Euphemios
Back to insights

Report

Is AI data-center spending a bubble? Evidence for and against

Colibrì shows that a large AI model can run on a personal computer, while stock returns around its milestones leave the case for an AI data-center spending bubble unresolved.

Euphemios10 min read

Executive summary

Colibrì, an open-source engine first published on GitHub on July 1, 2026, runs large open AI models on hardware people already own. If software like it reduced the need for data-center capacity, AI infrastructure stocks might weaken when it drew attention. We tested that around 18 dated milestones. A basket of 17 chip, data-center, power and cloud stocks did not fall against the broad market the day after; over five sessions it lagged slightly more often than usual (10 of 18 windows), and power companies were the weakest group. The pattern is weak, the sample is small and other news overlapped, so the evidence neither confirms nor rules out a spending bubble.

Key findings

  • The study retained 18 of 35 candidate dates. The AI-stock basket averaged +0.07 percentage points against the broad U.S. market over the first trading session, −0.19 over five, and −0.56 from five sessions before to five after the reference close. (Euphemios Colibrì event study, through November 19, 2026.)
  • The basket lagged the market in 10 of 18 five-session event windows, versus 96 of 218 five-session windows across the full price sample. The median event ranked at the 40.8th percentile of that baseline. (Euphemios Colibrì event study, through November 19, 2026.)
  • Power stocks averaged −1.71 percentage points versus SPY over five sessions, while the large cloud operators averaged +0.68. (Euphemios Colibrì event study, through November 19, 2026.)
  • After the July 22 v1.1.1 release, the basket fell 9.45% over five sessions and lagged SPY by 7.05 percentage points. The next retained release, July 28 v1.2.0, was followed by +5.39 percentage points of five-session excess. (Euphemios Colibrì event study, through November 19, 2026.)
  • GitHub’s aggregate history counted 52,312 Colibrì stars dated through November 19. A single July 9 posting date accounted for 937 of 939 reported Hacker News points. (GitHub aggregate star history and Algolia Hacker News search, through November 19, 2026.)
  • On one Mac, Colibrì generated about 1.7 tokens per second with a 35-billion-parameter model. Ollama generated 44.8 tokens per second with an 8-billion-parameter model on the same computer. (Euphemios Mac throughput measurement, October 21, 2026.)

What Colibrì changes about the question

Colibrì is an open-source engine for running AI models on hardware people already own. Some of its supported models use a design called a mixture of experts. Each piece of generated text activates only some parts of the model. Colibrì keeps frequently used parts in faster working memory and loads others from disk as needed. It can run on a computer's main processor alone. The project uses the Apache 2.0 license. Colibrì README · Apache 2.0 license

The October 21 run shows what local use can look like. A token is a chunk of text, often part of a word. A parameter is an internal value learned during training, and the parameter count gives a rough sense of model size. The tested larger model ran locally, at a pace that may suit a background task better than a conversation. The two runs used different models and software. Their speeds show the range observed on one computer and leave the effect of the engine itself unresolved.

Text generation speed on one Mac

Model and software
Colibrì · 35B model
Ollama · 8B model
Tokens per second
Use the arrow keys to explore chart values.
Text generation speed on one Mac (tokens per second)
Model and softwareOutput rate
Colibrì · 35B model1.7
Ollama · 8B model44.8

Observed on October 21, 2026. Model sizes and software differ.

Source · Euphemios Mac throughput measurement · Euphemios throughput test on one Mac, October 21, 2026 · As of

Stock prices moved in both directions

The study tracked what happened to an equal-weight basket of AI-linked stocks around public project milestones. It compared the basket and its subgroups with SPY, a fund that tracks the S&P 500. An excess return is the group return minus SPY's return over the same dates, expressed in percentage points.

Average excess return around Colibrì milestones

  • First session
  • Next five sessions
  • Five before to five after
Stock group
AI basket
Chips and hardware
Data-center property companies
Power
Large cloud operators
SPY
Mean return minus SPY (percentage points)
Use the arrow keys to explore chart values.
Average excess return around Colibrì milestones (percentage points)
Stock groupFirst sessionNext five sessionsFive before to five after
AI basket0.0747-0.1908-0.5592
Chips and hardware0.0346-0.2304-0.6167
Data-center property companies-0.24520.0543-0.4712
Power0.288-1.7116-2.6531
Large cloud operators0.13080.67910.7424
SPY000

Each series averages 18 events. SPY is the broad-market comparison and has zero excess by definition.

