NVIDIA Primer for the Chip Warranty / Reverse-Logistics / Financialization Pitch

Prepared per RDI methodology. Surfaces evidence; does not conclude. Every non-trivial claim is source-labeled. Anchored to the 2026-06-17 Greg + Lonny meeting and the 2026-06-19 bullet-doc deadline.

Outline changes (read first)

The Phase 0 outline was approved with edits — collapse old §5+§6 into one merged section, keep §7 as forward-looking closer. That structure is preserved. Additional revisions made during Phase 1:

§1 — added Run:ai / Mission Control / DGX Cloud detail as a competitive-overlap dimension. The “what NVIDIA already owns in fleet management” question turned out to matter for the pitch: NVIDIA acquired Run:ai for $700M (closed 2024-12-30) and folded it into Mission Control, which is now NVIDIA’s branded forward-deployed fleet-health product. Our pitch sits in the reverse layer Mission Control deliberately does not cover. This is a real surprise worth surfacing.
§2 — split “upstream supplier stack” from “reverse-flow stack” within the same section rather than interleaving. The forward stack (TSMC → HBM → ABF substrate → OSAT → ODM) and the reverse stack (Foxconn/Wistron repair, Expeditors 3PL, parallel ecosystem) are different sets of partners with different liabilities; clearer to treat them as two columns.
§4 — added concrete RFP-rival shape. We confirmed Greg/Lonny named Palantir, Accenture, and build-internal as the three finalists [Interview: Greg + Lonny, 2026-06-17 — debrief]. The broader competitive map (Servigistics, ServiceMax, etc.) is treated as the category NVIDIA could turn to next, not as current RFP finalists. Don’t confuse the two.
§5 — moved the WarrantyWeek $8.22B vs. NVIDIA 10-K $2.81B correction up front in the balance-sheet half. This is the single most consequential factual correction in the file [[reverse-logistics-warranty-tam-2026-05-29]] and the pitch deck needs to land on the filed number.
§6 — sharpened the exec-mapping path. Joe Malchow’s Shanker Trivedi offer is high-value but Trivedi retired April 2026 [Public: SDxCentral]. He is “fresh outside the company” rather than active inside — useful as warm intro to current EVP Sales / Debora Shoquist’s ops org, not a current insider. Pitch language should reflect that.

§1 — Revenue lines, segment economics, and where reverse-logistics pain concentrates

BLUF: Data-center silicon is 90% of NVIDIA’s revenue, the explicit warranty driver in the 10-K, and the layer where reverse-logistics dollars compound — which is why Greg + Lonny’s pain is real and at the level of a board-material problem, not an ops-team annoyance.

Segment split and FY2026 financials

NVIDIA reports two operating segments: Compute & Networking (data-center accelerators, networking, AI software, automotive) and Graphics (GeForce gaming, Quadro/RTX workstation) [Public: NVIDIA FY2026 10-K, accession 0001045810-26-000021, filed 2026-02-25].

Line	FY2026	FY2025	YoY
Total revenue	$215.9B	~$130.5B	+65%
Data Center revenue	$193.7B	$115.2B	+68%
Data Center share of total	89.7%	88.3%	—
Gaming revenue	~$16.0B	~$11.4B	~+40%
Total operating income	$60.4B	—	—
Compute & Networking segment OI	absorbed $4.5B H20 inventory/purchase-obligation charge in Q1 FY26 (export-controls hangover)	—	—
Blackwell architecture share of data center	”majority of data center revenue” in FY26	—	—

[Public: NVIDIA FY2026 Q4 earnings release; ServeTheHome FY26 Q4 recap; NVIDIA FY2026 10-K filed 2026-02-25]

Why this matters for the pitch. Lonny said “we’re not even armed” when talking to customers [Interview: Greg + Lonny, 2026-06-17]. The “customer” here is overwhelmingly a data-center buyer of GB200/GB300 systems — the same population that is also driving the warranty curve explicitly named in the 10-K. The cost center we’re trying to address is not a corner of NVIDIA’s P&L. It is the cost center on top of the segment that is the company.

Product mix in data-center

H100 / H200 / B100 / B200 / GB200 are the relevant SKUs Greg + Lonny live with. GB200 NVL72 entered mass production by early 2026; NVIDIA had already begun the transition to GB300 NVL72 (Blackwell Ultra) [Public: IntuitionLabs GB200 supply-chain analysis; 36Kr 2026]. The B200 carries 192 GB HBM3e (+140% over H100’s 80 GB) [Public: EnkiAI 2026 HBM analysis] — a fact that matters in §5 because more HBM stacks per package means more interfaces that can fail.

Customer-concentration footnote (it is severe)

From the FY26 10-K customer-concentration disclosure and FY26 earnings color:

Two direct customers ≈ 36% of FY26 revenue. [Public: DCD reporting on NVIDIA 10-K; Daloopa]
Four direct customers each >10%; collectively 61%. [Public: NVIDIA FY26 10-K customer-concentration footnote, via Motley Fool / Daloopa]
Top single customer ≈ 22% of revenue. [Public: NVIDIA FY26 10-K]
Top 5 hyperscalers ≈ “a little over 50%” of revenue per CFO Colette Kress on Q4 FY26 earnings call [Public: NVIDIA Q4 FY26 earnings call transcript].
Jensen at GTC 2026: “about 40% of revenue comes from outside the top 5 hyperscalers” — i.e., the non-hyperscaler share grew materially as sovereign / neocloud (CoreWeave, xAI) absorbed allocation [Public: GTC 2026 keynote].

NVIDIA does not name the customers. Public reporting and Alex Zhu’s call in May 2026 jointly identify the population: Microsoft, Google, Meta, Amazon (the four hyperscalers); CoreWeave, Oracle (neoclouds); xAI, OpenAI (frontier labs); plus sovereigns [Interview: Alex Zhu, 2026-05-27; Public: DCD/CIO Dive 2026]. CoreWeave specifically has a $6.3B capacity agreement with NVIDIA through 2032 (NVIDIA agreed to buy any unused capacity), and OpenAI doubled CoreWeave compute commitments to $22.4B [Public: CoreWeave press release 2025; IO Fund analysis].

Where warranty / RMA cost concentrates: Compute & Networking, explicitly

The FY26 10-K product-warranty footnote attributes additions in product warranty liabilities “primarily related to Compute & Networking segment” — i.e., data-center, not consumer GPUs [Public: NVIDIA FY26 10-K product-warranty footnote; cross-checked in [[reverse-logistics-warranty-tam-2026-05-29]]]. The tech-press narrative blaming consumer RTX 50-series 16-pin connector melt is a separate story; the dollars in NVIDIA’s own filing are data-center dollars [Synthesis].

This converges sharply with Greg + Lonny’s hyperscaler examples — the Phoenix five-truck event, the “16 people on the reply-all,” the 30-day SLA, the “200 units” Lonny used as the example return tranche — all describe data-center RMA, not gaming [Interview: Greg + Lonny, 2026-06-17].

