When Intelligence Becomes Cheap, Why Are We Still Training So Many Models?
Model Narratives, Venture Capital, and the Economics of Real-World AI
Read and discuss with AI · ChatGPT / Claude
Open this article in ChatGPT or Claude to discuss it. If the model cannot access the page, copy the prompt and paste the article.
Technical progress ≠ Product value ≠ Business revenue ≠ Capital returns.
01. The Model-First Startup Narrative
Over the past few years, a recognizable pattern has emerged in AI startups: raise money, purchase compute, collect data, train a model, climb a benchmark, release a polished demonstration, and use technical momentum to raise the next round. The pattern does not invalidate genuine model research. Many of today's advances have required exactly this kind of patient and expensive experimentation. But it also creates a troubling shortcut in how we talk about technology businesses: a higher benchmark score can begin to stand in for a better product, stronger demand, and even a more valuable company.
Imagine a startup that takes an open-weight foundation model, performs task-specific supervised fine-tuning or reinforcement learning, and reports a substantial improvement in a legal, financial, coding, or robotic benchmark. It has demonstrated a technical result. But what does that result mean for a customer who can already access an extensively trained frontier model, combine it with retrieval, tools, memory, and a suitable agent harness, and complete the same workflow?
I am not arguing that post-training is useless. It can improve task behavior, reliability, latency, cost, privacy, and deployability. Nor am I arguing that training new foundation models has become unnecessary. My concern is that we too often collapse three separate questions into one: Can training improve a model? Does that improvement create additional customer value? Can the company retain enough of that value to earn a long-term return on its investment? The answers need not be the same.
02. The Right Baseline Is a Product, Not a Base Model
Technical papers frequently compare a base model, supervised fine-tuning, reinforcement learning, and the proposed new method. This is useful for causal attribution: it tells us whether a training intervention helped under controlled conditions. It is not necessarily the comparison a buyer cares about.
In an actual procurement decision, the alternative to a specialized model may be a frontier model with in-context examples, retrieval-augmented generation, access to trusted databases, an execution environment, an agent harness, and a verifier. Crucially, that frontier model has itself already undergone extensive general-purpose pre-training and post-training. Calling this a contest between “post-training” and “in-context learning” would therefore be conceptually misleading. The real comparison is additional domain-specific adaptation versus the best available inference-time and system-level adaptation, given already trained models.
Suppose the specialized system succeeds on 80 percent of tasks while a frontier model plus tools succeeds on 90 percent. The specialized system might still win economically if its inference cost is much lower or if it can run locally under strict privacy requirements. Conversely, a small benchmark lead might be worthless if its acquisition required costly training and it disappears with the next frontier release. Performance, total cost per successful outcome, latency, reliability, generalization, and deployment constraints all matter.
There are thus two distinct research questions. The scientific question holds the base model and environment fixed to measure the incremental benefit of updating weights. The commercial question compares the actual systems a customer can choose. We should not confuse the answers.
03. Finance, Law, and Robotics Are Different Adaptation Problems
The phrase “vertical AI” can conceal enormous differences between industries. Financial and legal workflows often depend on data access, document provenance, jurisdiction, authorization, and accountable decisions. A model that lacks yesterday's regulatory update does not necessarily need another training run; it may need reliable retrieval from the appropriate authoritative source. A system that produces plausible but uncited legal analysis may need better evidence checking, workflow constraints, and professional review before it needs additional weight adaptation.
Nevertheless, additional training may be justified when a model systematically fails to execute a specialized task or when a smaller, cheaper model can reliably handle high-volume work. Financial risk assessment, repetitive document extraction, and structured legal drafting can produce stable distributions that reward specialization. The relevant question is not whether the domain is large, but where the bottleneck actually lies.
Embodied intelligence adds a different challenge. Continuous actions, contact dynamics, feedback control, sensor timing, and physical safety impose constraints that a general-purpose conversational model does not automatically satisfy. A separately trained action policy, world model, or low-level controller may be essential. Yet this does not imply that every robotics company must pre-train an entire new multimodal foundation model. A frontier reasoning system coupled to reliable specialized skills could be the more economical architecture for many tasks.
The choice among retrieval and tools, task-specific post-training, continued pre-training, and training from scratch should follow the capability gap. A huge market alone cannot justify a huge training budget.
