AI Bubble Index: what would make it burst
Is the AI stock market in a bubble?
Heated, but not at bubble-peak levels, by our measure. As of November 17, 2026, the AI Bubble Index reads 56 out of 100, in the Heated band (51 to 70). It has never reached Overheated (71 to 85) or Euphoria (86 to 100) since the series began on August 20, 2026.
The most extreme reading is how the buildout is being paid for. AI chip makers led by NVIDIA and Broadcom have put up loans, guarantees, and lease backstops for their own customers worth 51.5% of the group's revenue, most of it not yet drawn. Lucent's figure in 2000 was 24%. The calmest component is momentum, at 13 out of 100.
A dashboard of market temperature from 0 to 100, computed from 9 components
The AI Bubble Index reads 56 out of 100, in the Heated band, as of November 17, 2026 at 2:35 p.m. ET.
Multiple indicators elevated together. The index measures speculative heat, not the odds of a crash, so it can fall during a selloff even while credit stress is rising. The score recalculates four times each weekday. 5 of its 9 components move with the market; the other 4 move only when a utility, regulator or company files something new. Every raw value, score and anchor sits below.
Why 56?
Each component's score times its weight. The contributions sum to the composite, so the longest bar is what is actually driving the reading, not just what is elevated. The previous reading was 57, down 1. Recalculates at 9:35am, 12:35, 3:35 and 6:35pm ET each weekday.
| Component | Score | Weight | Contribution |
|---|---|---|---|
| Commitment gap | 62 | 18% | 11.16 |
| Circular financing | 100 | 10% | 10.00 |
| Insider selling | 77 | 10% | 7.70 |
| Contracted cover | 64 | 11% | 7.04 |
| Cost of money | 60 | 11.5% | 6.90 |
| Counterparty concentration | 63 | 10% | 6.30 |
| Valuation | 29 | 13.5% | 3.92 |
| Momentum | 13 | 10% | 1.30 |
| Complacency | 21 | 6% | 1.26 |
| Composite | sum of contributions | 55.57 → rounds to 56 | |
History
The series began August 20, 2026 and holds 67 daily readings; the chart shows the last 30. A ringed point marks a methodology change recorded in the changelog, so a component addition is never mistaken for a market move. Nothing before the first point is reconstructed.
What goes into the number
| Component | Reading | Score | How it maps |
|---|---|---|---|
| Commitment gapquarterlyIs the demand behind the buildout real?Weight 18% | 65.41% of announced demand unbacked median of 7 independent screens run by grid operators, utilities and regulators | 62 | 50 percent maps to 40 because roughly half of announced data center projects have historically failed to arrive on schedule, so that much attrition is the baseline rather than a warning. 95 percent maps to 100. Anchors: 0 to 0, 50 to 40, 75 to 75, 95 to 100 |
| ValuationdailyHow much is priced in?Weight 13.5% | 17.4median forward P/E of the AI basket median of 10 forward P/E readings across 10 symbols | 29 | 16 is the long-run S&P 500 median forward multiple and maps to 25. 60 maps to 100, in the region of dot-com peak multiples. Anchors: 16 to 25, 25 to 50, 40 to 80, 60 to 100 |
| Insider sellingdailyAre the people inside selling more than usual?Weight 10% | 2.4762discretionary selling against its own trailing baseline $972M of discretionary selling across 7 of 25 AI infrastructure names in the last 90 days, against a baseline of $392M from 4 earlier windows. NVDA is 97 percent of the current window. | 77 | The last 90 days of discretionary selling against the median of the prior four 90-day windows, so the basket is judged against itself rather than an outside benchmark. A reading of 1.0 is its own normal rate and maps to 40. Rule 10b5-1 plans and tax withholding are excluded. Anchors: 0.4 to 10, 1 to 40, 2 to 70, 4 to 100 |
| Counterparty concentrationquarterlyHow few counterparties is this resting on?Weight 10% | 44top direct customers as % of revenue 44% of NVIDIA's revenue came from three direct customers in the first half of fiscal 2027 (NVIDIA 10-Q). Vendor financing is now scored separately under Circular financing. | 63 | 15 maps to 20 because that is what Lucent's largest customer reached before its collapse, per its fiscal 2000 annual report. Vendor financing moved to its own component, Circular financing, on October 10, 2026. Anchors: 15 to 20, 30 to 45, 54 to 75, 70 to 100 |
| MomentumdailyHow far above trend is the price?Weight 10% | 0.915basket price divided by its 200-day average median price to 200-day average across 10 symbols | 13 | A basket trading at its 200-day average maps to 30. Twenty percent above trend maps to 75. Anchors: 0.9 to 10, 1 to 30, 1.1 to 50, 1.2 to 75, 1.4 to 100 |
| ComplacencydailyIs any of this priced as a risk?Weight 6% | 25.58VIX, inverted CBOE Volatility Index, last regular-session print | 21 | Inverted on purpose: a low VIX means the market is pricing no stress, which raises the reading rather than lowering it. 20 is near the long-run average and maps to 35. 45 maps to 0, the level reached at the dot-com bust peak. Anchors: 10 to 100, 12 to 85, 15 to 65, 20 to 35, 30 to 10, 45 to 0 |
| Contracted coverquarterlyHow much disclosed contract coverage supports the valuation?Weight 11% | 23.37contracted revenue as % of market value $3372B of disclosed GAAP obligations against $14428B of market value across 21 companies | 64 | Inverted: cover falls as the market pays more for demand nobody has signed for, so a low reading raises the score. 100 percent, where signed contracts equal the whole market value, maps to 0; 10 percent, where nine tenths of the price rests on uncommitted revenue, maps to 100. Only 10-Q and 10-K figures count. Anchors: 10 to 100, 25 to 60, 50 to 25, 100 to 0 |
