Better Dishwashers

Paradoxically, profit for frontier labs will occur when they run out of resources.

Under the assumption that scaling doesn't plateau, excess margins from revenue-driving activities will be reinvested into training larger models until the marginal dollar allocated to training isn't worth the accompanying capability gains. When there's significantly more demand than available resources (consider "resources" as short-hand for the combination of compute & energy (Fig 1)), the allocation shifts to inference (Fig 2). As long as that inference is targeted at their high willingness-to-pay customers in markets where they've established short-lived monopolies, the high margins associated with their services will finally allow the labs to accrue revenue in excess of their total costs.

The labs need high-margin revenue streams to survive this mix-shift and accrue enough free cashflow in excess of their fully-loaded costs when resource supply runs low. To create these revenue streams, the labs will need to transition to repeated cycles of short-lived monopolies across different domains. This transition will establish the labs as the bottleneck in the intelligence supply chain where they can shift pricing from markup on compute.

Josephine Cochrane was able to establish a short-lived monopoly in a single domain before it was cool. She patented the dishwasher in 1886. By 1893, with a little bit of luck, commercial demand had exploded. The dishwasher offered immense time & cost savings for restaurants and was such a superior good to dishwashing by hand that it established a new market.

Josephine had done everything right: She was the first mover, with state-backed IP protection, in a market devoid of competitors. Yet with all these advantages, the Garis-Cochran Dishwashing Company was still swallowed up by competitors a mere 30 years later.

Barnett's argument describes this same trap for the frontier labs: move too fast and you cannibalize yourself, move too slow and open source distills you. By assuming that AI improvements are incremental gains in one market, it's inevitable that they'll become subject to commoditization and price competition (Fig 3).

But frontier labs aren't building better dishwashers. The frontier labs are building new markets where they can identify demand and create differentiated supply, long before competitors can catch up. Gating access to their frontier products can allow the labs to allocate inference to high willingness-to-pay customers and works as a synthetic patent, preventing quick distillation from competitors, creating short-lived monopolies in those new markets. The business model transition will look like a shift from the pay-per-token, public API paradigm to labs providing verticalized, results-oriented services with differentiated products per domain and, in some cases, per customer.

Within 10 years of launching Prozac, Eli Lilly had found itself with one of the best-selling drugs in the history of the pharmaceutical industry. Their first-mover advantage in the easily-prescribable-anti-depressant market, combined with patent protection, had resulted in a transformative source of revenue for the company. However, the short-lived monopoly ended when their patent wore out. Without state-backed suppression of competitors, generic brands flooded the market and crushed the pricing power Eli Lilly had.

Yet, 30 years later, Eli Lilly has not been swallowed up. At their current $880 billion market cap, they are the 14th most valuable publicly traded company in the world. Their GLP-1 product, tirzepatide, was one of the first advanced weight-loss drugs to market and has established a duopoly with Novo Nordisk. Eli Lilly has made a business of establishing short-lived monopolies in new markets. In other words, the business model works (Fig 4).

While Eli Lilly derives their short-lived monopolies from first-mover advantage and state-backed IP protection, the frontier labs cannot rely on the state to prevent competitors. Instead, frontier labs will have to create synthetic patents through capability-driven monopolies where very few competitors can quickly catch up to their services.

The cases where repeated, capability-driven short-lived monopolies fail are cases when scaling plateaus. In cases where scaling plateaus before labs can create a differentiated product for a given market, they will fall victim to commoditization economics. Scaling reaching a plateau after labs have established footholds into a diverse set of markets is actually a bull case for quicker profits. If scaling stops then, labs can't reinvest margin into training larger models even if they want to, so surplus from revenue-driving inference becomes profit even earlier than it would when the compute cap forces it.

The short-lived monopoly lifecycle for frontier labs consists of 4 steps:

1. Stage 1, enter a new domain via public APIs, chat interfaces, or FDEs.
2. Stage 2, verticalize into the ecosystem (Claude Code, Codex).
3. Stage 3, scale capabilities past the ecosystem and gate access to the product/service (Mythos, Glasswing), effectively establishing a new market with their superior product.
4. Stage 4, convert to outcome-based pricing with deep integration into the client's business.

Then repeat across domains.

This short-lived monopoly lifecycle has already begun. Take coding/cybersecurity for example: Stage 1 was the advent of public APIs and the chat interface. Anthropic found tremendous demand and a high willingness to pay for coding augmentation use cases. This demand translated into Claude Code, or Stage 2, verticalizing them into the ecosystem. By gating Mythos, Anthropic has created the perfect conditions for a short-lived monopoly, where they are the only known producer at Stage 3.

The next logical step would be Stage 4, where Anthropic operates as the price maker, providing results-as-a-service and pricing based solely on the outcomes driven by using their product. In this scenario, the client never touches the model at all. Instead, deeply integrated FDEs could drive the model-harness, pricing based on security/reliability metrics such as uptime without security-related SEVs or elimination of hazardous zero-days (Fig 5).

While Stages 3 & 4 are currently happening in cybersecurity, Stage 1 is running for every domain. We are seeing the beginnings of Stage 2 in finance/general knowledge work (Claude Cowork, investments into deep integrations). Arriving at Stage 3 in just a few domains will establish Anthropic as the bottleneck in the intelligence supply chain. Combined with a world where resource demand further outpaces GPU or energy supply, they will not be able to reinvest all of their revenue into training and will be forced to turn a profit.