Opening verdict
“Bubble” can describe technology that does not work, spending ahead of use, financing dependent on easy capital, or securities priced for implausible outcomes. Those claims are often collapsed. They should not be.
The evidence shows reported users and a physical buildout. Organizations report broad use. Combined company-reported CMBU+CDBU revenue rose, while selected BLS occupations have divergent ten-year projections. But use somewhere inside a company is not full deployment, and deployment is not material profit.
The better answer is not “yes” or “no.” Technology can be real while parts of the capital cycle overshoot. The question is how much spending becomes durable cash flow, for whom and when.
Evidence snapshot
- Survey · organizational AI use
- 88% reported regular AI use in at least one business function in 2025.Survey
- Survey · enterprise EBIT
- 39% of respondents at AI-using organizations attributed any enterprise-level EBIT impact to AI; most reported less than 5%.Survey
- Private AI investment · 2025
- $344.7bn globally in the Stanford/Quid dataset, up 127.5% year over year.Dataset estimate
- Combined company-reported CMBU+CDBU revenue · FQ3 2026
- $25.293bn on Micron’s retrospectively recast business-unit basis.Historical actual · recast
- HBM market · 2025 estimate to 2028 forecast
- Approximately $35bn to $100bn—Micron’s market estimate and supplier forecast, not observed industry results.Supplier estimate / forecast
- US data-scientist employment · 2024–34
- +33.5% in the latest BLS ten-year projection.Projection
Adoption is broad; scaling is shallow
McKinsey’s 2025 survey captures both the reach and the limit of the current cycle. Eighty-eight percent of respondents said their organizations regularly used AI in at least one function, up from 78% in 2024. Generative-AI use rose from 71% to 79%. These are self-reported, weighted survey results from 1,993 respondents across 105 countries—not a census of companies.
Reported AI use is broad; organization-wide deployment is not
Two horizontal-bar panels. The first compares reported AI and generative-AI use in 2024 and 2025. The second shows 2025 deployment phases among organizations that use AI.
Reported use in at least one business function
Share of respondents; 2024 versus 2025
% of survey respondents| Measure | Value | Classification | Note |
|---|---|---|---|
| Any AI · 2024 | 78% | Survey | No additional note. |
| Any AI · 2025 | 88% | Survey | No additional note. |
| Generative AI · 2024 | 71% | Survey | No additional note. |
| Generative AI · 2025 | 79% | Survey | No additional note. |
Deployment phase in 2025
n=1,753 respondents whose organizations reported AI use; phases sum to 100%
% of AI-using organizations| Measure | Value | Classification | Note |
|---|---|---|---|
| Experimenting | 32% | Survey | No additional note. |
| Piloting | 30% | Survey | No additional note. |
| Scaling | 31% | Survey | No additional note. |
| Fully scaled | 7% | Survey | No additional note. |
The denominator matters. Scaling plus fully scaled equals 38% among AI-using organizations in the phase question, or roughly one-third of the full respondent base after accounting for the 88% adoption rate. Only 7% of respondents at AI-using organizations placed deployment in fully scaled. Separately, 39% of respondents at AI-using organizations reported any enterprise-level EBIT impact, and most of that group attributed less than 5% of EBIT to AI.
The survey therefore supports diffusion, not mature economics. It is possible for AI to be common, operationally useful and still early in proving enterprise-wide returns.
Capital is moving faster than demonstrated value
Stanford’s 2026 AI Index, using Quid data, estimates global private AI investment reached $344.7 billion in 2025, up 127.5% from $151.5 billion in 2024. Generative-AI investment reached $170.9 billion—49.6% of the total—and the United States accounted for $285.9 billion, about 83% of the global figure.
Those are dataset estimates, not audited totals. Quid tracks AI and machine-learning companies above a funding threshold; it excludes government capital and may understate markets where private transactions are less visible. Concentration also cuts two ways: it can reflect a genuine US ecosystem advantage, and it can make the cycle depend on a small set of funders, customers and assumptions.
Private investment more than doubled while survey-reported organizational AI use rose ten percentage points. Those measures have different units, populations and baselines, and the adoption percentage is already approaching a ceiling, so they do not belong on one axis. The comparison is directional: capital is accelerating much faster than reported diffusion. That does not prove a bubble. It raises the return threshold the industry must clear.
Micron’s revenue rose; the HBM endpoint is still a forecast
AI accelerators can process calculations faster than conventional memory can deliver data. High-bandwidth memory, or HBM, stacks specialized DRAM close to the processor and moves far more data in parallel. That makes memory bandwidth a real system bottleneck, not a narrative accessory.