Source · Euphemios Colibrì event study, daily closes through November 19, 2026 · As of

GroupFirst sessionNext five sessionsFive before to five after
AI basket+0.07 pp−0.19 pp−0.56 pp
Chips and hardware+0.03 pp−0.23 pp−0.62 pp
Data-center property companies−0.25 pp+0.05 pp−0.47 pp
Power+0.29 pp−1.71 pp−2.65 pp
Large cloud operators+0.13 pp+0.68 pp+0.74 pp
SPY0.00 pp0.00 pp0.00 pp

Source: Euphemios Colibrì event study, average excess return across 18 events, price sample through November 19, 2026.

The broader sample helps put the figures in context. Across all eligible five-session windows, the AI basket averaged +0.31 percentage points versus SPY. Power averaged −0.67 and cloud operators −0.20. Around the events, power lagged its own typical result and cloud operators beat theirs. Data-center property companies had a small positive event mean over five sessions, +0.05, despite a −0.98 median and 11 of 18 windows below SPY. A few positive returns can lift the average even when most windows are negative. (Euphemios Colibrì event study, through November 19, 2026.)

Five-session excess return at each retained event

Event type
Repository creation
Event 2026-07-01 · Sessions 2026-06-30 to 2026-07-08 · Return −2.85% · Excess −2.66 pp
Release
Event 2026-07-19 · Sessions 2026-07-17 to 2026-07-24 · Return 3.49% · Excess +4.07 ppEvent 2026-07-22 · Sessions 2026-07-22 to 2026-07-29 · Return −9.45% · Excess −7.05 ppEvent 2026-07-28 · Sessions 2026-07-28 to 2026-08-04 · Return 9.51% · Excess +5.39 ppEvent 2026-08-01 · Sessions 2026-07-31 to 2026-08-07 · Return 4.65% · Excess +1.14 ppEvent 2026-08-13 · Sessions 2026-08-13 to 2026-08-20 · Return −3.25% · Excess −1.29 ppEvent 2026-08-20 · Sessions 2026-08-19 to 2026-08-26 · Return 1.04% · Excess +1.43 ppEvent 2026-08-24 · Sessions 2026-08-24 to 2026-08-31 · Return 1.5% · Excess +1.03 ppEvent 2026-08-28 · Sessions 2026-08-28 to 2026-09-04 · Return 3.09% · Excess +2.98 ppEvent 2026-09-06 · Sessions 2026-09-04 to 2026-09-14 · Return −3.72% · Excess −2.51 ppEvent 2026-09-20 · Sessions 2026-09-18 to 2026-09-25 · Return 2.56% · Excess +1.29 ppEvent 2026-09-24 · Sessions 2026-09-24 to 2026-10-01 · Return −0.44% · Excess −0.02 ppEvent 2026-10-06 · Sessions 2026-10-06 to 2026-10-13 · Return −2.46% · Excess −2.32 ppEvent 2026-10-31 · Sessions 2026-10-30 to 2026-11-06 · Return −6.94% · Excess −4.23 pp
GitHub star surge
Event 2026-07-14 · Sessions 2026-07-13 to 2026-07-20 · Return −3.68% · Excess −2.73 ppEvent 2026-09-13 · Sessions 2026-09-11 to 2026-09-18 · Return −0.71% · Excess −0.61 pp
Hacker News peak
Event 2026-07-09 · Sessions 2026-07-08 to 2026-07-15 · Return 0.61% · Excess −0.65 ppEvent 2026-08-11 · Sessions 2026-08-10 to 2026-08-17 · Return 3.27% · Excess +3.31 pp
Basket return minus SPY (percentage points)
Five-session excess return at each retained event
EventEvent typeDateStartEndExcess (pp)Raw return (%)
Repository createdRepository creation2026-07-012026-06-302026-07-08-2.6638-2.8472
v1.0.0Release2026-07-192026-07-172026-07-244.07483.4882
v1.1.1Release2026-07-222026-07-222026-07-29-7.053-9.4546
v1.2.0Release2026-07-282026-07-282026-08-045.39419.5069
v1.4.0Release2026-08-012026-07-312026-08-071.14124.6524
v1.6.1Release2026-08-132026-08-132026-08-20-1.2876-3.2519
v1.7.0Release2026-08-202026-08-192026-08-261.4311.0435
v1.8.0Release2026-08-242026-08-242026-08-311.03031.4992
v1.9.0Release2026-08-282026-08-282026-09-042.97653.0857
v1.10.2Release2026-09-062026-09-042026-09-14-2.5114-3.7202
v1.12.0Release2026-09-202026-09-182026-09-251.29362.5618
v1.12.1Release2026-09-242026-09-242026-10-01-0.0242-0.4401
v2.0.0Release2026-10-062026-10-062026-10-13-2.3199-2.4586
v2.1.0Release2026-10-312026-10-302026-11-06-4.2349-6.9383
2026-07-14GitHub star surge2026-07-142026-07-132026-07-20-2.7301-3.6752
2026-09-13GitHub star surge2026-09-132026-09-112026-09-18-0.6136-0.7063
2026-07-09Hacker News peak2026-07-092026-07-082026-07-15-0.65080.6116
2026-08-11Hacker News peak2026-08-112026-08-102026-08-173.31353.2669

Hover or focus on a marker for event details.

