Goldman Sachs predicts $7.5 trillion in AI infrastructure spending over five years.
That number is not a forecast. It’s a headline designed to justify current market multiples.
Let’s run the audit.
// Hook
The number appears in a Crypto Briefing report. 7.5 trillion. Annualized: $1.5 trillion. Compare: the entire global semiconductor market is ~$600 billion per year. AI chips alone would need to absorb 50-60% of that spend — roughly $750 billion annually, or 1.25x the current chip market. That implies a 5x expansion of chip fabrication capacity in five years. Not impossible, but physically constrained. s heart.
// Context
Goldman’s thesis rests on two assumptions: Scaling Law holds — larger models yield proportional intelligence gains — and inference demand explodes as AI agents, autonomous driving, and enterprise automation hit mass adoption. The prediction made sense in the 2024 hype cycle. But 2026 is a bear market for crypto, and a correction cycle for tech. Investor survival matters more than growth narratives.
Crypto Briefing’s audience is blockchain-native. They want to know: does this $7.5 trillion wave lift AI tokens, or does it mask a different reality — one where centralized hyperscalers capture all value, leaving decentralized compute networks with scraps?
Based on my audit experience with DeFi interest rate models and AI-agent smart contract interfaces, I can tell you the answer is structural.
// Core: Systematic Teardown
1. The ROI Paradox
Assume $1.5 trillion invested annually. To earn a 10% return, the AI application layer must generate $150 billion per year in gross profit. Today, the entire AI SaaS market (OpenAI, Anthropic, Midjourney, etc.) is under $20 billion. Even with 5x growth, you’re short by $50 billion. The gap must be filled by non-software revenue — autonomous vehicles, robotics, military contracts. These sectors have longer deployment cycles. The Jevons paradox (efficiency increases total usage) helps, but not enough to close a 2.5x shortfall within five years.
2. The Power Wall
Each high-end AI chip (B200, 700W) requires roughly 6 MWh per year per chip. To deploy the implied 12.5 billion chips (at $3k each, 50% of budget), you need 75,000 TWh annually. The entire world generates ~30,000 TWh today. The prediction tacitly assumes 40% of global electricity goes to AI. That’s not an investment forecast; it’s a power grid fantasy. Even with 100% new nuclear and solar, build times exceed five years. The bottleneck is concrete and copper, not silicon.
3. Supply Chain Chokepoints
Advanced packaging (CoWoS) is already oversubscribed 18 months out. HBM memory production is likewise constrained. Expanding capacity requires new fabs, which take 3-5 years and consume hundreds of billions themselves. The $7.5 trillion number double-counts the cost of building the factories to build the chips. It’s a circular reference.
I saw this type of feedback loop failure during my Terra pre-mortem analysis. The seigniorage flow looked stable on paper but collapsed under high volatility. Here, the failure mode is geometric: constrained supply raises chip prices, which reduces effective units deployed, which lowers AI adoption, which kills ROI. The prediction assumes infinite elasticity. Markets don’t behave that way.
4. The AI Token Trap
Crypto Briefing’s framing serves a purpose: it funnels retail into AI-based altcoins (Render, Fetch, Akash, etc.). The logic: “AI needs decentralized compute, therefore AI tokens moon.” But the $7.5 trillion is almost entirely going to centralized hyperscalers — Azure, AWS, GCP. Their private networks and proprietary chips don’t feed public blockchain revenue. The narrative of “AI + blockchain synergy” is a manufactured narrative VCs use to push new products. I’ve seen it before in DeFi’s “liquidity fragmentation” myth. It’s the same playbook: identify a problem (compute centralization), propose a solution (tokenized GPU markets), and sell the solution before the problem is verified.
Real-world adoption of decentralized compute remains below 1% of total AI workload. Even if it captures 5% of $7.5 trillion, that’s $375 billion over five years — split among dozens of protocols. The unit economics don’t work unless a single network achieves monopoly. That’s not how permissionless systems function.
// Contrarian: What the Bulls Got Right
Bulls are correct that infrastructure spend will be massive. The $7.5 trillion is directionally plausible in a 10-year window, not 5. They are also right that inference will dominate — by 2028, inference likely accounts for 70% of AI compute. This favors edge devices and low-latency networks, which could benefit specialized hardware tokens (e.g., Helium’s wireless IoT, if adapted for AI).
Additionally, the prediction correctly identifies the importance of energy infrastructure. Nuclear and renewable projects will see sustained capital inflows. That’s positive for energy-backed tokens (e.g., Powerledger, or new tokenized nuclear funds). But the value accrues upstream, not to generic compute marketplaces.
// Takeaway
The $7.5 trillion number is not a forecast. It’s a psychological anchor. Anchors benefit those who set them — banks underwriters, hyperscaler shareholders, and influencers selling AI tokens.
For crypto natives, the critical question is: where does the value accumulate? Not in GPUs. Not in generic compute. In applications that solve real bottlenecks: energy distribution, supply chain financing, data labeling. Or in protocols that enforce security and sovereignty. Code is law, but only if the audit is real.
I’ll be watching the next NVIDIA earnings. If data center revenue growth drops below 150% YoY, the narrative breaks. Until then, trade narratives, not fundamentals. But know you’re trading.
s heart.