AI's Unprecedented Acceleration: Reading BOND Capital's 2025 AI Report

Mary Meeker’s BOND Capital released its “Trends — Artificial Intelligence” report in May 2025. Several technology cycles deserve careful reading on this site; this is one of them. The compounding scale of AI adoption, capital deployment, and infrastructure build-out is unfolding at a pace that compresses several decades of typical technology-curve dynamics into eighteen months. The cycle deserves a measured analytical treatment, not a hot take.

This is a reading of what the BOND report documents and what we think it means for capital allocation.

The order-of-magnitude shifts

A handful of numbers in the report stand out as structural signals rather than headline figures.

800 million weekly active users on a single LLM (ChatGPT), reached in seventeen months — an 8× growth on prior cycle benchmarks.
$212 billion in projected 2024 capital expenditure by the “Big Six” US tech companies, up 63% year-on-year, largely driven by AI infrastructure.
+448% growth in US AI-related job postings over seven years, against a −9% decline in non-AI IT roles. The labour-market reallocation has already happened.
99.7% drop in cost per million tokens for leading inference workloads over two years, alongside a >100,000× decline in energy per token for NVIDIA GPUs over the past decade. Inference is becoming infrastructure-cheap; training is becoming infrastructure-expensive.
25% of global data-center capital expenditure now flows through NVIDIA, a share that continues to rise.
+1,150% surge in large-scale multimodal AI model releases over two years, with China increasingly leading in open-source contributions.
122 days to construct xAI’s “Colossus” data center — against an average of 234 days for a single American home. The build-out has changed regime.

The dispersion across these numbers matters. We are not looking at one technology curve. We are looking at four or five technology curves stacked on top of each other, each compounding the others.

Where the inflections sit

A few themes from the report deserve to be carried into investment frameworks.

Pace that exceeds prior cycles. AI adoption, model release cadence, and infrastructure build-out are running faster than the equivalent stages of the internet or mobile transitions. Pattern-matching from earlier cycles, while useful, systematically underestimates the slope of the curve. The implication: assumptions about adoption lag built into traditional venture diligence frameworks need recalibration.

Compute economics bifurcate. Training costs continue to climb — frontier model training now requires hundreds of millions in compute, and rising. Inference costs are collapsing — by 99.7% over two years, with further declines structurally probable. The two curves push capital toward different categories: a small number of well-financed labs at the training frontier; a far larger universe of applications built on cheap inference.

The physical layer joins the digital one. Robotics, autonomous vehicles, scientific discovery (drug development, materials science), and defence systems — AI is no longer confined to consumer software interfaces. The investment surfaces that follow are physical: sensors, actuators, embedded compute, energy systems, supply chains.

The competitive landscape is bipolar at infrastructure, multipolar at application. US and Chinese efforts dominate compute, foundation models, and large-scale industrial deployment. The application surface — every higher-layer use of these primitives — is global and pluralistic. The geographic distribution of value creation will not be the same as the geographic distribution of value capture.

Work is being reorganised, not eliminated. Productivity gains documented in the BOND report are real but uneven. The strongest signal is that AI-fluent workers in nearly every category outperform their counterparts substantially, and the gap is widening. Capital should price this into operational diligence: portfolio companies that build effective AI workflows internally compound advantage; those that delay accumulate technical debt.

Reading

The compounding scale and the bifurcating economics are the two most important features of the cycle. Both reward investors with long enough horizons to sit through compute-capex absorption and patient enough conviction to underwrite category emergence on top of plummeting inference. Both penalise short-horizon trading of the headline numbers.

The most consequential capital allocations of the next five years will not be the ones reacting to next week’s model release. They will be the ones reading the structural curves correctly and positioning quietly, ahead of the next layer of build-out.