One dot per retained event. The horizontal distance from zero shows excess versus SPY. Hover for date, window and raw basket return.

Source · Euphemios Colibrì event study, daily closes through November 19, 2026 · As of

The July 22 and July 28 releases frame the difference between an event and an explanation. From the July 22 reference close to July 29, the basket fell 9.45% and SPY fell 2.40%. The following release's window uses some of the same price history and points in the other direction. The dot chart shows how much the result varied across markers, including releases and spikes in public attention. (Euphemios Colibrì event study, through November 19, 2026.) July 22 release · July 28 release

Attention grew in bursts

GitHub stars and Hacker News points measure different things. Stars signal that people bookmarked or endorsed the project. Hacker News points are scores displayed on matched items at collection, grouped here by the day each item was posted. The lines use separate scales and time intervals. Release markers show when the project published another version.

GitHub stars by week and Colibrì release dates

  • GitHub stars
  • Published releases
Stars in week
2026-07-192026-07-222026-07-222026-07-282026-07-292026-08-012026-08-052026-08-122026-08-132026-08-142026-08-202026-08-242026-08-282026-08-312026-08-312026-09-062026-09-132026-09-202026-09-242026-10-062026-10-31
Use the arrow keys to explore chart values.
Week starting Monday
GitHub stars by week and Colibrì release dates (stars)
Week starting MondayGitHub stars
2026-06-2923
2026-07-067054
2026-07-139096
2026-07-203176
2026-07-272528
2026-08-031290
2026-08-101678
2026-08-17777
2026-08-24522
2026-08-31489
2026-09-073438
2026-09-146187
2026-09-211405
2026-09-281654
2026-10-051756
2026-10-121494
2026-10-191197
2026-10-262100
2026-11-023221
2026-11-092338
2026-11-16889
Published releases
DateEvent
2026-07-19v1.0.0
2026-07-22v1.1.0
2026-07-22v1.1.1
2026-07-28v1.2.0
2026-07-29v1.3.0
2026-08-01v1.4.0
2026-08-05v1.5.0
2026-08-12v1.6.0
2026-08-13v1.6.1
2026-08-14v1.6.2
2026-08-20v1.7.0
2026-08-24v1.8.0
2026-08-28v1.9.0
2026-08-31v1.10.0
2026-08-31v1.10.1
2026-09-06v1.10.2
2026-09-13v1.11.0
2026-09-20v1.12.0
2026-09-24v1.12.1
2026-10-06v2.0.0
2026-10-31v2.1.0

Hover or focus on a marker for event details.

Stars still present at retrieval, summed by source dates. The first and last weeks are partial. Markers use release dates.

Source · GitHub Colibrì activity · GitHub star history for JustVugg/colibri, retrieved November 20, 2026 · As of

