Artificial intelligence is dominating most news media, with coverage oscillating among fears for humanity, announcements of breakthrough advances, and trepidation about the impending infrastructure investment that could dwarf every previous industrial and technological revolution. Anthropic and OpenAI, the corporate developers of the most advanced “frontier” AI systems, hope, audaciously, to raise capital that values each of them close to $2 trillion despite the headwinds.
There is no precedent for these valuations except the recent offering by SpaceXAI that plowed new financing ground, but its business is diversified across space, communications services, computing infrastructure, and AI software. The enthusiasm for OpenAI and Anthropic is entirely grounded in their meteoric speed of revenue growth, now approaching $50 billion in annualized terms. Sustaining such growth is risky business.
Disclosures of AI misbehavior keep surfacing, and public concern about AI safety, seeded by Anthropic’s and OpenAI’s own pronouncements and sparked by these incidents, has exploded. Fear, a primal emotion to begin with, has been amplified by stealthy online campaigns from U.S. adversaries and our domestic political mud fights. The most capable AI is also the most prone to misuse and unintended behavior, which will expose Anthropic and OpenAI to developmental slowdowns if the public pressure intensifies.
Other strategic challenges for Anthropic and OpenAI, discussed below, offer additional reasons for caution.
Scrutiny of spending. Customers’ return-on-investment calculations are constantly evolving, based on value generated, cost-effectiveness, and newly available alternative solutions. Corporate spending levels have become highly transparent, no longer lost in IT or departmental budgets.
Anthropic’s newly amended IPO filing lists about $895 billion of commitments, nearly all of it for computing and data centers, almost seven times the roughly $132 billion of cash and short-term investments the company held at the end of September. At the same time, the migration of applications to open-source models — those that can be downloaded for free, run on less expensive hardware, and customized to business requirements — is often a better fit for production applications than extensive experimentation and product development efforts.
Pricing opacity. The nature of AI models entails unpredictability in the amount of computing power needed to respond to any given request. Customer cost surprises are common. As a result, the current pricing model of charging for “tokens,” the diced-up bits of information that are the internal components of AI processing, is becoming fragile. In response, token-based prices are plummeting.
Sources of market growth. While there is no limit in sight for overall demand for computing intelligence, that does not automatically create a defensible niche at the frontier. As more applications enter production, a larger portion of the market will effectively become commoditized. The market for defense applications may split off as requirements and controls diverge from commercial markets. Some applications, such as automated software development, have diminishing headroom for improvement, which could diminish the competitive advantage of Anthropic and OpenAI in those spaces.
Datacenter revulsion. Creating new AI models depends on massive computing power, making the growing public resistance to data centers a strategic threat. Google, Meta, SpaceXAI, Nvidia, et al., are already heavily invested in computing infrastructure whose costs are partly amortized over their other product lines. Vertical integration across software, computing infrastructure, chips, and energy to reduce supplier dependencies and assure resource availability is underway. Anthropic and OpenAI are least prepared to similarly diversify their risk.
Continued push for regulation. The public perceives the fear of out-of-control AI but not the opportunity costs of impeding AI evolution. Government regulation therefore continues to loom large. This will accelerate the move to run models on in-house systems or on computers dedicated by a provider to a single customer. These private systems, inaccessible to the public, will justifiably be somewhat immune to regulation. Anthropic and OpenAI are not set up for private delivery of their services and therefore, contrary to their hopes, regulation may spell the loss of market share.
Future leadership. China is moving at full speed to capture the dominant position in the AI space. Their goal is to capture as many developers and applications of frontier AI as possible, dramatically increasing their leverage over the world economy and reducing the market for Anthropic, OpenAI, and other U.S. AI companies. Our current political and public zeitgeist is working in China’s favor.
Thus, the frontier intelligence business model is under severe pressure technically, competitively, economically, legally, and politically. Investors may be blinded by the torrid growth experienced by Anthropic and OpenAI and fail to sufficiently account for the strategic risks. Can these two companies walk a flimsy tightrope to sustained dominance with their current business model? The market will render its own judgment, but the long-term outlook is cloudy at best.
Andrew I. Fillat spent his career in technology venture capital and information technology. His M.I.T. master’s thesis pioneered relational and object database technology, and he holds an MBA from Harvard. Henry I. Miller, a physician and molecular biologist, is the Glenn Swogger Distinguished Scholar at the Science Literacy Project. They were undergraduates together at M.I.T.