CAL NEWPORT: Enterprise value from language models is proving slower and more bespoke than promised, which exposes the frontier labs to cheaper specialized competitors

The Gist

Newport says companies are trying AI tools, but getting real value is slower and harder than promised; Microsoft has had to rework Copilot. That opens the door for many cheaper, specialized tools that could undercut OpenAI and Anthropic, which are still losing billions. Some surveys do show returns growing, and OpenAI's revenue is rising fast. This packet steelmans the arguments made by Cal Newport and Ed Elson in their best light for analysis and is not an endorsement of their conclusions.

Conclusion

Language models will be widely used in enterprises, but through bespoke, domain-specific products that take longer to build than promised, and that slower, more specialized path favors many cheaper competitors over a few frontier labs whose business case depends on general-purpose models at premium prices.

Premises

1. Enterprises are spending heavily on language-model tools, but much of the spending is experimental and measured returns are uneven.
2. Microsoft's Copilot, a natural-language interface to office software, struggled enough that Microsoft merged its Copilot apps and retired several features in 2026 after an internal memo said the product needed to earn the right to exist.
3. Coding tools became the first major commercial success partly because their builders were themselves programmers, and other fields will need similar domain expertise and custom engineering.
4. Open-weight, cheaper, and on-device models combined with good agent frameworks can perform well enough for many specialized tasks.
5. Frontier labs are losing large sums, with Anthropic's 2025 operating loss about $8 billion, and price competition from open-weight and Chinese models limits their ability to raise prices.
6. Headline lab revenue figures are not consistently defined: OpenAI's widely reported run rate of nearly $70 billion was an investor estimate, while OpenAI's own net figure at the end of September 2026 was about $50 billion.

Assumptions

• Enterprise adoption is real and rising; the claim concerns the pace and form of value capture, not whether enterprises use these tools.
• Microsoft did not shut Copilot down; it merged its Copilot apps and retired several features in 2026 while continuing the product, which shows difficulty rather than abandonment.
• The widely reported near-$70 billion OpenAI run rate was an investor estimate on a gross basis; OpenAI's own net figure at the end of September 2026 was about $50 billion, with 77 percent run-rate growth in the third quarter.

Analysis

Argument type: Inductive.

Potential Fallacies

• Non sequitur (missing bridge premise) (From P1–P5 to the competitive claim in the conclusion) — Slow, bespoke adoption (P1 to P3) and lab losses (P5) do not by themselves show that value flows to cheaper rivals. The argument needs the further premise that specialized builders will choose cheaper models over frontier ones, and that labs cannot capture the specialized layer as suppliers or vertical builders.
• Hasty generalization (P2 and P3) — One product (Copilot) and one domain (coding) are used to characterize the whole enterprise market. Copilot's difficulties may reflect Microsoft-specific bundling, pricing, or integration choices rather than limits of general-purpose models. Coding may have succeeded because it is text-based, verifiable, and data-rich.
• Unsupported causal attribution (P3) — Coding's early success is attributed to builders being programmers, without ruling out rival explanations such as automatic verifiability, abundant training data, and tight feedback loops.
• Capability-to-market-outcome leap (P4 to the conclusion) — That cheaper models can perform 'well enough' on many tasks (P4) is a statement about capability. The conclusion needs a statement about market share and pricing. Procurement, security, compliance, switching costs, and the moving frontier all sit between the two.
• Analogy imported without checking its conditions (Conclusion (Christensen scenario)) — The Christensen disruption template assumes incumbents overshoot customer needs and cannot respond downmarket. Frontier labs already sell smaller, cheaper tiers, distill their own models, and move into vertical products, so the conditions are not shown to hold.
• Selective emphasis / losses read as weakness (P5 and P6/A3) — A large operating loss is presented as exposure, but growth-stage platform businesses often lose heavily before reaching profitability. The argument's own 77% run-rate growth figure points the other way and is not weighed against the thesis.
• Vague, hard-to-falsify terms and an unfixed baseline (Title, conclusion, and P4) — 'Slower than promised', 'bespoke', 'well enough', 'many', and 'favors' have no stated baseline, timeframe, or threshold. Almost any outcome could be read as consistent with the thesis, and the comparison to 'promised' is partly an artifact of hype rather than the historical norm for enterprise technology.