| Cost of moneydailyIs money getting expensive?Weight 11.5% | 60.33average of three sub-scores: the 12-month change in the 10-year Treasury yield, private credit funds' price to NAV, and AI borrowers' bond spread premium10-year Treasury yield, 12-month change: +98 bp (5.12% on 2026-11-16, against 4.14% on 2025-11-14) → scores 50 through its own anchors: -100 to 0, 0 to 25, 100 to 50, 209 to 85, 277 to 100Private credit funds, median price to NAV: 0.76x (OTF 0.56, TCPC 0.57, OBDC 0.75, MFIC 0.78, BXSL 0.92, ARCC 0.98) → scores 69 through its own anchors: 0.5 to 100, 0.7 to 80, 0.85 to 55, 1 to 30, 1.15 to 10AI borrowers' bond spread premium: +229 bp, the 5-day median premium (latest reading 2026-11-16: median AI issuer 457 bp vs median high-yield bond 236 bp; issuers: CoreWeave 1039, Galaxy Helios 781, Edged 709, SoftBank 522, Core Scientific 457, Applied Digital 360, Oracle 312, TeraWulf 264, Cipher 236) → scores 62 through its own anchors: -100 to 0, 0 to 20, 100 to 40, 250 to 65, 500 to 90, 800 to 100 Prices are the latest close; each fund's NAV is its most recent quarterly report, so the discount can move a few points when new NAVs are filed. | 60 | The raw reading is already a 0-100 score, the average of the three sub-scores, so it passes through unchanged. The rate half is calibrated to history: the 10-year yield's 12-month rise peaked at 209 basis points on January 18, 2000, two months before the Nasdaq top, which maps to 85, and at 277 basis points in November 2022, which maps to 100. A flat year maps to 25. The private credit reading is anchored arithmetically: a fund trading at exactly its net asset value maps to 30, and each 15 percent discount adds about 25. The AI borrowers' reading is how much more AI builders pay on their bonds than the typical high-yield issuer: zero premium maps to 20, 500 basis points maps to 90. Anchors: 0 to 0, 100 to 100 |
| Circular financingquarterlyAre suppliers paying for their own sales?Weight 10% | 51.55chip makers' loans, guarantees and lease backstops for their own customers, as % of revenueLoans, guarantees and lease backstops (Lucent's basis, scored): 51.6% ($251.0B of $486.9B revenue) → scores 100 through the same anchorsAdding equity stakes and capacity purchases (not scored): 71.9% ($349.9B) → scores 100 through the same anchorsExplicitly funded per filings (a minimum): 0.4% ($2.0B) → scores 2 through the same anchors 7 disclosed loans, guarantees and lease backstops: NVIDIA $128.5B, Broadcom $114.0B, AMD $8.5B, against $486.9B of trailing 12-month revenue across NVIDIA, Broadcom, AMD, Marvell, Qualcomm. Lucent's figure on the same basis was 24.0 percent. The filings do not show whether some commitments cover the same leases, so overlaps cannot be ruled out. Most of the total is undrawn or contingent. | 100 | Calibrated to the dot-com vendors' own 10-K disclosures: Cisco's customer financing was 11.5 percent of revenue in fiscal 2001 and maps to 50, Nortel's 17.1 percent at the end of 2000 maps to 70, and Lucent's 24.0 percent in fiscal 2000, the peak of the cycle, maps to 85. Twice Lucent's level maps to 100. Only loans, loan guarantees and lease backstops are scored, counted at their maximum, drawn or not, because those are the categories in Lucent's figure; equity stakes and capacity-purchase agreements are published beside the score but not in it. Anchors: 0 to 0, 11.5 to 50, 17.1 to 70, 24 to 85, 48 to 100 |
The five things that ended the last one
Each row pairs a mechanism that actually did the damage in 2000 with today's equivalent and its current reading. All five are now measured. The last three readings come straight from the live index components below, so they update every weekday.
| What ended the dot-com bubble | Today | Reading |
|---|---|---|
| Vendor financing reversedLucent committed $8.1 billion to help customers buy its own equipment, 24 percent of its $33.8 billion of fiscal 2000 revenue, much of it to carriers the banks had already refused. Nortel had about $4.1 billion of unfunded customer financing committed at the end of 2000 and offered terms reaching 130 percent of purchase price. Lucent's revenue fell from about $30 billion to $12 billion and it lost $16.1 billion in 2001 alone. | Circular financingAcross NVIDIA, Broadcom, AMD, Marvell, and Qualcomm, disclosed loans, guarantees, and lease backstops for their own customers come to about 42 percent of trailing 12-month revenue, against Lucent's 24 percent on the same basis; adding equity stakes and capacity purchases takes it to about 62 percent. The deal-by-deal ledger is below, and filings show very little of it funded yet. The largest single piece: Nvidia agreed on Aug. 17, 2026 to guarantee up to $105 billion behind an OpenAI data center in Pike County, Ohio. Against the $303.0 billion Nvidia earned in the twelve months to July 26, 2026, that is 35 percent, against Lucent's 24 percent. The direction is the interesting part: the commitment was discussed at up to $250 billion on July 27 and cut below $120 billion after investors objected to the exposure, a 58 percent reduction in three weeks. Four days later, on Aug. 21, Nvidia extended the same pattern into software: a $6 billion license of AI coding startup Poolside's model-development tools plus a separate roughly $1 billion equity stake at a $12 billion valuation, more than triple where Poolside was priced in 2024 against essentially no shipped revenue. That deal is a license and a minority stake, not a compute-purchase guarantee, so it is kept out of the reading rather than added to it, but it is the same supplier-funds-its-own-customer logic showing up a layer higher in the stack. | 51.6% of revenueof chip makers' revenue in loans and guarantees for their own customers |
