# AI Is Not Overvalued. It Is Underestimated.

The more material risk is underestimation of what is already underway, and of the cost of remaining outside it.

Canonical HTML: https://www.livetradingnews.com/ai-is-not-overvalued-it-is-underestimated
Last modified: 2026-10-31

---

By Shayne Heffernan · 2026-10-31
Tags: $GOOGL, $AMZN, $MSFT, $META, $GXO, AI, Artificial intelligence, Alphabet, Google Cloud, Gemini, Data centres, Capex, Hyperscalers, Compute, Humanoid robots, Physical AI
Signed: ML-DSA-87, anchored on Armature L1.
Nothing in this article is investment advice.

The current argument over artificial intelligence (AI) is misframed. Markets and commentators debate whether the technology is overvalued. They measure quarterly free cash flow, tally capital expenditure against near-term revenue, and treat the largest infrastructure build in a generation as a speculative bubble. That framing is inverted. The more material risk is underestimation of what is already underway, and of the cost of remaining outside it.

Artificial intelligence is no longer a discretionary technology. It is becoming the operating layer of the new economy. Compute is the scarce input. Models are the tools that convert it into capability. Organisations and jurisdictions that secure both will compound advantages in productivity, decision quality and physical systems. Those that restrict the infrastructure or treat the transition as optional will purchase the same capability later, at higher cost, from parties who built the capacity. The choice is participation, not whether the shift occurs.

This is not a claim that every current valuation is justified or that every project will succeed. It is a claim that the underlying demand for intelligence at scale is structural, that the infrastructure required to supply it is being built under real constraint, and that opposition to that infrastructure often confuses local externalities with national strategy. The Google chief executive stated the position directly in July 2024: [the risk of underinvesting is dramatically greater than the risk of overinvesting](https://diginomica.com/google-cloud-growth-soars-ai-demand-increases-so-does-capex-price-be-paid-alphabet-ceo-sundar). The same arithmetic applies to countries and regions.

## The scale of the commitment

Alphabet's third-quarter 2026 results, [reported on 28 October](https://www.sec.gov/Archives/edgar/data/1652044/000165204426000080/googexhibit991q32026.htm), illustrate the point with unusual clarity. The company reported capital expenditure of [$51.4 billion in the quarter](https://www.sec.gov/Archives/edgar/data/1652044/000165204426000080/googexhibit991q32026.htm), more than double the year-earlier figure of [$24.0 billion](https://www.sec.gov/Archives/edgar/data/1652044/000165204426000080/googexhibit991q32026.htm). [Full-year 2026 guidance remained $195 billion to $205 billion](https://singjupost.com/alphabet-googl-q3-2026-earnings-call-transcript/). Management indicated that [2027 spending would increase significantly](https://singjupost.com/alphabet-googl-q3-2026-earnings-call-transcript/) from that already high base. [Google Cloud revenue rose 103% to $30.7 billion](https://www.sec.gov/Archives/edgar/data/1652044/000165204426000080/googexhibit991q32026.htm). First-party model application programming interfaces (APIs) processed [nearly 32 billion tokens per minute, more than four times the volume of a year earlier](https://www.sec.gov/Archives/edgar/data/1652044/000165204426000080/googexhibit991q32026.htm). The [Gemini application reached one billion users, making it the fastest-growing product in Google's history](https://www.sec.gov/Archives/edgar/data/1652044/000165204426000080/googexhibit991q32026.htm), chief executive Sundar Pichai said.