Counterarguments

These challenges are not ranked.

• Conclusion (High impact) — Bespoke enterprise products are typically built on top of frontier models through APIs and fine-tuning, so specialization makes the labs suppliers to many builders rather than victims of them. A many-products outcome can increase demand for frontier models. The thesis holds only if builders switch to cheaper models in large numbers.
• Premise 3 (High impact) — Coding, the lead success case, is led by tools from the frontier labs themselves (such as Claude Code and Codex). This suggests labs can capture vertical value directly, by hiring domain experts, acquiring builders, or deploying engineers alongside customers.
• Premise 4 (High impact) — Frontier capability keeps moving, so the 'good enough' bar rises. Open-weight models may stay a generation behind on the hardest, highest-value tasks, and enterprises often pay premiums for reliability, liability coverage, and security. Cheaper models also partly depend on distillation from frontier research.
• Premise 5 (Medium impact) — Large losses are normal for capital-intensive platform businesses (cloud, ride-hailing, e-commerce) and reflect training and infrastructure investment. With 77% quarterly run-rate growth and deep funding access, losses need not become a binding constraint. Price cuts can also expand total revenue.
• Premise 2 (Medium impact) — Copilot's restructuring may reflect Microsoft-specific issues (bundling, pricing, permissions and data hygiene, bolting an assistant onto legacy apps) and normal product iteration. Strong seat and API growth for other enterprise offerings suggests the difficulty is not category-wide.
• Premise 1 (Medium impact) — Uneven early returns and experimental spending are the historical pattern for ERP, cloud, and electrification, and those waves ultimately rewarded platform leaders. Slow enterprise ROI therefore does not predict competitive displacement of the leaders.
• Conclusion (Medium impact) — Future agentic improvements could let general-purpose models handle domain work with little customization. That would remove the need for heavy bespoke engineering and reverse the thesis in favor of the labs.

Suggested Improvements

• Bridge the descriptive and competitive claims — State and defend the premise that bespoke builders will predominantly use cheaper or open-weight models rather than frontier APIs, and that frontier labs cannot capture the specialized layer through partnerships, acquisitions, or vertical products. This is the load-bearing hidden step between 'slow and bespoke' and 'bad for frontier labs'. Without it the headline conclusion is a possibility, not a forecast.
• Quantify P4 — Provide task-level benchmarks and cost-per-task comparisons showing where open-weight or on-device models match frontier models, and what share of enterprise spend those tasks represent. Disruption depends on adequacy in revenue-weighted tasks, not the long tail. Vague terms like 'well enough' and 'many' cannot carry the conclusion.
• Broaden the evidence base — Add named, methodologically described surveys with time series, evidence of enterprise customers switching from frontier to open-weight models, and market-share and pricing trends, instead of relying on Copilot and coding anecdotes. This would replace anecdote with rates and separate product-specific failures from category-wide ones.
• Reckon with the counter-evidence — Integrate the 77% run-rate growth, the lab-built coding tools, and the cases of frontier-model success in enterprises, explaining why these do not undermine the thesis. Evidence that cuts against the thesis is mentioned but not weighed, which gives the argument a one-sided appearance.
• Fix baselines and make claims falsifiable — Specify who promised what, compare against historical enterprise-technology adoption curves, and state a timeframe and threshold (for example, a frontier-lab share of enterprise spend falling below a given level by a given year). Without a baseline, 'slower than promised' is hard to test, and almost any outcome can be absorbed by the thesis.
• Model frontier-lab responses — Address how labs could respond through cheaper tiers, distillation, bundling, vertical products, forward-deployed engineering, or acquisitions, and show why those responses would fail. Treating the labs as static makes the disruption story look more inevitable than it is.

