Project TBD Memory Vault

From Compliance to Financialization — A Quarter of Research-Driven Inquiry into the Semiconductor Supply Chain

36 min read

From Compliance to Financialization

A Quarter of Research-Driven Inquiry into the Semiconductor Supply Chain

GSBGEN 390 Independent Study — First-Draft Report for Professor Jonathan Berk Dustin Ross & J Bliss Perry · Spring Quarter 2026 · drafted 2026-05-29

What this document is. This is a first-draft report assembled from our research vault — roughly 79 logged conversations and ~75 synthesis/primer memos compiled January–May 2026. It is written to be edited by us, not sent as-is. Per the research methodology we adopted this quarter, the document is built to surface evidence and end in questions; where it offers a recommendation, that recommendation is explicitly flagged as a tentative founders’ lean for us to adopt, revise, or cut. Every non-trivial claim carries a source label: [Interview: Name, Date] for something a real person told us, [Public: Source] for a public document, [Synthesis] for our own inference across sources, and [Speculation] for a hypothesis we have not yet grounded.

Outline changes vs. our working outline (logged for transparency). Our original outline (D. Ross, 2026-05-29) had three “where we go” branches: commodity dynamics, reverse supply chain, and insurance. In assembling this draft we (1) re-grouped the three branches around financial-product families (commodity/compute financialization; reverse supply chain + warranty; risk-transfer/insurance), (2) pulled “memory/DRAM as a distinct commodity track” up into its own treatment inside the financialization section because four independent interviewees converged on it, (3) added a prominent disconfirming-evidence subsection to each branch — most importantly the history of failed semiconductor futures, which no interviewee raised but which is the single strongest challenge to the financialization thesis, (4) drew out data-secrecy as the through-line connecting the abandoned compliance thesis to the new one, and (5) folded the AI-methodology story into a light-touch “means and methods” subsection rather than a standalone chapter. These are revisions to our outline, surfaced here so the change is visible rather than hidden.

0. Executive Summary

We began this study intending to compile a structured database of the downstream semiconductor supply chain from public filings and interviews, using AI agents to do extraction at a scale a two-person team could not otherwise reach, with export-compliance as the wedge that would justify the data-gathering. Over roughly four months and ~79 conversations, two findings forced us off that path. First, the data we needed is structurally, deliberately opaque — held behind NDAs and protected by intermediaries whose value is their proprietary relationships. Second, and more fatally, the commercial buyers we expected to pay for compliance are incentivized not to know: a chip company selling into a Singapore distributor that re-routes to China would rather not learn where its product ends up, because knowing only costs it sales. [Interview: Yisroel, 2026-05-08] [Interview: Josh, 2026-04-30] [Interview: David/Matt (Shield Capital), 2026-05-22] Compliance pain is real and acute — but only for the buyer that must care: the U.S. government and the defense primes that serve it. We chose not to pursue the defense market commercially, and so we abandoned the compliance wedge.

What we turned toward is the financialization of the semiconductor and compute supply chain — the family of instruments (futures/hedging, parametric insurance, warranty-risk transfer, inventory financing) that let operators hand risk they don’t want to specialists who do. The single organizing idea, owed to a finance primer we built mid-quarter, is that operators buy certainty and speculators sell it. [Synthesis] Three product families emerged from the interviews, each with a clear anchor and each with live disconfirming evidence:

  1. Commodity & compute financialization — GPU-hour futures (CME + Silicon Data launched the first such contracts on 2026-05-12 [Public: CME, 2026-05-12]; Pluto is building a CFTC-regulated exchange to launch this summer [Interview: Ronit Jain (Pluto), 2026-05-22]), and the older idea of a commodity-trading or “Glencore of chips” model concentrated on memory, the most commodity-like layer. The disconfirming evidence is strong: every prior attempt to build a physical semiconductor futures market has failed on non-fungibility (Enron’s DRAM forwards in 2001; earlier DRAM-futures attempts in 1989 and 2003) [Public: Felix Stocker; Synthesis], and even the builders concede the binding constraint is not whether compute is hedgeable but whether anyone will use the hedge — “no tech-company CFO has ever had to hedge their cost of goods sold.” [Interview: Spencer Powers (DRW), 2026-05-22]

  2. Reverse supply chain & warranty — the most concrete operational pain we found anywhere. NVIDIA runs a $5-trillion company’s returns operation “on email and spreadsheets,” is actively procuring outside tooling, and carries an $8.22B warranty reserve (≈20× its FY2023 level), a liability that AMD’s own trajectory ($310M → $1.05B) shows is not NVIDIA-specific. [Interview: Lonny Orona (NVIDIA), 2026-05-12] [Interview: Alex Zhu (NVIDIA), 2026-05-27] [Public: WarrantyWeek, 2026-04] Two opportunities sit here: an operational integration layer, and the financial transfer of the warranty liability itself.

  3. Risk transfer / parametric insurance — anchored by two reinsurance-side conversations. The analytical backbone is a four-part test (you need a metric, a trusted measuring agent, a loss model, and a market of reinsurers willing to write it); for man-made equipment failure in fabs and data centers, the measuring agent simply does not exist yet — which is either the opportunity or the reason it can’t be done. [Interview: Preston (Guy Carpenter), 2026-05-22] The strongest disconfirming voice is an insurance-software founder who told us flatly that “the parametric market is still small and it’s not worth it — people are just not comfortable with parametric triggers.” [Interview: Jeremy Jawish (Shift), 2026-05-22]

Two themes cut across all three. The first is that the data secrecy that killed compliance did not go away — it reappears as the obstacle to the data-asset moat behind every financialization wedge (“a single leakage would probably mean the end of PDF”). [Interview: Andrzej Strojwas (PDF Solutions), 2026-05-22] The second is market thickness: in a market with three memory makers and a handful of hyperscaler buyers, any intermediary’s margin can be “pounded down” by a counterparty that simply refuses to pay it. [Synthesis; financialization primer, 2026-05-29] The report ends on the question these two themes force: where, in a thin and secretive market, does a durable, defensible position actually come from?