Hacker News points by posting day and Colibrì release dates

  • Hacker News points
  • Published releases
Reported points
2026-07-192026-07-222026-07-222026-07-282026-07-292026-08-012026-08-052026-08-122026-08-132026-08-142026-08-202026-08-242026-08-282026-08-312026-08-312026-09-062026-09-132026-09-202026-09-242026-10-062026-10-31
Use the arrow keys to explore chart values.
UTC posting date
Hacker News points by posting day and Colibrì release dates (points)
UTC posting dateHacker News points
2026-07-010
2026-07-020
2026-07-030
2026-07-040
2026-07-050
2026-07-060
2026-07-070
2026-07-080
2026-07-09937
2026-07-100
2026-07-110
2026-07-120
2026-07-130
2026-07-140
2026-07-15Missing
2026-07-160
2026-07-170
2026-07-180
2026-07-190
2026-07-200
2026-07-21Missing
2026-07-220
2026-07-230
2026-07-240
2026-07-250
2026-07-26Missing
2026-07-27Missing
2026-07-280
2026-07-29Missing
2026-07-300
2026-07-310
2026-08-010
2026-08-020
2026-08-030
2026-08-040
2026-08-050
2026-08-060
2026-08-070
2026-08-080
2026-08-090
2026-08-100
2026-08-111
2026-08-120
2026-08-130
2026-08-14Missing
2026-08-150
2026-08-160
2026-08-170
2026-08-180
2026-08-19Missing
2026-08-200
2026-08-210
2026-08-220
2026-08-230
2026-08-240
2026-08-250
2026-08-26Missing
2026-08-270
2026-08-280
2026-08-290
2026-08-300
2026-08-310
2026-09-010
2026-09-02Missing
2026-09-030
2026-09-040
2026-09-050
2026-09-060
2026-09-070
2026-09-080
2026-09-090
2026-09-100
2026-09-110
2026-09-120
2026-09-130
2026-09-140
2026-09-150
2026-09-160
2026-09-170
2026-09-180
2026-09-190
2026-09-200
2026-09-211
2026-09-220
2026-09-230
2026-09-240
2026-09-250
2026-09-260
2026-09-270
2026-09-280
2026-09-290
2026-09-300
2026-10-010
2026-10-020
2026-10-030
2026-10-040
2026-10-050
2026-10-060
2026-10-070
2026-10-080
2026-10-090
2026-10-100
2026-10-110
2026-10-120
2026-10-130
2026-10-140
2026-10-150
2026-10-160
2026-10-170
2026-10-180
2026-10-190
2026-10-200
2026-10-210
2026-10-220
2026-10-230
2026-10-240
2026-10-250
2026-10-260
2026-10-270
2026-10-280
2026-10-290
2026-10-300
2026-10-310
2026-11-010
2026-11-020
2026-11-030
2026-11-040
2026-11-050
2026-11-060
2026-11-070
2026-11-080
2026-11-090
2026-11-100
2026-11-110
2026-11-120
2026-11-130
2026-11-140
2026-11-150
2026-11-160
2026-11-170
2026-11-180
2026-11-190
Published releases
DateEvent
2026-07-19v1.0.0
2026-07-22v1.1.0
2026-07-22v1.1.1
2026-07-28v1.2.0
2026-07-29v1.3.0
2026-08-01v1.4.0
2026-08-05v1.5.0
2026-08-12v1.6.0
2026-08-13v1.6.1
2026-08-14v1.6.2
2026-08-20v1.7.0
2026-08-24v1.8.0
2026-08-28v1.9.0
2026-08-31v1.10.0
2026-08-31v1.10.1
2026-09-06v1.10.2
2026-09-13v1.11.0
2026-09-20v1.12.0
2026-09-24v1.12.1
2026-10-06v2.0.0
2026-10-31v2.1.0

Hover or focus on a marker for event details.

Where matched items had no reported scores, the line has a gap. Markers use GitHub release dates.

Source · Hacker News via Algolia · Hacker News search via Algolia, retrieved November 20, 2026 · As of

The largest complete GitHub day was July 13, with 3,152 stars. The largest Hacker News date was July 9, with 937 reported points. The Hacker News search matched 3 stories and 15 comments. Comment scores were absent, which leaves gaps on dates with comments and no scored item. The study's timing rule kept the July 9 Hacker News date and excluded the July 13 star marker. (GitHub and Algolia Hacker News, through November 19, 2026.) GitHub star history · July 9 Hacker News discussion

The case for and against a spending bubble

For. Local generation gives some users another way to run AI. If enough workloads move to existing computers, central providers could see weaker demand than their investment plans assume. The comparatively weak five-session results for the basket and power stocks are compatible with that possibility. The study did not observe workload migration or test whether investors were reacting to it.

Against. The larger local model produced text slowly in the Mac trial. Training a model and serving many people or requests at once impose different demands from a single local run. Better local tools can also expand the number of people and tasks using AI. Cloud operators' positive average excess across all three event windows cuts against a broad market reaction to local substitution. Asset values, facility use, and returns on new spending remain open questions.

What would change the view. Tests using the same model, numerical precision, prompts, and output quality could show which workloads run locally at acceptable speed, cost, and energy use. Evidence that paid workloads are moving to local computers while data-center utilization, realized prices, or returns falter would strengthen the case for excess capacity. Sustained facility use and returns as capacity expands would weaken it.