Combined company-reported CMBU+CDBU revenue rose; the HBM endpoint is forecast
The first panel shows historical actuals on Micron’s retrospectively recast basis. The series is Combined company-reported CMBU+CDBU revenue. The second panel separately shows Micron’s 2025 HBM market estimate and 2028 supplier forecast.
Combined company-reported CMBU+CDBU revenue
Historical actuals; retrospectively recast after Micron’s FQ4 2025 reorganization
US$ billions per fiscal quarter| Measure | Value | Classification | Note |
|---|---|---|---|
| FQ3 2025 | $4.916bn | Historical actual · recast | No additional note. |
| FQ2 2026 | $13.436bn | Historical actual · recast | No additional note. |
| FQ3 2026 | $25.293bn | Historical actual · recast | No additional note. |
Global HBM total addressable market
Micron supplier estimate and forecast; not reported market revenue
US$ billions per year| Measure | Value | Classification | Note |
|---|---|---|---|
| 2025 estimate | ≈$35bn | Supplier estimate | No additional note. |
| 2028 forecast | ≈$100bn | Supplier forecast | No additional note. |
Combined company-reported CMBU+CDBU revenue rose from $4.916 billion to $25.293 billion year over year on that recast basis. These units include broader memory and storage activity, so the series is not HBM revenue, memory demand, data-center demand or an industry-demand measure. Changes reflect shipments, pricing and mix. Separately, Micron said price-and-volume agreements covered its planned calendar-2026 HBM supply, including HBM4. Contracted Micron capacity is evidence about one supplier, not proof that all industry capacity is sold out.
Micron forecasts the HBM market rising from about $35 billion in 2025 to about $100 billion in 2028. The endpoint arithmetic implies 41.9% annualized growth, broadly consistent with its “approximately 40%” wording. The 2028 value remains a supplier forecast. Micron’s own filing warns about uncertain long-term demand, customer concentration, power and water availability, construction delays, capital needs and the risk that HBM capacity could return to conventional DRAM and pressure prices. Strong demand has not abolished semiconductor cyclicality.
The evidence supports an active infrastructure cycle: rising company-reported revenue, Micron-specific HBM agreements and a supplier market forecast. Those are three different evidence classes—not one measure of HBM demand or proof of durable returns.
Selected US tech-occupation projections diverge
BLS projects different 2024–34 paths across the selected occupations, including growth in data, security, research and software-development categories and a decline in computer programmers. These are ten-year projections—not current hiring, observed employment changes or evidence that AI caused the differences.
Among the selected categories, only computer programmers decline. One declining category cannot establish that older technical roles broadly are weakening.
Selected US tech-occupation projections diverge
Diverging horizontal bars compare BLS 2024–34 projections across five selected occupations and the all-occupations benchmark; they do not show current hiring or observed employment changes.
BLS projected employment change, 2024–34
One national projection set; zero is shown on the axis
% change in employment| Measure | Value | Classification | Note |
|---|---|---|---|
| Data scientists | +33.5% | Projection | No additional note. |
| Information security analysts | +28.5% | Projection | No additional note. |
| Computer & information research scientists | +19.7% | Projection | No additional note. |
| Software developers | +15.8% | Projection | No additional note. |
| All occupations | +3.1% | Projection | No additional note. |
| Computer programmers | −6.0% | Projection | No additional note. |
Stanford reports that AI skills appeared in 2.56% of US postings in 2025, using Lightcast’s scraped and deduplicated postings; generative-AI skill mentions in AI postings more than doubled even as their share of AI postings slipped 5%. LinkedIn reported stronger AI-specific hiring and posting measures while its broad member-based US hiring measure in February 2026 remained 23% below February 2020. Job postings are advertisements, not hires. Lightcast measures recruitment advertisements. LinkedIn measures activity among its platform members and postings. Neither source represents the whole labor market.
Lightcast postings and LinkedIn platform measures add current but differently defined signals. Together they make labor re-sorting a limited hypothesis worth watching; they do not establish it. One declining programmer category cannot show that older technical roles broadly are weakening. This is not a present-market conclusion or a net-jobs verdict.
Where the bubble argument is strongest
The strongest case is not that AI is fictitious. It is that financing and valuation can outrun demonstrated economics. Private investment accelerated far faster than reported adoption, enterprise-wide EBIT effects remain shallow for most respondents, and the infrastructure case leans partly on supplier forecasts whose endpoints require sustained growth.