Adjacent software lines that touch chip telemetry / fleet health (the competitive-overlap dimension)

NVIDIA already monetizes fleet management at the forward layer. Important to know going in:

Product	What it does	Status
NVIDIA Mission Control	End-to-end AI factory infrastructure mgmt: provisioning, monitoring, error diagnosis. Branded fleet-health surface	GA on Blackwell systems `[Public: NVIDIA Mission Control docs; NVIDIA blog 2025]`
NVIDIA Run:ai	GPU workload orchestration, scheduling, utilization. Acquired Apr 2024, closed 2024-12-30 for ~$700M; open-sourced post-close	Folded into Mission Control to “boost GPU availability and utilization” `[Public: VentureBeat 2024; SiliconAngle 2024; Tom's Hardware 2024-12-30]`
NVIDIA Base Command Manager (BCM)	Cluster provisioning / management; ex-Bright Computing (acquired 2022-01-10)	GA, integrated with Mission Control `[Public: Bright/BCM docs]`
NVIDIA DGX Cloud	Hosted DGX-class compute, full-stack	Subscription; NIM tier free on DGX Cloud for prototyping; production needs AI Enterprise license `[Public: NVIDIA DGX docs]`
NVIDIA AI Enterprise	Software platform — NIM microservices, SDKs, GPU drivers, k8s operators, support SLA	Subscription `[Public: NVIDIA AI Enterprise docs]`
NVIDIA Fleet Command	Hybrid-cloud platform for edge AI deployment across dozens to thousands of devices	GA; open-sourced repo `[Public: github.com/NVIDIA/fleet-command]`
NVIDIA NIM microservices	Containerized inference microservices	Distributed under AI Enterprise `[Public: NVIDIA docs]`

The surprise: NVIDIA owns the forward fleet-management layer and runs it in production. The reverse layer is where Lonny’s “we’re not even armed” lives. Mission Control + Run:ai is what a hyperscaler runs while a unit is in their data center. Lonny’s portal is what NVIDIA itself wants for after a unit fails — RMA case state, ASN status, repair turnaround, escalation triggers. The reason these don’t overlap is structural: customers strip telemetry before returning chips, so Mission Control’s signals stop at the customer fence line [Interview: Greg + Lonny, 2026-06-17]. We are not competing with Mission Control. We are downstream of the boundary Mission Control hits. [Synthesis]

NVIDIA’s M&A pattern reinforces the read that the company will buy fleet/management tooling when it views the layer as strategic — Bright Computing (cluster mgmt, 2022), Excelero (high-performance storage, 2022), SwiftStack (multi-cloud data mgmt, 2020), Run:ai ($700M, 2024). Reverse logistics has not yet seen a comparable acquisition [Public: CIO Influence M&A roundup; DCD].

Sources for §1: Greg + Lonny @ NVIDIA, 2026-06-17; reverse-logistics-warranty-tam-2026-05-29; NVIDIA FY26 10-K; NVIDIA Mission Control / Run:ai materials.

§2 — Upstream supplier stack: who feeds NVIDIA, and who touches a chip on the way back

BLUF: NVIDIA owns design and the customer relationship; everyone else in the silicon path is a contracted partner. Under UCC §2-719 component-liability terms, when a chip fails in a hyperscaler’s data center, the cost flows back up to NVIDIA — not down into TSMC, HBM vendors, or system integrators. That is why §1’s warranty reserve concentrates here.

Forward stack — making a GB200

Layer	Named partner	2026 capacity / share	Source
Logic fab (B100/B200 die, Blackwell)	TSMC (Taiwan; AZ fab ramping)	N3 fully booked Q1 2026	`[Public: Silicon Analysts 2026]`
CoWoS advanced packaging	TSMC CoWoS-L (Blackwell) and CoWoS-S (legacy H100/H200)	Total CoWoS 75K → 120–130K wafers/mo by end-2026; NVIDIA secured ~60% of total CoWoS in 2026 (515K wafers, ~510K CoWoS-L). Demand exceeds supply by 40–50%	`[Public: TrendForce 2025-12; DigiTimes 2025-12-10; Astute Group; NextWaves Insight]`
HBM (high-bandwidth memory)	SK Hynix ~50–55% share; Samsung 35–40%; Micron 5–10%	Entire 2026 HBM production sold out at SK Hynix and Micron; SK Hynix has ~2/3 of HBM allocation for Vera Rubin (next gen after Blackwell)	`[Public: EnkiAI 2026; Silicon Analysts; SoftwareSeni 2026]`
ABF substrate (organic build-up)	Unimicron, Ibiden, Nan Ya PCB, Shinko Electric, AT&S — five-firm oligopoly, ~70% market share	Ajinomoto Build-up Film mandatory dielectric	`[Public: IntuitionLabs GB200 supply chain; SemiconductorX; Nikh substrate substack]`
Backend OSAT (assembly/test outside TSMC CoWoS)	ASE (Taiwan); Amkor (Korea/US — $7B Arizona facility); SPIL (ASE subsidiary)	ASE + Amkor + SPIL share ~80K wafers of CoWoS capacity for NVIDIA in 2026, mainly Vera CPU + automotive; Amkor handles US packaging of future Blackwell variants	`[Public: 36Kr 2026; CW Newsroom 2026; TEEPTRAK OSAT analysis 2027]`
System integration (rack-scale: GB200 NVL72)	Foxconn / Hon Hai (Houston), Wistron (Dallas/Fort Worth), Quanta, Inventec, Compal, Pegatron, Wiwynn	Foxconn 24% and Wistron 5% of NVIDIA’s 2023 global GPU server shipments per Morgan Stanley	`[Public: CommonWealth Magazine via Morgan Stanley; cross-ref [[reverse-supply-chain-research-2026-05-13]]]`

NVIDIA committed up to $500B in US AI infrastructure spend over four years with TSMC, Foxconn, Wistron, Amkor, SPIL [Public: Dallas Innovates; The Texan]. Wistron’s Fort Worth (15200 Heritage Parkway) plant: 323,700 sq ft, $150M, construction began Aug 2025, scheduled completion April 2026 [Public: Fort Worth Report 2025-07-10]. Public reporting describes this facility as production; Lonny described the Dallas site as a repair line going live July 2026 [Interview: Lonny Orona, 2026-05-12]. The two are likely co-located but distinct workstreams; [[reverse-supply-chain-research-2026-05-13]] flagged this exact divergence. Worth clarifying with Lonny. Update (Greg 2026-06-26): Dallas, Houston, and Guadalajara are all confirmed compute repair nodes — which Wistron building and whether the repair SOW is separate from production is still the open question.

Reverse stack — who actually touches a returning chip

NVIDIA does not run its own repair lines. Returns route through CMs (currently Wistron + Foxconn for NVIDIA-dedicated lines) under a playbook NVIDIA specifies: “we give them the playbook. Here’s how you’re going to diagnose it. Here’s how you’re going to repair it. And here’s your acceptance criteria” [Interview: Lonny Orona, 2026-05-12].