Methodology and definitions

  • Events. The manifest contains one repository creation, 21 published releases, and attention candidates from GitHub and Hacker News. GitHub candidates used the five largest complete-day star totals plus all days above the 95th percentile. Hacker News candidates combined up to three dates with the most matched items and up to three with the most reported points, under the thresholds documented in the study. A later marker was removed if its reference close was within two SPY sessions of the most recent retained marker. The final set contained one creation, 13 releases, two GitHub star surges, and two Hacker News peaks. (Colibrì event manifest, through November 19, 2026.)
  • Timing and prices. The reference is the last SPY close strictly before a timed event. Attention dates without a timestamp start at midnight UTC in the main run. The three windows go from the reference close to the next close, from the reference close to the fifth subsequent close, and from five closes before the reference to five after. Daily closing prices were adjusted for dividends. The price sample contains 223 trading sessions from January 2 through November 19, 2026. (Euphemios Colibrì event study, daily closes through November 19, 2026.)
  • Baskets. The 17 stocks each receive the same initial weight within the overall basket. Each subgroup uses equal weights among its own members. The study grouped NVDA, AVGO, AMD, MU, MRVL, ANET, and SMCI as chips and hardware, EQIX and DLR as data-center property companies, VST, CEG, and NRG as power, and MSFT, GOOGL, AMZN, META, and ORCL as the large cloud operators. The broad-market baseline includes every eligible price window, including event windows. The 40.8th percentile means the median event ranked above roughly two-fifths of that baseline. (Euphemios Colibrì event study, daily closes through November 19, 2026.)
  • Attention. Stars are those still present in GitHub's aggregate snapshot. The chart sums GitHub's source-date counts into Monday-start weeks. The first and last weeks are partial. The Hacker News series groups identified stories and comments by UTC posting date and retains scores that the API supplied. In both charts the release markers come from the full manifest. (GitHub and Algolia Hacker News, chart cutoff November 19, 2026.)

Limits

Daily closes hide intraday reactions. Earnings, company news, and economic releases can land inside the same windows. Subtracting a broad index leaves industry and company effects in the comparison. The study describes correlation. Its 18 events and overlapping windows leave too little independent evidence to attribute price moves to Colibrì. When the study moved attention markers to the end of their source days, 17 events remained and average excess became −0.03, −0.64, and −0.60 percentage points across the three windows. The first-session mean changed sign. (Euphemios Colibrì event study, day-end sensitivity through November 19, 2026.)

The Mac observation covered one machine and two distinct model workloads. The study contains no measures of data-center utilization, costs per unit of output, demand for model training, or the financial returns from new capacity.

GitHub's timestamped star request, GET https://api.github.com/repos/JustVugg/colibri/stargazers?per_page=100 with Accept: application/vnd.github.star+json, returned HTTP 404 in the original study. The aggregate-history request succeeded, although GitHub's date buckets may not follow UTC day boundaries. (GitHub API, retrieved November 20, 2026.)

GDELT's news volume and tone were unavailable: its API returned rate-limit errors (HTTP 429) or empty responses for every query. A complete log of the failed requests is available on request.

FAQ

What would make data-center spending a bubble?

The concern is that companies are committing more money to new capacity than future paid demand can support. Testing it requires evidence on how much capacity gets used, what customers pay, and whether the resulting revenue covers the investment.

What is local inference?

Inference is the work a trained model does when it generates an answer or performs another task. Local inference runs that work on hardware the user controls, such as a laptop or office computer.

Does Colibrì replace cloud computing?

Colibrì can run some supported models on a user's own machine. Cloud facilities handle other demands, including training and serving many users simultaneously. The Mac trial tested the feasibility of local generation for two particular workloads.

What does excess return mean?

It measures how much more or less the selected stocks returned than SPY over the same period. A negative excess means the group lagged that broad-market comparison.

Why track GitHub stars and Hacker News points?

They help identify days when the project drew public attention. They do not tell us how many people deployed the software or why anyone traded an infrastructure stock.

Does the study settle the bubble question?

The evidence shows that local operation is possible and that relative stock returns around project milestones were mixed. Whether current spending outruns future demand is a question for operating and financial data.

Company and network names are trademarks of their owners. They are mentioned as facts about public filings and data, not as a relationship or an endorsement.

Sources

  • Euphemios Colibrì event study
    Run Euphemios Colibrì event study, daily closes through November 19, 2026
    Data as of
  • GitHub Colibrì README and license
    Run Commit 0a5504a1dc64f86fcd0fe7db298b0448c730e7e6, retrieved 2026-11-20
    Data as of
  • GitHub Colibrì activity
    Run GitHub star history for JustVugg/colibri, retrieved November 20, 2026
    Data as of
  • Hacker News via Algolia
    Run Hacker News search via Algolia, retrieved November 20, 2026
    Data as of
  • Euphemios Mac throughput measurement
    Run Euphemios throughput test on one Mac, October 21, 2026
    Data as of
  • GDELT DOC 2.0
    Run GDELT DOC 2.0 requests, November 19-20, 2026 (no usable response)
    Data as of