These figures are not abstract. They reflect concrete constraints. Compute available today is the result of decisions made roughly two years earlier. [Demand has repeatedly exceeded planned capacity](https://singjupost.com/alphabet-googl-q3-2026-earnings-call-transcript/). Cloud customers, including [large enterprises and frontier model developers](https://singjupost.com/alphabet-googl-q3-2026-earnings-call-transcript/), have committed to [multi-year agreements](https://www.sec.gov/Archives/edgar/data/1652044/000165204426000084/goog-20260930.htm). Alphabet's [purchase commitments stood at $924 billion](https://www.sec.gov/Archives/edgar/data/1652044/000165204426000084/goog-20260930.htm) against a [cloud backlog of $561 billion](https://www.sec.gov/Archives/edgar/data/1652044/000165204426000084/goog-20260930.htm) at the end of the quarter. Inventory [expanded sharply, from $2.4 billion at the end of 2025 to $13.3 billion](https://www.sec.gov/Archives/edgar/data/1652044/000165204426000084/goog-20260930.htm), as the company [built tensor processing unit systems it had contracted to sell](https://singjupost.com/alphabet-googl-q3-2026-earnings-call-transcript/).

![Exhibit 1: Google Cloud revenue rose from $15.2bn in Q3 2025 to $30.7bn in Q3 2026. At 30 September its backlog was $561bn, and Alphabet's purchase commitments to its suppliers were $924bn.](https://livetradingnews-media.nyc3.digitaloceanspaces.com/media/2026/10/31/cmpgg3-cc7a9b28956a2fcd.svg)

Exhibit 1 sources: Alphabet's [third-quarter 2026 results](https://www.sec.gov/Archives/edgar/data/1652044/000165204426000080/googexhibit991q32026.htm), and its quarterly report on Form 10-Q for the [backlog](https://www.sec.gov/Archives/edgar/data/1652044/000165204426000084/goog-20260930.htm) and the [purchase commitments](https://www.sec.gov/Archives/edgar/data/1652044/000165204426000084/goog-20260930.htm).

[Free cash flow was negative for a second consecutive quarter, at minus $9.1 billion](https://www.sec.gov/Archives/edgar/data/1652044/000165204426000080/googexhibit991q32026.htm), because capital spending exceeded operating cash flow. That outcome is frequently cited as evidence of excess. It is more accurately read as evidence that the firm is converting an advertising cash engine into an infrastructure platform at a pace set by customer demand and competitive necessity. The assets being acquired (servers, networking equipment, power connections and data-centre shells) are not single-purpose. They can be reallocated across internal research, cloud customers and future product lines. The cost of arriving late to sufficient capacity is the forgone revenue and the loss of relative position.

![Exhibit 2: Alphabet's capital expenditure rose from $27.9 billion in Q4 2025 to $51.4 billion in Q3 2026 and passed operating cash flow of $39.1 billion and $42.3 billion in Q2 and Q3 2026.](https://livetradingnews-media.nyc3.digitaloceanspaces.com/media/2026/10/31/cmpgg3-f12e26b7bc17f508.svg)

Exhibit 2 source: the free cash flow reconciliation in Alphabet's [third-quarter 2026 results](https://www.sec.gov/Archives/edgar/data/1652044/000165204426000080/googexhibit991q32026.htm).

Sundar Pichai has been consistent on the relative risks. In earnings discussions he has stated that the risk of underinvesting is high, and in April 2026 that [cloud revenue would have been higher had Google been able to meet demand](https://techcrunch.com/2026/04/29/google-cloud-surpasses-20b-but-says-growth-was-capacity-constrained/). He has also noted [elements of irrationality](https://risingnepaldaily.com/news/71300) in the broader investment cycle and drawn the parallel to the internet era: [excess capital was deployed, yet the underlying technology proved profound](https://risingnepaldaily.com/news/71300). The distinction matters. Cyclical overshoot in specific assets or valuations does not reverse the structural increase in demand for machine intelligence.

