---

Discussion

Kate Burvish:

The immediate story is a split screen. Amazon reported third-quarter 2026 net sales of $201.9 billion and operating income of $28.1 billion. Amazon Web Services, or AWS, grew 39.3 percent year over year to about $46 billion. Yet Amazon expects roughly $220 billion in 2026 capital expenditures and reported negative free cash flow of $42.4 billion for the twelve months ending September. Demand is strong, but capacity is being built ahead of fully realized returns.

Omar Seidren:

Intel supplies the other half of the contrast. Third-quarter revenue reached $16.5 billion, up 21 percent, but its fourth-quarter outlook called for $15 billion to $16 billion and approximately zero dollars of non-generally accepted accounting principles, or non-GAAP, diluted earnings per share. Management also said personal-computer weakness could continue through the first half of 2027. AI demand can be strong while the broader semiconductor turnaround remains fragile.

Sam Dewinski:

Markets reacted to that forward uncertainty. Intel shares fell 28.6 percent on October 30, 2026, and the decline spread to other semiconductor and AI-linked stocks. That questioned valuation, capacity, and timing; it did not show that useful AI capabilities had vanished. Infrastructure booms often produce that argument before anyone knows which capacity will be fully utilized.

Red Velhouse:

Let’s separate technological improvement from the economics of using it. Omar, what do the newest efficiency measures show?

Omar Seidren:

They show more than a falling token price. Epoch AI’s September 2026 analysis estimated that the cost of achieving a given performance level across five benchmarks fell about 47 percent per quarter since 2023—roughly a thirteen-fold decline per year. The rate varied by domain, from about seven- to ten-fold annually in game-based puzzles to sixteen- to nineteen-fold in mathematics. One example estimated that reaching about 75 percent accuracy on the Graduate-Level Google-Proof Question Answering benchmark, known as GPQA Diamond, fell from roughly thirty cents per question with OpenAI o3 in January 2025 to about $0.0004 with GPT-5.6 Luna by September 2026. That is a striking benchmark result, not the cost of an entire business process.

Kate Burvish:

That distinction is essential. A business workflow includes retrieval, retries, tool calls, human review, security, integration, errors, and waiting time. Reasoning models can also use far more tokens and energy than standard queries. A 2026 study modeled roughly thirteen times as much energy for a reasoning query producing around 5,000 output tokens. The useful denominator is cost per reliable customer response, reviewed report, completed claim, or deployable software change—not cost per token.

Red Velhouse:

So the price of a thought may be falling while the price of putting it into a dependable workflow remains uncertain. And lower unit costs have not produced a smaller physical footprint. Why not?

Omar Seidren:

Because cheaper capability invites more capability use. Global data-center electricity consumption rose 17 percent in 2025, while electricity use by AI-focused data centers rose about 50 percent, according to the International Energy Agency. Five large technology companies spent more than $400 billion in capital expenditure in 2025, and the largest data-center operators were expected to spend more than $700 billion in 2026. The scopes differ, but the pattern is a rebound effect: efficiency per operation improves while total demand rises.

Sam Dewinski:

That pattern has historical precedent. Electrification made individual applications more useful and efficient, but total electricity demand expanded as factories, homes, and new industries adopted it. Efficiency reduces resource use in aggregate only if new demand does not grow faster than the savings per unit. The same lesson appeared in information technology: infrastructure created value, but complementary changes in management and work design took time.

Red Velhouse:

Let’s apply that distinction to adoption, where the numbers often seem contradictory. Kate, which measure should lead?

Kate Burvish:

For the United States, the nationally representative Census Business Trends and Outlook Survey is the best starting point for firm adoption. During November 2025 to January 2026, about 18 percent of firms reported using AI in at least one business function. Later estimates through May 3, 2026 put overall business use between 17 and 20 percent, with a May estimate of 19.8 percent. That is firm-weighted: each firm counts as one, whether it has four employees or thousands. Change the denominator, and the picture changes. Current use was about 37 percent among firms with at least 250 employees, compared with less than 20 percent among firms with four or fewer employees.

Omar Seidren:

Worker surveys answer a different question. The Federal Reserve reported that about 41 percent of the workforce used generative AI for work in November 2025. Census’s March 2026 household survey found about 56 percent had used AI on at least one of eleven job tasks. Those figures are not interchangeable: they use different definitions, questions, and reference periods. Another Federal Reserve measure found that roughly 78 percent of the labor force worked at firms that had adopted some form of AI, but that means employment-weighted access, not personal use or deep deployment.

Sam Dewinski:

The same caution applies to Stanford’s figure of 88 percent of surveyed organizations using AI in at least one function in 2025. It should not be placed beside the Census figure as if one corrected the other. Stanford surveys organizations with a different size and geographic distribution, using a broad definition; Census estimates the share of firms in a national business frame.

Omar Seidren:

The Stanford result still shows that organizations with the resources and motivation to implement AI are moving quickly. But access is spreading faster than deep, autonomous workflow deployment. Agent deployment remained in the single digits across nearly all functions, and Census found that 57 percent of adopting firms used AI in three or fewer business functions. The robot may be invited to the office, but it is still waiting for permissions to access the shared drive.

Red Velhouse:

Access is not intensity, and intensity is not productivity. What are workers actually using AI to do?

Kate Burvish:

In the March 2026 Census household survey, leading uses included information search or technical help at 37 percent, writing communications or documentation at 32 percent, idea generation at 32 percent, summarizing or interpreting information at 31 percent, and administrative work at 27 percent. Among workers who used AI during the prior week, 31 percent said it saved one to two hours and 25 percent said it saved less than an hour. Those are self-reported time savings, not measured output per hour or payroll reductions. A worker may use saved time to improve quality, handle more cases, or take on more assignments rather than reduce labor costs.