| The demand was never realCapacity was built against orders from companies that had no revenue and, in many cases, no customers. The orders were real on paper. The ability to pay for them was not. | The commitment gapSeven screens run by grid operators, utilities and regulators in four states each remove between 50 and 95 percent of the data center demand put to them. None of them coordinate and none uses the same method. Detail for every screen is below. | 65%of announced demand is unbacked |
| Nobody was pricing itThe VIX peaked at 45.08 during the entire dot-com collapse, a level since exceeded by a credit downgrade, a flash crash and a tariff announcement. Volatility markets did not see the biggest equity unwind in modern history while it was happening. | VolatilityThis is not a safety reading. It is a complacency reading. The market is pricing essentially no stress against a pipeline that is 65 percent air and a supplier guaranteeing 35 percent of its revenue. Note that VIX is a poor detector of slow deflation: it registers sudden credit and liquidity shocks, which is why 2008 and 2020 produced readings above 80 and the dot-com bust never did. | 25.58VIX, against a long-run average near 19 |
| Money got expensiveThe Federal Reserve raised the funds rate from 4.75 percent to 6.5 percent between June 1999 and May 2000. Speculative capital left first. | The cost of moneyThree readings, averaged. The first is how far the 10-year Treasury yield has risen over the past year. That rise peaked at 209 basis points on Jan. 18, 2000, two months before the Nasdaq topped, and at 277 basis points in November 2022. The second is what investors pay for the listed private credit funds run by Ares, Blue Owl, Blackstone, Apollo, and BlackRock, the five managers that capped withdrawals from their non-traded credit funds this year. A fund trading well below its net asset value is the market declining to believe the loan marks. Private credit lends directly to the buildout: funds managed by Blue Owl own 80 percent of the venture financing Meta's roughly $27 billion Hyperion data center in Louisiana, and Blue Owl's non-traded technology income fund drew withdrawal requests for 39 percent of its shares in the third quarter. The third is what AI builders pay to borrow: the bonds of CoreWeave, Oracle, Applied Digital, Cipher, TeraWulf, Core Scientific, Galaxy, and Edged, priced daily, measured against the typical high-yield bond. A widening gap means lenders are charging AI borrowers more than other risky companies. | +98 bp · 0.76x NAV · AI +229 bp10-year yield change, private credit funds' price to NAV, and AI borrowers' bond spread premium |
| Insiders sold firstLockup expirations released enormous insider supply into a market that had already stopped absorbing it. | Form 4 selling in AI infrastructureDiscretionary Form 4 selling across 25 AI infrastructure companies over the last 90 days, measured against the median of the four 90-day windows before it. Sales under Rule 10b5-1 plans and shares withheld for tax are excluded. | 2.48xdiscretionary selling against its own trailing baseline, where 1.0 is normal |
Circular financing, then and now
A supplier standing behind its customer's ability to pay for the supplier's own product. Commitments compared to commitments, because Lucent's widely quoted $8.1 billion was a ceiling: at the end of fiscal 2000 about $2.0 billion of it was outstanding, $1.3 billion of loans and $770 million of guaranteed debt, per Lucent's 10-K.
| Lucent, fiscal 2000 | Nvidia, 2026 | |
|---|---|---|
| Financing committed | $8.1 billion | $105 billion |
| Revenue | $33.8 billion | $303.0 billion |
| Commitment as share of revenue | 24% | 35% |
| Largest customers | top two, 23% | top three, 44% |
Three differences cut in Nvidia's favor. Lucent lent to carriers with no revenue that banks had refused, and booked the resulting purchases as revenue in the same period. Nvidia is guaranteeing a 20-year lease to one of the largest buyers of computing in the world. And Lucent expanded its lending as conditions worsened, where Nvidia cut its commitment by 58 percent because shareholders objected.
NVIDIA quarterly filings; Robert D. Atkinson, Who Lost Lucent, American Affairs (August 2020)
The money loop: who funds whom
Disclosed deals in which an AI supplier puts up money, credit, or a guarantee for a customer that then buys the supplier's own product. Amounts are maximum commitments as filed, so most are undrawn or contingent. Counting only loans, loan guarantees, and lease backstops, the same categories as Lucent's figure, chip makers have committed $251B, or 51.6 percent of their $487Bof trailing 12-month revenue, against Lucent's 24.0 percent at the dot-com peak. That is the scored figure. Adding equity stakes and capacity-purchase agreements brings it to $350B, or 71.9 percent, and including Amazon, Microsoft, and Alphabet, the loop totals $494B. The filings do not show whether some commitments cover the same leases, so overlaps cannot be ruled out. Filings explicitly identify about $2.0Bof the chip makers' total as funded, a disclosure-based minimum rather than a complete count of cash paid.