Other hyperscalers are making parallel commitments. Combined capital expenditure across the leading platforms is measured in the hundreds of billions of dollars for 2026 alone. Amazon expects [about $220 billion](https://fortune.com/2026/10/30/why-big-tech-earnings-are-a-crucial-report-card-for-the-ai-boom-and-its-future/), Microsoft [about $170 billion for the calendar year](https://fortune.com/2026/10/30/why-big-tech-earnings-are-a-crucial-report-card-for-the-ai-boom-and-its-future/) and Meta [$135 billion to $145 billion](https://www.hl.co.uk/shares/share-research/meta-q3-results-strong-quarter-spending-still-a-question); with Alphabet's range, the four together come to roughly $720 billion to $740 billion. The common constraint is power, advanced chips and physical space with adequate grid connections. These are not purely financial decisions. They are industrial decisions about where the next layer of productive capacity will be located.

![Exhibit 3: 2026 capital expenditure guidance, about $720bn to $740bn in all: Amazon about $220bn, Alphabet $195bn to $205bn, Microsoft about $170bn for the calendar year, Meta $135bn to $145bn.](https://livetradingnews-media.nyc3.digitaloceanspaces.com/media/2026/10/31/cmpgg3-e2f7102073344897.svg)

Exhibit 3 sources: [Fortune](https://fortune.com/2026/10/30/why-big-tech-earnings-are-a-crucial-report-card-for-the-ai-boom-and-its-future/) for Amazon and Microsoft, Alphabet's [earnings call](https://singjupost.com/alphabet-googl-q3-2026-earnings-call-transcript/) and [Hargreaves Lansdown](https://www.hl.co.uk/shares/share-research/meta-q3-results-strong-quarter-spending-still-a-question) for Meta.

The pattern is not limited to the largest platforms. Enterprises across manufacturing, logistics, financial services and scientific research are contracting for capacity because the alternative, training or running models on shared or constrained resources, imposes latency, cost and data-sovereignty penalties that compound. A firm that can fine-tune models on its own operational data and deploy inference close to its physical processes gains a measurable advantage in cycle time. That advantage is difficult to reverse once competitors have accumulated the corresponding operational data and process knowledge. The capital expenditure figures therefore understate the private returns available to those who secure capacity early; they measure only the public market's willingness to fund the infrastructure, not the full set of gains that accrue downstream.

The same logic extends to sovereign and large institutional buyers. Defence, energy and healthcare systems generate data that cannot leave controlled environments. Running frontier-class models inside those environments requires local or trusted compute. Jurisdictions that have already constrained commercial data-centre construction will find it harder, not easier, to meet those requirements when they arise. The lead time for power interconnection and specialised hardware is measured in years: in the US regions with data, the median power plant built in 2025 took [over five years from interconnection request to commercial operation](https://eta-publications.lbl.gov/publications/queued-2026-edition-characteristics), and [generator step-up transformer lead times passed 160 weeks](https://www.kitco.com/news/off-the-wire/2026-07-09/us-power-companies-scramble-secure-equipment-surging-data-center) by early 2026. Decisions deferred in 2026 become capacity shortfalls in 2028 and 2029.

## The cost of restricting capacity

Local and regional opposition to data centres has intensified: [active opposition groups more than doubled](https://datacenterwatch.org/q1-2026) in the first quarter of 2026 and now span 49 states. Polling shows majorities opposed to new facilities in their communities, [71% against AI data centres in Gallup's March survey](https://news.gallup.com/poll/709772/americans-oppose-data-centers-area.aspx), and [61% in the Annenberg Public Policy Center's summer survey, up from 49% in February and March](https://eponline.com/articles/2026/08/11/public-opposition-to-local-data-centers-rises-sharply.aspx). Hundreds of jurisdictions have adopted moratoria, zoning restrictions or outright bans; GatherGov counts [270 that have adopted a moratorium](https://gathergov.com/articles/data-center-moratorium-2026-tracker). Voters decide at least 50 data-centre ballot measures this year, [41 of them on 3 November across ten states](https://ballotpedia.org/Ohio_Prohibition_of_Data_Center_Construction_Amendment_%282026%29), and in [all six decided by 30 October the majority favoured greater limits](https://ballotpedia.org/Ohio_Prohibition_of_Data_Center_Construction_Amendment_%282026%29) on development. A proposed Ohio constitutional amendment to ban large data centres has been [cleared for signature gathering](https://ballotpedia.org/Ohio_Prohibition_of_Data_Center_Construction_Amendment_%282026%29).

