How far can we push computing?

Six major hyperscalers are on track to spend more than $700 billion this year on computing infrastructure, about twice their 2025 record, while data-center expansion strains the electrical grid. The article argues that transistor scaling and related performance trends are slowing, making efficiency gains and approaches such as reversible logic, optical interconnects, and DNA storage increasingly relevant. It describes large theoretical headroom to fundamental computing limits but emphasizes that practical implementation remains uncertain.

Analysis

The investable implication is not that compute demand is ending, but that returns may migrate from transistor shrinkage toward performance per watt at the system level. Power and interconnect constraints raise the strategic value of optical links, networking, packaging, and power delivery; if efficiency gains lower the cost of a unit of compute, they may also stimulate enough usage to sustain infrastructure demand (a Jevons-style rebound), rather than simply reduce capex. That makes an abrupt “AI buildout is over” inference premature.

Broadcom (AVGO) and NVIDIA (NVDA) have exposure to optical interconnects, but the article does not establish revenue or earnings materiality. Treat this as a medium-term product-mix question, not an immediate earnings signal. Intel (INTC) faces a relative risk if investors continue to value system-level efficiency more than conventional CPU scaling, but the article alone does not establish a near-term competitive loss. Reversible logic, DNA storage, and fundamental-limit computing remain research options, not credible near-term earnings drivers for IBM or Texas Instruments.

The contrarian point: slowing historical scaling curves do not imply a hard performance ceiling. System scaling and specialized architectures can extend demand, while physical-limit concepts may take years to commercialize. The near-term constraint is practical—power, cooling, interconnect economics, and execution—not proximity to fundamental limits. Over 1–3 months, watch hyperscaler capex and vendor disclosures on optical attach rates and energy per bit; over 6–18 months, watch whether efficiency gains translate into better margins or merely higher system complexity. Falsify the efficiency thesis if optical adoption stalls, cost per bit fails to improve, or hyperscalers materially cut infrastructure plans despite persistent compute demand.

Key Decisions for Investors

• No trade on the article alone: it is a technology thesis, not a new company-specific earnings catalyst. Avoid treating it as evidence that hyperscaler capex is about to peak.
• Set a 1–3 month alert for AVGO and NVDA earnings commentary on optical interconnect adoption, customer qualification, and energy per bit. Upgrade the theme only if disclosures show measurable product traction or improved economics; the article provides no basis to size that impact today.
• For a conditional relative-value expression, consider long AVGO versus short INTC only after confirming that optical/networking growth is translating into stronger relative estimates and that Intel’s competitive position is weakening. Invalidate on relative estimate revisions reversing or evidence Intel is gaining share in relevant workloads.
• Track hyperscaler capex plans alongside data-center power availability and interconnect spending. A capex slowdown without improved efficiency economics would challenge the infrastructure thesis; lower cost per unit of compute with stable or rising total spend would support it.