1. Where We Started: The Compliance / Downstream-Data Thesis

Our GSBGEN 390 petition (Spring 2026) proposed to “undertake the compilation of a database of the downstream semiconductor supply chain based on publicly available information.” [Public: GSBGEN 390 Petition, 2026] The intuition — which traces to early conversations with Professor Berk — was that the upstream of the chip industry (fabrication, equipment, materials) is heavily studied, while the downstream (who buys the chips, through which distributors and integrators, under what terms, to what end markets) “is not well mapped in any single, structured resource.” [Public: GSBGEN 390 Petition, 2026] The plan drew on four source types — 10-K filings, expert interviews, industry reports, and academic literature — and leaned heavily on a fifth enabler: AI agents to extract and cross-reference data across hundreds of filings at a scale not previously practical for a two-person team.

The wedge that made this commercially interesting (rather than merely academic) was export-compliance. The thesis: chipmakers face mission-critical, high-penalty ship/no-ship decisions under EAR/ITAR; a platform that automated export classification and licensing would (a) sell into a real pain and (b) accrete the proprietary relationship and transaction data that would, over time, become the downstream supply-chain map. Compliance was the Trojan horse for the data asset. [Synthesis; thesis evolution log]

The case for compliance pain was genuinely strong on the enforcement side, and it is worth stating plainly because it is not the reason we abandoned the thesis. Enforcement had gone blockbuster: Applied Materials was fined $252M and Cadence $140M, against a backdrop of continuous AI/chip EAR rulemaking through 2024–25. [Synthesis: compliance-wedge-takeaways] An advisor we trust, Ann Miura-Ko, even endorsed the sequencing — “focus on compliance data collection now, worry about derivatives and insurance later” — and the image of becoming, over time, “the JP Morgan of the industry.” [Interview: Ann Miura-Ko, 2026-03-06] We note this because our eventual pivot directly inverts that advice, and intellectual honesty requires owning that.

2. What We Did, and What We Learned

2.1 Means and methods (a light-touch note on the research itself)

Between mid-January and late May 2026 we logged ~79 conversations in our research vault, alongside ~33 synthesis briefs and ~40 primer documents. [Synthesis; vault file counts, 2026-05-29] The cadence accelerated sharply once our transcript-ingestion pipeline came online in late April (≈3 conversations in January rising to ≈42 dated files in May). By rough perspective bucket, the corpus spans industry operators (NVIDIA reverse-logistics and strategy staff, Renesas, PDF Solutions, ex-Samsung memory), finance/insurance/trading (DRW, Pluto, Guy Carpenter reinsurance, Shift), investors (Sequoia’s Roelof Botha, Floodgate’s Ann Miura-Ko, Jillian Manus, Mo Islam), and academics/advisors (Professor Berk, Steve Blank, Richard Dasher, plus several STRAMGT class sessions). [Synthesis] The composition reflects the methodology we adopted — deep, slow learning through conversation, deliberately seeking diverse vantage points and treating surprises and contradictions as the highest-value output rather than confirmation of what we already believed.

On the AI dimension that the petition foregrounded, candor is warranted. We did not build the automated 10-K extraction pipeline as originally conceived. What we built instead is an AI-assisted research operating layer: a version-controlled markdown “memory vault,” a set of Claude Code skills (for deep-dive research, interview synthesis, transcript ingestion, and cross-vault pattern-finding), semantic search over everything we’ve heard, and an auto-publishing web layer. The “scale not previously possible” turned out to be scale in processing and synthesizing conversations — turning ~79 raw transcripts into queryable, cross-referenced knowledge — rather than scale in parsing filings. We flag the gap between promise and practice honestly; it is itself a finding about where AI currently adds the most leverage in qualitative industrial research (synthesis of primary conversation), and where the harder problem (extracting reliable structured relationships from disclosure documents) remains. [Synthesis]

2.2 Why we abandoned the compliance / data-modeling thesis

Two findings did it. Neither is “the pain wasn’t real”; both are about structure.

Finding 1 — The data we needed is deliberately, structurally held. This is not incidental friction that better tooling overcomes; opacity is the business model of the people who hold the data. Josh framed it directly: in this chain you usually “just care about the last step,” there are “300,000-plus components” to track, and the value-chain intermediaries who could make the chain transparent have “no incentive” to do so — “relationships beat data.” [Interview: Josh, 2026-04-30] PDF Solutions, which sits on some of the densest manufacturing data in the industry, was blunt about why it will never be sold or pooled: “a single leakage would probably mean the end of PDF.” [Interview: Andrzej Strojwas (PDF Solutions), 2026-05-22] Independent corroboration came from multiple directions: an operator noted that operations teams want to share but “legal teams kill initiatives even when operational teams see value,” and a prior multi-million-dollar academic effort to assemble this kind of data reportedly failed to crack acquisition. [Interview: David/Matt (Shield Capital), 2026-05-22] [Synthesis]