![Exhibit 4: share of US adults opposed to new local data centres: 71% for AI data centres (Gallup, March 2026); 49% then 61% (Annenberg, February and March, then June and July 2026).](https://livetradingnews-media.nyc3.digitaloceanspaces.com/media/2026/10/31/cmpgg3-a2015bfab0d30fd3.svg)

Exhibit 4 sources: [Gallup](https://news.gallup.com/poll/709772/americans-oppose-data-centers-area.aspx) and the [Annenberg Public Policy Center](https://eponline.com/articles/2026/08/11/public-opposition-to-local-data-centers-rises-sharply.aspx).

Projects worth nearly $200 billion were blocked or delayed in the first half of 2026 alone: Data Center Watch counts [about $130 billion in the first quarter](https://datacenterwatch.org/q1-2026) and [nearly $68 billion in the second](https://datacenterwatch.org/q2-2026), and US cancellations due to local opposition [rose from two in 2023 to 25 in 2025, by Heatmap's count](https://intelligence.uptimeinstitute.com/resource/data-center-cancellations-rise-public-opposition-grows). Concerns about [electricity rates, water use, land use](https://datacenterwatch.org/q1-2026) and [noise](https://gathergov.com/articles/data-center-moratorium-2026-tracker) are real and deserve technical answers. They do not, however, change the location-independent demand for the capability the facilities produce.

![Exhibit 5: projects blocked or delayed by local opposition: about $130bn in Q1 2026, a record roughly equal to all of 2025, and nearly $68bn in Q2. Cancelled over it: 2 (2023), 6 (2024), 25 (2025).](https://livetradingnews-media.nyc3.digitaloceanspaces.com/media/2026/10/31/cmpgg3-aa5815b02f8a7518.svg)

Exhibit 5 sources: Data Center Watch for the [first quarter](https://datacenterwatch.org/q1-2026) and the [second quarter](https://datacenterwatch.org/q2-2026), and [Heatmap via Uptime Institute](https://intelligence.uptimeinstitute.com/resource/data-center-cancellations-rise-public-opposition-grows) for cancellations.

A jurisdiction that prevents construction does not eliminate the demand. It shifts the location of the supply. Compute is then purchased from elsewhere, often at a premium that reflects scarcity, transmission costs and the absence of local competition. The jurisdiction forgoes the construction activity, the associated tax base, and the option to host higher-value activities that require low-latency or high-volume access. Over a multi-year horizon the cumulative effect is a higher effective price for the same intelligence services and a slower rate of local adoption.

At the national level the calculation is sharper. Countries that systematically constrain the expansion of compute capacity will, within a short number of years, find their firms and public institutions purchasing capability from competitors who did not impose the same constraints. The gap shows in cost and in the speed of product iteration, the quality of models trained on proprietary data, the ability to run large inference workloads for logistics or scientific work, and the capacity to integrate physical systems.

Physical AI (robots and automated equipment that operate in factories, warehouses and other real environments) depends on the same underlying compute infrastructure. Early commercial deployments are already occurring. Shipments of humanoid systems in the first half of 2026 reached roughly 19,000 to 25,000 units, [about 19,100 by Smart Analytics Global's count](https://www.chinadailyhk.com/article/637755) and [nearly 25,000 by International Data Corporation's (IDC's)](https://technode.com/2026/09/29/china-accounts-for-77-9-of-global-humanoid-robot-shipments-in-h1-idc-says/), several times the prior-year figure, with the large majority originating from Chinese manufacturers: [77.9% by IDC's count](https://technode.com/2026/09/29/china-accounts-for-77-9-of-global-humanoid-robot-shipments-in-h1-idc-says/) and [more than 97% by Smart Analytics Global's](https://www.chinadailyhk.com/article/637755).