04. Training Is an Investment With a Half-Life
The economics of domain adaptation can be expressed in a deliberately simplified inequality. Assume a frontier-based solution and a specialized solution meet the same quality and safety requirements. Let N be the lifetime number of tasks, cF and cS their expected cost per task, CT the initial training and deployment cost, and CM the added maintenance cost. The specialized route is attractive when:
Here N is the lifetime task volume, c denotes the expected cost per task, and the right-hand side includes training and ongoing maintenance. This deliberately assumes equal quality. In practice one must also account for failure, retries, human review, legal data access, deployment, depreciation, and the cost of capital. Moreover, cF changes as frontier models improve and API prices fall. A training advantage that disappears before break-even is a poor investment, however impressive the static benchmark gain.
A larger market does not on its own justify pre-training from scratch. Before that decision, the company needs evidence of a fundamental capability gap, legal access to training data, transfer across tasks, and benefits that cannot be achieved more efficiently with continued pre-training, task-specific post-training, or specialized modules.
05. Anthropic: Market Validation Is Not the Same as Capital Return
Anthropic's November 2026 public-market debut offers a remarkable real-world test of the frontier-model business thesis. Its shares were priced at $83 and closed the first session at $97.17. Reuters reported a fully diluted closing valuation of roughly $1.77 trillion; the approximately $2 trillion headline referred to an intraday level, not the closing valuation.[1]
The company's IPO disclosures reported approximately $33.24 billion in revenue in the first nine months of 2026, while also describing extensive compute and infrastructure commitments.[2] These figures are evidence of meaningful demand. They are not proof that maintaining a frontier position will generate shareholder returns commensurate with the capital required.
It matters to distinguish inference contribution margin, operating profitability, free cash flow, and return on invested capital. Positive operating cash flow does not by itself account for the entire future infrastructure burden. Nor should long-term contractual commitments be mechanically treated as expenses already paid. A rigorous valuation requires the timing and nature of those obligations, the durability of revenue, reinvestment needs, and the cost of capital.
There is also a value-capture problem. If AI allows a law firm to save a million dollars, the underlying model vendor may receive only a small fraction of the savings. The customer, the law firm, application providers, or consumers may capture the rest. Artificial intelligence can create tremendous social value while individual suppliers or investors realize much less. Technological success and capital success are separate propositions.
06. Cursor and the Battle for the Work Interface
Cursor began from a different point on the stack: the developer's editor, codebase context, tools, and agent workflow. It could use progress from several upstream models rather than needing to own the best model at every moment. This creates an appealing strategy: gain distribution and identify real workflow bottlenecks, then choose which model capabilities are worth developing internally.
That strategy is not without risk. The upstream supplier can become the downstream competitor; its pricing or availability can change; users can migrate to a new interface. Cursor's acquisition by SpaceX in August 2026 and OpenAI's subsequent announcement concerning its model-supply arrangement illustrate how strategically consequential model access can become.[3][4] The specific dispute does not establish that every frontier lab intends to cut off independent applications.
Cursor's revenue scale or transaction valuation, meanwhile, does not reveal its customer-level profitability. A fixed subscription may be attractive for a light user and expensive to serve for someone running long autonomous coding sessions. True contribution margins depend on model routing, token usage, caching, inference discounts, support, and product costs.
The stronger lesson is not “applications always beat models.” It is that the order of investment matters. Starting from customers gives a company evidence about where its proprietary model investment can actually improve quality or cost. Starting from a model creates a different burden: finding a customer need that the model can serve better than the alternatives.
07. Why a Model Story Can Be Easier to Finance
An investor once put the question to me plainly: “If you already have orders and cash flow, and you do not need to burn money training a model, why do you need venture funding—unless you want to scale?” It is a revealing question because a company's economic attractiveness is not identical to its fit with a venture fund's return profile.
A highly profitable, capital-efficient services or vertical-software business may be a great business without requiring a huge round. A frontier-model company can more easily describe large and immediate capital needs: accelerator clusters, proprietary data pipelines, researchers, and another generation of training. For investors seeking very large outlier outcomes, this narrative can also appear to offer a larger platform opportunity.
There is nothing inherently irrational about that preference. Some genuinely novel technologies must be financed long before a market exists. At the same time, model training can become both a technical plan and a reason to raise more capital. The risk is a feedback loop in which the ability to attract capital is mistaken for proof that the chosen technical route creates a durable commercial moat.
Technical complexity, access to financing, and sustainable pricing power are different assets. A startup should raise capital when deploying that capital produces sufficient additional long-term enterprise value—not merely because expensive training makes a compelling presentation.
08. Robotics: A Factory Pilot Is Not an Economic Breakthrough
Embodied-AI startups face a particularly demanding commercial test because they enter markets already served by mature automation. Industrial arms, automated conveyors, mobile robots, and dedicated machines can weld, pick, transport, sort, and inspect with excellent throughput in well-structured settings. A humanoid or general-purpose mobile manipulator does not create incremental value merely by performing the same action in a promotional video.