Finding 2 — Commercial buyers are incentivized not to know. This is the sharper, more decision-relevant finding. The compliance product assumed customers wanted visibility into where their parts came from and went to. In commercial semis, the opposite is true. Yisroel put it cleanly: “if I know it’s going to China now I can’t sell it anymore — you’ve done nothing good for me.” [Interview: Yisroel, 2026-05-08] Dustin reached the same conclusion talking to investors: “these commercial semiconductor companies don’t want to know if what they’re selling is going to China … because that’s just sales they’d be getting otherwise.” [Interview: David/Matt (Shield Capital), 2026-05-22] In the May 8 session, the framing crystallized as “I’m selling to a firm in Singapore — who knows what they’re doing with it.” [Interview: Jonathan Berk session, 2026-05-08] Two further data points sealed it: a semiconductor professional we spoke to had never heard of UFLPA (the forced-labor statute we’d assumed was a live compliance driver), and an NVIDIA strategy contact observed that Qualcomm books a large share of commodity-memory revenue into China “with minimal scrutiny” — “the horse has left the barn.” [Interview: Josh, 2026-04-30] [Interview: Nicole (NVIDIA), 2026-05-01]

The one place compliance does work is precisely the place where the buyer cannot choose not to know: the U.S. government and defense primes, who “pay a 10× markup for China-free supply chains” and for whom traceability is “table stakes.” [Interview: Josh, 2026-04-30] [Interview: Yisroel, 2026-05-08] We made a deliberate decision not to build a defense-first company in this study, and so compliance — alive and valuable in that one niche — was killed for our commercial purposes.

The lesson we carry forward. The abandoned thesis left us with a structural fact, not just a dead end: secrecy and concentration are permanent features of this industry, not bugs to be engineered away. As Section 6 argues, that fact is exactly what threatens the data-asset moat behind the financialization wedges we turned toward — which is why we treat it as a through-line rather than a closed chapter.

3. Where We Go (1): Commodity & Compute Financialization

3.1 The frame: operators buy certainty, speculators sell it

The intellectual engine of the pivot is an analogy we pressure-tested all quarter: semiconductors (and compute) as the new oil. An airline has no edge predicting jet-fuel prices and a 50% spike can wipe out its year, so it pays a trader to take that price risk off its hands — giving up the windfall of cheap fuel in exchange for certainty. That single trade — operators buying certainty, speculators selling it — is the seed of every instrument in this report: futures, hedging, parametric insurance, and the warranty transfer of Section 4. [Synthesis: financialization primer, 2026-05-29] The structural question, in Dustin’s framing of the two tests, is always (1) materiality — is the cost big and distinct enough that buyers care — and (2) volatility — does the price move both ways, because hedging only has value against two-sided uncertainty. [Synthesis]

There are, per the same framing, three layers you can financialize, top to bottom: the token ($/token of model output), the compute layer ($/GPU-hour or $/GPU), and the physical chip layer (the silicon itself, where memory is the clearest commodity case). [Synthesis; Interview: Adhi (5CC Capital), 2026-05-27] The crucial tension, which recurs below, is that the layer that behaves most like a commodity (memory) is also the layer most controlled by a handful of suppliers — which is exactly what historically kills exchanges.

3.2 The live market: compute futures (CME, Silicon Data, Pluto, DRW)

This space went from thesis to live product during our study. On 2026-05-12, CME Group and Silicon Data announced the first compute futures — cash-settled contracts referencing Silicon Data’s daily benchmark indices for on-demand GPU rental rates (H100, H200, and successors), pending regulatory review. [Public: CME, 2026-05-12] [Public: CNBC, 2026-05-12] Silicon Data is backed by DRW, the Chicago proprietary-trading firm. The U.S. policy backdrop is unusually permissive: the federal AI Action Plan explicitly recommends developing “a spot and forward market for GPU compute,” lowering the political cost of listing such contracts. [Public: Dave Friedman, 2026]

We spoke with two of the people building this market:

  • Spencer Powers (DRW) traces the thesis to a 2023 observation by DRW founder Don Wilson — “the financial and risk infrastructure that oil has that compute doesn’t” — which DRW has been operationalizing for ~2 years across four bets: Silicon Data (the data/index layer), Compute Exchange (a spot/auction market for reserved compute, founded by Wilson), Vast.ai (“Airbnb for GPUs”), and SF Compute (cluster bursts for smaller startups). [Interview: Spencer Powers (DRW), 2026-05-22] His unit of commoditization is $/GPU-hour, deliberately — it bundles the power cost, sidesteps NVIDIA’s monopoly pricing on the chip itself, and matches how neoclouds and customers already talk. [Interview: Spencer Powers, 2026-05-22]

  • Ronit Jain (Pluto) is building what he describes as a CFTC-designated derivatives exchange and clearinghouse for compute — the structural point being that, unlike CME/ICE partnering with index providers, Pluto intends physical settlement capability, which he frames as the durable edge. He cites a ~15-month regulatory process and a launch targeted for summer 2026. [Interview: Ronit Jain (Pluto), 2026-05-22] (We have not independently verified Pluto’s CFTC designation status against public CFTC filings; treat the regulatory specifics as the founder’s representation pending confirmation. [Speculation])

Three use cases emerged, and they matter because they identify who buys first — the gap our own working outline left open:

  1. Hedging compute COGS. AI products, unlike classic 99%-margin SaaS, “actually do have cost of goods sold in their inference cost,” and that cost is volatile. [Interview: Spencer Powers, 2026-05-22]
  2. GPU collateralization for lending. A forward price curve lets a lender treat GPUs as collateral over a 3–5-year window — underwriting the asset value rather than a startup’s creditworthiness, exactly as commercial real estate underwrites the building more than the tenant. Both Spencer and Ronit independently named debt financiers to neoclouds as the beachhead buyer, because they are sophisticated enough to hedge collateral they’re exposed to. [Interview: Spencer Powers, 2026-05-22] [Interview: Ronit Jain, 2026-05-22]
  3. GPU price-depreciation insurance. This is the hardest revenue datapoint in the entire corpus: Pluto reports having sold ~$60M of H200 depreciation coverage, structured as a put option and operated as a swap dealer (not an insurance carrier), covering new-model releases, hardware advances, and geopolitical events including a Taiwan invasion; its head of trading is a former UBS swaptions director. [Interview: Ronit Jain, 2026-05-22] Note that this product sits across our financialization and insurance branches — a useful reminder that the boundaries we draw for exposition are softer in practice.