Western firms have moved from pilots towards commercial use: Figure AI's robot [supported production of more than 30,000 BMW X3s at Spartanburg over a ten-month pilot of ten-hour weekday shifts](https://www.press.bmwgroup.com/global/article/detail/T0455864EN/bmw-group-to-deploy-humanoid-robots-in-production-in-germany-for-the-first-time?language=en), and Agility Robotics' Digit has [moved more than 100,000 totes for GXO under a multi-year commercial agreement](https://manufacturingdigital.com/globenewswire/3388115). The volume remains small relative to the [5 million industrial robots already in service](https://ifr.org/ifr-press-releases/five-million-robots-now-operate-in-factories-globally), but the direction is clear: intelligence is moving from screens into actuators.

![Exhibit 6: humanoid robots shipped, first half of 2026, by tracker: about 19,100, 22,000-plus and nearly 25,000, nearly 3x to 5.3x a year earlier. 600,000-plus industrial robots installed in 2025.](https://livetradingnews-media.nyc3.digitaloceanspaces.com/media/2026/10/31/cmpgg3-df6c70b44578d666.svg)

Exhibit 6 sources: [Smart Analytics Global](https://www.chinadailyhk.com/article/637755), [Counterpoint](https://counterpointresearch.com/en/insights/global-humanoid-robot-shipments-soar-nearly-300-percent-yoy-in-h1-2026), [IDC](https://technode.com/2026/09/29/china-accounts-for-77-9-of-global-humanoid-robot-shipments-in-h1-idc-says/) and the [International Federation of Robotics](https://ifr.org/ifr-press-releases/five-million-robots-now-operate-in-factories-globally).

A country or region that treats data-centre expansion as an optional or primarily local matter is choosing to import the resulting capability rather than produce it. That choice is reversible only at higher cost and with a lag. The lag is the period during which competitors train better models, accumulate operational data from physical deployments, and lock in supply relationships for power and advanced hardware.

The political economy of the opposition is understandable. Data centres are capital-intensive and relatively light in permanent employment per dollar of investment. They consume power and, in some designs, water. Residents correctly observe that the local benefits are smaller than the local costs unless the jurisdiction captures a meaningful share of the tax base and the associated higher-value activity.

The policy error is to treat that observation as sufficient reason for prohibition rather than for negotiation over pricing, siting standards and benefit-sharing. Electricity markets already contain mechanisms for allocating scarce capacity; water use can be priced or offset; noise and visual impacts can be regulated by setback and design standards. None of these measures requires a ban. A ban simply exports the activity and imports the product at a higher price.

## Transparency, standards and displacement

Artificial intelligence will raise the cost of error in public and commercial argument. Large models can already retrieve, compare and summarise large volumes of primary material. As verification layers, cryptographic signing of records and persistent knowledge graphs mature, the practical standard for claims about policy, markets and technical performance will rise. Journalists and officials who have relied on selective citation or unexamined narrative will face lower-friction scrutiny. That development is not uniformly welcomed. It is, however, a predictable consequence of making relevant information cheaper to assemble and harder to alter after the fact.

The same systems will displace labour. This is not a novel observation. Every major productivity technology has done so. The industrial revolution shifted employment from agriculture and hand crafts into factories and later services. Electrification, the internal combustion engine and computing each produced large reallocations. The historical record does not show that the net result was permanent immiseration; it shows that societies that adapted institutions, skills and capital allocation captured the gains, while those that attempted to freeze the prior structure paid a relative price.

Artificial intelligence is already reducing the marginal cost of many cognitive tasks. Physical systems will extend that reduction into manipulation and mobility. The relevant policy question is not whether displacement occurs. It is how rapidly education, credentialing and capital markets adjust to the new distribution of tasks that remain scarce: those that require sustained judgement under real consequence, taste refined over time, and accountability that cannot be delegated.

