The Scaling Bet

The scaling bet, stripped to its core, holds that more compute and more data will keep delivering significant improvements in AI capability; that the businesses making the bet will build defensible positions on those improvements; and that society will accept the businesses that result.

The financial stakes riding on the bet are now measured in trillions. The four largest publicly traded US hyperscalers — Amazon, Microsoft, Google, and Meta — have spent roughly $1.3 trillion on capital expenditure cumulatively from 2024 through 2026, with the 2026 figure alone reaching about $700 billion and roughly three-quarters of that 2026 spend going to AI infrastructure. Analysts at the St. Louis Fed have attributed up to 39 percent of marginal U.S. GDP growth over the last year to AI-related investment in software, specialized processing equipment, and data-center buildout. The country is short more than 530,000 construction workers in 2026, a shortage industry analysts attribute in part to industrial-scale construction demand, data centers among it. Three electricians under 30 at a single Plano, Texas data center were reported by Fortune in March 2026 to be earning between $240,000 and $280,000 a year.

The scaling bet and the economy’s growing reliance on it are not all that is staked. The vision the labs sell is, in their own words, utopian: a hundred years of biomedical progress compressed into a decade; “all disease” addressed within Hassabis’s lifetime; and automation of the most onerous knowledge work. Dario Amodei, in a February 2026 interview, has discussed the possibility of an AI-productivity-fueled global economic boom, potentially up to 15 percent GDP growth — territory not seen in modern economic history.

The same powerful AI carries risks the labs are sometimes quieter about. Large-scale displacement of work, beginning with coding tasks and customer support, and potentially moving into a wide variety of knowledge work over time. The concentration of wealth and decision-making power across a small number of trillion-dollar corporations and governments. The accrual of power by the AI itself — and the question of what it chooses to align that power to.

The named voices in this issue do not agree on whether the bet pays. They agree on something stranger.

Three roads

The named voices weigh in on the bet from three different directions.

The first road is to keep scaling. Dario Amodei, the chief executive at Anthropic, was a coauthor on the 2020 paper at OpenAI that established the scientific foundation the scaling bet rests on: how model performance scales with compute, data, and parameter count. Since then, he has argued that scaling will continue for at least another decade. Sam Altman, OpenAI’s chief executive, laid out the case for continued investment in Three Observations, his February 2025 essay. The intelligence of an AI model improves roughly with the log of the resources used to train and run it; the cost of any given level of capability falls by roughly a factor of ten each year. Demand grows on both axes at once — users buy more intelligence because each unit gets better, and again because each unit costs less. The case for continued capital deployment, in Altman’s framing, follows from arithmetic on the two curves rather than from any single technological forecast. Demis Hassabis, the chief executive at Google DeepMind, holds the same destination and may run one of the largest scaling programs outside OpenAI, alongside the largest world-models program inside Google DeepMind — even as Anthropic competes with Google DeepMind for his own company’s compute. At Google I/O 2026 Hassabis described the industry as standing in “the foothills of the singularity,” with artificial general intelligence reachable “as soon as 2030.” Noam Brown, the research scientist at OpenAI whose work on game-playing AI produced the reinforcement-learning lineage behind a new category of reasoning models — OpenAI’s o-series among them — has argued since September 2024 that scaling has a second axis: the compute a model spends thinking at the moment it answers, separate from the compute that trained its weights. Post-training rewards the model for reasoning step by step, so that when it answers it goes through smaller steps — a process that resembles reasoning. It is, once again, more compute buying more capability — only now spent when the model answers, not when it trains. The four voices on this road operate inside the same paradigm. They differ on where the scaling happens — training compute, post-training scaffolds, inference compute — and agree that scaling something within the paradigm continues to deliver.

The second road is a different way to train. Richard Sutton, the reinforcement-learning pioneer at the University of Alberta and Keen Technologies, is the road’s senior voice. His 2019 essay The Bitter Lesson, frequently cited as the doctrine behind the first road, was about reinforcement learning: systems that learn from experience by trying actions and observing the consequences, rather than systems trained on a corpus assembled in advance. Speaking to Dwarkesh Patel in September 2025, Sutton said the large language models that dominate the field today “are learning from training data; they are not learning from experience.” His position is not against scaling. It is against scaling supervised pretraining on internet text when the lesson he wrote about specifically was about scaling something else. Ilya Sutskever, who co-founded OpenAI and now runs Safe Superintelligence, makes a narrower version of the recipe argument: the pretraining-on-internet recipe specifically is exhausted because the public internet is finite, and the curve continues only on a different recipe. Sutskever does not publicly name the recipe. He says it exists. Gary Marcus, the cognitive scientist and longtime skeptic of the current paradigm, argues that current systems fail categorically on problems they have not seen during training, and that scaling supervised pretraining will not fix the failure mode. His prescription is to enrich the training mix with the structured-reasoning operations that an earlier generation of AI was built around: symbolic logic, taxonomies, compositional rules, combined with the statistical pattern-matching of current models. Marcus has placed a series of public bets that market leaders will not achieve certain performance benchmarks by certain dates. The road is unified by a single claim. Change what gets trained, or how, before adding compute.