3.3 The physical-chip track: memory as the “Glencore of chips” candidate

The May 8 session with Professor Berk generated the second strand: not financializing services but trading the physical commodity, on the model of Glencore in oil. Berk’s framing — “Glencore prints money because it’s a physical option holder at industrial scale” — set up the structural comparison we then researched in depth. [Interview: Jonathan Berk, 2026-05-08]

The comparison is instructive precisely because it is mostly negative. Glencore’s edge rests on four things chips largely lack: (1) a storage moat (oil storage is capital-intensive; “anyone can store semiconductors — there’s no storage advantage,” as Berk put it [Interview: Jonathan Berk, 2026-05-08]); (2) information asymmetry from physical flow (Glencore handles ~4.2M barrels/day; in chips the valuable information sits inside fabs, accessed via “channel checkers in Taiwan,” not via intermediaries [Public: S&P Global; Interview: Nihar, 2026-05-06]); (3) deep liquid spot+futures markets (no semiconductor futures market has ever survived — see §3.4); and (4) fungibility (a barrel of Brent is substitutable; “DRAM is not perfectly fungible — it’s not like if Compaq doesn’t want it you can push it to Dell” [Public: Felix Stocker / Samsung VP]). The existing analogues, Arrow and Avnet, demonstrate the ceiling: at $30.9B and $22.2B in revenue they earn only ~1.9% and ~1.1% net margins, and they do not speculate on inventory — they hold it on consignment for customers. [Public: company filings] [Interview: Holly Rawlins (Renesas), 2026-04-29] During the greatest dislocation in the industry’s history (the 2020–22 shortage, ~$200B in lost auto revenue), Arrow’s net income roughly tripled and then fell below its pre-shortage level by 2024 — it captured some volatility but could not hold it. [Public: MacroTrends; Synthesis]

The one segment where the commodity thesis genuinely has legs is memory (DRAM/NAND), and four independent sources converged on it: Berk (“Glencore of memory chips”), Nihar (hedge-fund investor), Minseok Kim (ex-Samsung), and Ronit Jain all pointed to memory as the most oil-like layer. [Interview: Jonathan Berk, 2026-05-08] [Interview: Nihar, 2026-05-06] [Interview: Minseok Kim, 2026-05-05] [Interview: Ronit Jain, 2026-05-22] The evidence is on their side: JEDEC standardization creates real fungibility, a transparent spot market exists (TrendForce/DRAMeXchange), and prices swing like a commodity — DRAM contract prices rose ~90–95% QoQ in Q1 2026 and are forecast +63% in Q2; NAND +55–60% then up to +75% — part of an AI-driven “memory supercycle.” [Public: TrendForce, 2026] The scale is enormous (OpenAI’s Stargate reportedly contracted for up to ~900,000 DRAM wafers/month — on the order of 40% of global output). [Public: TrendForce/industry, 2026]

But memory also embodies the central tension in its sharpest form, and the disconfirming evidence here is as strong as the confirming. The market is a three-supplier oligopoly (Samsung, SK Hynix, Micron ≈ 95% of DRAM), feeding a handful of hyperscalers — Nihar’s “3 × 5 = 15 bilateral relationships covering ~80% of demand,” which may simply be too concentrated for any intermediary to insert itself. [Interview: Nihar, 2026-05-06] The oligopolists prefer opaque bilateral pricing because it protects margin — which is exactly why memory makers killed the futures market the last time it was tried (§3.4). And the fastest-growing, highest-value memory segment, HBM, is moving in the opposite direction: co-designed with NVIDIA under long-term contracts, it behaves “more like a specialty chemical” than a commodity. [Public: TrendForce; Synthesis] The commodity thesis may apply to a shrinking share of memory.

3.4 Disconfirming evidence (foregrounded)

Per the methodology that surprises are the signal, we lead the skeptical case rather than burying it.

(a) The graveyard of failed semiconductor futures. This is the most important fact in this section and no interviewee raised it — it surfaced only in our own research. Physical semiconductor futures have been attempted at least three times and failed every time: the Pacific Stock Exchange proposed DRAM futures in 1989, Enron launched DRAM forward contracts in 2001 (which died with Enron), and SGX proposed chip futures in 2003. [Public: Felix Stocker; Synthesis: market-sizing-grand-slam] The structural reason is non-fungibility plus product churn: the very unit of sale keeps changing (256KB in 1989 → 128MB in 2001 → multi-GB today), defeating contract standardization. The live CME/Pluto products implicitly bet that pricing at the $/GPU-hour service layer finally routes around this problem — but that is a bet, not a proof, and it should be stated as one.

(b) The adoption / “COGS-blindness” problem. Both builders told us the binding constraint is demand, not the instrument. “No CFO of a tech company has ever had to hedge their cost of goods sold,” and “a company isn’t exactly waking up every day thinking about how to hedge its GPU costs.” [Interview: Spencer Powers, 2026-05-22] [Interview: Ronit Jain, 2026-05-22] Ronit calls the core work “engineering the consumer behavior that’s necessary for this market to work.” The risk is not that compute is un-hedgeable; it’s that the natural buyer doesn’t yet feel the pain.

