What Will Be AI’s Biggest Bottleneck? Coinbase CEO Gives His Take

Coinbase CEO Brian Armstrong argues that energy and compute infrastructure, not model quality, will define the upper limits of artificial intelligence growth.

Armstrong made the observation in reply to a post by investor Tommy Shaughnessy, who outlined how metered API pricing is pushing enterprise AI spend well beyond what flat-rate subscriptions led companies to expect.

Demand for Intelligence Is Near Infinite

The Coinbase CEO’s core argument is that the appetite for AI-generated intelligence has no practical ceiling.

However, he expects the market to divide sharply within 12 to 18 months. Around 80% of workloads, he predicts, will migrate to models priced up to 99% below current top-tier options.

Good takeMy guess is– demand for intelligence is near infinite– but 80% of workloads will be running on 99% cheaper models within 12-18 months– 20% of workloads will still run on latest gen models where IQ maxing is important (scientific breakthroughs, higher level… https://t.co/gAFtYjorRh

— Brian Armstrong (@brian_armstrong) June 8, 2026

The remaining 20%, covering use cases where peak performance matters, such as scientific research or high-level orchestrator agents, will continue running on frontier models.

Armstrong compared the split to consumer hardware, noting that most buyers skip maxed-out specs on MacBooks and gaming PCs, even as prices fall faster than Moore’s Law would predict.

He added that this compression will not resolve the scarcity problem. As model costs drop and cheap alternatives proliferate, the bottleneck simply shifts upstream. It moves to the power and silicon required to run any model at scale.

Projected U.S. data center power demand, Source: Statista, McKinsey, Gartner, IDC, Nvidia company filings. As of April 2025. 

Coinbase’s Routing Strategy

Coinbase is already applying this logic in practice. Armstrong said the exchange routes prompt to lower-cost models where appropriate, keeping AI spend roughly flat even as token usage grows exponentially.

His Coinbase AI-native restructuring earlier in 2026 signaled a broader shift toward efficient, agent-driven workflows. His stance against AI overregulation reflects confidence that the technology’s trajectory should not be constrained by policy.

That approach speaks directly to the pressure Shaughnessy described. He cited Uber exhausting its full 2026 AI budget by April as one example of how fast enterprise AI cost overruns can accelerate.

Shaughnessy also noted that open-source models such as DeepSeek V4 perform within the range of top proprietary systems at roughly one-thirtieth the cost, placing a ceiling on what frontier labs can charge.

Energy as the Binding Constraint

Armstrong’s conclusion is that model quality will converge while cheaper alternatives close the performance gap. The real limit, he says, will be the physical infrastructure powering every tier of AI deployment.

That view aligns with capital flows already visible in the market. AI venture funding in Q1 2026 reached $242 billion globally, yet data center capacity is already stalling against demand.

Armstrong’s point is not which model prevails, but whether energy and computing infrastructure can keep pace with demand that, by his own assessment, has no natural ceiling.