. It is the foundation of a cost advantage that could become more important as AI usage scales.
The Real Test Is Whether AI Becomes A Commodity
The disagreement exposes two competing visions of the AI industry. Wilson’s thesis assumes that model capabilities will converge quickly enough that the model itself becomes difficult to monetize. Under that scenario, customers will have little reason to remain locked into expensive proprietary systems when cheaper alternatives are available.
Khosla’s thesis assumes that the leading laboratories will maintain meaningful technological advantages while simultaneously gaining control over the infrastructure underneath their models. That could allow them to reduce costs, improve performance, and retain pricing power.
The outcome depends heavily on the pace of commoditization.
If models become interchangeable, the enormous capital commitments made by OpenAI and Anthropic could become problematic. The companies would have to generate enough revenue to service infrastructure commitments even as customers demand lower prices.
If frontier models remain materially better, however, the infrastructure race could reinforce their lead. Training sophisticated models requires enormous amounts of computing power, and the companies with access to the largest pools of capital could maintain advantages that smaller competitors cannot easily reproduce.
There is also a middle scenario in which the model layer becomes cheaper while value migrates elsewhere.
AI customers may ultimately care less about which laboratory produced the underlying model and more about applications, proprietary data, distribution, workflow integration, and reliability. In that world, model providers could face falling margins even as the broader AI economy continues to grow.
The implication is expected to be significant for companies valued on the assumption that model development itself will capture a large share of the industry’s eventual profits.
OpenAI And Anthropic Face A Different Test From Their Investors
The debate comes at a critical point for both companies. OpenAI has already demonstrated that investors are willing to assign enormous value to its position at the center of the AI ecosystem. Anthropic, meanwhile, is expected to go public next month, putting its financial model under much closer scrutiny from public-market investors.
Then, the question will not simply be whether the companies can produce increasingly capable models. It will be whether those models can generate enough recurring revenue and gross profit to justify the capital required to develop and operate them.
For now, the market has largely rewarded the spending race. The biggest laboratories have attracted extraordinary amounts of capital, while chipmakers, cloud providers and data-center operators have benefited from the resulting infrastructure buildout.
But a large infrastructure footprint is only an advantage if demand and margins grow quickly enough to absorb it.
Wilson’s SpaceX investment exemplified the power of being early to a technological transformation. His current argument is effectively that investors may be making the opposite mistake with AI: paying enormous prices for companies before the economics of the industry have been established.
Khosla believes those economics are precisely what the infrastructure race will establish.