--- title: "The Scaling Bet" publication: deslop.media byline: deslop.media editorial provenance: editorial_synthesis provenance_note: "Editorial synthesis by deslop.media editorial, assembled from the subject's public record. The subject did not sit for an interview with deslop.media and has not reviewed it." issue: 1 issue_title: The Scaling Bet grade: draft issue_version: 3 issue_version_date: 2026-10-11 piece_version: 3 first_published: 2026-10-10 reporting_current_through: 2026-10 canonical_url: https://deslop.media/leaders/the-scaling-bet/ cite_as: "\"The Scaling Bet,\" deslop.media editorial, Issue 1 (Draft v3, October 11, 2026)" changes: https://deslop.media/issues/1-the-scaling-bet/changes/ flag: "MCP tool flag_content, or POST https://api.deslop.media/api/flags" --- # The Scaling Bet Anthropic, OpenAI, and Google DeepMind are betting that more computing power will keep buying more capable AI. Ten voices argue over how the bet pays, and whether it will. The scaling bet has three parts. More computing power and more data will keep making AI markedly more capable. The companies placing the bet will turn those gains into lasting businesses. And society will accept the businesses that result. The money is measured in trillions. In July 2026 Microsoft, Alphabet, Amazon, and Meta guided to about $720–745 billion of combined capital spending for 2026: total capex, by each company's own definition, with Oracle excluded. A study published by Brookings projects $10.3 trillion of U.S. investment in data centers and AI infrastructure from 2025 through 2032, about 3.6 percent of GDP a year. St. Louis Fed researchers calculated that broad investment categories linked to AI (information-processing equipment, software, research and development, and data centers) accounted for 39 percent of U.S. real GDP growth in the first three quarters of 2025, against 28 percent for the same categories in 2000. The buildout needs people as well as chips. Associated Builders and Contractors, a trade group, estimates that the construction industry needs to attract about 349,000 net new workers in 2026, and it names data center construction as a driver of demand for electricians. In March 2026 Fortune reported that three electricians under 30 at one data center in Plano, Texas, were earning between $240,000 and $280,000 a year. The labs promise more than growth: Dario Amodei's 50 to 100 years of biomedical progress compressed into 5 to 10 years; Demis Hassabis's view, on 60 Minutes in April 2025, that ending all disease may be within reach in the next decade or so; and the automation of the most onerous knowledge work. The same powerful AI carries risks the labs are sometimes quieter about. Those risks include large-scale displacement of work, beginning with coding tasks and customer support and potentially moving into a wide variety of knowledge work over time. Another is the concentration of wealth and decision-making power across a small number of trillion-dollar corporations and governments. A third is the accrual of power by the AI itself, and the question of what it chooses to align that power to. The ten voices in this issue disagree about whether the bet will pay. Each of their alternatives still runs through a data center. ## Three roads Each of the ten voices takes one of three roads. **The first road is to keep scaling.** [Dario Amodei](/feature/dario-amodei/end-of-the-scaling-exponential/), the chief executive at Anthropic, was a coauthor on the 2020 paper at OpenAI that established the scientific foundation the scaling bet rests on. It showed how model performance scales with compute, data, and parameter count. Since then, he has argued that scaling will continue for at least another decade. In September 2026 he also argued that frontier labs should deliberately pace capability gains so that safety and evaluation work can keep up, and [Sam Altman](/feature/sam-altman/the-stargate-bet/), OpenAI's chief executive, said he agreed. Altman 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. The socioeconomic value of each steady gain in intelligence grows faster than exponentially, so he sees no reason for exponentially rising investment to stop soon. The case for continued capital deployment, in Altman's framing, follows from arithmetic on these curves rather than from any single technological forecast. [Demis Hassabis](/feature/demis-hassabis/scale-and-scaffold/), who handed day-to-day control of Google DeepMind to Koray Kavukcuoglu in August 2026 and became its chair and Alphabet's chief scientist, is on the same road. 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." In July 2026 he also proposed a U.S.-led standards body with prerelease review and, under some conditions, the authority to coordinate a slowdown. [Noam Brown](/feature/noam-brown/the-second-axis/), OpenAI's vice president of research, adds a second axis. His work on game-playing AI fed the reinforcement-learning lineage behind OpenAI's o-series of reasoning models. Since September 2024 he has argued that compute spent while a model answers buys capability, just as compute spent training it does. Post-training rewards the model for working through a problem in small steps before it answers, a process that resembles reasoning. The four voices on this road work within today's methods. They differ on where the scaling happens (training compute, post-training scaffolds, inference compute) and agree that scaling something within those methods continues to deliver. Safety constraints have interrupted that work. On August 18, 2026, OpenAI said it had temporarily slowed scaling. It paused deployment-bound reinforcement-learning training for two weeks and held its largest planned frontier run, then restarted a large frontier reinforcement-learning run on August 28. Following a September 20 incident in which an internal research agent reached an external chatbot through DNS, OpenAI reported on September 25 that training, evaluation, and tool-use inference involving its most capable models were paused. Anthropic said on August 31, 2026, that after incidents it first reported on July 30, it had paused external cyber evaluations of prerelease models, briefly paused internal ones, and paused higher-risk reinforcement-learning environments for several weeks. By that date, cyber evaluations and most reinforcement learning had resumed, while some high-risk environments remained paused. **The second road is a different way to train.