Function	Confirmed (NVIDIA-specific)	Adjacent ecosystem (capability exists, NVIDIA-attribution unconfirmed)
Repair line — compute	Foxconn + Wistron: Dallas, Houston, Guadalajara (Greg 2026-06-26); back-office Hong Kong, warehouse Taiwan; European compute returns to the US	Jabil/Retronix (component reclamation, acquired 2023-11); Celestica/NCS Global (ITAD/ITAM, $56M acq. ~2024); Flex Aftermarket Services; Sanmina Global Services (23 logistics/repair centers, OEM warranty mgmt)
Repair line — networking	Separate flow (Greg 2026-06-26): Vietnam, Israel (post-Mellanox), India — switches carry no GPUs	—
Failure analysis	NVIDIA frontline triage (Lonny’s org US team + Israel/Mellanox team) + failure-analysis lab for unknown failure codes (Greg 2026-06-26)	EAG Laboratories; Thermo Fisher; Infinita Lab; Intech Technologies
Reverse 3PL	Expeditors Dallas (replacing Omni in the US); Omni now owned by Forward Air, possibly retained in Asia (Greg 2026-06-26)	DHL Supply Chain; FedEx Logistics; UPS Supply Chain Solutions; DB Schenker
Component reclamation / chip salvage	NVIDIA salvages GPU dies “a very, very high percentage of the time,” reuses on other assemblies	Reconext; PanurgyOEM; Green Wave; Ingram Micro Lifecycle
Secondary GPU market / certified-refurb	Not directly run by NVIDIA	Dell Recertified, HPE Renew, Supermicro refurb; ALTA Technologies (used H100); ITAMG; SK tes

[Interview: Lonny Orona, 2026-05-12; Alex Zhu, 2026-05-27; Public: Jabil/Celestica/Flex/Sanmina IR; cross-ref [[reverse-supply-chain-research-2026-05-13]] and [[independent-distributors-research-2026-05-13]]]

Liability allocation — why cost flows back to NVIDIA

Standard semiconductor supplier terms cap liability at the purchase price of the component (sometimes as little as ~2%) and exclude consequential/indirect damages (enforceable under UCC §2-719); supplier remedy is repair/replace/refund, not downstream system value or the customer’s lost compute revenue [Public: Law Insider warranty-cap clause library; Stevens & Bolton; Lexology, accessed 2026-05; cross-ref [[reverse-logistics-warranty-tam-2026-05-29]]]. NVIDIA owns the consignment inventory at its CMs [Interview: Alex Zhu, 2026-05-27] and runs an Advance Replacement RMA (ARMA) where NVIDIA ships the replacement before receiving the defective unit and pays shipping both directions [Public: NVIDIA Enterprise Support User Guide]. The customer must return within 10 days [Public: NVIDIA Enterprise Support docs].

The whole-system OEMs (Dell, HPE, Supermicro) do not absorb this:

Vendor	Latest FY warranty reserve	YoY direction
Dell	$450M (FY26)	Essentially flat $467M → $426M → $424M → $450M `[Public: Dell FY26 10-K accn 0001571996-26-000008]`
HPE	$284M (FY25)	Declining $318M → $301M → $284M `[Public: HPE FY25 10-K accn 0001645590-25-000130]`
Supermicro	$17.0M (FY25)	Flat, ~0.2–0.3% accrual rate on >$20B revenue (pure-play AI server) `[Public: SMCI FY25 10-K accn 0001375365-25-000027]`

That is, the most AI-exposed integrator (Supermicro) carries a $17M reserve while NVIDIA’s is **$2.8B** [Public: NVIDIA FY26 10-K product-warranty footnote]. The cost is funneling up to NVIDIA, not distributing down [Synthesis; cross-ref [[reverse-logistics-warranty-tam-2026-05-29]] §3].

Sources for §2: 2026-06-17-logistics-catch-up-wgreg-lonny-nvidia; 2026-05-12-lonny-orona; Alex Zhu, 2026-05-27; reverse-supply-chain-research-2026-05-13; substrate-research-2026-04-17; reverse-logistics-warranty-tam-2026-05-29; primary 10-Ks; trade press cited inline.

§3 — Customer relationships: allocation, account coverage, and the buyers with leverage

BLUF: The customer set is small, named, and concentrated. The “buyers with leverage” — Microsoft, Meta, Google, CoreWeave — are also the customers Lonny said NVIDIA “comes to the conversation not armed” against. The buying center for our deal sits in service ops; the customer side of the leverage equation is hyperscaler procurement and SRE, and we should track exec-coverage org accordingly.

Named customers — who has the allocation

The constrained-supply environment makes allocation the central commercial fact. Identifiable populations:

Tier 1 hyperscalers (>10% of revenue each per FY26 10-K) — Microsoft, Meta, Google, Amazon [Public: NVIDIA FY26 10-K customer concentration; Daloopa 2026; CIO Dive 2026]. Lonny’s call cited Meta explicitly as the example data-center customer he had worked across the table at while at Meta and now at NVIDIA [Interview: Lonny Orona, 2026-05-12].

Tier 2 neoclouds / frontier labs — CoreWeave ($6.3B capacity agreement with NVIDIA through 2032, $2B NVIDIA investment Jan 2026, $22.4B OpenAI commitment, $14.2B Meta deal through Dec 2031, 1.7 GW active power target by end-2026); xAI, OpenAI [Public: CoreWeave press releases 2025–26; IO Fund]. Alex Zhu named Meta, Google, Core Wave, xAI, OpenAI as the major CSPs his reverse-logistics org services [Interview: Alex Zhu, 2026-05-27].

Tier 3 sovereigns — UK ($11B AI factories, 120K GPUs target by 2026); various national programs [Public: trade-press summary]. Lonny: “we’re accelerating into the government entity market” and was on a call about it earlier that week [Interview: Greg + Lonny, 2026-06-17]. Flag for the human (not for synthesis): government/defense is the killed-compliance-wedge territory per CLAUDE.md — surface, do not reframe.

Allocation mechanics under constraint

CoWoS demand exceeds supply 40–50% even as TSMC quadruples capacity by end-2026 [Public: NextWaves Insight; TrendForce 2025-12-08]. HBM is sold out at SK Hynix and Micron for 2026 [Public: EnkiAI 2026]. In that environment, who gets the next 100K GPUs is decided by long-term commitments and account leverage — not by a published price list. The four tier-1 hyperscalers locked multi-billion-dollar forward orders for GB200/B200 through end-2026 and into 2027 [Public: SoftwareSeni 2026]. CoreWeave, Oracle, xAI, OpenAI, Meta have moved ~$120B of AI infrastructure debt off balance sheet using SPVs funded by Wall Street [Public: Quartz 2025].

Implication for the reverse-flow pitch. The same customers fighting for forward allocation are the ones holding NVIDIA’s feet to the fire on RMA SLA. Lonny: “these hyperscalers, they’re different animals. I come from one. They are not shy to pound for what they want. They spend a lot of money” [Interview: Greg + Lonny, 2026-06-17]. Aggressiveness on returns is the same aggressiveness that secures their next 100K GPU allocation. The two negotiations run in parallel.