Scenario Tests

• Frontier labs sell APIs to thousands of vertical startups and enterprise builders, and most bespoke products keep calling frontier models. (Challenges) — Slow, bespoke adoption would still be a real finding, but it would channel revenue to the labs as upstream suppliers. The competitive half of the conclusion would fail.
• Open-weight and on-device models match frontier models on most high-volume enterprise workloads, and enterprises migrate to them to save cost with acceptable security and compliance. (Supports) — Price pressure would erode frontier margins, as the thesis predicts. The labs would then depend on retaining a capability premium in a shrinking set of high-end tasks.
• Agentic capabilities improve sharply, so general-purpose models handle domain tasks with little customization. (Challenges) — The 'bespoke and slow' premise would weaken, and general-purpose frontier models would gain share. The thesis would reverse.
• Lab revenue keeps growing at rates near the reported 77% quarterly run-rate growth while enterprise ROI surveys remain mixed. (Challenges) — Demand would be outpacing the 'slow adoption' narrative, and losses would look like investment rather than structural weakness. Even so, mixed enterprise ROI could still support the 'bespoke' claim.
• Capital markets reprice lab valuations, funding tightens during a prolonged slow-adoption period, and labs are forced to cut prices or consolidate. (Supports) — Losses would become a binding constraint, giving the 'exposed' claim real force. This path depends on financing conditions that the premises do not establish.
• Microsoft's Copilot recovers after redesign, with strong seat growth and retention. (Challenges) — This would show the 2026 consolidation was normal product iteration and weaken the use of Copilot as a proxy for enterprise value.

Coherence & Relevance

The argument is coherent in its first half. Wide but uneven enterprise use, a slower path, and bespoke engineering fit together and are scoped carefully by the stated assumptions. The second half is much less secure. The step from 'slow and bespoke' to 'favors many cheaper competitors over a few frontier labs' relies on an unstated bridge, an unexamined analogy to disruption theory, and a single vague premise about cheaper models being good enough. The thesis leaves out the layered-ecosystem possibility, in which frontier models are the inputs to bespoke products, and the argument's own data on coding and revenue growth partly support that alternative. The conclusion is plausible and worth taking seriously but not established, and it is better read as one scenario among several than as a forecast.

• Enterprises are spending heavily on language-model tools, but much of the spending is experimental and measured returns are uneven. (Strong) — Directly supports the 'slower than promised' element. It is silent on who captures the eventual value, and the pattern is equally consistent with an ordinary adoption J-curve.
• Microsoft's Copilot, a natural-language interface to office software, struggled enough that Microsoft merged its Copilot apps and retired several features in 2026 after an internal memo said the product needed to earn the right to exist. (Moderate) — Illustrates the difficulty of general-purpose interfaces, but one bundled product does not stand for the category. It also concerns the application layer more than the labs' competitive position.
• Coding tools became the first major commercial success partly because their builders were themselves programmers, and other fields will need similar domain expertise and custom engineering. (Moderate) — Supports the 'bespoke' mechanism, but the causal story is untested against alternatives. The labs' own leadership in coding tools points toward the opposite competitive conclusion.
• Open-weight, cheaper, and on-device models combined with good agent frameworks can perform well enough for many specialized tasks. (Strong) — The key link to the competitive claim. It is vague, unquantified, and says nothing about whether builders actually choose cheaper models or whether the frontier keeps a premium on high-value tasks.
• Frontier labs are losing large sums, with Anthropic's 2025 operating loss about $8 billion, and price competition from open-weight and Chinese models limits their ability to raise prices. (Moderate) — Establishes financial strain and pricing pressure. It does not show that strain translates into competitive loss, since losses are compatible with future dominance. Pricing pressure may also vary by segment.
• Headline lab revenue figures are not consistently defined: OpenAI's widely reported run rate of nearly $70 billion was an investor estimate, while OpenAI's own net figure at the end of September 2026 was about $50 billion. (Weak) — A data-quality caveat with no stated inferential path to the conclusion. It supports skepticism toward headline numbers, and the associated 77% growth figure cuts against the weakness narrative.

View this argument on LogicFirst.ai