| Who funds whom | Type | Up to | Status | Source |
|---|---|---|---|---|
| NVIDIA → OpenAI (data-center leases at its Ohio site) | Guarantee | $105B | Signed | sec.gov |
| Amazon → OpenAICustomer commits to buy up to $138B | Equity stake | $50B | Funded | sec.gov |
| Broadcom → Anthropic (residual value backstop on the $63B lease vehicle) | Lease backstop | $43B | Signed | sec.gov |
| Broadcom → AnthropicCustomer commits to buy up to $125B | Loan | $42B | Signed | sec.govpartly press-reported |
| Alphabet (Google) → AnthropicCustomer commits to buy up to $121B | Equity stake | $40B | Partly funded | cnbc.compartly press-reported |
| NVIDIA → Unnamed AI cloud partners | Capacity backstop | $36B | Signed | sec.gov |
| NVIDIA → OpenAI | Equity stake | $30B | Signed | openai.com |
| Broadcom → Anthropic (through an Apollo-led lease vehicle) | Guarantee | $29B | Signed | sec.gov |
| NVIDIA → Unnamed third parties (leases to be reassigned) | Lease backstop | $20B | Signed | sec.govpartly press-reported |
| Amazon → Anthropic | Financing facility | $20B | Partly funded | sec.gov |
| Microsoft → OpenAICustomer commits to buy up to $250B | Equity stake | $13B | Partly funded | sec.gov |
| Amazon → AnthropicCustomer commits to buy up to $100B | Equity stake | $13B | Funded | sec.gov |
| NVIDIA → Anthropic | Equity stake | $10B | Signed | anthropic.compartly press-reported |
| AMD → Undisclosed commercial partners | Lease backstop | $8.5B | Signed | sec.gov |
| NVIDIA → CoreWeave | Capacity backstop | $6.3B | Signed | sec.gov |
| Microsoft → AnthropicCustomer commits to buy up to $45B | Equity stake | $5.0B | Signed | anthropic.com |
| AMD → AnthropicCustomer commits to buy up to $20B | Equity stake | $5.0B | Announced | ir.amd.compartly press-reported |
| NVIDIA → Unnamed AI cloud partners | Lease backstop | $3.5B | Signed | sec.gov |
| Alphabet (Google) → Anthropic | Equity stake | $3.0B | Funded | cnbc.compartly press-reported |
| NVIDIA → Meta | Equity stake | $3.0B | Announced | nvidianews.nvidia.com |
| NVIDIA → IREN | Equity stake | $2.1B | Signed | sec.gov |
| NVIDIA → CoreWeave | Equity stake | $2.0B | Funded | sec.gov |
| NVIDIA → Nscale | Equity stake | $2.0B | Partly funded | sec.gov |
| NVIDIA → SB Energy (landlord of OpenAI's Ohio site) | Equity stake | $1.5B | Announced | sec.gov |
| NVIDIA → Naver | Equity stake | $1.0B | Partly funded | bloomberg.compartly press-reported |
| CoreWeave → OpenAICustomer commits to buy up to $22B | Equity stake | $0.3B | Funded | sec.govpartly press-reported |
Not counted in the score: 8warrant grants that let customers buy supplier stock for little or nothing as they buy (AMD's two, to OpenAI and Meta, are worth about $199B at recent prices, all unvested), because their value moves with the share price; and 7 chip-backed loans from outside lenders such as Apollo and Blackstone, because the lender is not the supplier. Deals still in talks or reported without a filing are left out until they are confirmed.
Who carries the risk. The same structure can land on either side of the line. In June, an Apollo-led vehicle arranged $35 billion to buy Broadcom-designed racks and lease them to Anthropic, and Broadcom backstops up to $29 billion of what Anthropic owes under that lease. That backstop is why the deal counts here. SpaceX's planned $40 billion vehicle would buy Nvidia chips with money from bond investors led by Apollo, and none of the reports on the deal describe a guarantee or investment from Nvidia. Nvidia would be paid for its chips, and the lenders would hold the risk, priced in the open: earlier talks put the lowest-ranking debt at a yield in the mid-8 percent range before turmoil in AI credit reopened pricing, Bloomberg reported. Nvidia's $105 billion OpenAI deal runs the other way. Nvidia covers the shortfall if OpenAI stops paying for its Ohio data center leases, starting as each lease begins, in 2028 at the earliest. The vehicle is meant to keep the debt off SpaceX's balance sheet and protect its investment-grade ratings. But SpaceX's earlier sale-leasebacks with Valor Equity Partners did not count as sales, and $12.9 billion of that financing sat on its books as debt on Sept. 30.
Warrants and chip-backed debt (15)
- AMD → OpenAI: warrants, $99B (Signed) sec.gov
- AMD → Meta: warrants, $99B (Signed) sec.gov
- Cerebras → OpenAI: warrants, $5.6B (Signed) sec.gov
- Marvell → Google: warrants, $4.0B (Signed) sec.gov
- Qualcomm → Amazon: warrants, $0.4B (Signed) sec.gov
- HPE → Oracle: warrants, $0.3B (Signed) sec.gov
- Generac → Amazon: warrants, $0.0B (Signed) sec.gov
- Akamai → Anthropic: warrants, $0.0B (Signed) sec.gov
- Banks and a Blackstone-led tranche → Anthropic (Broadcom-designed chips): chip-backed debt, $63B (Announced) investors.broadcom.com
- Apollo-led investors (Goldman Sachs arranging) → SpaceX (through a vehicle managed by Apollo-backed Atlas SP): gpu finance lease, $40B (Signed) sec.gov
- Apollo-led SPV (Blackstone, banks) → Anthropic (lessee of Broadcom-designed racks): chip-backed debt, $35B (Partly funded) apollo.com
- Lenders led by MUFG and Morgan Stanley → CoreWeave: chip-backed debt, $14B (Funded) sec.gov
- Outside investors (SPV) → Amazon: chip-backed debt, $8.0B (Announced) ft.com
- Blue Owl, PIMCO and IG private-placement lenders → IREN: chip-backed debt, $6.4B (Partly funded) sec.gov
- Apollo, Goldman Sachs (SPV) → Oracle: chip-backed debt, n/d (Announced) wsj.com
Dot-com comparison, from the vendors' own annual reports: Lucent 24.0% (fiscal 2000); Nortel 17.1% (end of 2000); Cisco 11.5% (fiscal 2001). Ledger as of November 13, 2026.
What the market is pricing
The dot-com collapse never produced an extreme volatility reading. It peaked at 45.08 while the Nasdaq fell 78 percent, a level since exceeded by a credit downgrade, a flash crash and a tariff announcement. Whatever the VIX measures, it is not bubbles.
VIX methodology changed in 2003, so pre-2003 figures are back-calculated from the old VXO, which ran on the S&P 100 and tends to read higher. If anything the true comparable figure for 2002 was lower than 45.08.
