没有基准的资产类别，怎么度量

Measuring an Asset Class That Has No Benchmark: Issuance Spreads of AI-Infrastructure ("Neocloud") Debt

Working paper v0.1 · November 2026

Brian Tian¹ · SpreadBench

¹ Brian Tian, SpreadBench, Seoul. Correspondence: hi@spreadbench.com. The index and dataset described here are published under CC BY 4.0. Replication code accompanies the deposit. Comments welcome; this is a first draft circulated for discussion, not a peer-reviewed publication.

JEL classification: G12, G14, G23, G32

Keywords: corporate bond issuance spreads, asset pricing, market microstructure, data infrastructure, AI infrastructure, neocloud, credit spreads, index construction

---

Abstract

A new borrower class entered the U.S. dollar corporate bond market between 2024 and 2026: companies that lease GPU compute and operate AI data centers, commonly called neoclouds. They now carry more than $28 billion of public bullet debt, yet no replicable benchmark exists for the credit spread at which they borrow. Ratings cover some issuers but aggregate issuers rather than instruments; secondary-market indices depend on licensed transaction data that most researchers cannot obtain; and vendor indices are not constructed from publicly verifiable inputs. This paper builds the missing measurement. Using only public filings and the U.S. Treasury par yield curve, we construct NCSX-IS, a face-value-weighted index of issuance-date credit spreads for every publicly issued U.S. dollar bullet bond from AI-infrastructure issuers meeting five eligibility conditions and no exclusion criteria. As of 4 November 2026 the index stands at 432.3 basis points across 13 bonds from 6 issuers ($28.06 billion face value), with a secured tranche at 391.2 bp and an unsecured tranche at 540.1 bp — a 148.9 bp structure premium. Three features of the cross-section are notable. First, the entire universe is short-dated: every bond has an original tenor between 4.99 and 6.08 years, so the index is effectively a single point on the curve rather than a term structure. Second, dispersion is wide — 173.9 bp to 580.8 bp, a 406.9 bp range on nominally similar instruments — and is only partly explained by seniority. Third, quarterly fund marks from SEC Form N-PORT imply a spread 129.8 bp wider than issuance pricing on the same seven bonds, indicating rapid repricing after issuance rather than a stale benchmark. We argue that the measurement itself is the contribution: an asset class that cannot be measured cannot be studied, and the scarcity of public pricing observations — thirteen bonds is the population, not a sample — is precisely what makes the series informative. All inputs, intermediate steps, and the index arithmetic are released so that any reader can reproduce 432.3 bp from the raw files in under a minute.

---

1. Introduction

Asset classes become research objects in a predictable order. First a market appears, then a benchmark is agreed upon, then the benchmark is quoted, and only afterwards does a literature accumulate around its properties. Real estate had no pricing index until Case, Shiller and Weiss (1993) built one from repeat sales; the index made housing a measurable asset class and its construction became a subject of study in itself (Shiller, 1981). Corporate credit followed a similar path: transaction-level data (Bessembinder, Maxwell and Venkataraman, 2006; Edwards, Harris and Piwowar, 2007; Dick-Nielsen, 2009) turned bond liquidity from an assumption into a measurable quantity.

The AI-infrastructure borrowers are at step one. Between May 2025 and July 2026, six U.S.-listed companies whose business is GPU compute and AI data centers came to the dollar bond market and raised $28.06 billion in thirteen bullet issues. Their credit quality is not obviously comparable to anything in the existing taxonomy: they are capital-intensive, lease-heavy, customer-concentrated, and their revenue depends on the resale price of compute, which is itself volatile and thinly observed. Market participants compare them to high-yield issuers, to data-center REITs, and to project finance, and none of these comparisons is exactly right.