Account coverage org — mapping for the deal

The customer-facing exec coverage org is in flux:

Role	Person	Status	Relevance to our deal
EVP Enterprise Sales / Enterprise Business	Shanker Trivedi	Retired April 2026 after 17 years (joined NVIDIA 2009; SVP 2016; ex-SAP CMO) `[Public: SDxCentral 2026; LinkedIn]`	Joe Malchow’s warm intro offer `[Interview: Joe Malchow, 2026-06-17]`. Now on NPH board. SAP background is directly relevant given NVIDIA’s $2M+ SAP planning deal in flight `[Interview: Alex Zhu, 2026-05-27]`. Useful as warm outside voice with current-team access; not an active inside sponsor.
EVP Operations	Debora Shoquist	Active; joined NVIDIA 2007; oversees manufacturing, foundry ops, supplier management, CM management, supply planning, logistics, quality	`[Public: NVIDIA Newsroom bio; The Org]`
EVP CFO	Colette Kress	Active; joined NVIDIA 2013; product-warranty reserve sits on her balance sheet	`[Public: NVIDIA Newsroom]`
Senior Director, Service Team	Greg Dalcello	Active; our exec sponsor; ex-Cisco 22 yr; ex-My Next Power 2.5 yr	`[Interview: Greg + Lonny, 2026-06-17]`
Customer Service Operations lead (compute science frontline)	Lonny Orona	Active; ex-Meta 8 yr; 8 weeks in as of 2026-05-12	`[Interview: Lonny Orona, 2026-05-12]`
Reverse supply chain operations	Alex Zhu	Active; 2 yr at NVIDIA; ex-Apple, ex-consulting	`[Interview: Alex Zhu, 2026-05-27]`
Head of service-supply-chain org (above Greg + Lonny)	“Manu”	Named 2026-06-26; Greg + Lonny are peers reporting to him; his org has four functions (customer service, reverse logistics + warehousing, repair ops, planning) + Greg as horizontal “fifth wheel.” Sits under Shoquist per 2026-06-25 in-person debrief. Spelling/LinkedIn ID being confirmed via Ali	`[Interview: Greg, 2026-06-26; 2026-06-25 in-person debrief]`

Org-chart implication: Greg + Lonny sit inside customer service ops, which in NVIDIA’s structure rolls up under Operations (Shoquist) given the supplier/CM/quality scope of her remit [Public: NVIDIA Newsroom bio; Synthesis]. The decision authority for a $2M+ SaaS deal (Alex Zhu’s SAP comparison) is plausibly at the SVP / EVP layer above Greg.

How customer-service ops interfaces with the front-line account team

This is the friction Lonny named most directly: customer requests come in from all directions — through the SFDC portal, through account managers, through whoever the customer happened to deal with last on a different ticket. Greg: “that entry into NVIDIA could be coming from all different types of tentacles, so different quality engineers, leadership… if there’s not a consistent way in which we’re looking at the same data and able to communicate a consistent status, you get death by a thousand emails” [Interview: Greg + Lonny, 2026-06-17]. The dashboard we’re scoping is, in part, an exec-coverage tool — it gives whoever the customer calls a place to look first instead of restarting the email chain [Synthesis].

Sources for §3: 2026-06-17-logistics-catch-up-wgreg-lonny-nvidia; Greg en-route call 2026-06-26; 2026-05-12-lonny-orona; 2026-05-27-nvidia-reverse-logistics-supply-chain-discussion; Joe Malchow, 2026-06-17; NVIDIA FY26 10-K; SDxCentral; CoreWeave IR.

§4 — Software procurement: who NVIDIA already buys from, and what we’d be competing against

BLUF: NVIDIA’s current ops stack is SFDC + SAP + Baxter + Expeditors with $2M+ already committed to SAP for planning automation. Three RFP finalists are named (Palantir, Accenture, build-internal). The broader category — Servigistics, ServiceMax, Syncron, ReverseLogix, Optoro/Blue Yonder — exists but is not in the current RFP. The path NVIDIA has historically used to absorb startups (Inception → acquisition) is a credible alternative end-state and worth referencing in the bullet doc.

Named incumbents in NVIDIA’s reverse-ops stack

System	Function at NVIDIA	Status / spend	Source
Salesforce (SFDC)	Front-end customer portal + case management	In production; “everything downstream happens outside the system via email and spreadsheets”	`[Interview: Greg + Lonny, 2026-06-17]`
SAP	Material planning, ERP back end	$2M+ planning automation deal in flight	`[Interview: Alex Zhu, 2026-05-27]`
Baxter Planning (BaxterPredict)	Service-parts demand planning, failure-rate forecasting	In use; founded 1993 by ex-TI service-parts planner; Marlin majority 2024; customers manage $11B+ inventory across 35K locations / 120 countries	`[Interview: Lonny Orona, 2026-05-12; Public: Baxter Planning corporate; PRNewswire/Marlin 2024]`
Expeditors International	3PL, replacing Omni	Non-asset-based; 340+ locations / 100+ countries; reverse logistics + warehousing	`[Interview: Lonny Orona, 2026-05-12; Public: Expeditors corporate]`
WMS + TMS	Warehouse + transportation management — rolling out now, including into 3PL partners	In rollout; mirrors Cisco’s 2007–09 ERP transformation per Greg	`[Interview: Greg + Lonny, 2026-06-17]`
Ernst & Young	Workshop facilitator that defined the two big initiatives (this dashboard is one of them)	Workshop ran ~2 weeks before the 2026-06-17 meeting	`[Interview: Greg + Lonny, 2026-06-17]`

Procurement authority chain

Lonny’s stated procurement posture is unambiguous and the strongest buying-signal language we’ve heard from any NVIDIA contact: “We have no time for in-house tooling. If these tools don’t generate revenue, why would we spend time building them?” and “No way we can afford to spend any time developing something ourselves — it’s just not scalable” [Interview: Lonny Orona, 2026-05-12].

The authority chain on this specific deal:

Lonny owns customer-service ops (Lonny’s pillar in the four-pillar org). [Interview: Lonny Orona, 2026-05-12]
Greg owns systems integration and transformation across all four pillars. Senior Director. [Interview: Greg + Lonny, 2026-06-17]
An unidentified exec sponsor above Greg/Lonny holds the budget. Per Joe Malchow’s framing, the deal closes when the bullet doc reaches that level and lands [Interview: Joe Malchow, 2026-06-17].

Entry paths that exist but are not the path:

NVIDIA Inception — free startup program; co-marketing, exclusive pricing, VC exposure. Joe: “companies Joe has worked with have had good experiences, but it’s a formalized process (a dozen companies per year, YC-style)… we already have what Inception companies hope to get: a hair-on-fire problem with an executive sponsor” [Interview: Joe Malchow, 2026-06-17]. Recommendation per Joe: keep warm, don’t spend time on it.
NVentures — NVIDIA’s corporate VC arm. Not engaged here.

Confirmed RFP rivals on this specific deal

Per Lonny + Greg debrief: Palantir, Accenture, build-internal [Interview: Greg + Lonny debrief, 2026-06-17]. Bliss came from Palantir. The pitch logic Joe surfaced is that we beat Palantir’s SaaS-platform pitch and Accenture’s billable-hours pitch by being structurally different — “two solutions people + arm’s-length independent data repository + milestone economics” — rather than by being a better version of either [Interview: Joe Malchow, 2026-06-17; [[nvidia-engagement-bullets-2026-06-17]]].