The third road is a different machine. Albert Gu, chief scientist at the audio-AI startup Cartesia and assistant professor at Carnegie Mellon, argues that the architectural primitive at the heart of current AI is the wrong target of the 2026 hyperscaler capital expenditure. The primitive is called attention: the mechanism by which a transformer-based language model compares each new token to every previous token in its context, with a memory cost that grows as the square of the input length. Gu’s alternative is the state-space model, an architecture that maintains a fixed-size compressed summary of what it has seen and updates the summary as it goes, instead of caching every previous token. Hybrid models that combine state-space layers with a smaller number of attention layers are shipping in 2026 at seven independent labs — NVIDIA, IBM, Tencent, the Technology Innovation Institute, AI21, Microsoft, and Mistral. The closed-source United States frontier flagships have not, as of May 2026, publicly confirmed any state-space component. Yann LeCun, who left Meta in November 2025 and founded Advanced Machine Intelligence Labs in January 2026, makes a deeper architectural argument: predicting the next token, the objective every large language model is trained against, is the wrong objective to be optimizing. AMI Labs closed a $1.03 billion seed round in March 2026, the largest in European history, at a $3.5 billion pre-money valuation, with strategic participation from NVIDIA, Samsung, Toyota Ventures, Temasek, and Jeff Bezos’s investment vehicle. LeCun’s proposed substrate, the Joint Embedding Predictive Architecture, learns to predict future states of the world in a learned representation, and has shipped competitive vision systems at roughly half the parameter count of comparable transformers. François Chollet, the creator of the Keras deep-learning library who left Google to co-found Ndea, has spent six years arguing that scaling pretraining produces what he calls crystallized intelligence — memorized expertise — but cannot produce fluid intelligence, the capacity to solve a problem from a small number of examples without prior exposure. The Abstraction and Reasoning Corpus (ARC) benchmark family he designed to measure that gap is unique in that AI models often score poorly while humans, even children, often score perfectly. On the AI Startup School stage in 2025 Chollet said: “Even after a 50,000× scale-up of pre-trained base models, performance on ARC stayed near zero. We can decisively conclude that fluid intelligence does not emerge from scaling up pre-training.”

The road is unified by the claim that what compute runs on must change.

What every road runs through

Every road to powerful AI runs through a data center.

Amodei wants more compute against the paradigm he helped measure. Sutton wants more compute applied to a different training methodology. Gu wants more compute against a different model structure. The road varies. The data center does not. Even Marcus, the field’s most articulate critic of current systems, has spent the past year arguing on Substack that the improvement of Anthropic’s Claude Code is real, and that the improvement is real because Anthropic is incorporating the neurosymbolic-style techniques he has been calling for. The field is no longer arguing about whether compute delivers. It is arguing about which kind of compute delivers, on what substrate, against what recipe, measured against what benchmark. We are all scalers now.

The deeper question is one of weighting, not category. Every lab on the roster does some of both: research that produces new ideas and scaling that deploys them at production volume. The disagreement is about the mix. Amodei’s bet is weighted heavily toward scaling what he helped measure, with research absorbed inside the deployment as needed. LeCun’s bet is weighted heavily toward research into a substrate the field does not yet have, with scaling deferred until that substrate exists. The other voices occupy positions in between, each one a different weighting of how much of the marginal AI dollar should go to scaling what is known and how much should go to inventing what is not.

The shared objective is AI that does a broad array of economically valuable work. The shared input is the big computers. The shared bet is that the big computers keep producing the powerful AI.

What hangs on the agreement

The agreement that compute is the binding input and powerful AI is the destination raises the macro stakes the disagreement does not. Three strands carry the bet’s weight. The first is whether the science continues, whether the algorithmic and recipe progress the named voices fight about delivers significant improvements in model capability. The second is whether the inputs of chips, power, data, capital, and talent remain available at the pace the science requires. The third is whether the current paradigm produces useful agents — systems that require fewer humans in the loop. It is one thing to make a human more productive. It is another to do the job without one.

How the science, inputs, and agent questions resolve drives the scenarios — mapped voice by voice in AI Futures by deslop.media.

Just add ideas

The scientific ability to convert compute into powerful AI will continue at pace — not because today’s recipe will keep working, but because the field shares so many ideas, so freely — and the flagships with the big computers keep absorbing them. That is this publication’s view on the public record through June 2026, and it is a claim about the field’s habits as much as about its science.

The pattern is visible in the dissent itself. LeCun may be known for world models; Google DeepMind absorbed world models into Genie 3, the interactive environment it shipped in 2025, drawing the idea inward from the very critics — LeCun, Sutton — who pressed it from outside the dominant paradigm. Brown’s inference-time-compute axis, argued from outside the pretraining orthodoxy, has shipped at three frontier labs in eighteen months: OpenAI’s o-series, Anthropic’s Extended Thinking, Google DeepMind’s Deep Think. Gu’s state-space hybrids are in production at seven labs. Even Marcus, ever the skeptic, saw in Claude Code an absorption of his preferred neurosymbolic logic. The field continues producing — and sharing — ideas. The closed flagships continue absorbing them. The absorption is expected to continue. And on the science, the scaling bet keeps delivering more powerful AI.

continue — pick a path

Choose a position 3 camps

Keep scaling n=4

More of the existing paradigm — more compute, more data, more capital, all aimed at what already works.

01 The bet at the end of the curve · Dario Amodei stay tuned

02 What $1.4 trillion is supposed to buy · Sam Altman stay tuned

03 Scale and scaffold · Demis Hassabis stay tuned

04 The second axis · Noam Brown stay tuned

Different training n=3

Same compute, different inputs and methodology — change what gets trained or how it gets trained.

01 The bull case quotes Sutton; Sutton no longer reads it that way · Richard Sutton stay tuned

02 The recipe is exhausted; the curve is not · Ilya Sutskever stay tuned

03 Pattern Matching, Not Reasoning · Gary Marcus stay tuned

Different machine n=3

Same compute, different model substrate — change what the compute actually runs on.

01 The architecture that capex forgot · Albert Gu stay tuned

02 The bet against the language model · Yann LeCun stay tuned

03 Scaling isn't enough · François Chollet stay tuned

Originally published in Issue 1 · The Scaling Bet