(c) Price-signal integrity. A futures product is only as good as its reference index, and the reference is currently distorted by company psychology rather than market clearing. NVIDIA strategically underprices (a Blackwell outputs ~30× the tokens of a Grace Hopper but costs only 70% more); closed-model makers subsidize tokens ($5,000 of compute billed at ~$200); and listed neocloud rates are “totally unreliable,” sometimes double the negotiated price. [Interview: Ronit Jain, 2026-05-22] [Interview: Mo Islam, 2026-05-22] Mo Islam’s question — “what is the index for compute?” — is the unsolved prerequisite.

(d) Obsolescence vs. storability. Steve Blank’s structural objection to the oil analogy: oil can be stored strategically and its forward curve is shaped by storage economics; semiconductors “obsolete in 9 months.” [Interview: Steve Blank, 2026-01-22] A physical-inventory hedge may be impossible for a fast-obsoleting good — which weakens the Glencore-of-chips model specifically.

(e) Demand is not infinite. A useful corrective from two sources: “anytime you think demand is infinite, all you know is it’s not infinite” [Interview: Tim (Etched), 2026-05-22], and a contrarian read that data-center buildout is “saturated” with edge AI as the next cycle [Interview: Josh, 2026-04-30]. A demand plateau is itself the un-hedged downside the whole thesis assumes away.

3.5 The business-model question

A recurring, unresolved sub-question is what form a TBD financialization business would take. Three candidates appeared, often from the same mouths: (1) software/data (be Silicon Data — the index/measurement layer); (2) the market itself (be Pluto/CME — the exchange or the risk-taker); or (3) a capital-markets advisory / “investment bank for compute” that finds companies with exposure, structures the hedge, and lays the risk off to a market maker. Spencer raised the advisory model unprompted (“it basically sounds like an investment bank,” Dustin replied), while warning it’s “not the sexiest startup.” [Interview: Spencer Powers, 2026-05-22] Tim at Etched independently pushed the idea down the stack to component-level financialization — speculating on the cost of “each individual component of each chip.” [Interview: Tim (Etched), 2026-05-22] We flag this as an open design question, not a settled choice.

4. Where We Go (2): The Reverse Supply Chain & Warranty

4.1 The anchor: NVIDIA’s returns operation

The most concrete operational pain we encountered all quarter came from two NVIDIA insiders. Lonny Orona, weeks into running compute-science frontline support, described a “$5-trillion company running on email and spreadsheets,” with reverse logistics split across siloed systems (Salesforce for tickets, SAP for material planning, Baxter for demand planning, Expeditors for 3PL) and manual hand-offs at every seam. NVIDIA is standing up dedicated repair lines (Dallas, going live ~July 2026, run by Wistron and Foxconn) and — critically for us — is actively procuring outside tooling: “we have no time for in-house tooling.” [Interview: Lonny Orona (NVIDIA), 2026-05-12] The scaling math is stark: a single hyperscaler (Meta) holds ~100K GPUs today and wants ~1M within five years; NVIDIA already struggles with hundreds of returns, and “thousands will break the system.” [Interview: Lonny Orona, 2026-05-12]

Alex Zhu, who leads parts of NVIDIA’s reverse supply chain, supplied the financial scale: NVIDIA carries roughly $8B against warranty liabilities, a balance-sheet item that “has grown 20 times in the last year,” repairs are currently free to customers, and of every 100 units returned only ~60 are economically repairable — the other ~40 are filled from new inventory, with the candid aside that “new buy is all Jensen cares about.” [Interview: Alex Zhu (NVIDIA), 2026-05-27]

4.2 Reconciling the numbers (so the report can be cited safely)

Two figures circulate internally and deserve precision before publication:

  • Warranty. NVIDIA’s warranty reserve balance was $8.22B at the end of FY2025, up from ~$416M in FY2023 — almost exactly the “grown 20×” Alex described. The single-year accrual addition was $2.59B (vs. ~$1.75B for the entire rest of the U.S. semiconductor industry), and claims paid jumped ~1,000% to $894M (from $81M). Additions relate “primarily to the Compute & Networking segment” — i.e., data-center GPUs. [Public: WarrantyWeek, 2026-04; NVIDIA 10-K FY2025] One nuance worth making explicit for a finance audience: a warranty reserve is an accrued accounting liability, not necessarily a segregated pile of cash — which sharpens, rather than softens, the “is this capital being managed efficiently?” question. [Synthesis]
  • Failure rate. Our sources cite both ~4% (“4% of NVIDIA GPUs fail upon reaching data centers” [Interview: Tim (Etched), 2026-05-22]) and ~9% (annualized, from Meta’s Llama-3 run: 16,384 H100s, one failure every ~3 hours, ~80% hardware-related [Public: Meta Engineering, 2025]). These are likely different denominators — early-life/arrival failure vs. annualized operational failure — and the report should present both with that caveat rather than picking one.

4.3 Does it generalize beyond NVIDIA? (an outstanding question, with evidence)

For generalization: AMD’s warranty trajectory mirrors NVIDIA’s at smaller scale — reserves $310M (2023) → $597M (2024) → $1.05B (FY2025), claims $110M → $238M, claim rate 0.43% → 0.68%. [Public: WarrantyWeek, 2026-04] The failure mode is structural to advanced packaging (HBM stacks bonded via CoWoS; thermal stress at 1,400W Blackwell parts), not a quirk of NVIDIA’s execution. [Public: Synthesis]

Against generalization: failures concentrate specifically in data-center AI accelerators. Intel server CPUs show near-zero recorded failures; server DRAM 0.2–0.27%. [Public: Puget Systems, 2025] And for the largest custom-ASIC accelerator vendor (Broadcom), we found no public warranty-reserve spike, leaving open whether the warranty burden for custom silicon sits with the vendor or the hyperscaler customer. [Public: Broadcom filings; Synthesis] The honest read: this is a large and fast-growing niche (AI accelerators), not “all semiconductor reverse logistics.”