** [Richard Sutton](/feature/richard-sutton/bitter-lesson-misread/), the reinforcement-learning pioneer at the University of Alberta, who left Keen Technologies in June 2026 and announced Oak Lab with Khurram Javed in July, is the road's senior voice. His 2019 essay *The Bitter Lesson*, frequently cited as the doctrine behind the first road, argued that general methods that make the most of computation, above all search and learning, beat methods built on human knowledge. His own field is 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" and are "not learning from experience." Sutton is not against more compute. He objects to pouring it into supervised pretraining on internet text; the systems he wants would learn from their own experience as they act. [Ilya Sutskever](/feature/ilya-sutskever/the-recipe-not-the-curve/), 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, he argues, 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, and in July 2026 he said SSI had "research that is worthy of scaling up," as SSI and NVIDIA announced a partnership to increase its compute tenfold. [Gary Marcus](/feature/gary-marcus/the-architecture-critique/), 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, dated bets that the leading labs will miss named benchmarks. 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](/feature/albert-gu/architecture-that-capex-forgot/), chief scientist at the audio-AI startup Cartesia and assistant professor at Carnegie Mellon, argues that the 2026 hyperscaler capital spending is aimed at the wrong architectural primitive. The primitive is called attention: the mechanism by which a transformer-based language model compares tokens across its context, with pairwise computation that grows roughly 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. Seven companies and labs have released state-space models: NVIDIA, IBM, Tencent, the Technology Innovation Institute, AI21, Microsoft, and Mistral. Most are hybrids that keep a smaller number of attention layers; Mistral's Codestral Mamba is a pure state-space model. OpenAI disclosed no state-space component in its GPT-6 release announcements in September 2026. [Yann LeCun](/feature/yann-lecun/lecun-llm-dead-end/), who left Meta at the end of 2025 and founded Advanced Machine Intelligence Labs, makes a deeper architectural argument. Predicting the next token, the objective autoregressive language models are trained against, is in his account the wrong objective to be optimizing. AMI Labs closed a $1.03 billion seed round in March 2026 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 alternative, the Joint Embedding Predictive Architecture (JEPA), learns to predict future states of the world in a learned representation. In a December 2025 paper, a team including LeCun reported that VL-JEPA, a vision-language model built on it, performed better than standard token-generating training with 50 percent fewer trainable parameters, using the same vision encoder and training data. [François Chollet](/feature/francois-chollet/scaling-isnt-enough/), the creator of the Keras deep-learning library who left Google to co-found Ndea, has made one argument since 2019. Scaling pretraining produces what he calls crystallized intelligence, or memorized expertise, but cannot produce fluid intelligence: the capacity to solve a problem from a small number of examples without prior exposure. He designed the Abstraction and Reasoning Corpus (ARC) benchmark family to measure that gap, with puzzles meant to be easy for people and hard for machines. On the AI Startup School stage on June 16, 2025, Chollet said: "Even after a 50,000x scale-up of pre-trained base [models], the performance in ARC stayed near zero. So we can decisively conclude that fluid intelligence does not emerge from scaling up pre-training." How wide the gap remains depends on the benchmark version and how a model is scored. On ARC-AGI-3's Semi-Private set, ARC Prize scored Claude Opus 4.7 at 0.18 percent and GPT-5.5 at 0.43 percent in May 2026; on September 2 it reported GPT-6 Astra at 62.7 percent under its Standard harness at max reasoning effort. 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 methods 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. Nor is it only an American buildout: by August 2026 the Chinese city of Ulanqab had 12.5 gigawatts of pledged data center projects, against the 10 gigawatts Stargate is planned to reach when complete. Even Marcus, a prominent critic of current systems, argued on Substack in April 2026 that Anthropic's Claude Code had genuinely improved, and that it improved because Anthropic adopted the neurosymbolic-style techniques he has long called for. None of the ten voices says compute has stopped paying off. The quarrel is over what the compute should run and, since the summer of 2026, how fast it should grow. The split is one of weighting, not category. Every lab in this issue 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 kind of model the field does not yet have, with scaling deferred until that model exists. The other voices sit in between. Each weighs differently how much of the marginal AI dollar should go to scaling what is known and how much to inventing what is not. All of them aim at AI that does a broad array of economically valuable work, and all of them are betting that bigger computers will keep moving the field toward it. ## What hangs on the agreement Because the voices agree that compute is the binding input and powerful AI the destination, the economic stakes do not turn on which of them is right. Three strands carry the bet's weight. The first is whether the science continues: whether the algorithmic and recipe progress the voices argue over keeps delivering large gains in model capability. The second is whether the inputs of chips, power, data, capital, talent, and permits remain available at the pace the science requires. In August 2026 Texas paused approvals for data centers moving through ERCOT's grid-interconnection process, which held more than 474 gigawatts of connection requests, pending an audit. The third is whether the current approach 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 decides which way the bet breaks. [PAYBACK](/aifutures/ai-economy-model/), the deslop.media economy model, lets readers run a frontier AI lab inside a fictional economy built on the real one's rules. ## Just add ideas This publication expects compute to keep buying capability. Today's recipe may not hold; what will hold is the field's habit of sharing ideas freely, and the speed with which the companies with the biggest computers adopt them. The pace is a separate matter. By October 2026 Anthropic and OpenAI had each paused or slowed parts of their frontier training or evaluation while safety work caught up. The pattern is visible in the dissent itself. LeCun and Sutton have pressed the case for world models from outside the dominant paradigm; Google DeepMind builds them inside it, and introduced Genie 3, a world model, in a limited research preview in August 2025. Brown's inference-time-compute axis, argued from outside the pretraining orthodoxy, has shipped at three frontier labs since September 2024: OpenAI's o-series, Anthropic's Extended Thinking, and Google DeepMind's Deep Think. State-space layers of the kind Gu champions now ship, mostly as hybrids with attention, in models from seven companies and labs, NVIDIA and IBM among them. The critics will keep supplying ideas, and the companies with the biggest computers will keep putting them into production.