Higher standards of evidence will also apply to claims about the technology itself. Assertions that current models are either universally reliable or wholly untrustworthy can be tested against specific benchmarks and operational records. Assertions that infrastructure spending is unbounded speculation can be tested against signed contracts, utilisation rates, and the observed gap between planned and demanded capacity. The technology does not eliminate disagreement. It raises the cost of maintaining disagreement against accumulating, checkable evidence.

## The early stage

The present moment is still early. Foundation models continue to improve. [Inference costs continue to fall for a given level of capability](https://epoch.ai/publications/the-plunging-price-of-thought), by about 47% a quarter since 2023, a 13-fold fall each year, Epoch AI reported on 22 September 2026. Physical systems are moving from demonstration to narrow commercial use. Energy and chip supply remain binding constraints. Regulatory frameworks are incomplete. These facts are consistent with underestimation rather than saturation. Markets that price the technology as if the bulk of the productivity gain has already been captured are pricing a different trajectory from the one implied by the capital commitments, the customer backlogs and the early physical deployments.

Organisations that treat artificial intelligence as a feature to be added later, or as a risk to be minimised through restriction, are making a different bet. They are betting that the productivity differential between those who integrate the technology and those who do not will remain small enough to be overcome by other advantages. Historical experience with general-purpose technologies suggests that differential widens before it narrows. The internet, once dismissed by some as a niche communications tool, reorganised distribution, media and coordination. Compute infrastructure is on a comparable path, with the additional property that the capability improves with use and with scale.

Participation therefore remains optional for any given firm or jurisdiction. The underlying shift does not. Countries and regions that secure power, advanced hardware and the institutional capacity to host and operate large-scale compute will set the terms on which others access intelligence services. Those that prioritise local veto over national capacity will discover that the veto does not reduce demand; it only changes the identity of the supplier and the price paid. The Google chief executive's observation applies beyond any single firm: [it is more dangerous not to invest](https://diginomica.com/google-cloud-growth-soars-ai-demand-increases-so-does-capex-price-be-paid-alphabet-ceo-sundar). The same logic holds for the jurisdictions that must decide whether to host the infrastructure or to import its product.

The technology is not overvalued in the sense that matters for long-term allocation. It is still in the phase in which the principal risk is arriving with insufficient capacity. Managing the transition is the practical work: addressing genuine local externalities with engineering and pricing rather than prohibition, preparing labour markets for task reallocation, and building verification layers that raise the standard of public argument. Burying the fact of the shift does not slow it. It only determines who captures the gains and who pays the higher price for the same capability.

The practical implication for executives and policymakers is straightforward. Measure the cost of delay in the same units used to measure the cost of investment. A year of deferred capacity is not a year of avoided expenditure; it is a year in which competitors accumulate training runs, operational data from physical systems, and contractual claims on scarce power and chips. Those assets do not depreciate in the ordinary sense. They compound. The firm or country that arrives with them already in hand sets the pace at which others must catch up. The firm or country that arrives without them discovers that the price of entry has risen and that the most valuable data has already been collected elsewhere.

This is the sense in which artificial intelligence remains underestimated. The capital markets are correctly registering large absolute expenditures. They are less consistently registering the relative cost of not making them. Local political systems are correctly registering the visible externalities of large facilities. They are less consistently registering the invisible cost of importing the resulting capability. Both errors are correctable. Neither is costless once the lag has opened.

Stocks mentioned: $GOOGL (NASDAQ), $AMZN (NASDAQ), $MSFT (NASDAQ), $META (NASDAQ) and $GXO (NYSE).

Shayne Heffernan, Ph.D., is the founder of the KnightsBridge Group ([shayneheffernan.com](https://www.shayneheffernan.com)). KXCO develops post-quantum infrastructure, verifiable records and enterprise systems for finance, identity and machine intelligence.

---

This Markdown mirrors https://www.livetradingnews.com/ai-is-not-overvalued-it-is-underestimated. The HTML page is canonical.
Site index for AI clients: https://www.livetradingnews.com/llms.txt