Inside the commitment gap
The seven screens behind the 65 percent figure, ranked by how much each removes.
Pennsylvania, counted by the state inside Executive Order 2026-05.
| Who is screening | Asked for | Survived | Cut | Test applied |
|---|---|---|---|---|
| Pennsylvania DEPPennsylvania · Announced projects against fully permitted projectsAs of August 18, 2026 Disclosed by the state inside Executive Order 2026-05, which also bars DEP from issuing any data center permit before the project secures local approval. The intermediate stages are 58 projects engaged with the agency and 15 that have filed a permit application. Commonwealth of Pennsylvania, Executive Order 2026-05 (August 18, 2026) | 100in publicly sourced databases, at least | 5holding every first-phase permit | 95%20x more asked for than got through | Permitted |
| Georgia PowerGeorgia · Large load pipeline against signed commitmentsAs of March 31, 2026 Filed quarterly with the Georgia Public Service Commission, which makes this the most auditable series in the set. Of the committed 12,400 MW, 8,500 MW has physically broken ground and 10 projects are already online. PSC staff testimony separately records 33 projects totalling 11,332 MW removed from the pipeline since 2023. Georgia Power, Quarterly Large Load Economic Development Report, Q1 2026, Georgia PSC Docket 56002 | 76,200 MWtotal pipeline through the mid-2030s | 12,400 MWcommitted, across 31 customers | 84%6.1x more asked for than got through | Signed agreement |
| AEP OhioOhio · Inquiries before and after a minimum-take tariffAs of February 13, 2026 The closest thing in this data to a controlled experiment. PUCO approved a tariff on July 9, 2025 requiring data centers above 25 MW to pay for at least 85 percent of the energy they subscribe to whether or not they use it, over a 12-year term with an exit fee and proof of financial viability. Four in five megawatts of stated interest did not survive being asked to pay for itself. AEP Ohio (February 13, 2026) | 30,000 MWof inquiries, more than 50 customers across more than 90 sites | 5,642 MWafter binding financial commitments | 81%5.3x more asked for than got through | Money posted |
| PPL Electric UtilitiesPennsylvania · Data center pipeline against signed service agreementsAs of June 30, 2026 PPL reports a third figure below the two here: 6,500 MW actually under construction. Read across all three, the pipeline converts to signed agreements at about one in three, and to steel in the ground at about one in five. PPL Electric Utilities, second quarter 2026 disclosure | 31,800 MWof pipeline in Pennsylvania | 11,000 MWunder signed electric service agreements | 65%2.9x more asked for than got through | Signed agreement |
| ERCOTTexas · Contracted load against officer-letter attestations, 2030As of December 19, 2025 ERCOT splits its own forecast into load carrying signed interconnection agreements and load backed only by a transmission provider officer letter, which ERCOT defines as an attestation of confidence rather than a signed agreement. By 2030 the larger share of the forecast is the weaker category. ERCOT, Report on the Capacity, Demand and Reserves in the ERCOT Region, 2026 to 2030 (December 19, 2025) | 47,783 MWof new large load in the 2030 planning forecast | 19,560 MWwith a signed interconnection agreement | 59%2.4x more asked for than got through | Signed agreement |
| ERCOTTexas · Batch Zero interconnection screenAs of July 28, 2026 ERCOT labels this preliminary. Final determinations were due Aug. 7, 2026, with results communicated by Sept. 2. The Batch Zero deadline for posting financial security and site control is planned for June 2027, which is the point where this number gets tested again. ERCOT, Assessing the Grid, presented to the Texas Senate Committee on Business and Commerce (July 29, 2026) | 498 GWrepresented in the screen, 688 projects | 205 GWstudy-eligible, 326 projects | 59%2.4x more asked for than got through | Accepted for study |
| NERCTexas · Requested load against what NERC will countAs of January 1, 2026 NERC counts a data center project only once it has advanced from speculative and exploratory stages into development commitments. On Texas it states that new data center demand projections are reduced by almost 50 percent of their original requested load level. NERC, 2025 Long-Term Reliability Assessment | 100%of the originally requested load level | 50%after NERC screening, approximately | 50%2.0x more asked for than got through | Accepted for study |
What survives when someone has to pay
Real capacity is being built at very large scale. These two figures are the cleanest evidence of it anywhere in the data, and they belong here rather than in a footnote.
AEP told investors it had collected nearly $2 billion in cash or collateral for load commitments in ERCOT, covering all of the required credit support for the full 45 GW in its Batch Zero filing. Posted collateral is the cleanest available filter against one project being counted in several places at once. AEP, second quarter 2026 results and earnings call (July 30, 2026)
Concrete is the one test no press release can pass. Ten of these projects are already online and serving load. Georgia Power, Quarterly Large Load Economic Development Report, Q1 2026
Which stocks this actually touches
The ones already in the ground. A rule that slows every new entrant does not hurt a project that is already sited, permitted and locally supported. If anything it widens the moat. These are the names with real Pennsylvania exposure that mostly clear the new bar.
The most exposed name on this page. It reports 31,800 MW of Pennsylvania data center pipeline, 11,000 MW under signed service agreements and 6,500 MW under construction. Its growth story lives at the top of exactly the funnel this order narrows.
Sold its 960 MW Cumulus campus next to the Susquehanna nuclear plant to Amazon for $650 million, then expanded to a 1,920 MW power agreement running to 2042. An operating nuclear site with decades of local relationships clears a local-approval test easily.
Restarting Three Mile Island Unit 1 as the Crane Clean Energy Center. Same logic as Talen: an existing licensed site, not a greenfield fight.