There are several distinct milestones: a strategic partnership, a technical proof of concept, a paid order, delivery and acceptance, sustained use in production, and repeat purchases driven by measured return on investment. One should not be substituted for another. A genuine sale can fund an experiment rather than an economically superior production system; that does not make the sale fraudulent.
The most promising incremental market may be tasks where traditional automation is too costly to reconfigure: high-mix low-volume manufacturing, frequent line changes, irregular handling, and some household chores. These are settings where the ability to adapt quickly may be worth more than peak speed on a single fixed task. Even then, a wheeled manipulator, conventional arm with improved perception, or a redesigned fixture may beat a humanoid robot.
Three measurements should become standard: cost of adaptation to a new task, complete cost per successful operation, and evidence of repeat buying based on actual results. Without these, “robot deployment” remains an incomplete description of economic value.
09. Simulated Ability Is Not Yet Deployable Ability
The gap between apparent robot intelligence and usable production capability appears in research evaluations as well. NVIDIA's Astranomy study compared different ways to deploy a general-purpose frontier model in the simulated RoboLab-120 tasks. The report found 66.3 percent success for a direct policy configuration and 74.8 percent for an agentic one, with the latter using approximately 30.9 million tokens per episode on average at a reported estimated cost of $36.40.[5]
This result is significant, but not for the simplistic reason that a general model “defeated” all specialized robotics systems. The additional success came with substantial inference expenditure. The simulation also paused while the model reasoned, whereas a physical robot must live with real latency, contact, and timing constraints. Average cost for successful simulated episodes is not the full cost per successful industrial task.
Benchmark validity matters as well. If physical parameters fail to reach the simulator, joint limits are unrealistic, or a score awards success without the intended contact or manipulation, the reported capability may overstate the real one. Before drawing economic conclusions, researchers need to audit control interfaces, physics and scoring, log failures and human intervention, and test at least a subset of predictions on hardware.
The fundamental issue is not choosing between a VLA, a world model, an action policy, and an agent harness as labels. It is deciding which competencies should live in weights, which in specialized control, and which in inference-time interaction—and what that combination costs in a real environment.
10. What Remains Scarce When Intelligence Gets Cheaper?
As models grow more capable and affordable, the bottleneck may move toward three resources: execution, verification, and distribution. Execution means access and permission to act on real systems, budgets, machines, and workflows. Verification means the ability to establish that a result is correct, safe, and useful. Distribution means access to customers, trust, and the place where work actually begins.
None is an automatic moat. Tools can be copied; validators can be open-sourced; interfaces change. What matters is whether a company accumulates a system of capabilities and relationships that grows more useful as upstream intelligence improves.
I would therefore apply three tests to AI companies. Model substitution: how much value survives switching to a comparably capable model? Model improvement: does the company get stronger or become redundant when frontier models improve dramatically? Interface migration: if users stop working through today's editor or SaaS screen, what durable assets remain?
This is why I am more interested in the economic boundary of general-purpose robotic intelligence than in yet another isolated capability ranking. Can adaptability make tasks economical that previously were not worth automating? And can we measure that through engineering time, demonstration requirements, inference costs, failures, and real task output?
Conclusion: From Model-First to Value-First
My conclusion is not that model training should stop. It is that model training should be treated as one possible investment in solving a customer problem—not the business objective itself. In finance, law, coding, and robotics, the correct answer could be a more capable foundation model, more specialized weights, better tools, a verifier, cheaper inference, or a more effective work system.
The research question that follows is precise: When does additional domain-specific weight adaptation create durable value beyond an already post-trained frontier model? Answering it requires controlled comparisons and real cost data; an opinion essay cannot establish the result in advance.
Five years from now, when today's leading checkpoint is no longer leading, what will the company still own? A customer base, execution rights, trusted outcomes, adaptive systems, and distinctive feedback may outlive a checkpoint. That, to me, is a more durable way to think about where the value of intelligence ends up.
Sources and Evidence Boundaries
- Reuters, Anthropic market debut, Nov. 19, 2026 — closing and intraday valuation are different.
- Anthropic SEC prospectus amendment, Nov. 17, 2026.
- Cursor, Joining SpaceX, Aug. 2026.
- OpenAI, decision on Cursor following its acquisition.
- NVIDIA, Astranomy, 2026 — simulation performance, not factory ROI.
- PostTrainBench — automated post-training research, not a direct frontier-ICL economic comparison.
Analysis and hypotheses are distinguished from verified financial and benchmark disclosures. This is a personal industry perspective, not investment advice or a report of new controlled experiments.