Project TBD Memory Vault

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

29 min read

From Compliance to Financialization

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

GSBGEN 390 Independent Study — Second Draft for Professor Jonathan Berk Dustin Ross & J Bliss Perry · Spring Quarter 2026 · revised 2026-05-31

Assembled from ~79 logged conversations and ~75 synthesis/primer memos (Jan–May 2026). Per our methodology, the document surfaces evidence and ends in questions; recommendations are flagged as a tentative founders’ lean. Source labels: [Interview: Name, Date], [Public: Source], [Synthesis], [Speculation].

0. Executive Summary

We set out to map the downstream semiconductor supply chain using AI-assisted extraction, with export-compliance as the commercial wedge. Two structural facts killed it. The data is deliberately opaque — secrecy is the business model of every intermediary who holds it. [Interview: Andrzej Strojwas (PDF Solutions), 2026-05-22] And commercial buyers are incentivized not to know — a chipmaker selling into a Singapore distributor that re-routes to China would rather not learn where its product ends up; knowing only costs sales. [Interview: Yisroel, 2026-05-08] [Interview: David/Matt (Shield Capital), 2026-05-22] Compliance pain is acute, but only for the buyer who must care: the U.S. government and the defense primes. We chose not to build defense-first, so we killed the wedge.

We turned to the financialization of the semiconductor and compute supply chain under one frame: operators buy certainty; speculators sell it. [Synthesis] Three product families emerged, each with a live anchor and disconfirming evidence: compute futures — CME and Silicon Data launched the first GPU-hour contracts on 2026-05-12, Pluto is building a CFTC-regulated exchange targeted for summer 2026 [Public: CME, 2026-05-12] [Interview: Ronit Jain (Pluto), 2026-05-22]; reverse supply chain & warranty — NVIDIA runs a $5T returns operation “on email and spreadsheets,” carries an $8.22B warranty reserve (≈20× FY2023), and is actively procuring outside tooling [Interview: Lonny Orona (NVIDIA), 2026-05-12] [Interview: Alex Zhu (NVIDIA), 2026-05-27]; and parametric / structured risk transfer for fab and data-center failure, anchored by Guy Carpenter reinsurance and tested against an insurance-software CEO who calls the parametric market “still small … people are not comfortable.” [Interview: Preston (Guy Carpenter), 2026-05-22] [Interview: Jeremy Jawish (Shift), 2026-05-22]

The two structural features that killed compliance return as the lens for every wedge. Secrecy — the same opacity that killed compliance threatens every data-asset moat behind financialization (“a single leakage would probably mean the end of PDF” [Interview: Strojwas]). Thinness — three memory makers and ~five hyperscalers may be too concentrated for any intermediary to exist, and intermediary margin can be “pounded down” by counterparties who refuse to pay it. [Synthesis: financialization primer] Each wedge section returns to those two tests; §5 reads the results. The report ends in the question they force: where, in a thin and secretive market, does durable position come from?

1. The Starting Thesis and Why We Abandoned It

Our GSBGEN 390 petition proposed compiling a database of the downstream semiconductor chain from public filings and interviews, using AI agents to extract at a scale a two-person team could not otherwise reach. [Public: GSBGEN 390 Petition, 2026] The intuition — seeded by Professor Berk — was that the upstream (fabrication, equipment, materials) is heavily studied while the downstream (who buys, through which distributors, under what terms, to what end markets) “is not well mapped in any single, structured resource.”

The commercial wedge was export-compliance. Chipmakers face mission-critical, high-penalty ship/no-ship decisions under EAR/ITAR; a platform that automated classification and licensing would (a) sell into real pain and (b) accrete the proprietary transaction data that would become the downstream map. Compliance was the Trojan horse for the data asset. [Synthesis] Enforcement was strong — Applied Materials fined $252M, Cadence $140M, against continuous AI/chip EAR rulemaking through 2024–25. [Synthesis] Ann Miura-Ko endorsed the sequencing: “focus on compliance data collection now, worry about derivatives and insurance later” — toward becoming “the JP Morgan of the industry.” [Interview: Ann Miura-Ko, 2026-03-06] Our eventual pivot inverts that advice; intellectual honesty requires owning it.