4.4 Chip repair feasibility (the naive questions, answered)

Dustin’s outline framed this with deliberately naive questions, which turn out to be the right ones:

  • “Why don’t they just make the chips so they don’t break?” — At hyperscale, failure is statistical, not a defect: with ~9% annualized failure across 100K+ GPUs, a 16K cluster has a mean-time-to-failure of ~1.8 hours. You cannot engineer this to zero; you manage the flow. [Public: Meta/Jason Hoffman, 2026]
  • “Why don’t they just throw out the chips?” — Because the unit economics are large (DGX-class units cost “millions” [Interview: Alex Zhu, 2026-05-27]) and a structured secondary market exists (used A100 80GB ~$12–18K; CoreWeave rebooking 2022 H100s at ~95% of original price). [Public: Introl, 2025]
  • Is repair actually feasible? — Board/system-level repair: yes, and economically sensible (NVIDIA’s playbook-driven CM repair lines; ~60% repairable per Alex). Die/package-level repair: largely no — once HBM is bonded to the GPU die via CoWoS, a failed stack or microbump generally scraps the whole module, and chiplet designs push the chain further toward replace-and-scrap. [Synthesis: packaging literature; reverse-supply-chain brief]

4.5 The two opportunities here

  1. An operational integration layer for semiconductor reverse logistics — the seamless flow “from case opening to shipping to customer to receiving back” that no incumbent (ServiceMax, Baxter, IFS) cleanly owns, sold into a buyer (NVIDIA) who is actively procuring. This is the clearest “someone is trying to give us money” signal in the corpus. [Interview: Lonny Orona, 2026-05-12] [Synthesis]
  2. Warranty-risk transfer — the financial mirror of the same problem. A specialist assumes NVIDIA’s warranty obligation for a premium; NVIDIA wins by (a) shedding an operational nightmare and (b) redeploying capital that compounds far faster in GPU R&D than it earns sitting against the reserve; the specialist wins on underwriting margin plus investment income on the float. This is the bridge to Section 5 — and to Pluto’s depreciation product in §3.2 — and Max Mirgoli independently suggested studying exactly “NVIDIA’s warranty claim size vs. revenue and the potential to reinsure that warranty risk.” [Interview: Max Mirgoli, 2026-05-22] [Synthesis: financialization primer] Caveat: no interviewee is yet paying to transfer this risk — it remains our inferred opportunity, not a validated willingness-to-pay.

5. Where We Go (3): Risk Transfer — Parametric Insurance & Structured Products

5.1 The anchor and the analytical backbone

Our richest insurance conversations were with Preston, a facultative reinsurance professional at Guy Carpenter (Marsh McLennan). He mapped the full stack (insured → retail broker → carrier → reinsurance broker → reinsurer → retrocession → capital markets) and, unprompted, proposed the structure that organizes this whole section: “structure an ILS product with a parametric trigger and go straight to the capital markets.” [Interview: Preston, 2026-05-07; 2026-05-22]

Parametric insurance pays a pre-agreed amount the instant a measurable parameter crosses a threshold — “if the temperature of one of those machines gets above a certain threshold, then I get a $100 million paycheck, because that’s just codified” — rather than reimbursing assessed losses. [Interview: Preston, 2026-05-22] Its advantage is speed and objectivity; its hazard is basis risk (the trigger fires but you had no loss, or you had a loss the trigger missed). Preston’s four-pillar test is the cleanest diagnostic we found for whether any such product can exist: you need a metric, a trusted third-party measuring agent, a loss model, and a market of reinsurers willing to write it. For natural catastrophes all four exist. For man-made equipment failure (a fab overheating, a GPU process breakdown) the three non-market pillars are missing — there is no agreed metric, no trusted measuring agent, and no actuarial model. [Interview: Preston, 2026-05-22] That absence is simultaneously the opportunity (build the measuring agent) and the reason it may not be buildable.

There is real precedent that the structure works: a U.S. company with a Philippines supplier triggered a tropical-cyclone CBI parametric and was paid in 1–2 weeks, with funds held in escrow and tiered sublimits by supplier tier. [Interview: Preston, 2026-05-22] And on the demand side, a Lloyd’s/WTW survey found 88% of 100+ semiconductor risk professionals consider supply-chain insurance “mission-critical,” while 81% cite a lack of available risk-transfer solutions — plus a documented case of a semiconductor company buying a parametric earthquake policy keyed to magnitude and distance from its supplier’s fab. [Public: Lloyd's/WTW; market-sizing-grand-slam]

5.2 The data angle and the MGA path

If the missing pillar is a trusted measuring agent, the natural question is who has the telemetry to be one. The candidate is PDF Solutions, whose fault-detection systems run in “every TSMC fab” and which owns dense per-wafer characterization data across hundreds of equipment-connectivity clients. [Interview: Andrzej Strojwas (PDF Solutions), 2026-05-22] The model to emulate is Munich Re’s performance-warranty reinsurance for batteries via TWAICE/Hithium — a data/monitoring partner enabling a device-triggered insurance product — and Coalition, the data-advantaged cyber MGA (~$3.5B). [Public: market-sizing; mga-intelligence] An MGA (managing general agent) structure would let TBD underwrite on a proprietary model using a reinsurer’s capital without becoming a carrier.

5.3 Disconfirming evidence (foregrounded)

(a) The buyer-side skeptic. Our strongest counter-signal came from Jeremy Jawish, CEO of Shift Technology (insurance AI): “the parametric market is still small and it’s not worth it — people are just not comfortable with parametric triggers,” and customers choose “best price over simplicity.” He also notes claims processing is only ~15% of premium cost, capping the value of payout-speed innovation. [Interview: Jeremy Jawish (Shift), 2026-05-22] That the two strongest insurance voices disagree — Preston (sell-side) bullish, Jawish (closer to buyers) bearish — is itself the finding.