Supplying seven gas turbines to the Homer City Energy Campus, the largest project in Pennsylvania at 4.5 gigawatts and about $10 billion. The turbine order is already placed, though environmental groups are crowdfunding a legal challenge to the site.
Announced an agreement in principle to supply the gas for Homer City.
Owns the Susquehanna campus outright and has confirmed talks to take capacity at Homer City, with no deal signed.
The ones that will move anyway. These names have no Pennsylvania exposure we can find. They tend to trade on data center policy headlines regardless of geography, which is a sentiment reaction rather than a change in their business. Worth separating the two before treating a move as information.
There is a live precedent for that split. When New York paused state data center permits on July 14, 2026, TeraWulf finished as the biggest decliner among the AI infrastructure names we track and Nebius was close behind, despite Nebius having nothing to do with New York.
The long-run base rate
Of all generation and storage capacity that entered a US interconnection queue between 2000 and 2020, 13 percent reached commercial operation. Attrition did not stop at the paperwork either: 41 percent of the capacity that signed an interconnection agreement between 2000 and 2022 had withdrawn by the end of 2025. The median project reaching operation in 2025 took 61 months from its request, up from 36 months in 2015.
This measures generation and storage, not data center load. Lawrence Berkeley National Laboratory states plainly that large loads sit in separate queues its report does not cover. It belongs here as the long-run base rate for how queue capacity turns into working plant, not as a data center figure. Lawrence Berkeley National Laboratory, Queued Up: 2026 Edition (June 2026)
This is deliberately not a count of blocked or cancelled projects. That figure is quoted everywhere and comes almost entirely from one organisation whose clients are the AI industry, with recent project data behind a paywall and approved projects still sitting in its delayed column. Announcement froth is also normal: companies have always announced more than they build, so counting the froth alone tells you nothing.
Two things cut against a bearish reading. Local defeats often do not stick, as when Saline Township, Michigan rejected a $16 billion campus, was sued two days later, settled for about $14 million in community benefits and saw ground broken within two months. And friction runs both ways, since West Virginia has preempted local ordinances that limit large data center projects.
What is coming next
- Data center debt, priced directly. The cost-of-money reading uses Treasury yields and listed private credit funds because no free daily series prices data center bonds or loans themselves. Spreads on data center asset-backed securities would be the sharper read.
Related coverage
- A thermal drone found 45 unpermitted gas generators running at a Microsoft AI site
- Nvidia cut its OpenAI backstop to $105 billion, and the Lucent parallel
- Four states now publish how much announced data center demand they expect
- Pennsylvania will not permit a data center until the town says yes
- ECB researchers say an AI stock market correction is likely
- AI infrastructure backlog leaderboard
Why this index exists
The AI Bubble Index scores nine readings from grid filings, SEC filings, bond prices, and market data, publishes every raw value behind the score, and keeps a dated daily history you can download.
No one can predict the moment a bubble will burst. Any website that says otherwise is trying to sell you something. A better approach is less ambitious and more practical: identify what brought the previous bubble to an end, map those forces onto today's market, and track them. The dot-com boom did not collapse simply because investors realized stocks were expensive. It collapsed when funding dried up and much of the apparent demand proved illusory. Both conditions can be measured today.
Methodology
The method is fixed until a scheduled review after the third-quarter 10-Q filings in late November 2026. Any change after that will be dated in the changelog below.
The AI Bubble Index is a weighted composite of 9 components. Each one is scored from 0 to 100 against published anchor points, then weighted and averaged. Every raw reading, weight and anchor on this dashboard is printed here, with the source it came from and the date that reading was taken.
How the score is calculated
- Measure. Each of the 9 components is read from a filing, a grid operator docket, or a market price.
- Score. Each reading is placed on a 0 to 100 scale by drawing a straight line between its published anchor points. Example: Circular financing anchors Lucent's 24.0 percent at 85, so today's 51.5 percent scores 100.
- Weight and average.Index = sum of (weight × score) ÷ sum of weights. A component without a reading drops out and the others are reweighted.
- Band. 0 to 30 Cold, 31 to 50 Building, 51 to 70 Heated, 71 to 85 Overheated, 86 to 100 Euphoria.
Today: 62×0.18 + 100×0.1 + 77×0.1 + 64×0.11 + 60×0.115 + 63×0.1 + 29×0.135 + 13×0.1 + 21×0.06 = 55.6, rounded to 56.
The published series begins August 20, 2026 and now holds 67 daily readings. No earlier reading has been reconstructed and none will be: two of the components rest on grid screens and forward estimates that were never published for earlier periods, so any back-history would be a guess wearing a date. The composite recomputes four times each weekday.