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

Finding 1 — The data is deliberately, structurally held. Opacity is the business model of the people who hold the data. Josh framed it directly: you mostly “just care about the last step,” with “300,000-plus components” to track, and intermediaries who could make the chain transparent have “no incentive” — “relationships beat data.” [Interview: Josh, 2026-04-30] PDF Solutions, on the densest manufacturing data in the industry: “a single leakage would probably mean the end of PDF.” [Interview: Strojwas, 2026-05-22] Operations teams want to share but “legal teams kill initiatives”; a prior multi-million-dollar academic effort to assemble this data reportedly failed at acquisition. [Interview: David/Matt (Shield), 2026-05-22]

Finding 2 — Commercial buyers are incentivized not to know. Yisroel: “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] Investors echoed: commercial semis “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, 2026-05-22] Two sealers: a semiconductor professional we spoke to had never heard of UFLPA; 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 works is 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.” [Interview: Josh] [Interview: Yisroel] We chose not to build defense-first, so compliance — alive in that niche — was killed for our commercial purposes.

The structural facts that killed it become the lens for everything that follows. Secrecy and concentration are permanent features of this industry, not bugs to be engineered away. They threaten every data-asset moat behind every financialization wedge below, which is why we test each one against them.

On the AI premise. The petition foregrounded AI agents extracting structured relationships from 10-Ks at scale. We did not build that pipeline. We built an AI-assisted research operating layer instead — a version-controlled memory vault, Claude Code skills for synthesis and transcript ingestion, semantic search across 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, not parsing filings. We flag the gap honestly; it is itself a finding about where AI currently adds leverage in qualitative industrial research (synthesis of primary conversation) and where the harder problem (reliable structured extraction from disclosure documents) remains.

2. Wedge 1: Commodity & Compute Financialization

Pre-test on the two structural features. Compute pricing is more public than fab data — the secrecy obstacle is index integrity, not data access. Compute is the least thin layer in the chain (hundreds of cloud buyers, dozens of neoclouds). The physical-chip variant — memory commodity trading — fails both tests harder: oligopolists keep pricing opaque to protect margin, and the market is famously thin. The live products cluster at GPU-hour, not DRAM, for these reasons.

2.1 The frame

Operators buy certainty; speculators sell it. An airline pays a trader to take jet-fuel price risk off its hands — giving up the windfall of cheap fuel in exchange for certainty. That trade seeds every instrument in this report. [Synthesis: financialization primer, 2026-05-29] The two structural tests, in Dustin’s framing: (1) materiality — is the cost big enough that buyers care; and (2) volatility — does the price move both ways, since hedging only has value against two-sided uncertainty. [Synthesis]

Three layers can be financialized: token ($/token of model output), compute ($/GPU-hour), and physical chip (memory the clearest case). [Synthesis; Interview: Adhi (5CC Capital), 2026-05-27] The recurring tension: the layer that behaves most like a commodity (memory) is also most controlled by a handful of suppliers — what historically kills exchanges.

2.2 The live market: compute futures

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

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

  • Ronit Jain (Pluto) is building a CFTC-designated derivatives exchange and clearinghouse, with physical settlement capability as the durable edge versus index-only competitors. ~15-month regulatory process; summer 2026 launch targeted. [Interview: Ronit Jain, 2026-05-22] (CFTC designation status unverified against public CFTC filings; treat as founder representation. [Speculation])

Three use cases — and who buys first:

  1. Hedging compute COGS. AI products, unlike 99%-margin SaaS, “actually do have cost of goods sold in their inference cost,” and that cost is volatile. [Interview: Spencer Powers]
  2. GPU collateralization for lending. A forward curve lets a lender treat GPUs as collateral over 3–5 years — underwriting the asset value rather than startup creditworthiness, as commercial real estate underwrites the building more than the tenant. Spencer and Ronit independently named debt financiers to neoclouds as the beachhead buyer. [Interview: Spencer Powers] [Interview: Ronit Jain]
  3. GPU price-depreciation insurance. The hardest revenue datapoint in the entire corpus: Pluto reports ~$60M of H200 depreciation coverage sold, structured as a put option, operated as a swap dealer (not insurance carrier), covering new-model releases, hardware advances, and geopolitical events including a Taiwan invasion; head of trading is a former UBS swaptions director. [Interview: Ronit Jain] This product sits across §2 and §4 — a useful reminder that the wedge boundaries are softer in practice.

