The $180b question

Biosecurity

Biosecurity is underfunded and unlikely to be well-capitalized by others, especially since solutions need to be built globally. If AI-enabled bioweapons never produce a pandemic, natural ones will still occur, so the baseline risk alone justifies spending. The most impactful measures are ambient, permissionless, and pathogen-agnostic.

My priority would be a continuous metagenomic surveillance system for sewage, which would flag any novel pathogen within days. SecureBio, the Institute for Progress, and the FY2026 budget all estimate the cost of a national system at $50m/year. Scaled globally, including the cost of infrastructure for less developed regions, it would be ~$25b over a decade.

Another biosafety control, DNA synthesis screening, is voluntary and failing: 36 of 38 providers shipped 1918 flu fragments to a fake lab in 2024. I’d permanently endow IBBIS, the nonprofit running the open-source screening tool, to keep it current and ahead of AI. I’d also fund tamper-responsive, on-device screening for benchtop synthesizers, engineered so that stripping it out destroys the machine.

I’d make additional R&D bets, like funding research into Far-UVC, a potentially human-safe light that kills airborne pathogens. Altogether, these biosecurity programs would require at most $35b, with most of the allocation going to surveillance.

Compute verification

Compute verification is a rare bet that targets misalignment without needing a law in place. Hardware-Enabled Governance Mechanisms (HEMs) are tamper-proof add-ons to AI chips; they prove what a chip did without exposing the model and enable the enforcement of rules that constrain dangerous training.

The instrument is buildable now, but legislation that would force its usage doesn’t exist yet. When it’s passed, this hardware can be immediately deployed. And if HEMs are never required by law, the same work protects against model weight theft, which labs want anyway.

The few real attempts at compute verification (FlexHEG, TamperSec) are undercapitalized and early-stage. Serious funding across competing teams could run up to ~$10b over a decade, mostly because of the limited talent that can do frontier hardware work.

Regulation

The highest-leverage move doesn’t clear my filter, because it involves developer constraints. In a world where OAIF prioritizes safety over commercial interest, the biggest catalyst would be lobbying for mandated insurance as a precondition of operating. It’s upstream of almost all safety work: if labs are legally liable for what their systems do, then alignment, compute governance, and metrology become their problem to fund. Insurers would demand safety practices as a condition of coverage.

This is how cyber insurance spread, except it wasn’t compulsory; regulators required it in some sectors, and counterparties demanded it. But frontier labs have more control than their customers, and they’re capitalized enough to self-insure—they would need a legal mandate.

Funding this is cheap relative to other safety measures. Pharma, the largest industry lobby, spends ~$390m a year, or about $4b a decade. OAIF could put $10b behind a single issue, more than double the biggest sustained lobby in U.S. history, and still spend under 6% of its capital. If it worked, it would unlock the entire developer-constraint bucket, mobilizing hundreds of billions in lab spending by changing incentives. However, OAIF would never fund a lobby against itself, and despite being more friendly to regulation, the Anthropic founders likely wouldn’t either.

Deployment

OAIF shouldn’t run like a perpetual endowment. The things worth building take years and must exist at full capacity as soon as they’re needed. Fund spend-down should be anchored to that lead time, not AGI projections.

Spray-and-pray grantmaking also works against this goal. Outsized returns in capital deployment most often come from venture, and I’d apply the same strategy here: find founder-like missionaries for each problem above, give them a nine-or-ten-figure mandate and total autonomy, and hold them to short-term, verifiable milestones. This is very similar to the entrepreneurial, metric-driven FRO model that groups like Convergent Research have proven on a smaller scale.

But even with ten Altman clones at the helm, there’s not much more OAIF can or will do, because money isn’t the bottleneck for better AI outcomes. Good bets are limited, the best bet is infeasible, and no bet can deploy all $180b. The smartest move is to build defenses that pay off regardless, allocate some leftovers to non-catastrophic problems where money converts and the market won’t, and hold the rest for when the field develops further.