- Commitment gap. Is the demand behind the buildout real? Weight 18 percent, measured as % of announced demand unbacked, updates quarterly.Source: grid operator, utility and regulator interconnection screens · as of August 19, 2026
- Valuation. How much is priced in? Weight 13.5 percent, measured as median forward P/E of the AI basket, updates daily.Source: Yahoo Finance, AI large-cap basket · as of November 17, 2026Universe (10): NVDA, AMD, AVGO, MSFT, GOOGL, META, AMZN, ORCL, MU, TSM
- Insider selling. Are the people inside selling more than usual? Weight 10 percent, measured as discretionary selling against its own trailing baseline, updates daily.Source: SEC Form 4 filings, rolling 90-day window · as of November 17, 2026Universe (25): NVDA, AMD, AVGO, MU, MSFT, GOOGL, META, AMZN, ORCL, TSM, CRWV, NBIS, IREN, WULF, CIFR, CLSK, CORZ, VRT, GEV, ETN, PWR, POWL, TLN, CEG, PPL
- Counterparty concentration. How few counterparties is this resting on? Weight 10 percent, measured as top direct customers as % of revenue, updates quarterly.Source: NVIDIA 10-Q for the quarter ended Jul 26, 2026 · as of July 26, 2026
- Momentum. How far above trend is the price? Weight 10 percent, measured as basket price divided by its 200-day average, updates daily.Source: Yahoo Finance, AI large-cap basket · as of November 17, 2026Universe (10): NVDA, AMD, AVGO, MSFT, GOOGL, META, AMZN, ORCL, MU, TSM
- Complacency. Is any of this priced as a risk? Weight 6 percent, measured as VIX, inverted, updates daily.Source: CBOE VIX via Yahoo Finance · as of November 17, 2026
- Contracted cover. How much disclosed contract coverage supports the valuation? Weight 11 percent, measured as contracted revenue as % of market value, updates quarterly.Source: AIStockWire backlog tracker, audited figures only · as of November 15, 2026Universe (21): AMAT, AMZN, ANET, AVGO, CAT, EME, EQIX, ETN, GEV, GOOGL, IBM, IREN, MSFT, MU, NBIS, ORCL, SHAZ, SMCI, SNDK, VRT, WYFI
- Cost of money. Is money getting expensive? Weight 11.5 percent, measured as average of three sub-scores: the 12-month change in the 10-year Treasury yield, private credit funds' price to NAV, and AI borrowers' bond spread premium, updates daily.Source: U.S. Treasury daily par yield curve (to 2026-11-16); fund prices via Yahoo Finance against reported NAV; AI bond prices from State Street's JNK and SPBO daily holdings · as of November 17, 2026Universe (6): ARCC, BXSL, MFIC, OBDC, OTF, TCPC
- Circular financing. Are suppliers paying for their own sales? Weight 10 percent, measured as chip makers' loans, guarantees and lease backstops for their own customers, as % of revenue, updates quarterly.Source: Company 10-Q, 10-K and 8-K filings; deal ledger below · as of November 13, 2026Universe (5): NVIDIA, Broadcom, AMD, Marvell, Qualcomm
The composite recomputes on the cadence of its fastest inputs, and each component carries its own as-of date so a quarterly reading is never presented as though it were current. No proprietary forecasts are used: every number traces to a public filing, a grid operator docket or a market series, and forward consensus estimates and preliminary grid projections appear only where they are identified as such. There is no login, no paywall and no proprietary input, and because every reading, weight and anchor is printed on this page, anyone can recompute the score by hand and get the same number. Nothing about it is tunable by the reader. The index measures how much today looks like the conditions that ended the last bubble. It does not predict when this one ends, and it is not investment advice.
Notes on the anchors
- Insider selling was first tried as a raw sell-to-buy ratio and discarded. Open-market buying by insiders at companies this size is close to zero, so the ratio pinned at the top of the scale and could never move. Measuring the basket against its own trailing baseline fixed that. Its baseline covers about a year; windows before that are excluded because the Form 4 pollers were still ramping and record a collection gap rather than an absence of selling.
- Contracted cover is anchored arithmetically rather than historically. No comparable contracted-revenue series was published for the 2000 cycle, so unlike most of the others this component is not calibrated against that bust. Inventing a parallel would be worse than saying so.
- Cost of money combines three readings in different units, so each is scored through its own anchors and the component is their average. The rate reading is calibrated to the 12-month rise in the 10-year yield in 2000 and 2022. The private credit reading is anchored arithmetically, and its NAVs are the funds' latest quarterly reports. The AI borrowers' reading prices the bonds of CoreWeave, Oracle, Applied Digital, Cipher, TeraWulf, Core Scientific, Galaxy, Edged, and SoftBank from State Street's daily JNK and SPBO holdings, takes each issuer's median spread over Treasuries, and subtracts the median spread of every high-yield bond priced the same way, so it moves only when AI debt gets riskier than junk debt in general. It uses a five-day median because these bonds trade thinly.
- Circular financing scores only loans, loan guarantees, and lease backstops, counted at their maximum, drawn or not, because those are the categories and the basis of the Lucent, Nortel, and Cisco figures it is anchored to. Equity stakes and capacity-purchase agreements are published beside it but not scored, and so are warrants and chip-backed loans from outside lenders. Very little of the total is explicitly funded, which is why the funded minimum is shown next to it.
Does the score survive different assumptions?
The weights are judgment calls, so here is the same data under different ones. Published weights give 56; equal weights give 54; removing any single component and renormalising leaves it between 51 and 60. The reading is a property of the inputs, not of the weighting.
- Without Commitment gap: 54
- Without Circular financing: 51
- Without Insider selling: 53
- Without Contracted cover: 55
- Without Cost of money: 55
- Without Counterparty concentration: 55
- Without Valuation: 60
- Without Momentum: 60
- Without Complacency: 58
Changelog
- October 14, 2026. Cost of money's AI borrowers' bond spread reading now includes SoftBank Group's US dollar notes due 2030, 2032, and 2034, held in State Street's JNK. They came from SoftBank's record September 2026 junk bond sale, which funds its OpenAI investment. On the October 12 holdings, adding SoftBank (a 379 basis point median spread) lifted the AI issuer median from 327 to 369 basis points and the premium over the median high-yield bond from 102 to 144, which scores 47 instead of 40. The reading is the median of the last five days, so the change phases in as older readings roll off. At full effect it adds about 0.3 points to the composite.
- October 10, 2026. Added Circular financing as a ninth component at 10 percent: loans, guarantees and lease backstops that NVIDIA, Broadcom, AMD, Marvell, and Qualcomm have disclosed for their own customers, against those companies' trailing 12-month revenue, the same categories as Lucent's figure (equity stakes and capacity purchases are shown but not scored), anchored to Lucent, Nortel, and Cisco's 10-K customer-financing figures. The vendor-financing half of Counterparty concentration moved into it, so that component now measures customer concentration alone and NVIDIA's $105 billion OpenAI guarantee is counted once. Cost of money gained a third sub-score, AI borrowers' bond spread premium, priced daily from State Street's JNK and SPBO holdings. The other weights were rescaled by 0.9 and rounded.
