Romeo Dean (Harvard / IAPS) (@romeo_dean)
Co-author of the AI 2027 Compute & Security Supplements and Fellow at the Institute for AI Policy and Strategy (IAPS) tracking semiconductor supply chains and FlexHEGs.
Harvard CS/Physics • AI 2027 Compute Co-Author
Immutable Forecast Track Record (0)
Authored & Cited Arguments (4)
US export controls on CoWoS and EUV lithography rationally force Beijing to nationalize and pool its fragmented domestic compute into a single Centralized Development Zone ( ~12% global share).
While US hyperscalers control ~70% of global AI compute across 4–5 competing frontier labs, China's ~12% domestic + smuggled chip stock is initially split across dozens of commercial and provincial clusters. Centralizing 75%+ of that capacity at a nuclear-powered hub (like Tianwan) allows DeepCent to match the single-run training scale of a top US lab.
Confidential GPU TEEs + optical air-gapping can bring a Stargate-scale cluster to RAND SL5 by 2028 (41.5% FAR.Labs expert poll).
Modern Blackwell/Rubin confidential compute modes encrypt weights in HBM and across NVLink domains; pairing hardware attestation with physical air-gapping and two-person cryptographic authorization gates makes weight theft infeasible even with a compromised host OS.
DPA Compute Consolidation (50% Global Share) Makes the Slowdown Branch (36%) Strategically Viable Without Unilateral Disarmament
The Slowdown branch does not require unilateral disarmament: because US labs collectively hold ~70% of global compute vs. China's ~12%, invoking the Defense Production Act to merge the top 5 US labs into an audited English-CoT project (Safer-1) preserves a 5:1 hardware advantage that more than offsets the 2.5x Neuralese capability tax.
In ~32% of Plan B scenarios, credible interdiction of illicit chip smuggling and power substations brings reluctant states back to a hardware-verified accord by 2032.
Once a rival state realizes it cannot secretly assemble a 5-GW cluster without triggering supply-chain interdiction or cyber sabotage, accepting mutual on-chip FlexHEG verification becomes strictly preferable to unilateral exclusion.