What is missing is not opinion but measurement. Three sources could in principle supply it, and each fails for a different reason. Rating agency reports cover issuers, not instruments, and cannot be aggregated into a spread. Secondary-market indices rest on TRACE, whose full transaction record is licensed; a researcher without a license cannot verify a quoted level. Commercial index providers publish levels but not construction details, which makes replication impossible and comparison across providers hazardous. None of them begins from the one price that is both public and unambiguous: the spread fixed at issuance.

This paper constructs that benchmark. Our contribution is fourfold. (i) We define an eligibility rule set for the asset class that is exhaustive rather than selective, and we show that the resulting thirteen bonds are the entire population of qualifying public issuance, not a sample. (ii) We construct a face-value-weighted issuance-spread index using only SEC filings and the Treasury par yield curve, with interpolation rules that do not extrapolate. (iii) We document the shape of the resulting cross-section: a single-tenor cluster, wide dispersion, a large structure premium, and substantial post-issuance repricing detectable in fund marks. (iv) We release the data, the data dictionary, and the replication code so that the number can be checked rather than trusted.

A note on what this paper is not. It does not propose a pricing model, estimate a risk premium, or test an asset-pricing hypothesis. It measures a quantity and makes the measure auditable. In a literature where data construction is often the binding constraint, we take that to be a legitimate first step — and, for this particular asset class, a prerequisite for everything that follows.

2. Why issuance spreads

Three candidate measurement bases exist for the credit spread of a bond, and the choice among them determines what can be verified.

Secondary-market transaction spreads are the scholarly standard. They reflect the price at which capital actually changes hands, they update continuously, and they are the basis of most empirical work on corporate bond liquidity and credit risk. Their limitation is access: TRACE dissemination is public in truncated form, but the transaction record needed to compute instrument-level spreads consistently is a licensed product, and redistribution is restricted. An index built on licensed data cannot be audited by a reader, which defeats the purpose of a public benchmark.

Model-implied spreads (from dealer quotes, fund marks, or structural models) are available more widely but embed the model. Two providers can report different spreads for the same bond without either being wrong, and the difference cannot be resolved without the model. For a new asset class, where the model itself is unvalidated, this circularity is fatal.

Issuance spreads — the coupon fixed at pricing minus the Treasury yield at the matching maturity on the same date — have one great virtue: they are a market-clearing price recorded on a specific date, and every input is in a public document. A bond's coupon is in the prospectus supplement or the 8-K exhibit; its size and maturity are in the same document; the Treasury curve for that date is published by the U.S. Department of the Treasury. Anyone can recompute the number.

The cost is frequency. An issuance spread is observed once per bond, so the index is event-driven: it changes when a bond is priced, not every day. We regard this as a feature for a benchmark whose purpose is to establish what the market charged, and a limitation for anyone who wants a daily series. Section 8 is explicit about this boundary; resolving it requires a TRACE license, which we do not have.

3. Data and eligibility

3.1 Eligibility rules

A bond enters the universe only if all five conditions hold:

1. Issuer activity. The issuer's primary business is AI compute infrastructure — GPU cloud, AI data centers, or compute hosting — as evidenced in its SEC filings, not merely a announced intention to enter the sector.
2. Instrument type. U.S. dollar-denominated bullet corporate bonds. Amortizing structures, convertibles, and bank facilities are excluded (see 3.2).
3. Currency. USD only. Non-USD issuance by the same issuers is out of scope.
4. Verifiability. The bond's terms are documented in an SEC filing that can be retrieved and cited, line by line.
5. Size. Face value of at least $500 million, so that the issue is a benchmark-scale placement rather than a private placement.

