Mihir Rao

Disclaimer. These views are entirely my own and do not reflect those of my current or past employers.

Knowns (as of July 2026)

A. The lab duopoly has hit recursive self-improvement. This will yield an exponential increase in the pace of general models climbing domain-specific evals via RLVR. The labs can curate these evals both internally and via external vendors like Mercor or even legacy/incumbent organizations within verticals (e.g. pharma for drug development evals). In summary, the labs own general purpose models which are sufficiently capable at specific domains.

B. Although varying in their technical and operational approaches, third-party AI service providers like Palantir, Databricks, and Applied Compute bet on the existence of valuable customer-specific "tribal knowledge" (information, processes) and thus unique evals and corresponding hill-climbs which justify "forward-deploying" into customers in order to compound the latent value sitting within the customer (for the customer). We can call this phenomenon customization ops. This is instead of being entirely product-driven, which suffers from a lack of specificity which prohibits effective compounding around unique hill-climbs of value to the individual customer.

C. There are focused, vertical, and even subdomain-specific, companies that are internally AI-native, and aim to compound on their own expertise. Examples include Formation Bio, Cambium AI, Anduril, and SpaceX. They have verticalized, owning the full stack, and operate such that they can defend against the labs infiltrating their domain. They achieve this via moats in, amongst others, the following categories: infrastructure, operations, data, relationships, legal. These are real, high-activation-energy moats against digital AI from the labs, not a weak "moat" in the way that Palantir used to have an operational moat with the FDE model (which is now commoditized).

Possible Futures

Let's throw a few, mutually-inclusive, possibilities on the table.

1. Wide, expert models, multi-million dollar and week- to month-long agent runs on high $-ROI problems/use-cases. In this reality, there will be labs, and verticalized firms serving the labs in the form of evals for RLVR, perhaps through intermediate eval middlemen.
   • Think about one-shot designing of a brain-penetrant Alzheimer's small molecule by reviewing all available literature and data, automatically setting up an RL loop to evolve from a seed population of known but suboptimal AD drugs, and then automatically dispatching orders to remote CROs to validate candidate molecules in the lab, all done within a provided finite $ and time budget.
   • In order for the agent to be effective at this task, it needs to have hill-climbed a drug development eval, perhaps even a CNS-specific eval. This eval would either need to be curated within the lab, or provided to the lab via an eval contractor or through a partnership with a biotech/pharma. This specific eval is subject to the difficulty of developing verifiable rewards in a problem space with expensive trajectories (clinical trials are more slow and more expensive than human feedback on prompt/response pairs) and low grindability (for the same cost and time reasons, it is hard to run many similar but slightly perturbed trials in parallel).
   • Players: Labs, eval providers, verticalized firms
   • What remains valuable: physical tool calls, evals to improve models in labs, ideas for what problems/creations to run long-horizon agents on
2. Customization ops persists as valuable, and labs and customizers compete for forward deployment contracts. Labs may choose to verticalize in some domains rather than competing on customization ops (less competition, higher reward if successful; customization ops may also be failing in a vertical i.e. life sciences). This is already happening today (see Anthropic Life Sciences). Customizers may also choose to verticalize, or get acquahired by labs.
   • Players: Labs, customizers, verticalized firms
   • What remains valuable: customer access and visibility into processes, domain expertise to make harder use cases more tractable, an ability for customizers to do things that a team within a lab cannot (e.g. Valkai v.s. Anthropic LS FDEs). Everything that remains valuable in possible future #1.
3. SQUEEZE. Customization ops fails as an ecosystem layer. On one side, due to a lack of control on underlying intelligence and the labs' ability to reproduce customization ops effectively internally (perhaps via acquahire). On the other side, due to a lack of customization ops efficacy on customer use cases. This is clearly not the case at least on operational use cases given Palantir's success. There are, however, limitations to customization ops as use cases become more technical and demand much more subject-matter expertise.
   • Players: Labs, verticalized firms, sufficiently efficacious customizers
   • What remains valuable: the ability to build, improve and customize models, domain-specific moats (infra, people, data, IP, legal, financial). Everything that remains valuable in possible future #1.