The risks also stack. Investment is geographically concentrated. Supplier revenue is exposed to a small number of large customers. Data centers require power, water, permits, construction labor and capital on schedule. High margins can attract capacity just as demand forecasts soften. A project can be technologically useful and still destroy value if it is financed too aggressively or built for utilization that never arrives.
That is classic repricing territory: not necessarily fraud or technological failure, but optimistic terminal assumptions embedded across a connected capital chain.
Where the bubble argument is weakest
The weakest version says there is no underlying activity. Survey use is widespread. Combined company-reported CMBU+CDBU revenue rose, HBM exists to solve an identifiable bandwidth constraint, and another major supplier, SK hynix, reported that 2025 HBM revenue more than doubled. Selected platform measures also show AI-related recruitment activity.
None of that establishes the correct valuation for a company, security or data-center project. It does establish that the cycle has users, physical constraints and supplier revenue. The labor evidence remains narrower: selected BLS projections and platform measures support only a hypothesis worth watching. The analogy to a boom built around a nonexistent product does not fit the present record.
Buteon view
Separate technology risk, infrastructure-cycle risk, company-economics risk and valuation risk. The evidence in this note weighs against the strongest “technology with no use” version of the first. The other three remain open: physical capacity can overshoot, business models can fail to capture value, and prices can assume outcomes that even successful technology cannot deliver.
Our classification is “Structural boom; valuation risk unresolved.” It is deliberately not a market call. It describes the evidence boundary: adoption and infrastructure are real; durable returns, broad value capture and the terminal scale implied by current financing are not yet settled.
What to watch next
- Value depth: whether organizations move from pilots to redesigned workflows and report material, repeatable productivity or EBIT gains—not just tool access.
- Infrastructure utilization: whether accelerator and data-center use remains durable after the current buildout, rather than capacity arriving ahead of workloads.
- Memory normalization: whether HBM supply catches up with demand, and whether pricing or margins reverse sharply as capacity expands.
- Labor diffusion: whether AI demand spreads beyond a narrow group of firms and roles into broader deployment, domain and operating work.
- Financing resilience: whether projects still clear return thresholds when power, construction and capital cost more than planned.
Stalled value depth, underused capacity, a sharp memory reversal, persistently narrow hiring, or projects that fail at higher costs would strengthen the bubble case. Durable workflow redesign, inference demand, utilized capacity, distributed hiring and resilient financing would strengthen the structural-boom case.
Sources and methodology
- McKinsey — The State of AI 2025 (opens in a new tab)
Survey methods, adoption, deployment phase and reported EBIT impact.
- Stanford AI Index 2026 — Economy chapter (opens in a new tab)
Investment synthesis using Quid data and labor synthesis using Lightcast data, both with coverage and revision limits.
- Micron Q3 FY2026 Form 10-Q (opens in a new tab)
Source for Combined company-reported CMBU+CDBU revenue, operating drivers, HBM context and risk disclosures.
- Micron 2025 Form 10-K (opens in a new tab)
FQ4 business-unit reorganization and retrospective adjustment of prior segment amounts.
- Micron Q1 FY2026 earnings deck (opens in a new tab)
HBM TAM supplier forecast and Micron’s 2026 HBM agreement statement.
- SK hynix FY2025 results (opens in a new tab)
First-party cross-check that a second major supplier reported HBM revenue more than doubling in 2025.
- US Bureau of Labor Statistics — 2024–34 projections (opens in a new tab)
National occupational projections on one consistent basis.
- Lightcast — job-posting methodology (opens in a new tab)
Posting collection, deduplication and classification limits; postings are not employment.
- LinkedIn Economic Graph — April 2026 AI Labor Market Update (opens in a new tab)
Platform-based AI hiring and entry-level comparisons; member and job-posting data are not a labor census.
- LinkedIn Economic Graph — US Monthly Insights, March 2026 (opens in a new tab)
Member-based February 2026 US hiring versus pre-pandemic; not a labor census.
Product bridge
Buteon does not monitor AI adoption, HBM markets, jobs or the datasets in this note live. The connection is methodological: separate what was observed from what was estimated, forecast or inferred, then keep the evidence boundary visible.
Disclaimer
Buteon is a research tool. This note is educational, uses public reports and datasets reviewed through 23 July 2026, and does not provide investment advice or buy, sell or short recommendations. Supplier forecasts, surveys and projections may change. Nothing here is a prediction, valuation opinion or performance claim.