Broader competitive category (if NVIDIA expands the search)

Not currently in the RFP, but if NVIDIA broadens, these are the credible category players:

Vendor	What they do	Status / scale
PTC Servigistics	Service-parts planning & optimization, “98.5% part availability with 30% less inventory” — the most direct overlap with Baxter	Acquired by PTC 2012; Servigistics 14.0 launched 2025 with industrial-AI focus `[Public: PTC IR; PTC blog 2025]`
PTC ServiceMax	Field service execution & dispatch, work orders, technician scheduling	Acquired by PTC 2023 for $1.46B; still PTC-owned (correcting the prior assumption in reverse-logistics-warranty-tam-2026-05-29 that PTC sold ServiceMax) `[Public: PTC FY25 8-K]`
Syncron	Aftermarket service lifecycle (parts, pricing, warranty)	~$144.5M revenue 2024, ~$433.5M valuation `[Public: GetLatka 2024 — estimate]`
ReverseLogix	End-to-end returns management SaaS	~$25M revenue, ~$20M raised, ~$75M valuation `[Public: GetLatka 2024 — estimate]`
Optoro	Returns SaaS	$434M raised; acquired by Blue Yonder Aug 2025 at undisclosed/likely modest price — cautionary comp `[Public: Optoro/Blue Yonder 2025]`
SAP Industry Cloud / Salesforce Industry Cloud	Vertical modules layered on existing ERP/CRM contracts	The stealth alternative — could subsume the workflow inside contracts NVIDIA already has

No vendor in this list specifically solves the reverse-logistics + repair workflow + service-parts planning + 3PL/logistics integration for semiconductor / data-center hardware as a unified surface [Synthesis; [[reverse-logistics-warranty-tam-2026-05-29]] §5]. The gap Lonny described — “integrating these pipelines of data into a representation of a portal and a view that customers can go quickly see the state of their business” — is real and not solved by any single one of these.

Startup-to-NVIDIA deal velocity (how NVIDIA absorbs the right startups)

NVIDIA’s acquisition pattern on adjacent software is fast when the strategic fit is clean:

Target	Year	Purpose	Notes
Run:ai	Apr 2024 announce, closed Dec 30, 2024, ~$700M	GPU workload orchestration	EU antitrust held it up; open-sourced post-close `[Public: VentureBeat; Tom's Hardware 2024-12-30; SiliconAngle 2024]`
OmniML	2023	Edge ML model compression	NVIDIA’s biggest 2023 acquisition `[Public: DCD 2023]`
Excelero	Mar 2022	High-performance storage	Israel-based, founded 2014 `[Public: GreyB]`
Bright Computing	Jan 10, 2022	Cluster management → Base Command Manager	Direct progenitor of BCM `[Public: GreyB]`
SwiftStack	Mar 6, 2020	Multi-cloud data mgmt	`[Public: GreyB]`
Deci	2024	AI model optimization	Reported ~$300M; Israel-based; second-biggest 2024 acquisition after Run:ai `[Public: Motley Fool; CIO Influence]`

The pattern: the median fit is a 50–200 person team with a specific platform, technical IP NVIDIA wants to fold into its software stack, and a clear post-close integration path (e.g., Run:ai → Mission Control). Acquisition prices range from undisclosed-small to $700M. That is the implicit ceiling Joe pointed to when he said “the opportunity to build is more valuable than over-optimizing IP ownership on an AI-era codebase” [Interview: Joe Malchow, 2026-06-17].

Sources for §4: 2026-06-17-logistics-catch-up-wgreg-lonny-nvidia; 2026-06-17-lonny-greg-debrief; 2026-05-12-lonny-orona; 2026-05-27-nvidia-reverse-logistics-supply-chain-discussion; 2026-06-17-joe-malchow-bliss-n-dustin-advice-re-nvidia-design-partnersh; reverse-logistics-warranty-tam-2026-05-29; reverse-supply-chain-research-2026-05-13; PTC/SAP/Salesforce IR; trade press cited inline.

§5 — The RMA workflow and warranty as a balance-sheet problem (two views of the same dollars)

BLUF: The same problem viewed at the workflow layer and the balance-sheet layer. At the workflow layer, NVIDIA has a 30-day SLA it can’t measure, a 60-of-100 repair-vs-replace ratio, and customer telemetry that stops at the fence. At the balance-sheet layer, NVIDIA’s product-warranty reserve hit $2.81B in FY26 (accrual rate climbing from 0.46% FY24 to 1.15% FY26, primarily Compute & Networking). These are the same dollars on different ledgers — and the gap between them is the wedge.

Workflow half — end-to-end RMA flow

The stack as Greg, Lonny, and Alex described it. A failed unit at a hyperscaler customer triggers:

Detection (customer side). Meta runs three detection systems — Fleetscanner (micro-benchmarks every 45–60 days), Ripple (alongside active workloads), Hardware Sentinel (kernel-space exception analysis, outperforming testing-based methods by 41%). Over 66% of training interruptions stem from component failures in SRAMs, HBMs, and network switches [Public: Meta Engineering Blog, Jul 2025]. Customers strip telemetry before returning: “You’re not gonna get telemetry off their devices. The chips are virtually wiped before being returned” [Interview: Greg + Lonny, 2026-06-17]. Government customers strip even more aggressively. The forward telemetry layer is structurally walled off.
Case open + three-gate approval (NVIDIA side). First, identify who you’re dealing with — the “customer” and the “sold-to” can differ: Google buys direct (customer = sold-to), while xAI is the customer but Dell/Quanta is the sold-to and integrator. Who initiates the RMA and where it ships to are the two key identifiers [Interview: Greg, 2026-06-26]. Then: customer web portal → SFDC case → Lonny’s frontline team validates serial number, warranty entitlement, advance-replacement vs. standard [Interview: Lonny, 2026-05-12]. Greg’s 2026-06-26 detail makes this a three-gate gauntlet that runs before the unit ships: case management (Lonny: warranty/failure-log/serial checks) → quality (known failure code = approve; unknown → failure-analysis lab) → finance (approves) → approval notice → customer ships. <1% of RMAs have ever been rejected — the gates are latency, not a filter, and known-defect batches (“6,000 H100s from this date range”) should be auto-fast-tracked [Interview: Greg, 2026-06-26].
Downstream-of-SFDC dropoff. Everything after the SFDC ticket happens outside the system — email and spreadsheets. Greg: “the spreadsheet gets updated manually for delivery notices, tracking numbers, serial numbers, ship-to info. Customers reply-all, add more people; threads balloon to 16+ people; customer version and Nvidia version of the data diverge” [Interview: Greg + Lonny, 2026-06-17].
ASN to warehouse. NVIDIA is rolling out automated Advance Shipping Notification (EDI). Today, mostly manual [Interview: Alex Zhu, 2026-05-27].
CM repair w/ consignment + turnkey materials — routed by category and product type. Repair facility uses consignment inventory (NVIDIA-owned high-value components — chips, boards) plus turnkey materials (CM-procured cables, third-party parts from Taiwan/China). Tracked today on “suspect sheets” — “think early ’90s spreadsheets” [Interview: Alex Zhu, 2026-05-27]. NVIDIA has a material manager reconciling because “they could be using more but not reporting it purposely.” Geography (Greg 2026-06-26): compute (GPUs, trays, baseboard assemblies, SXM modules) routes to Dallas, Houston, or Guadalajara (all Foxconn/Wistron); networking (switches — no GPUs) routes entirely separately to Vietnam, Israel, or India; European compute ships back to the US — no proximate compute repair there. Routing rule: whoever built it usually repairs it [Interview: Greg, 2026-06-26].
Repair vs. replace vs. scrap.
- ~60 of 100 repairable; 40 from new inventory — “this 40 we have to skew from new buy from resin… new buy is all what Jensen cares about because new buy is basically revenue” [Interview: Alex Zhu, 2026-05-27]. Greg confirms the revenue-side tension directly: reman competes with new production on the same assembly line (a joint, negotiated decision, revenue wins ~9/10), and warranty gets filled from new units when repaired stock is short (“we try to limit the number of new buys in our service depots”) [Interview: Greg, 2026-06-26].
- Three repair categories (Greg 2026-06-26 corrects the earlier 2-category split): reman (current-revision, known defect + fix, same live mass-production line), repair / sort-and-repair (N-minus-one, separate dedicated line at the same site), refurbishment (one-off, unknown condition, diagnosed on arrival, may be salvaged for parts). ~90% is reman + repair combined; refurb is the tail; reman + repair “are really recalls” [Interview: Greg, 2026-06-26].
- All repairs free to customer. No charge model [Interview: Alex Zhu, 2026-05-27].
- The GPU is rarely the problem. Greg: “most of our returns the chip is rarely the problem. It’s usually something else on the assembly. We salvage a very, very high percentage of our chips, either repair in place or pull the GPU off and reuse on a different assembly. NVIDIA does not repair GPUs themselves; repairable components are typically network switches and other board-level assemblies” [Interview: Greg + Lonny, 2026-06-17].
- Specific failure modes Greg surfaced: bent connector pins in cable cartridges; back-plane / slot damage from over-aggressive insertion (“gorilla grip”); heat-sink issues [Interview: Greg + Lonny, 2026-06-17]. Independent corroboration: CoWoS-L thermal/CTE mismatch at 1400W TDP for Blackwell causes warping; microbump failures in HBM PHY can render the entire chip inoperable; once CoWoS-bonded, individual chiplets and HBM stacks can’t be replaced [Public: SemiEngineering; Chiplet Summit 2025; proteanTecs].
Refurb into services pool → ship back to next customer. Customer gets like-for-like (same feature/functionality, different serial number) [Interview: Greg + Lonny, 2026-06-17].
SLA spec: 30 days from RMA arrival in warehouse → replacement unit at customer [Interview: Greg + Lonny, 2026-06-17]. Today, not measured. The bullet doc target is “95%+ SLA attainment on standard RMAs, measured for the first time” [[nvidia-engagement-bullets-2026-06-17]]. Real cycle-time numbers (Greg 2026-06-26): ~60 days end-to-end today, ~30 best case; 27–40 days is warehouse-receipt → repaired-unit-back-in-stock (the only stretch tracked weekly); the approval gauntlet adds ~1–2 untracked weeks because the clock starts only at warehouse arrival. Two decoupled clocks — the customer gets a replacement from stock in ~1–2 weeks, while their own serial takes 30+ days to repair; NVIDIA wants to measure both [Interview: Greg, 2026-06-26].
Where it breaks: the Phoenix five-truck anecdote. NVIDIA sent five trucks of “hundreds of millions of dollars of equipment” to a customer receipt location with no advance coordination. Customer couldn’t take them. Trucks turned around to a security yard, equipment sitting in the sun in Phoenix [Interview: Greg + Lonny, 2026-06-17]. Stranded-shipment incidents → zero is one of the four KPIs [[nvidia-engagement-bullets-2026-06-17]].
Telemetry boundary. Lonny: “you’re not gonna get telemetry off their devices. The access to any of that is just, it’s never going to happen.” Even Meta’s logs come post-fault, never proactive [Interview: Greg + Lonny, 2026-06-17]. Greg’s prior company (Panera/photonics) ran a “crawl program” that scanned customer networks for degradation patterns — but customers would “more times than not… let it fail, do your SLA with a 4-hour replacement” rather than coordinate maintenance windows. This is exactly why Joe’s “third-party independent data repository” argument is structural, not stylistic [Interview: Joe Malchow, 2026-06-17].

Volume — does the math close? Meta’s published Llama 3 16,384-H100 cluster: 466 interruptions over 54 days, ~78% hardware-related (~363 failures), GPU faults 30.1%, HBM3 17.2% — one failure every ~3 hours, ~9% annualized failure rate [Public: Meta Llama 3 paper, 2024; Tom's Hardware]. At Meta’s 100K GPUs today → ~9,000 failures/yr (~750/mo). At Meta’s target 1M GPUs over 5 years → ~90,000 failures/yr (~7,500/mo). Lonny’s “hundreds today, thousands soon” is consistent with the math [[reverse-supply-chain-research-2026-05-13]]; Synthesis. Hyperscalers are formalizing the spec they want NVIDIA’s products to meet: OCP GPU & Accelerator RAS Requirements v1.7 (2025-10-23) standardizes error reporting, crash dumps, RCA, error containment, and Redfish/IPMI SEL/APEI formats — i.e., the hyperscalers are writing the serviceability spec rather than accepting OEM defaults [Public: OCP GPU & Accelerator RAS Requirements v1.7]. That spec lives at the forward layer, but it sets the bar reverse-flow tools have to meet for hyperscaler buy-in.

Balance-sheet half — warranty as a balance-sheet problem

The product-warranty rollforward (NVIDIA filed 10-K, authoritative).

FY	Reserve (end, $M)	Claims paid ($M)	Accruals ($M)	Accrual rate
FY23 (Jan 29, 2023)	82	109	145	—
FY24 (Jan 28, 2024)	306	54	278	~0.46%
FY25 (Jan 26, 2025)	1,290	219	1,203	~0.92%
FY26 (Jan 25, 2026)	2,807	957	2,474	~1.15%

[Public: NVIDIA FY26 10-K product-warranty footnote, accession 0001045810-26-000021]

Reserve +118% YoY. Claims paid +337% YoY. The 10-K explicitly attributes additions “primarily related to Compute & Networking segment” — i.e., data-center, not consumer [Public: NVIDIA FY26 10-K].

