Andrew Ng, the AI pioneer who co-founded Coursera and led Google Brain, has been telling audiences that artificial general intelligence is not arriving anytime soon, placing it decades out rather than years. At the same time, he has argued that the real speculative excess in AI is not in applications or inference, but in the enormous capital being poured into training infrastructure. That distinction matters because it reframes where financial risk actually sits in the current AI boom, and it challenges the assumption that simply spending more on bigger training runs will keep producing proportional gains.
Scaling Laws and Their Diminishing Returns
The intellectual foundation for the current training buildout traces back to a 2020 paper by OpenAI researchers titled Scaling Laws for Neural Language Models, which established predictable power-law relationships between model performance and three variables: compute, data, and parameters. That research gave the industry a seemingly simple formula: spend more on training and models get better. It also gave investors a reason to treat GPU purchases and data-center construction as direct bets on capability gains. But power laws carry a catch. Each doubling of compute yields a smaller absolute improvement than the last, meaning the cost of the next increment of performance grows faster than the benefit it delivers.
DeepMind’s 2022 Chinchilla paper sharpened that picture. The researchers showed that many of the largest models at the time were undertrained relative to their size, meaning labs had spent heavily on parameters while skimping on training tokens. The compute-optimal tradeoff the paper identified implies that a significant share of prior training budgets was misallocated. If labs were already misspending at smaller scales, the risk of waste only grows as budgets climb into the billions. Ng’s warning about a training bubble echoes this logic: the industry is scaling an input whose marginal returns are falling, and the financial commitments are locking in before the economics have been proven out.
Big Tech’s Infrastructure Binge
The scale of capital flowing into AI training hardware is visible in public filings. Nvidia’s most recent Form 10-K discusses demand drivers tied to both AI training and inference workloads, with data-center GPU sales forming the core growth engine. The filing also lists risk factors that hint at concentration: when a handful of hyperscale customers account for the bulk of orders, any pullback in their training ambitions ripples through the entire supply chain. That dependency is exactly the kind of structural fragility Ng’s bubble framing highlights.
On the buyer side, Amazon’s annual report filed with the SEC contains capital expenditure and AI infrastructure disclosures showing the company’s aggressive expansion of AWS capacity. Microsoft’s 10-K filings reflect a similar posture, with data-center buildouts and AI partnership commitments driving spending upward. Alphabet’s own annual disclosures confirm that Google Cloud and internal AI research are absorbing a growing share of capital budgets. All four companies are betting that training demand will justify these outlays. If Ng is right that AGI remains distant, the payoff window for these investments stretches far longer than current valuations seem to assume.
Why AGI Timelines Shape Financial Risk
The gap between Ng’s decades-long AGI estimate and the more aggressive timelines popular in Silicon Valley is not just an academic disagreement. It carries direct financial consequences. If AGI, or something close to it, were arriving within a few years, then massive training infrastructure spending would look like a rational land grab. Companies that locked in GPU capacity and data-center leases early would hold an advantage. But if the timeline is measured in decades, those same commitments start to look like overbuilt capacity chasing a capability ceiling that keeps receding. The foundational scaling research hosted at institutions like Cornell University through its arXiv preprint server provides the technical basis for this skepticism: power-law gains do not accelerate, they flatten.
A separate analysis from the Council on Foreign Relations projected that if current capability trajectories hold, 2026 could see AI systems capable of autonomously executing week-long human projects. That is an impressive benchmark, but it describes narrow task automation, not general intelligence. The distinction matters because investors and executives often conflate incremental automation gains with progress toward AGI, inflating expectations for what training spend will deliver. Ng’s argument is that the industry should separate these two tracks: practical AI applications that generate near-term revenue, and speculative AGI research that may not pay off for a generation.
The Coming Shift From Training to Inference
If training budgets do hit diminishing returns before AGI materializes, the logical reallocation is toward inference, the process of running trained models at scale for end users. Inference workloads have different hardware requirements, favoring efficiency and throughput over the raw floating-point power that training demands. Nvidia’s own 10-K acknowledges both training and inference as demand drivers, but the current revenue mix is heavily weighted toward training-class GPUs. A mid-decade correction (where labs slow their training expansions and redirect spending toward deploying existing models more cheaply) would reshape the semiconductor supply chain and potentially compress margins for companies that tooled up for a training-first world.
This reallocation would not necessarily destroy value. Many of the same chips used for training can be repurposed for high-end inference, and cloud providers can amortize sunk costs by offering more competitively priced AI services. But it would change who captures that value. Instead of chip vendors and construction firms at the top of the cycle, the winners would be software companies and platforms that can translate existing model capabilities into reliable products. For Ng, whose focus has long been on applied machine learning and education, this is where the durable opportunity lies: not in chasing ever-larger models, but in systematically deploying today’s systems across industries that still run on spreadsheets and manual workflows.
More from Morning Overview
*This article was researched with the help of AI, with human editors creating the final content.