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

Berk’s May 8 framing — “Glencore prints money because it’s a physical option holder at industrial scale” — set up the comparison. [Interview: Jonathan Berk, 2026-05-08] The comparison is mostly negative. Glencore’s edge rests on four things chips largely lack:

  1. Storage moat — oil storage is capital-intensive; “anyone can store semiconductors.” [Interview: Berk]
  2. Information from physical flow — Glencore handles ~4.2M barrels/day; in chips the valuable information sits inside fabs, accessed via “channel checkers in Taiwan.” [Public: S&P Global] [Interview: Nihar, 2026-05-06]
  3. Deep spot+futures markets — no semiconductor futures market has ever survived (see §2.4).
  4. Fungibility — “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]

Arrow and Avnet demonstrate the ceiling: $30.9B / $22.2B revenue at 1.9% / 1.1% net margins, and they don’t speculate — they hold consignment for customers. [Public: company filings] [Interview: Holly Rawlins (Renesas), 2026-04-29] During the 2020–22 shortage (~$200B in lost auto revenue), Arrow’s net income roughly tripled then fell below pre-shortage level by 2024 — captured some volatility but couldn’t hold it. [Public: MacroTrends]

The one segment where the commodity thesis has legs is memory (DRAM/NAND), and four independent sources converged: Berk, Nihar (hedge fund), Minseok Kim (ex-Samsung), Ronit Jain. [Multiple] 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, forecast +63% in Q2; NAND +55–60% then up to +75% — part of an AI-driven “memory supercycle.” [Public: TrendForce, 2026] OpenAI’s Stargate reportedly contracted up to ~900,000 DRAM wafers/month, on the order of 40% of global output. [Public: TrendForce, 2026]

But memory embodies the central tension in its sharpest form. 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,” likely too concentrated for any intermediary to insert itself. [Interview: Nihar] Oligopolists prefer opaque bilateral pricing because it protects margin — exactly why memory makers killed futures markets the last three times. And the fastest-growing memory segment, HBM, is moving in the opposite direction: co-designed with NVIDIA under long-term contracts, behaving “more like a specialty chemical” than a commodity. [Public: TrendForce] The commodity thesis may apply to a shrinking share of memory.

2.4 Disconfirming evidence (foregrounded)

Per our methodology — surprises are the signal — we lead the skeptical case.

(a) The graveyard of failed semiconductor futures. Physical semiconductor futures have been attempted at least three times and failed every time: Pacific Stock Exchange DRAM futures, 1989; Enron DRAM forwards, 2001 (died with Enron); SGX chip futures, 2003. [Public: Felix Stocker] No interviewee raised this — it surfaced only in our own research, which is itself the finding. Structural reason: non-fungibility plus product churn — the unit of sale keeps changing (256KB → 128MB → multi-GB), defeating contract standardization. The live CME/Pluto products implicitly bet that pricing at the $/GPU-hour service layer finally routes around this problem. It’s a bet, not a proof.

(b) COGS-blindness. 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 waking up every day thinking about how to hedge its GPU costs.” [Interview: Spencer Powers] [Interview: Ronit Jain] Ronit calls the core work “engineering the consumer behavior that’s necessary for this market to work.”

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

(d) Obsolescence vs. storability. Steve Blank’s structural objection: oil can be stored strategically, with the forward curve 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 — weakens the Glencore model specifically.

(e) Demand is not infinite. “Anytime you think demand is infinite, all you know is it’s not infinite” [Interview: Tim (Etched), 2026-05-22]; 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 the un-hedged downside the whole thesis assumes away.

2.5 The business-model question

What form would a TBD financialization business 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 — exchange or risk-taker); or (3) a capital-markets advisory / “investment bank for compute” that finds exposed companies, structures the hedge, and lays risk off to a market maker. Spencer raised the advisory model unprompted (“it basically sounds like an investment bank”) while warning it’s “not the sexiest startup.” [Interview: Spencer Powers] Tim at Etched independently pushed down the stack to component-level financialization — speculating on “each individual component of each chip.” [Interview: Tim (Etched), 2026-05-22] Open design question.