⚠️ Critical correction the pitch deck must reflect

The widely-circulated $8.22B warranty reserve figure is wrong. WarrantyWeek’s April 9 2026 analysis published it and the tech press repeated it (TweakTown, TechPowerUp, BGR, WCCFTech, Overclock3D, Guru3D, AOL). But:

It conflicts with NVIDIA’s own 10-K (~$2.81B FY26) [Public: NVIDIA FY26 10-K].
It conflicts with WarrantyWeek’s own later industry-aggregate report, which put the entire US-semi-industry reserve at $3.80B for calendar 2025 [Public: WarrantyWeek 23rd Annual Report, 2026-04-16].
WarrantyWeek’s reserve series cannot be reproduced from its own accruals and claims (~$3.9B unexplained jump) [arithmetic check in [[reverse-logistics-warranty-tam-2026-05-29]]].

Use the filed ~$2.81B. The “$8B cash for warranty liabilities” line Alex Zhu didn’t confirm in May likely also traces to the WarrantyWeek figure [Interview: Alex Zhu, 2026-05-27; [[reverse-logistics-warranty-tam-2026-05-29]] §4]. The pitch is more credible at $2.81B + ~1.15% accrual rate growing ~100% YoY than at a number that doesn’t tie to the filing.

Meta 1M-GPU scale implication

Meta has 100K GPUs today and wants 1M over 5 years [Interview: Lonny Orona, 2026-05-12]. If NVIDIA’s accrual rate stays at ~1.15% of revenue and revenue keeps tracking GPU volume, the reserve grows roughly with shipped units. The single-customer-22% concentration footnote [Public: NVIDIA FY26 10-K] plus the explicit Compute & Networking attribution means a meaningful fraction of NVIDIA’s growing warranty reserve is allocated against returns from a handful of named hyperscaler customers [Synthesis]. This is what makes the workflow-half and balance-sheet-half the same dollars: the cost concentrates with the same customers Lonny “comes to the conversation not armed against.”

Risk-transfer / parametric / captive analogs — what’s been done elsewhere

The cleanest existing template for transferring an OEM warranty reserve to a third-party balance sheet is Munich Re + TWAICE for lithium-ion battery performance warranties (2019) — the world’s first such product, with Munich Re as underwriter on top of TWAICE analytics, covering repair/maintenance and extendable to lost-revenue downtime [Public: Munich Re / TWAICE, 2019].

Adjacent context that does not yet cover data-center hardware:

Munich Re aiSure — parametric-style insurance for AI model performance (bias, privacy, IP, accuracy). Written since 2018 (general AI), 2019 (LLM-specific). Mosaic + Munich Re partnership Feb 2026 brings up to €/$/C$15M initial coverage for AI vendors. Covers AI output errors, not AI hardware warranty. [Public: Munich Re aiSure; Reinsurance News; Mosaic Feb 2026]
Extended-warranty market overall: ~$147–161B (2025) at ~8.5% CAGR. Players: Assurant, Asurion, Allstate, AIG, AXA [Public: Mordor / Grand View 2025].
Gap: no public example of a specialist insurer writing warranty-liability risk transfer for data-center GPU/server hardware [[reverse-logistics-warranty-tam-2026-05-29]] §6.4]. The battery analog proves the structure exists; nobody has done it for compute yet.

CoreWeave / Burry / CITP — depreciation mismatch creates a hedging opening

Two related but distinct arguments, often conflated:

Argument A — Depreciation mismatch (CITP / Burry / Barclays). GPUs at 60–70% utilization survive 1–3 years; companies book-depreciate over 5–6 years [Public: Princeton CITP 2025-10-15; Tom's Hardware 2025]. Michael Burry estimates the five largest hyperscalers will understate depreciation by ~$176B over 2026–2028; Barclays cut AI-firm earnings forecasts up to 10% on realistic depreciation; CoreWeave extended useful life 5→6 years [Public: CNBC 2025-11-14]. This is a capital-stock mismatch: GPUs as consumables, books as capital equipment.

Argument B — Capex funding gap (Bain 2025 Global Tech Report). $2T in annual revenue is needed by 2030 to fund AI compute buildout; even with full cloud migration and AI productivity-driven reinvestment, there’s an $800B annual revenue shortfall [Public: Bain 6th Global Tech Report, 2025-09; PR Newswire; Tom's Hardware]. This is a capital-raise gap, not a depreciation gap. Keep these distinct in the pitch.

Why both matter to the wedge: Argument A directly creates demand for warranty/insurance/hedging products that align hyperscaler effective-life assumptions with cash-out reality. Argument B is the macro environment the pitch lands in — capital is scarce relative to ambition, and CFOs are receptive to anything that frees float [Synthesis; [[financialization-primer-2026-05-29]] §7 on time value of money / float].

Sketch of the financialization layer (per the pitch)

Restating the trade per [[financialization-primer-2026-05-29]] §7 with corrected numbers: a specialist (insurer or structured-finance buyer) assumes NVIDIA’s warranty obligation; NVIDIA transfers part of its reserve as premium; specialist holds the premium as float, services the warranties for less than premium received, and earns investment income on float. At the corrected ~$2.8B reserve with ~$957M annual claims, the operative figure is claims spend, not reserve balance — the float is large but the throughput of warranty claims is what gets repriced. The bridge to the workflow product: nobody underwrites this risk without the operating data Lonny and Alex are sitting on. Whoever runs the data layer is the natural counterparty for the underwriter [Synthesis].

Sources for §5: 2026-06-17-logistics-catch-up-wgreg-lonny-nvidia; Greg en-route call 2026-06-26; 2026-05-12-lonny-orona; 2026-05-27-nvidia-reverse-logistics-supply-chain-discussion; reverse-logistics-warranty-tam-2026-05-29; financialization-primer-2026-05-29; reverse-supply-chain-research-2026-05-13; NVIDIA FY26 10-K; WarrantyWeek (use w/ correction); Meta Engineering Blog; OCP RAS v1.7; Munich Re aiSure; Bain 2025; CITP / CNBC / Burry.

§6 — Engagement positioning: how this deal closes and what we walk in with

BLUF: The pitch closes if it lands two sequenced ideas — the dashboard the RFP is for, then the warranty financialization layer on top — and clears Greg/Lonny’s unidentified exec sponsor with a structure that is not Palantir’s SaaS pitch and not Accenture’s billable-hours pitch. The path runs through Thursday’s in-person meeting, Friday’s bullet doc, and Shanker Trivedi’s warm intro via Joe Malchow. Tier 1 open questions below are for the human, not for synthesis.

This section is forward-looking and lighter on external citation by design. It is grounded in the four internal sources captured 2026-06-17 plus Joe Malchow’s structural framing.

