Paul Bricman

Geopolitics of Computation

Habitable space is a finite resource, there is only so much of it at any given point in time. It is also one from which economic value can be extracted with different levels of efficiency, through labor, agriculture, manufacturing, etc. It is also one whose ownership is exclusionary, with land in principle belonging to one particular geopolitical unit, such as countries, city states, etc. It is also one which prompts defenses against invasions, coups, etc. It is also one which prompts trade, with the fruit of specialized land enabling mutually beneficial transactions. It is also one which prompts international politics and diplomacy, with alliances of geopolitical entities collaborating on tackling global issues affecting all their members, such as climate change.

Much of this can also be said about compute. It is also a finite resource, as there is only so much of it at any given point in time. It is also one from which economic value can be extracted with different levels of efficiency, through automated knowledge work, mining cryptocurrency, sustaining valuable products, etc. It is also one whose ownership is exclusionary, as we can’t both run workloads on the same processor thread at the same time. It is also one which warrants security against breaches, cyberweapons, etc. It is also one which prompts trade, with compute specialized into different arrangements enabling entire ecosystems of services. However, it may not appear to call for a politics of its own at first glance. Yet it may be that reconceptualizing compute governance as a full-blooded geopolitics of computation may make it easier for international alliances or corporate consortia to collaborate on tackling global issues affecting all their members, such as advanced artificial intelligence.

Why do embassies exist? Why don’t representatives of nation states simply drop by for the scheduled meetings, and then head back home? It appears that embassies directly serve multiple strategic functions. First, they provide means for sending nations to understand and monitor the political, economic, and social climates in host countries through direct observation and local contacts. Second, they provide means for securely communicating information to and from the sending nation from within the host nation. Third, they provide means for enacting immediate responses to various types of crises, due to the highest possible proximity. Fourth, they provide means for providing further consular services to citizens abroad, especially concerning identity documents, notary attestation, etc. Finally, through all these direct functions, embassies indirectly enable higher-level coordination between the two states.

The following coincidentally documents popular features of trusted platform modules (TPMs), dedicated chips specialized in cryptographic operations. First, they provide means to probe different parts of the system they are embedded in, such as the bootloader used by a virtual machine. Second, they provide means for securely communicating information to and from an external entity from within the host system, using on-chip private keys to facilitate asymmetric encryption. Third, they provide means for immediately inputting signals into the host system depending on arbitrary properties of its state through communication across chips. Fourth, they provide lower-level cryptographic services concerning verifiable signatures, digital certificates, etc. However, the potential of TPMs to act as veritable “digital embassies” is far from being actualized.

There is also something to be said about the minimal invasiveness of embassies. It is certainly not the case that sending nations get too deeply involved with the general functioning of the host nation. Kissinger might phrase it as “the inadmissibility of intervention in the domestic affairs of other states.” It is not the case that diplomats afford to reorganize the entire host nation by directly participating in its own internal governance. Nevertheless, they may well be the most invasive arrangement which is politically palatable. Similarly, cryptographic interventions which require entire computations to be reorganized in their image, such as homomorphic encryption requiring polynomials in a finite ring or zero-knowledge proofs requiring arithmetic circuits, may be less politically palatable. Labs will want to reign freely over their resources, especially when involving vital processes such as models trained, architected, and served in special ways.

Besides the above, trusted platform modules are only one tool in a growing menagerie of hardware-enabled and cryptographic primitives, which further includes trusted execution environments, hardware security modules, secure elements, digital emblems, physical unclonable functions, security coprocessors, smart contracts, memory encryption, etc. Many of these appear to similarly map to social, political, and diplomatic constructs in productive ways, and so could likely facilitate coordination if arranged appropriately. It is at once disheartening and encouraging to realize that the majority of accelerators presently used by the frontier labs to advance AI are increasingly equiped with hardware features which could enable such efforts, yet their average utilization remains low. Examples of such accelerators include Nvidia’s latest Hopper and Blackwell generations of GPUs. It feels as if the puzzle pieces are simply waiting to be connected.

For better or worse, AI development is embedded in a complex geopolitical context, and efforts which model it as a spherical cow in a void may not prove as effective as they can be. A full-blooded geopolitics of computation may offer a palatable way out of a non-cooperative zeitgeist.