3. Wedge 2: Reverse Supply Chain & Warranty

Pre-test on the two structural features. Reverse-logistics data is currently trapped inside one company (NVIDIA), but NVIDIA is actively buying access to a third party who can operate it — which converts secrecy from a wall into an entry path. Thinness cuts the other way: with NVIDIA and (eventually) AMD as effectively the only buyers, intermediary margin is structurally vulnerable to “we don’t want to pay your margin.” The wedge is unique in that secrecy is passable but thinness remains.

3.1 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, ~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] Scaling math: 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.”

Alex Zhu, who leads parts of NVIDIA’s reverse supply chain, supplied the financial scale: NVIDIA carries roughly $8B in 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: “new buy is all Jensen cares about.” [Interview: Alex Zhu (NVIDIA), 2026-05-27]

3.2 Reconciling the numbers

  • Warranty. NVIDIA’s warranty reserve balance was $8.22B at end of FY2025, up from ~$416M in FY2023 — almost exactly the “20×” Alex described. Single-year accrual addition was $2.59B (vs. ~$1.75B for the entire rest of the U.S. semiconductor industry combined). 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] A warranty reserve is an accrued accounting liability, not necessarily segregated cash — which sharpens, rather than softens, the “is this capital being managed efficiently?” question. [Synthesis]
  • Failure rate. Sources cite both ~4% (“4% of NVIDIA GPUs fail upon reaching data centers” [Interview: Tim (Etched)]) and ~9% (annualized, from Meta’s Llama-3 run: 16,384 H100s, one failure every ~3 hours, ~80% hardware-related [Public: Meta Engineering, 2025]). Likely different denominators — early-life/arrival vs. annualized operational — present both with caveat.

3.3 Does it generalize beyond NVIDIA?

For: 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.

Against: 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] For the largest custom-ASIC vendor (Broadcom), we found no public warranty-reserve spike — open whether the warranty burden for custom silicon sits with the vendor or the hyperscaler customer. [Public: Broadcom filings] The honest read: this is a large and fast-growing niche (AI accelerators), not “all semiconductor reverse logistics.”

3.4 Chip repair feasibility (the naive questions, answered)

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

3.5 The two opportunities

  1. 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) actively procuring. The clearest “someone is trying to give us money” signal in the corpus. [Interview: Lonny Orona]
  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 against the reserve; the specialist wins on underwriting margin plus float. This is the bridge to §4 — and to Pluto’s depreciation product in §2.2 — and Max Mirgoli independently suggested studying “NVIDIA’s warranty claim size vs. revenue and the potential to reinsure that warranty risk.” [Interview: Max Mirgoli, 2026-05-22] Caveat: no interviewee is yet paying to transfer this risk — our inferred opportunity, not validated willingness-to-pay.

4. Wedge 3: Risk Transfer — Parametric Insurance & Structured Products

Pre-test on the two structural features. The measuring agent that any parametric product requires would need access to fab/data-center telemetry — the most guarded data in the industry. Reinsurance-market thickness is fine; semi-fab market thinness (a handful of U.S. fabs) means a focused book may not sustain a carrier. Secrecy is the binding obstacle here.

4.1 The anchor and the diagnostic

Preston (Guy Carpenter / Marsh McLennan), our richest insurance source, mapped the full reinsurance stack (insured → retail broker → carrier → reinsurance broker → reinsurer → retrocession → capital markets) and, unprompted, proposed the structure that organizes this 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.” Its advantage is speed and objectivity; its hazard is basis risk (trigger fires but no loss, or loss but no trigger).

Preston’s four-pillar test is the cleanest diagnostic we found: any parametric product needs 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, no actuarial model. [Interview: Preston] That absence is simultaneously the opportunity (build the measuring agent) and the reason it may not be buildable.

Real precedent that the structure works: a U.S. company with a Philippines supplier triggered a tropical-cyclone CBI parametric, paid in 1–2 weeks, with funds held in escrow and tiered sublimits by supplier tier. [Interview: Preston] On demand, 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]

4.2 The data angle and the MGA path

If the missing pillar is the measuring agent, who has the telemetry to be one? Candidate: PDF Solutions, whose fault-detection runs in “every TSMC fab” with dense per-wafer characterization across hundreds of equipment-connectivity clients. [Interview: Strojwas] Model to emulate: 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 lets TBD underwrite on a proprietary model using a reinsurer’s capital, without becoming a carrier.