3.2 Exclusions

Five categories are excluded, each for a stated reason:

• Convertible bonds. The coupon embeds an equity option; its spread is not comparable to that of straight debt. Four qualifying issuers in the sector (IREN, Nebius, Super Micro, Bitdeer) have issued only convertibles and therefore do not appear in the index at all.
• Bank loans and term loans. Floating-rate, secured, covenant-laden; a different instrument class.
• Non-USD debt. Currency basis contaminates the spread.
• Privately placed debt held by a single lender. No observable market price.
• Amortizing project-finance SPV debt. Secured by a single project, amortizing on a schedule, and governed by debt-service coverage covenants. Its average life is materially shorter than its stated maturity, so its spread is not comparable to a bullet of the same tenor. Two Hut 8 issues ($3.25 billion and $4.25 billion, both 2042 maturities, both amortizing) fall here; Section 7 quantifies what happens if one ignores this.

3.3 Sample size is the population

Because the five conditions are exhaustive and the exclusions are instrument-type based rather than discretionary, the resulting universe is not a sample. We verified issuer by issuer that the thirteen bonds are the complete set of qualifying public USD bullet issuance as of the as-of date. The peer check is itself informative: the four convert-only issuers and the two amortizing-SPV issues are the reason the count is thirteen. Firms such as Crusoe, Lambda, Nscale, Fluidstack and GMI Cloud have no SEC reporting obligation and therefore no public pricing observation at all.

This matters for how the index should be read. A small N is usually a weakness. Here it is a description of the sector's financing structure: public credit pricing observations in AI infrastructure are scarce, and that scarcity is the information.

3.4 Sources

Coupons, maturities, issue sizes and pricing dates come from SEC EDGAR documents; each bond in the released universe file carries the URL of its source filing. Treasury yields come from the Treasury's Daily Par Yield Curve. Fund marks come from Form N-PORT filings. No licensed, private, or estimated inputs are used anywhere.

4. Construction

For bond i with coupon c_i (percent), pricing date t_i, maturity T_i, and issue size s_i (face value in millions, the weight):

tau_i     = (T_i - t_i) / 365                 # years from pricing to maturity, not stated tenor
y_i        = interp( ParCurve(t_i), tau_i )    # Treasury par yield on the bond's OWN pricing date
spread_i   = (c_i - y_i) * 100                 # basis points
NCSX-IS    = SUM(spread_i * s_i) / SUM(s_i)    # face-value weighted across the 13 bonds

Three construction details carry most of the replication risk, and each is enforced in code rather than left to prose.

Own-date curve. y_i is interpolated on the curve for t_i — the bond's own pricing date — not on the as-of date. Using a single as-of curve for all bonds is the most common replication error and produces a materially different series.

Actual tenor, not stated tenor. tau_i is the actual time from pricing to maturity in years. A "5-year" bond priced on 19 May 2025 and maturing 1 June 2030 has tau = 5.03 years, not 5.00.

No extrapolation. Interpolation is linear between the two bracketing curve points. If tau_i falls outside the published curve range, the code raises an error rather than extrapolating; an extrapolated Treasury yield is a fabricated number.

The index is face-value weighted, on the view that the benchmark should reflect where capital was actually committed. An equal-weighted variant is computable from the released file and is reported in Table 2 for transparency, but it is not the index.

Precision of the Treasury input is flagged per bond (prec): day where the published curve has a point at the required tenor, month where the curve for the pricing month is interpolated at the target tenor. Two of thirteen bonds carry day; the remainder month.

5. The cross-section

Table 1. Index level and tranches (as of 4 November 2026, methodology v0.2.1-real)

Note: tranche levels are computed from unrounded per-bond spreads. Recomputing from the released CSV (one decimal per bond) yields 391.1 bp for the secured tranche against the published 391.2 bp — a 0.1 bp rounding budget. The all-bond level of 432.3 bp is identical under both.

Table 2. Distribution of issuance spreads

Table 3. Composition

Four features are worth stating plainly.

The index is a point, not a curve. Every bond in the population has an original tenor between 4.99 and 6.08 years. The sector borrows exclusively at the short end of the investment-grade/high-yield boundary, and the index therefore measures one maturity region. This is a property of the market, not a construction choice, and it means NCSX-IS should not be compared to a broad high-yield index whose duration is materially different.