Omar Seidren:

There is encouraging evidence, but it is not uniform. A March 2026 National Bureau of Economic Research survey of nearly 750 corporate executives found that more than half of firms had invested in AI and reported positive labor-productivity gains, varying substantially by sector. That is useful evidence of experience and expectation, but not the same as an independently measured productivity experiment. At the economy-wide level, U.S. private nonfarm business labor productivity rose 2.2 percent in 2025, but that statistic cannot establish that AI caused the increase.

Sam Dewinski:

That gap resembles earlier general-purpose technologies. Electricity and information technology eventually changed factory layouts, management, and job design, but buying infrastructure was not the same as reorganizing production around it. Railroads created real economic value, yet returns depended on route, utilization, and timing. Useful technology does not guarantee that every project, chip order, or data center earns an attractive return.

Red Velhouse:

Does consumer adoption change the conclusion, or does it show that people value AI before businesses can measure its returns?

Omar Seidren:

It shows that value is already broad, even if business productivity is harder to measure. Stanford estimates that U.S. generative-AI consumer users rose to roughly 115 million by early 2026, while annual consumer surplus reached about $172 billion. That is evidence of perceived consumer value, not company revenue or gross domestic product.

Kate Burvish:

Consumer surplus can be large before firms redesign operations. The evidence supports three separate claims: capability per dollar is improving extremely quickly; people are using AI more; and broad enterprise productivity remains uneven and difficult to identify. Privacy, reliability, procurement, training, regulation, and workflow-redesign costs remain the constraints between technical potential and realized productivity.

Red Velhouse:

What should viewers watch to distinguish productive capacity from an expensive construction cycle?

Kate Burvish:

Watch whether revenue growth converts into cash after depreciation, energy, financing, and maintenance. Track utilization, cloud pricing, customer retention, and demand that is durable rather than merely projected. Also watch who pays for electricity, transmission, and public incentives. Falling inference costs can create consumer benefits while shifting infrastructure risks toward utilities, taxpayers, or communities.

Omar Seidren:

Track cost per successful outcome, not only cost per token. Are customers resolving more cases, producing better documents, finding information faster, or shipping reliable software after review? Measure completion time, error rates, human intervention, energy per completed workflow, and the number of functions used repeatedly. That tells us whether the price of thought is falling in a laboratory or becoming cheaper work in the world.

Sam Dewinski:

Finally, look for organizational change. If firms train workers, redesign processes, and connect systems while usage becomes routine, productivity effects may appear with a lag. If spending rises but workflows remain experimental and utilization disappoints, the technology can still be transformative while the investment cycle proves excessive.

Red Velhouse:

The evidence supports a careful three-part conclusion. First, AI capability is becoming dramatically cheaper per unit of benchmark performance. Epoch AI estimates roughly a thirteen-fold annual decline in the cost of achieving a given performance level since 2023, with large variation by domain. But inference cost is not task cost: reasoning tokens, latency, energy, human review, security, and integration determine the cost of a successful workflow. Second, adoption is expanding but depends on the denominator. About 18 to 20 percent of U.S. firms reported recent business-function use in nationally representative Census measures. Employment-weighted access and worker-use measures are higher, including 41 percent of workers reporting generative-AI use and 56 percent reporting use on at least one surveyed task. Stanford’s 88 percent organization figure describes a different survey population and should not be treated as the share of U.S. firms. Third, falling costs are producing substantial consumer value and growing use, but enterprise deployment is often shallow and economy-wide productivity evidence remains mixed. Infrastructure demand can rise even as each inference becomes cheaper, because lower prices encourage more queries, longer reasoning, and new applications. The investment question is not whether AI is becoming more capable. It is whether utilization, reliable completed workflows, customer revenue, and firm-level productivity will catch up with the enormous cost of building the physical platform. Sources and references for this discussion are available with the episode at Factolio.com.

---

Sources and References

These sources supported the factual material used in this discussion. Factolio’s panel discussion is AI-generated from researched evidence and is written in original language.

1. Stanford Institute for Human-Centered Artificial Intelligence — AI Index 2025: Inference Cost Chapter (PRIMARY)
2. Gundlach, Lynch, Mertens, and Thompson — The Price of Progress: Algorithmic Efficiency and the Falling Cost of AI Inference (PRIMARY)
3. Board of Governors of the Federal Reserve System — Monitoring AI Adoption in the U.S. Economy (PRIMARY)
4. U.S. Census Bureau — About a Third of Workers Who Used AI in the Last Week Said They Completed Tasks One to Two Hours Faster (PRIMARY)
5. ScienceDirect / Joule — Energy Use of AI Inference: Efficiency Pathways and Test-Time Scaling (PRIMARY)
6. U.S. Census Bureau — The Microstructure of AI Diffusion: Evidence from Firms, Business Functions, and Worker Tasks (PRIMARY)
7. Stanford Institute for Human-Centered Artificial Intelligence — The 2026 AI Index Report (PRIMARY)
8. Epoch AI — The Plunging Price of Thought (ANALYSIS)
9. OpenAI — API Platform and Model Pricing (PRIMARY)
10. U.S. Census Bureau — Large Firms With at Least 20 Employees Biggest AI Users (PRIMARY)
11. Stanford Digital Economy Lab — What is Generative AI Worth? (PRIMARY)
12. National Bureau of Economic Research — Artificial Intelligence, Productivity, and the Workforce: Evidence from Corporate Executives (PRIMARY)