4.3 Disconfirming evidence (foregrounded)

(a) The buyer-side skeptic. 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”; customers choose “best price over simplicity.” Claims processing is only ~15% of premium cost, capping the value of payout-speed innovation. [Interview: Jawish, 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 measuring agent, modeler, and insurer — “you’re incentivized to have the model output a certain result.” [Interview: Preston] 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) Precision vs. marketability. 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] Low basis risk and reinsurer acceptance pull in opposite directions.

(d) Soft market and a thin book. Commercial property rates down 25–30% over 2–3 years, so a novel structure cannot win on price [Interview: Preston]; 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.

(e) No validated willingness-to-pay. Every WTP signal 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.”

4.4 Market sizing (low confidence, stated as such)

Order-of-magnitude estimates from analogy, not bottom-up: parametric supply-chain insurance TAM ~$19–21B (→$48–64B by 2035), semi-specific SAM ~$1–3B, realistic Year-1–3 SOM ~$5–20M of gross written premium. The trading/benchmark opportunity (“Platts of chips”) sizes smaller: TAM ~$3–5B, SAM ~$200–800M, SOM ~$1–5M. [Synthesis: market-sizing-grand-slam] Treat accordingly.

5. Synthesis: The Two Tests, the Lean, and What Could Make Us Wrong

5.1 Three through-lines

  1. Secrecy is the constant. The opacity that killed compliance (§1) is the same opacity that threatens the data-asset moat behind financialization (§2) and the measuring-agent role behind insurance (§4). “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.
  2. Every wedge is the same move: stop tying up capital against a risk; transfer it to someone better suited to hold it. Futures, parametric, warranty transfer are three expressions of one idea — operators buying certainty.
  3. Thinness is a double threat. A market with three memory makers and ~five hyperscalers may be too concentrated for an exchange or intermediary to exist (oligopolists prefer opaque bilateral pricing), and can compress the margin of any intermediary that does — “if NVIDIA says ‘I don’t want to pay your margin,’ our margin gets pounded down.” [Interview: Nihar] Defensibility, not opportunity, is the scarce thing.

5.2 Convergences and divergences

Convergences:

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

Divergences (flag prominently):

  • Parametric: viable vs. “not worth it” — Preston (sell-side) vs. Jawish (buyer-adjacent).
  • Failure rate 4% vs. 9% — Etched vs. Meta; 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” cuts the other way (license the pivot; doubt any data/regulatory moat). [Interview: 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]

5.3 A tentative founders’ lean

⚠️ A recommendation, which our methodology says an agent should not make. Included at the authors’ request as a starting point — explicitly overwritable. Sections above are the evidence; this is one reading of it.

Forced to sequence today, the evidence points to entering through the reverse-supply-chain / warranty pain (§3) rather than leading with a compute exchange (§2) or a fab insurance carrier (§4). The argument:

(1) It is the only wedge with a named buyer actively trying to spend money (NVIDIA procuring tooling), solving the cold-start that killed compliance.

(2) It generalizes (AMD’s trajectory mirrors).

(3) Defensibility comes from data-by-operation, not data-by-acquisition. Operating the workflow is the most plausible legitimate way to earn proprietary failure/usage data that the rest of the industry guards — turning the secrecy through-line from obstacle into entry path. The defensible position is not the field-service layer (a contract manufacturer can bundle that) but the underwriting layer above it: earn data, model failure, price warranty-risk transfer that the operational players cannot themselves write.

(4) That data is exactly what a warranty-risk-transfer / reinsurance product (the §3.5 / §4.2 bridge) would need, giving a credible path from a services beachhead to a financial product with better margins.

Compute financialization (§2) is a market to participate in, not to found — CME, DRW, and Pluto already hold structural advantages (index data, balance sheet, regulatory designation) a two-person team is unlikely to out-build; the advisory role (§2.5) remains open.

Biggest risks we cannot retire: (a) no one has actually paid to transfer warranty risk — it’s our inference; (b) thinness could compress margins regardless of where we sit; and (c) NVIDIA could route reverse logistics through CMs — Wistron/Foxconn already operate the new Dallas line — bundling tooling with manufacturing and capturing the underlying data themselves, requiring us to partner with CMs rather than displace them. Each is a question we can pursue. Held loosely.