Dispersion is large and only partly structural. The 406.9 bp range between the cheapest and most expensive bond is not explained by seniority alone: CIFR's secured 6.000% 2031 printed at 173.9 bp while CRWV's secured 9.750% 2031 printed at 580.8 bp. Both are senior secured, both are 2031 maturities, and they differ by 406.9 bp. Whatever the market was pricing — customer concentration, contract structure, hedging, or simply the sequence in which the sector's credits were discovered — it was not pricing a single sector curve.

Structure commands a large premium. Unsecured bonds price 148.9 bp wider than secured bonds on a face-weighted basis. Since the two tranches are issued by overlapping issuers, this is close to a within-issuer comparison and is the cleanest single number in the cross-section.

The sector prices between single-B and CCC. Against the ICE BofA U.S. High Yield index (OAS 332 bp on the same date), NCSX-IS carries a 100.3 bp premium, sitting above the B bucket (351 bp) and well below CCC (1,297 bp). The comparison is indicative rather than econometric — different durations, different weighting, different instruments — but it locates the sector in the rating landscape without requiring a rating.

6. Post-issuance repricing: an external check

An issuance-spread index invites an obvious objection: if the number is fixed at pricing, it may be stale by the time it is published. Fund marks provide a partial answer. Registered funds report per-bond fair values quarterly on Form N-PORT; taking the median across funds for each bond, inverting the resulting price to a yield, and subtracting the interpolated Treasury yield gives a comparable spread on the same instrument.

For the seven index bonds with reporting-fund marks at 31 August 2026 ($15.37 billion of face value), the implied spread is 616.4 bp, against an issuance spread of 486.6 bp on exactly the same seven bonds: +129.8 bp of widening. For CoreWeave's four bonds — the largest and most widely held — the face-weighted widening is 231.4 bp, with individual bonds between +191.5 and +257.7 bp.

Two conclusions follow. First, the issuance benchmark is not a stale proxy for the current level; it is a benchmark of a different quantity, and the gap between the two is itself measurable and economically large. Second, the direction of the gap is consistent with a sector that was repriced wider after its first wave of issuance, which is exactly the kind of fact that a benchmark makes visible.

The comparison has limits that should be stated: N-PORT marks are quarterly, reported with a lag, Level 2 fair values rather than transactions, and the median across a small number of holders may not represent the market's marginal price.

7. What difference do the exclusions make?

Exclusions are the most discretionary part of any index and therefore deserve a counterfactual. If the two Hut 8 amortizing SPV issues ($7.50 billion combined) are mechanically added to the universe, the face-weighted spread falls to 370.4 bp — a 61.9 bp reduction, with the secured tranche at 323.1 bp.

We do not adopt that universe, for a reason that is about comparability rather than about the issuer: an amortizing bond with a stated 2042 maturity and a scheduled principal amortization has an average life far shorter than sixteen years, and its spread compensates for a different set of risks (construction, DSCR covenants, single-tenant lease) than a bullet bond of the same stated maturity. Mixing the two would produce a number that is neither a bullet-bond benchmark nor a project-finance benchmark.

The counterfactual is published alongside the index, not hidden. A reader who prefers the broader definition can compute it; a reader who wants a bullet-only benchmark can use the index. What is not acceptable is publishing one definition and quietly computing the other.

8. Limitations

Small N, but not a sample. Thirteen bonds, six issuers. We argue this is the population of qualifying public issuance, but the fact remains that any single new issue moves the index materially: a $2 billion bond at 500 bp would move the face-weighted level by roughly 20 bp.

Single tenor. As noted, all bonds are 5-year paper. The index carries no information about the term structure of neocloud credit.

Cross-date mixing. Bonds were priced between May 2025 and July 2026. Face-value weighting means the index is a weighted average of spreads fixed at different times, not a snapshot of a single day's market. This is intrinsic to issuance-spread measurement and should not be described as a daily level.

No secondary or derivative layer. No TRACE-based daily series, no CDS. The N-PORT layer is quarterly and partial. A v0.3 series with licensed transaction data is the natural next step and is out of reach without a data license.

Issuer concentration. CoreWeave alone accounts for 27.6% of face value. Its repricing dominates both the index level and any inference about sector credit trends.

Definition sensitivity. The exclusion of amortizing SPVs alone moves the level by 61.9 bp. Anyone citing the index should cite its version and methodology, and treat the exclusion set as part of the definition.

9. Replication

The deposit contains the universe file (one row per bond, with the SEC filing URL for each), the Treasury curves for all pricing dates, the as-of curve, the N-PORT marks layer, the sector reference layer, and recompute.py, which depends only on the Python standard library. Running python recompute.py reproduces Table 1 and asserts equality with the published index level; a mismatch raises an error rather than passing silently.

Version DOI: 10.5281/zenodo.23983572 (v0.2.1-real). Concept DOI: 10.5281/zenodo.23977166. Data are released under CC BY 4.0; code under MIT. Version history is public, including the one erratum to date (a maturity correction that moved a single bond by 0.2 bp and the index by 0.02 bp).

We regard reproducibility as the minimum condition for a benchmark to be cited. A number that cannot be recomputed is an assertion, not a measurement.

10. Conclusion

The neocloud borrowers now carry $28 billion of public debt with no benchmark to price it against. This paper supplies the first one that is public, replicable, and defined by an exhaustive rule set: 432.3 bp on 4 November 2026, with a 148.9 bp structure premium and a 406.9 bp dispersion that no single-factor story explains.

The larger point is about measurement as a public good. An asset class with no observable price cannot be studied, cannot be compared across time, and cannot be argued about with evidence. Building the observation is unglamorous work — reading filings, fixing a curve convention, writing a script that fails loudly — but it is the work that makes everything else possible. We release it in that spirit, and we invite replication, criticism, and co-authorship from researchers working on credit markets, market microstructure, and the economics of AI infrastructure.

---

References

Bao, J., Pan, J., & Wang, J. (2011). The illiquidity of corporate bonds. Journal of Finance, 66(3), 911–946.

Bessembinder, H., Maxwell, W., & Venkataraman, K. (2006). Market transparency, liquidity externalities, and institutional trading costs in corporate bonds. Journal of Financial Economics, 82(2), 251–288.

Case, K. E., Shiller, R. J., & Weiss, A. N. (1993). Index-based futures and options markets in real estate. Journal of Portfolio Management, 19(2), 83–92.

Dick-Nielsen, J. (2009). Liquidity biases in TRACE. Journal of Fixed Income, 19(2), 43–55.

Edwards, A. K., Harris, L. E., & Piwowar, M. S. (2007). Corporate bond market transaction costs and transparency. Journal of Finance, 62(3), 1421–1451.

Gürkaynak, R. S., Sack, B., & Wright, J. H. (2007). The U.S. Treasury yield curve: 1961 to the present. Journal of Monetary Economics, 54(8), 2291–2304.

Shiller, R. J. (1981). Do stock prices move too much to be justified by subsequent changes in dividends? American Economic Review, 71(3), 421–436.

---

Data availability

Tian, B. (2026). Neocloud Credit Spread Index — NCSX-IS, v0.2.1-real (Version 0.2.1) [Data set]. Zenodo. https://doi.org/10.5281/zenodo.23983572

Disclosure

Brian Tian is the founder of SpreadBench, which publishes the index described in this paper and offers commercial licensing of related data services. The dataset is released under CC BY 4.0 and is free for academic use; the index level, methodology, and replication code are public without restriction. Any error found should be reported to hi@spreadbench.com and will be corrected through a versioned erratum.

# 下载 download/dataset-v1.1/ 后
python recompute.py      # 输出逐券利差与分档水位，断言等于 432.3