Buying center (observed) and exec-mapping path (proposed)

What we observed in the 2026-06-17 meeting:

Lonny (compute science frontline support, customer-service ops): the operational pain owner
Greg (senior director, service team / transformation): the systems lead and executive proxy
Unidentified exec sponsor above Greg/Lonny: budget authority — not yet named to us
EY workshop ~2 weeks before the meeting defined two initiatives; the dashboard is one of them, surface evidence the exec sponsor is bought into a defined work program with budget

The exec-mapping path:

Through Greg upward: Greg is one rung from the exec sponsor. The bullet doc + Friday meeting puts the engagement language in Greg’s hands; he socializes upward [Interview: Joe Malchow, 2026-06-17].
Through Joe Malchow ↔ Shanker Trivedi: Trivedi retired from EVP Enterprise Sales April 2026 (Joe just added him to the NPH board). Trivedi is outside the company but fresh — useful as a warm outside voice with sitting-team relationships, particularly given his SAP CMO background and NVIDIA’s $2M+ SAP planning deal in flight. Per Joe, intro should land after we have something on paper, not before [Interview: Joe Malchow, 2026-06-17]. Surface to Bliss/Dustin: clarify with Joe whether Trivedi can still introduce inward to Shoquist’s ops org or whether his retirement means he’s now an advisor rather than an active connector.
Through ops org upward to Shoquist / Kress: Debora Shoquist (EVP Operations) is the org line Greg/Lonny most plausibly roll up under. The warranty-financialization layer specifically requires CFO buy-in (Colette Kress) — that is a sequencing question for Phase 2, not Phase 1 [Synthesis from [[nvidia-engagement-structure-2026-06-17]]].

Why “two solutions people + arm’s-length independent data repository + milestone economics” beats Palantir’s SaaS pitch or Accenture’s billable-hours pitch

The structural argument is Joe Malchow’s, restated tightly:

Against Palantir (SaaS): Palantir sells a platform NVIDIA buys into. NVIDIA becomes the customer. The problem with that for this specific reverse-flow problem is that the data NVIDIA needs (customer-side telemetry, CM repair records, 3PL shipment events, distributor inventory state) lives outside NVIDIA’s four walls. NVIDIA cannot collect it directly; customers strip telemetry because they treat NVIDIA as a partial competitor. An independent third party with structurally clean access can collect what NVIDIA’s own platform cannot [Interview: Joe Malchow, 2026-06-17; cross-ref §5 telemetry boundary]. Palantir’s pitch is solving the wrong problem at the wrong layer.
Against Accenture (billable hours): Accenture builds and walks. NVIDIA gets a code drop and an ops org with the same data problem six months later. The reverse-flow data layer needs to be operated, not shipped, because the data quality compounds over time and the operating company is the one that signs the next NDA with the next CM. Joe: “it’s us building and operating a standalone data repository that enables you to meet your business objectives around returns” [Interview: Joe Malchow, 2026-06-17].
Why ours: arm’s-length, milestone-paid, IP-clean. We carry the risk phase to phase (NVIDIA can exit cleanly if not satisfied); NVIDIA’s data is perpetually NVIDIA’s; the operating company is ours to build and run. The economics naturally lead to a long-term data + solution relationship [[nvidia-engagement-structure-2026-06-17]].

Cross-references to §4 incumbents and §5 telemetry boundaries are direct: the incumbents in §4 either can’t collect the data (SAP, Salesforce — bounded inside NVIDIA’s perimeter) or won’t (Servigistics, ServiceMax — sell platforms, not operating relationships). §5’s telemetry boundary is the structural argument for the independent operator.

The two-pitch shape: sequenced, not bundled

Phase 1 — the dashboard (the RFP scope, the “shiny object”):

Per [[nvidia-engagement-structure-2026-06-17]]: NDA close Wednesday; 14–30 day discovery sprint (6–10 NVIDIA interviews, 4–6 customer interviews, 3–4 CM repair partner, 3–4 3PL); 14–45 day prototype sprint with Bliss embedded FDE-style with Lonny’s team; problem-understanding doc as the through-line.
KPIs: 95%+ SLA attainment measured for the first time; 30%+ hyperscaler RMA cycle-time reduction; customer-initiated escalations per open RMA cut in half; Phoenix-style stranded-shipment events to zero [[nvidia-engagement-bullets-2026-06-17]].

Phase 2 — warranty financialization layer (the bigger idea on top):

The dashboard makes the data layer operational. Once that data layer exists, the same operator becomes the natural data partner for an underwriter (Munich Re, Swiss Re, Mosaic, Assurant) to write risk-transfer against NVIDIA’s product-warranty reserve.
This is the layer §5’s balance-sheet half describes: $2.81B reserve, $957M annual claims, ~100% YoY growth, explicit Compute & Networking attribution. With corrected numbers, the precise float economics shift — pitch deck must reflect filed figures.
Do not bundle these in the bullet doc. Joe’s structural framing is Phase 1’s dashboard scope, milestone-paid, satisfaction trigger; the long-term operating relationship language is the seed that lets Phase 2 grow without committing now [Interview: Joe Malchow, 2026-06-17].

Path to closing

Thursday 8am in-person meeting with Lonny + team at NVIDIA HQ Santa Clara [Interview: Greg + Lonny debrief, 2026-06-17].
Friday bullet doc to NVIDIA — per Joe’s format guidance, 4–5 descending-specificity bullets, problem-first, KPI-anchored. Joe will review draft Thu/Fri [Interview: Joe Malchow, 2026-06-17; [[nvidia-engagement-bullets-2026-06-17]]].
Monday clickable dashboard prototype — Bliss building the “shiny object” [Interview: Greg + Lonny debrief, 2026-06-17].
NDA close target Wednesday next week [[nvidia-engagement-structure-2026-06-17]].
Shanker Trivedi warm intro post-paper, via Joe [Interview: Joe Malchow, 2026-06-17].
Discovery sprint 14–30 days, prototype 14–45 days, Phase 2 commercial conversation by day 60.

Open questions for the human (tiered)

Tier 1 — Shanker Trivedi or Joe Malchow can answer:

With Trivedi retired April 2026, does the warm-intro path land at the current EVP Enterprise Sales (whoever replaced him) or at the EVP Operations (Shoquist)? Service ops more plausibly sits under Operations; the deal sponsor is plausibly there, not under Sales.
Did Trivedi work with any of the SAP/Salesforce ecosystem at NVIDIA, given his ex-SAP CMO background? If yes, he can short-circuit the implicit competitive dynamic with the $2M+ SAP planning deal.

Tier 2 — Greg or Lonny can answer: 3. Who is the unidentified exec sponsor above Greg/Lonny? Plain ask in the Thursday meeting. 4. Is the Dallas Wistron facility primarily production (per public reporting) or repair line (per Lonny)? The two interpretations have different implications for what Phase 1 has access to [[reverse-supply-chain-research-2026-05-13]] divergence]. 5. Of the two EY workshop initiatives, is the *other one* a customer-facing dashboard sibling or an internal-ops effort? Affects what we have to integrate against in Phase 1. 6. Concrete per-RMA-cycle cost for a data-center GPU — the weakest input in the TAM Method D reverse-logistics-warranty-tam-2026-05-29 §5`.

Tier 3 — Alex Zhu (deferred to follow-up after he’s back from PTO): 7. Does the consignment material manager already reconcile against a clean source of truth, or is that reconciliation a candidate Phase 1 module? 8. Among “repair vs. replace vs. scrap,” what’s the GPU salvage rate exactly? Greg said “very, very high” — turning that into a number lets us put a dollar on §5’s balance-sheet half.

Tier 4 — external (Munich Re aiSure, Mosaic, Assurant, Swiss Re, