Goldman Sachs’ baseline forecast of $7.6 trillion in capital spending on AI ultimately depends on how long AI silicon remains useful. Decentralized networks promise great cost efficiency but continue to struggle with latency issues, and experts say their long-term viability will hinge on prioritizing verifiability over raw performance.
-
Key takeaways:
- Goldman Sachs points to spending $7.6 trillion by 2031, depending on whether the chips last more than 3 years.
- StealthEX and Cysic experts warn that DePIN’s response time limits decentralized AI in assembling tasks via live chat.
- Onchain companies like Maple may bridge the $5 million to $50 million credit gap for Tier 2 data centers by 2028.
Baseline: $7.6 trillion
A recent report from Goldman Sachs shifts the discussion from whether the demand for AI exists to the supply-side factors that will determine the actual cost of construction. The report forecasts $7.6 trillion in capital spending on AI as a baseline, but stresses that this number is highly sensitive to “fluctuating variables,” including the useful life of AI silicon.
Longevity is seen as the most important factor because rapid innovation can make standard chips – which typically last four to six years – obsolete in three years, driving up costs. Conversely, a “scaled model” where older chips are reused for simpler tasks, such as inference, could stabilize costs.
Data center complexity and elasticity of computing demand are other variables that will likely impact the amount of capital spent on AI infrastructure in the next five years. Shortages in power grid capacity, specialized labor, and electrical equipment are also seen as factors prolonging construction.
On the other hand, a separate report argues that this staggering infrastructure spending forms the cornerstone of the emerging “machine economy.” In this model, AI agents become the primary economic actors, executing high-frequency transactions and autonomously managing resource allocation. The report’s authors contend that legacy financial systems, characterized by slow settlement cycles and stringent know-your-customer (KYC) frameworks, are fundamentally ill-equipped for the speed of proxy commerce.
Decentralized infrastructure and the latency trade-off
Thus, it positions cryptocurrency protocols and decentralization as the essential, permissionless “economic rails” required to facilitate this transformation. However, skeptics remain cautious, questioning whether decentralized physical infrastructure networks (DePINs) can truly mitigate the bloated capital requirements of AI.
Vadim Taziki, Head of Growth at StealthEX, points out that while decentralized networks can provide significant cost savings, they face physical limitations. While a decentralized provider like Akash might rent an H100 GPU for $1.48 per hour compared to $12.30 on Amazon Web Services, the trade-off is speed.
“Big cloud providers can do this [fast work] “Because their GPUs are located next to each other in one building, connected by special cables that transmit data in milliseconds,” Tasecki explained, decentralized networks, which bring GPUs together across different countries over the public Internet, add milliseconds of delay. Latency makes this decentralized format competitive for batch functionality and fine-tuning but is unsuitable for serving large-scale live chatbots where the user experience relies on near-instantaneous responses.
Liu Fan, founder of Cysic, echoed these sentiments, insisting that decentralized inference is not suitable for low-latency workloads. However, Fan argued that latency is the wrong standard for comparing decentralized platforms and hyperscalers like AWS.
“The hard problem is not distributed computing, but discovery, scheduling and attestation,” Fan said. “The wedge is not the price per token, but verifiability.” He noted that Trusted Execution Environments (TEEs) and Zero-Knowledge (ZK) certificates allow decentralized networks to compete in sectors where trust and verification are more important than “response time.”
Onchain credit and financing gap
Beyond computing, the focus is shifting to how to finance these capital-intensive projects. While traditional private credit has ample capital, it often overlooks smaller or non-standard deals. Onchain credit offers distinct benefits, such as allowing retail investors to share in data center revenues that were previously limited to institutional limited partners. Furthermore, platforms like Maple and Centrifuge can raise loans worth between $5 million and $50 million, a segment that companies like Apollo often ignore due to higher underwriting costs compared to fees.
Finally, onchain credit enables new “pay-per-inference” models, where revenue fluctuates with GPU usage. Such models fit more naturally into nominal revenue-sharing structures than traditional strict 20-year leases.
Despite this potential, experts identify four “gates” that remain closed to institutional adoption: legal enforceability in bankruptcy courts, lack of tamper evident infrastructure for servicing covenants, regulatory uncertainty for multibillion-dollar tranches, and non-standardized tax and accounting products.
Consensus suggests a realistic timeline of 12 to 24 months for mid-sized syndicated deals to gain traction on-chain, with majority mezzanine debt likely to be on-chain three to five years away. The first breakthroughs are more likely to come from tier 2 players rather than industry leaders like Coreweave.