(b) The moral-hazard separation problem. You cannot simultaneously be the measuring agent, the modeler, and the insurer — “you’re incentivized to have the model output a certain result.” [Interview: Preston, 2026-05-22] This directly complicates the elegant “digital twin underwrites its own products” vision: the very integration that would be our edge may be structurally disallowed.

(c) The precision-vs-marketability tension. The most accurate semiconductor triggers are multi-metric, but “the more simple you make it, the more backing you actually have from the marketplace.” [Interview: Preston, 2026-05-22] Low basis risk and reinsurer acceptance pull in opposite directions.

(d) Soft market and a thin book. Commercial property rates are down 25–30% over 2–3 years, so a novel structure cannot win on price [Interview: Preston, 2026-05-07]; and the limited count of U.S. fabs may be too small a book to sustain a focused insurance business — carriers diversify across sectors for exactly this reason. [Interview: Preston, 2026-05-22]

(e) No validated willingness-to-pay. Every WTP signal here is sell-side or survey-level; we have not yet heard a fab CFO or risk manager say “I would buy this at price X.” [Synthesis]

5.4 Market-sizing (low confidence, stated as such)

For calibration, our internal sizing work (explicitly low confidence) put parametric supply-chain insurance at a TAM of ~$19–21B (growing toward $48–64B by 2035), a semiconductor-specific SAM of ~$1–3B, and a realistic Year-1–3 SOM of ~$5–20M of gross written premium. The trading/benchmark opportunity (a price-reporting-agency-style business — “Platts of chips”) sized smaller: TAM ~$3–5B, SAM ~$200–800M, SOM ~$1–5M. [Synthesis: market-sizing-grand-slam] These are order-of-magnitude estimates from analogy, not bottom-up build-ups, and should be treated accordingly.

6. Synthesis: Cross-Cutting Tensions, a Tentative Lean, and Outstanding Questions

6.1 Three through-lines

  1. Secrecy is the constant. The opacity that killed compliance (§2) is the same opacity that threatens the data-asset moat behind financialization (§3) and the measuring-agent role behind insurance (§5). “A single leakage would probably mean the end of PDF” is not a compliance fact — it’s an industry fact. Any wedge that depends on aggregating proprietary data inherits this problem. [Interview: Andrzej Strojwas, 2026-05-22] [Synthesis]
  2. Every wedge is the same move: stop tying up capital against a risk; transfer the risk to someone better suited to hold it. Futures, parametric insurance, and warranty transfer are three expressions of one idea — operators buying certainty. This is the conceptual unity of the pivot, and it is worth stating as such for Professor Berk. [Synthesis]
  3. Thinness is a double threat. A market with three memory makers and ~five hyperscalers (a) may be too concentrated for an exchange or intermediary to exist (the oligopoly prefers opaque bilateral pricing), and (b) can compress the margin of any intermediary that does exist — “if NVIDIA says ‘I don’t want to pay your margin,’ our margin gets pounded down.” [Interview: Nihar, 2026-05-06] [Synthesis: financialization primer] Defensibility, not opportunity, is the scarce thing.

6.2 Convergences and divergences

Convergences (multiple independent sources agree):

  • Memory is the most commodity-like layer — Berk, Nihar, Minseok, Ronit, plus public spot-market data. [Multiple]
  • The warranty/reverse-logistics burden is real, large, and growing — Lonny + Alex internally; NVIDIA and AMD filings externally. [Multiple]
  • The binding risk for compute futures is adoption, not hedgeability — Spencer and Ronit, independently. [Multiple]

Divergences (flag prominently):

  • Parametric: viable vs. “not worth it” — Preston (sell-side) vs. Jawish (buyer-adjacent).
  • Failure rate 4% vs. 9% — Etched vs. Meta data; likely different denominators, unreconciled.
  • Advisor guidance — Ann Miura-Ko said defer derivatives/insurance and lead with compliance data; the pivot inverts this. Botha’s “companies have multiple founding moments” and “AI will be the biggest drainer of corporate moats in history” cut the other way (license the pivot; doubt any data/regulatory moat). [Interview: Ann Miura-Ko, 2026-03-06] [Interview: Roelof Botha, 2026-04-24]
  • The pivot is not internally settled — as of the May 20 summer-strategy session, Dustin placed the probability of staying in semiconductors below 50%. [Interview: Summer Strategy, 2026-05-20]

6.3 A tentative founders’ lean (clearly flagged — for us to adopt, edit, or cut)

⚠️ This subsection is a recommendation, which our research methodology says an agent should not make. It is included at the authors’ request as a starting point for our own judgment, and it is explicitly overwritable. The sections above are the evidence; this is one reading of it.

If forced to sequence today, the evidence points toward entering through the reverse-supply-chain / warranty pain (§4) rather than leading with a compute exchange (§3) or a fab insurance carrier (§5). The reasoning: (a) it is the only wedge with a named buyer actively trying to spend money (NVIDIA procuring tooling), which solves the cold-start problem that killed compliance; (b) it generalizes (AMD’s numbers); (c) operating the workflow is the most plausible legitimate way to earn the proprietary failure/usage data that the rest of the industry guards — turning the secrecy through-line from an obstacle into an entry path; and (d) that data is exactly what a warranty-risk-transfer / reinsurance product (the §4.2 / §5.2 bridge) would need, giving a credible path from a services/tooling beachhead to a financial product with better margins. In this reading, compute financialization (§3) is a market to participate in, not to found — CME, DRW, and Pluto already hold structural advantages (index data, balance sheet, regulatory designation) that a two-person team is unlikely to out-build, though the advisory role (§3.5) remains an open question. The biggest risks to even this lean are the two we cannot yet retire: no one has actually paid to transfer warranty risk (it’s our inference), and market thickness could compress margins regardless of where we sit. We hold this loosely.

6.4 Outstanding questions (tiered, with people who could answer them)

Tier 1 — could change the direction:

  1. Does $/GPU-hour actually solve the non-fungibility problem that killed DRAM futures three times, or just relocate it? → Spencer Powers (DRW), Ronit Jain (Pluto), Simon Moores (Benchmark Mineral Intelligence, as a PRA-building precedent).
  2. Will the natural buyer actually use a compute hedge — and who is first? → debt financiers to neoclouds; the lender Spencer offered to introduce (Joseph at Compute Exchange).
  3. Is anyone willing to pay to transfer the NVIDIA-style warranty liability? → Greg/Gregory and Greg DeLoccio (NVIDIA, introductions offered by Alex Zhu and Lonny); a structured-finance/reinsurance buyer.
  4. Is parametric fab/data-center insurance a real market or a structurally un-writable one? → Preston (Guy Carpenter) and his parametric specialist; a fab CFO or corporate risk manager (the missing buyer-side voice).

Tier 2 — sharpens the picture: 5. Does the reverse-logistics pain generalize to AMD/Broadcom operations, or is NVIDIA’s just poor execution? → contacts at AMD/Broadcom; an Expeditors account manager who sells across the industry. 6. How concentrated is the memory market really — what share trades spot vs. under LTA? → TrendForce/DRAMeXchange analysts; a memory procurement lead. 7. Could PDF Solutions’ data ever be the measuring agent for an insurance product without “leaking”? → John Kibarian (PDF CEO) and Kimon (product), introductions offered by Andrzej Strojwas. 8. Where does our defensible margin come from in a thin market? → the “Andrew Auerbach strategy” Dustin referenced (needs definition); Professor Berk.

6.5 Confidence summary

ClaimConfidenceBasis
Compliance died on buyer incentives + data opacity (not weak enforcement)HighMultiple converging interviews + enforcement data
Memory is the most commodity-like semiconductor layerHigh4 independent interviews + public spot-market data
NVIDIA reverse-logistics pain is real and NVIDIA is actively procuringHighTwo direct NVIDIA insiders, consistent with warranty filings
Warranty burden generalizes (at least to AMD)Medium-HighNVIDIA + AMD filings; Broadcom unclear
Compute futures are live and growingHighCME launch + DRW/Pluto interviews + policy backdrop
Adoption (not hedgeability) is the binding constraint for compute hedgesMedium-HighBoth builders independently
Parametric fab insurance is a viable wedgeLow-MediumSell-side bullish, buyer-side bearish; no WTP
Anyone will pay to transfer warranty riskLowOur inference; no validated buyer
A two-person team can found (vs. join) a compute exchangeLowIncumbents hold structural advantages

6.6 What would make the core thesis wrong?

The financialization thesis is wrong if (a) compute price volatility proves one-directional (prices only fall), removing the two-sided uncertainty hedging requires; (b) the non-fungibility problem reasserts at the GPU-hour layer as model generations churn, as it did for DRAM; (c) the markets stay thin enough that oligopolists keep pricing bilateral and refuse to feed any exchange or pay any intermediary’s margin; or (d) the warranty “inefficiency” turns out to be rational — NVIDIA keeps the reserve because no specialist can actually run its reverse chain better, collapsing the “peace of mind” half of the trade. Each is a question we can pursue with the contacts above, which is the point.

Sources

Internal (vault):

  • interviews/2026-05-08-390-jonathan-bliss-dustin.md (Berk session — anchor)
  • interviews/2026-05-12-lonny-orona.md; interviews/2026-05-22-alexbliss.md; inbox/2026-05-27-nvidia-reverse-logistics... (NVIDIA reverse logistics / Alex Zhu)
  • interviews/2026-05-22-45-mins-dustin-ross-and-spencer-powers.md (DRW); interviews/2026-05-22-30-min-meeting-between-ronit-jain-and-dustin-j-ross.md (Pluto); interviews/2026-05-22-word-vomit-post-etched-mtg.md (Etched)
  • interviews/2026-05-07-prestondustinbliss.md, interviews/2026-05-22-prestonbliss.md (Guy Carpenter); interviews/2026-05-22-intro-dustin-jeremy-ross-jeremy-jawish.md (Shift); interviews/2026-05-22-mtng-w-andrzej-strojwas-pdf...md (PDF Solutions)
  • interviews/2026-05-08-...yisroel.md, interviews/2026-04-30-josh-bliss-dustin.md, interviews/2026-05-01-nicole-x-bliss-x-dustin.md, interviews/2026-05-22-davidmattblissdustin.md (compliance kill)
  • interviews/2026-05-06-nihardustinbliss.md, interviews/2026-05-05-minseok-kim-bliss-dustin.md, interviews/2026-05-22-mo-islam.md (memory/compute)
  • interviews/2026-04-roelof-botha.md, interviews/2026-03-ann-miura-ko.md, interviews/2026-05-20-summer-strategy.md (advisor/strategy)
  • synthesis/glencore-of-semiconductors-2026-05-13.md, synthesis/reverse-supply-chain-research-2026-05-13.md, synthesis/market-sizing-grand-slam.md, synthesis/data-centers-research-2026-05-24.md, primer/dram-market-deep-dive.md, primer/financialization-primer-2026-05-29.md, primer/semis-risk-financial.md, primer/mga-intelligence.md
  • docs/GSBGEN_390_Petition_Answers (1).docx (original proposal)

Public:

Graph View