AI at the Speed of Light

AI at the speed of light

This is not a hype tour. It is a piece about pace — and the gap that pace creates. Frontier AI capability is now compounding on a clock measured in months, while the institutions and enterprises meant to absorb it still move in years. That divergence, not the technology itself, is where both the risk and the missed value now concentrate.

The thesis is simple to state and uncomfortable to act on: for most leaders, the binding constraint is no longer access to capability — it is the capacity to absorb it. The models are arriving faster than your org chart, your controls, and your regulators can metabolize them.

The curves are real, and they compound

Start with the evidence, because the "speed of light" framing has to be earned, not asserted.

• Compute. Epoch AI's dataset of notable models shows frontier training compute growing roughly 4–5x per year — doubling about every 5–6 months since 2020. For comparison, Moore's Law doubled transistor density roughly every two years. We are watching something move four times faster than the curve that defined the last half-century of computing.
• Efficiency. Independent of raw hardware, Epoch estimates the compute needed to reach a fixed level of capability halves about every eight months. So the two effects stack: hardware scales up while the software needed to hit a given bar scales down, simultaneously.
• Cost. Capability-adjusted token cost has fallen roughly 10x per year. GPT-4-level performance dropped from about $30–60 per million tokens in early 2023 to cents on that dollar — and Epoch notes the steepest declines arrived only after January 2024. This is an acceleration, not a steady glide.

These are not one trend dressed three ways. They are compounding curves, and compounding is precisely what makes the human intuition fail.

What changed in 2025: a second axis

The structural shift of 2025 was not bigger models — it was a new place to spend compute. Letting a model "think" longer at inference time, via extended chain-of-thought, delivered step-changes that pure scaling had stopped producing.

The numbers are stark. Inference-time reasoning lifted AIME math-competition accuracy from roughly 15% to 71% on a single attempt — and to ~87% with majority voting. On Google's line, Gemini 2.5 pushed AIME 2025 from 17.5% to 88% and GPQA Diamond to 86% in a single generation, while SWE-bench Verified — real GitHub bug-fixing — reached roughly 75%.

The economic consequence matters more than the leaderboard. Cost is migrating from one-time training to per-query inference: OpenAI's 2024 inference spend reportedly hit ~$2.3B, dwarfing the training cost of GPT-4.5. The unit economics of any AI feature now hinge on inference governance, not training budgets.

Two clocks, diverging

The two clocks are not just moving at different speeds — they are diverging. The wider that gap grows, the more both exposure and forgone advantage accumulate inside it.

The stakes are concrete, not abstract

The upside is real and measured. A randomized study of 5,172 customer-support agents found a 14% average productivity gain, rising to 34–35% for the least-experienced workers. A BCG field experiment with 758 consultants found tasks completed 25% faster with higher quality. And AI is compressing science itself: DeepMind's GNoME predicted 2.2 million stable crystal structures, AlphaFold3 extended structure prediction to molecular interactions, and the underlying work earned the 2024 Nobel Prize in Chemistry.

The cost is equally real. Stanford's analysis of payroll records found a 13% relative decline in employment for early-career workers (ages 22–25) in the most AI-exposed occupations since late 2022 — with no comparable decline for workers 35 and over in the same roles. And the infrastructure is concentrating: the US holds roughly 75% of global AI supercomputing power, while Africa holds under 1% of data-center capacity despite 18% of the world's population.

The pattern is unavoidable: AI widens gaps as fast as it closes them. Distribution is a leadership choice, not an automatic outcome.

The counterintuitive part: value isn't being captured

On timelines: plan against scenarios, not a date

Be honest about what is genuinely contested. Anthropic's Dario Amodei argues "powerful AI" could arrive as early as 2026; DeepMind's Demis Hassabis holds roughly a 50% chance of AGI by 2030; aggregate forecasters on Metaculus cluster near 2028; and Yann LeCun, now raising over $1B for world-model research, argues today's LLMs are an "off-ramp" from human-level AI entirely.

Under the hood, pretraining scaling is slowing as high-quality data is exhausted — but reasoning and test-time compute keep the gains coming. Epoch AI concludes hardware can keep scaling to roughly 2030, with power and capital, not data, as the binding constraints (single training runs drawing 4–16 GW).

The spread itself is the signal: even insiders disagree by years. The correct posture is scenario planning over date-betting — and acting on the measurable trends (agent autonomy, inference cost, energy exposure) rather than the contested ones.

The leadership mandate

If the constraint is absorption, then the agenda follows directly:

• Front-load investment in workflow integration and back-office use cases, where ROI is demonstrably highest.
• Govern inference cost and maintain vendor optionality; treat energy and supply-chain exposure as core to any build strategy, not an afterthought.
• Build internal governance that assumes safeguards lag capability — pre-deployment evaluation, not post-incident response.
• Plan against a range of trajectories while acting decisively on the trends that are already empirically grounded.

The organizations that win the next three years will not be the ones with the best model. They will be the ones that closed the gap between capability pace and their own.