Goldman Sachs' latest report forecasts global AI capital expenditure could reach $7.6 trillion by 2031, but the figure hinges on the useful life of dedicated AI silicon. Rapid innovation could render standard chips obsolete in three years, driving costs higher; a tiered model repurposing older chips for inference could stabilize spending. The report flags that legacy financial systems—with slow settlement cycles and rigid KYC frameworks—are ill-suited for the speed of agentic trading, positioning crypto and decentralized protocols as essential “permissionless economic rails”.
DePIN's Latency Tradeoff and the ZK Breakthrough
Decentralized physical infrastructure networks (DePIN) offer dramatic cost savings—Akash rents an H100 GPU for $1.48/hour versus $12.30 on AWS—but latency remains the Achilles' heel. Vadim Tashitsky, Head of Growth at StealthEX, explains that stitching together GPUs across continents over the public internet introduces milliseconds of delay, making decentralized orchestration competitive for batch jobs and fine-tuning but unsuitable for live chatbot inference at scale.
Cysic founder Leo Fan counters that latency is the wrong metric. “The hard problem isn't decentralized compute; it's discovery, scheduling, and attestation. The wedge isn't price-per-token; it's verifiability,” he says. By leveraging trusted execution environments (TEEs) and zero-knowledge (ZK) proofs, decentralized networks can compete in sectors where trust and verification outweigh tail latency. ZK proofs enable users to cryptographically verify AI inference results without relying on any single node, a feature centralized cloud providers struggle to offer.
Onchain Credit: Closing the $5-50M Funding Gap
Beyond compute, capital formation is a bottleneck. Traditional private credit often ignores smaller, non-standard deals of $5-50 million. Onchain credit platforms like Maple and Centrifuge can syndicate loans in this range, enabling retail investors to participate in data center revenues previously limited to institutional LPs. New “pay-per-inference” models emerge, where revenue fluctuates with GPU utilization, fitting tokenized revenue-share structures more naturally than rigid 20-year leases.
Despite potential, four gates remain closed for institutional adoption: legal enforcement in bankruptcy courts, lack of tamper-evident oracle infrastructure serving covenants, regulatory uncertainty for billion-dollar financing tranches, and non-standard tax/accounting products. Consensus suggests 12-24 months for medium-sized syndicated deals to gain onchain traction, with mostly onchain mezzanine debt 3-5 years away. First breakthroughs are expected from tier-2 operators rather than industry leaders like CoreWeave.
Meta Commits Up to $27B to Nebius in AI Infrastructure Frenzy
Meta has pledged up to $27 billion in cloud computing power with the Nebius Group, underscoring the accelerating AI arms race. This investment highlights both the insatiable demand for specialized infrastructure and the ongoing tension between centralized and decentralized provisioning.