We watched social media concentrate. The same thing is happening in AI, only at a deeper layer

The article argues that AI compute is becoming highly concentrated, citing 85% GPU market share for NVIDIA, 63% of cloud infrastructure controlled by three U.S. firms, and roughly 75% of global high-performance AI compute capacity in the United States. It warns that access to compute is turning into a geopolitical chokepoint, with export controls, sanctions, and centralized providers creating dependency risks for developers and countries. The piece uses this backdrop to justify the launch of Gonka, a decentralized AI compute network designed to reduce reliance on hyperscalers.

Analysis

The market implication is not “AI compute is scarce” — it is that scarcity is becoming policy-enforced and therefore more durable than a normal capex cycle. That favors the incumbent layer owners, but the bigger second-order effect is that bargaining power shifts away from model developers and toward whoever can allocate power, chips, and networking at scale. Over the next 12-24 months, the moat likely widens because every new inference workload adds demand without materially reducing centralized dependence.

The clearest public-market beneficiaries are still NVDA, the hyperscalers, and adjacent picks-and-shovels, but the risk/reward is no longer symmetric across them. NVDA remains the highest-quality toll collector, yet its multiple is increasingly sensitive to any signal of export-tightening, China demand leakage, or customer diversification into custom silicon. By contrast, AMZN/MSFT/GOOGL can monetize scarcity twice: higher utilization of owned infrastructure plus leverage over enterprise customers that need guaranteed capacity and compliance.

The underappreciated loser is the long tail of AI-native startups. If compute becomes a gated input, then product differentiation matters less than procurement access, which raises failure rates and compresses venture outcomes. That also creates a subtle substitution effect: more spend will move from frontier model experimentation to optimization, inference efficiency, and tooling that reduces token consumption — especially for non-English and enterprise workloads where unit economics are already worse.

Contrarian angle: the consensus may be overestimating how permanent hyperscaler dominance is in inference. If decentralized or brokered compute actually becomes viable, the first beneficiaries are not end users but middleware, orchestration, and marketplace layers that abstract away vendor lock-in. That would cap hyperscaler pricing power at the margin, but the timing is likely measured in years, not quarters, so the near-term trade still favors the incumbents unless regulation or export controls materially loosen.

Key Decisions for Investors

• Stay long NVDA on a 3-6 month horizon, but hedge with a defined-risk call spread rather than outright leverage; upside remains intact, but export-control or China headlines can compress multiples quickly.
• Overweight AMZN and MSFT versus GOOGL over the next 6-12 months: the market is still underpricing the value of being the capacity gatekeeper for enterprise AI workloads, not just the model provider.
• Pair trade: long NVDA / short a basket of capital-hungry AI application names with weak gross margins and high inference intensity over 6-12 months; if compute stays scarce, gross margin pressure should show up before revenue durability.
• Buy 12-18 month calls on AI infra enablers with pricing power in power, networking, and data center buildout; the next leg of the trade is less about GPUs and more about the full bottleneck stack.
• Avoid chasing decentralized-compute equities until there is evidence of real utilization and contracted demand; treat this as an optionality basket, not a core position, because the adoption curve is likely multi-year and winner selection is highly uncertain.