5.4 Outstanding questions (tiered)

Tier 1 — could change 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 — PRA precedent).
  2. Will the natural buyer actually use a compute hedge — and who is first? → debt financiers to neoclouds; Joseph at Compute Exchange (intro from Spencer).
  3. Is anyone willing to pay to transfer NVIDIA-style warranty liability? → Greg/Gregory and Greg DeLoccio at NVIDIA (intros from Alex Zhu and Lonny); a structured-finance / reinsurance buyer.
  4. Is parametric fab/data-center insurance a real market or structurally un-writable? → 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, or is NVIDIA’s just poor execution? → AMD/Broadcom contacts; Expeditors account manager. 6. How concentrated is memory 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 without “leaking”? → John Kibarian (PDF CEO) and Kimon (intros from Strojwas). 8. Where does defensible margin come from in a thin market? → “Andrew Auerbach strategy” Dustin referenced (needs definition); Professor Berk.

5.5 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, NVIDIA actively procuringHighTwo NVIDIA insiders, consistent with warranty filings
Warranty burden generalizes (at least to AMD)Med-HighNVIDIA + AMD filings; Broadcom unclear
Compute futures are live and growingHighCME launch + DRW/Pluto + policy backdrop
Adoption (not hedgeability) is binding constraint for compute hedgesMed-HighBoth builders independently
Parametric fab insurance is a viable wedgeLow-MedSell-side bullish, buyer-side bearish; no WTP
Anyone will pay to transfer warranty riskLowOur inference; no validated buyer
Two-person team can found (vs. join) a compute exchangeLowIncumbents hold structural advantages

5.6 What would make us wrong

The financialization thesis fails if (a) compute price volatility proves one-directional (prices only fall), removing the two-sided uncertainty hedging requires; (b) non-fungibility reasserts at the GPU-hour layer as model generations churn, as it did for DRAM; (c) 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.

Sources

Public:

Internal (vault) — anchor files:

  • Berk session anchor: interviews/2026-05-08-390-jonathan-bliss-dustin.md
  • NVIDIA reverse logistics: interviews/2026-05-12-lonny-orona.md, interviews/2026-05-27-nvidia-reverse-logistics-...md (Alex Zhu)
  • Compute futures: interviews/2026-05-22-...spencer-powers.md (DRW), interviews/2026-05-22-...ronit-jain...md (Pluto), interviews/2026-05-22-word-vomit-post-etched-mtg.md
  • Insurance / data: interviews/2026-05-07-prestondustinbliss.md, interviews/2026-05-22-prestonbliss.md, interviews/2026-05-22-...jeremy-jawish.md, interviews/2026-05-22-...andrzej-strojwas-pdf...md
  • Compliance kill: 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
  • Memory / compute: interviews/2026-05-06-nihardustinbliss.md, interviews/2026-05-05-minseok-kim-bliss-dustin.md, interviews/2026-05-22-mo-islam.md
  • Advisor / strategy: interviews/2026-04-roelof-botha.md, interviews/2026-03-ann-miura-ko.md, interviews/2026-05-20-summer-strategy.md
  • Synthesis briefs: 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
  • Original proposal: docs/GSBGEN_390_Petition_Answers (1).docx

Pressure-Test Log

Pressure-tested: 3 passes.
Pass 1→2: Merged the redundant §1 (starting thesis) + §2 (why abandoned) into one section;
          cut the "outline changes" preamble and the long methodology preface;
          compressed the executive summary from a numbered list to three through-line
          paragraphs (hostile-investor lens — body was repeating itself in the summary).
Pass 2→3: Sharpened the lean's defensibility argument (data-by-operation vs.
          data-by-acquisition; underwriting layer vs. field-service layer); added the
          contract-manufacturer-bundling scenario as an explicit retire-able risk;
          restored a brief candid note on the AI-premise gap (Professor-Berk lens —
          wanted defensibility teeth and methodology honesty for the audience he is).
Pass 3:   Convergence-gate scenario — "NVIDIA routes reverse logistics through
          Wistron/Foxconn (already happening on the Dallas line), CMs capture the
          underlying data" — survives: the underwriting layer is data-derived but
          doesn't require owning operations; partnering with CMs is a viable fallback,
          and the risk is now named explicitly in §5.3. No material changes → converged.

Graph View