- October 9, 2026. Added Cost of money as an eighth component at 13 percent, the mechanism the page had listed as untracked since launch. It averages two sub-scores: the 12-month change in the 10-year Treasury yield from the Treasury's daily par yield curve, and the median price-to-NAV of six listed private credit funds run by Ares, Blue Owl, Blackstone, Apollo, and BlackRock. The other seven weights were rescaled proportionally and rounded, and their order of importance is unchanged.
- August 21, 2026. Counterparty concentration now shows its two readings separately, customer concentration and vendor financing against revenue, each scored through the component's own anchors; the composite still uses their average and did not move. Every market-data component now publishes its exact ticker universe. Contracted cover's question was reworded to say disclosed contract coverage rather than implying the price itself is under contract. Added the contribution breakdown, a history strip and a sensitivity check.
- August 21, 2026. Added Contracted cover as a seventh component at 14 percent, measuring the basket's disclosed GAAP remaining performance obligations against its market value, and rescaled the other six proportionally so their order of importance is unchanged. Every component now publishes its own as-of date and source, and this changelog begins.
- August 20, 2026. First composite published, on six components. The series starts here; no earlier reading has been reconstructed.
For the demand side of the same question, our AI infrastructure backlog and RPO tracker lists what these companies have actually sold under contract, which is the one number a valuation cannot argue with.
Embed the index on your site
Questions people ask
Is there an AI bubble?
No single number can settle that, and the index does not give a yes-or-no answer. It scores how hot conditions are from 0 to 100 by measuring today's version of each thing that ended the dot-com bubble, from vendor financing to unbacked data center demand. The current reading and its band are at the top of this page. Nothing here is investment advice.
How is AIStockWire's AI Bubble Index different from other AI bubble indexes?
Much of it comes from filings, not stock prices. Four of its nine components move only when a company, utility, or regulator files something new. One is the commitment gap: announced data center capacity against the capacity backed by a signed interconnection agreement, a permit, or posted collateral. Another is circular financing: chip makers' loans and guarantees for their own customers, measured against what Cisco, Nortel, and Lucent disclosed in 2000 and 2001. The other five components move with the market, and the score updates four times each weekday. Every raw value is published, the history downloads as a CSV, and the method stays fixed until a scheduled review in late November 2026.
Why can the AI Bubble Index fall when AI stocks sell off?
Because it measures speculative heat, not crash risk. Three of its nine readings, momentum, valuation, and complacency (the VIX, inverted), cool when prices fall and fear rises, so the score drops. Stress can build at the same time, for example in AI borrowers' bond spreads, which the index smooths over a rolling median of up to five days.
When will the AI bubble pop?
Nobody knows, and the historical record is unkind to anyone claiming otherwise. The VIX peaked at 45.08 during the entire dot-com collapse, a level since exceeded by a credit downgrade, a flash crash and a tariff announcement. Volatility markets did not register the largest equity unwind in modern history while it was happening. This page therefore tracks mechanisms rather than timing.
What actually caused the dot-com crash?
Two things above all. The Federal Reserve raised rates from 4.75 percent to 6.5 percent between June 1999 and May 2000, which made speculative capital expensive. And vendor financing reversed: equipment makers had been lending customers the money to buy their equipment, with Lucent committing $8.1 billion and Nortel about $4.1 billion, often to carriers banks had already refused. Lucent's revenue fell from about $30 billion to $12 billion and it lost $16.1 billion in 2001.
What is the AI commitment gap?
The difference between data center capacity that has been announced or requested and the capacity that survives a test requiring a binding commitment: a signed interconnection agreement, a permit, posted collateral, or concrete in the ground. Across the seven screens tracked here the gap runs from 50 percent to 95 percent, with a median near 65 percent.
Which number here is the strongest evidence?
AEP Ohio. Its data center inquiries stood at about 30,000 MW. After Ohio regulators approved a tariff requiring data centers above 25 MW to pay for at least 85 percent of the energy they subscribe to whether or not they use it, expected load fell to 5,642 MW. Same market, same period, one variable changed: developers had to pay for what they had asked for.
Does this mean AI data centers are not being built?
No, and the counter-evidence is on the page deliberately. AEP has 45 GW of contracted Texas load backed by nearly $2 billion in posted cash or collateral, and Georgia Power has 8,500 MW that has physically broken ground with 10 projects already online. Real capacity is being built at very large scale. The gap is between that and the far larger announced figures that reach headlines.
Do Nvidia's price increases change the index?
Not the score, because no component here measures input costs. But it is a live reading on the cost mechanism. Bloomberg reported on August 23, 2026 that Nvidia warned major customers of price increases of at least 15 percent on Grace Blackwell and Vera Rubin server systems shipping from early 2027, as memory prices surge. Memory already makes up about 25 percent of a high-end AI rack's bill of materials, and server DRAM contract prices roughly doubled in the first quarter of 2026. Nvidia has not confirmed the report. Rising cost per unit of compute squeezes the buildout the same way expensive credit did in 2000: every committed dollar builds less than it did when the commitment was made.
Use this data
Our figures are free to reuse in articles, newsletters, and research with attribution and a link back to this page. How every number is computed: methodology. Current table as a file: download CSV.
Suggested citation: AIStockWire, “AI bubble index,” aistockwire.com/ai-bubble-index. Questions about the data: support@aistockwire.com.
Grid, utility and regulator evidence last updated August 19, 2026. A record of what those bodies and companies have published about their own exposure, not a forecast and not investment advice. Most of these bodies file quarterly at best, so this page updates as filings land rather than live.