What are ZK proofs and why do they matter for AI infrastructure?

Zero-knowledge (ZK) proofs are cryptographic tools that allow systems to verify information without revealing the underlying data. In decentralized AI infrastructure, experts believe ZK proofs could help prove that AI computations, workloads and transactions are trustworthy and tamper-resistant.

Why are experts linking ZK proofs to DePIN growth?

Industry executives argue that decentralised physical infrastructure networks (DePINs) cannot always compete with hyperscalers on speed, so verifiability may become their biggest advantage. ZK proofs and trusted execution systems could make decentralized AI networks more attractive for enterprise and financial applications where trust matters.

What did Goldman Sachs say about AI infrastructure spending?

Goldman Sachs projected a baseline scenario of $7.6 trillion in AI infrastructure spending over time. The bank said the final cost could vary significantly depending on chip lifespan, energy constraints and how efficiently older AI hardware is reused.

Why is AI chip lifespan becoming a major issue?

AI hardware evolves rapidly, raising concerns that expensive chips could become obsolete faster than traditional servers. Goldman Sachs said shorter hardware life cycles could sharply increase AI infrastructure costs if companies need to replace chips sooner than expected.

How do DePIN networks compare with Amazon Web Services for AI workloads?

Vadim Taszycki from StealthEX said decentralised networks can be much cheaper than Amazon Web Services for some workloads. However, centralised providers still dominate real-time AI applications because their GPUs operate in tightly connected data centres with lower latency.

Why is latency a challenge for decentralised AI networks?

DePIN systems often rely on GPUs distributed across multiple countries using the public internet. That setup introduces delays measured in milliseconds, making decentralised systems less suitable for applications like live AI chatbots that require near-instant responses.

Which AI workloads could still work well on decentralised infrastructure?

Experts said decentralised AI networks may perform well for batch processing, AI model training and fine-tuning tasks where ultra-fast response times are less important. These workloads can tolerate higher latency while still benefiting from lower infrastructure costs.

What did Cysic founder Leo Fan say about decentralised AI?

Cysic founder Leo Fan said the long-term opportunity is not simply cheaper compute but verifiable infrastructure. He argued that discovery, scheduling and attestation are more important technical challenges than raw processing speed.

ZK proofs seen key to DePIN AI growth

Written by Mahathir BayenaPublished May. 14 2026

Goldman Sachs said a baseline forecast for artificial intelligence infrastructure spending could reach $7.6 trillion, although the final cost will depend heavily on factors including how quickly AI chips become obsolete and whether older hardware can be reused for simpler inference tasks.

The report said shortages in power infrastructure, skilled labour and electrical equipment could also slow AI data centre expansion as companies race to build the computing capacity required for large-scale AI systems.

Vadim Taszycki, head of growth at StealthEX, said decentralised physical infrastructure networks can offer substantial cost savings compared with hyperscalers such as Amazon Web Services but remain disadvantaged in low-latency workloads because globally distributed GPUs cannot match the microsecond-level speeds of centralised data centres.

Taszycki said decentralised networks may still compete effectively in batch processing and AI fine-tuning tasks where speed is less critical, despite latency limitations that make them unsuitable for real-time chatbot applications requiring near-instant responses.

Leo Fan, founder of Cysic, argued that verifiability rather than raw performance will ultimately determine whether decentralised AI infrastructure gains long-term traction in enterprise and financial applications.

“The hard problem isn’t distributed compute but discovery, scheduling, and attestation. The wedge isn’t price-per-token; it’s verifiability,”

Fan said.

Industry participants also highlighted growing interest in onchain credit markets as a funding mechanism for AI infrastructure, with platforms such as Maple and Centrifuge potentially enabling smaller syndicated AI infrastructure loans that are often overlooked by large private credit firms.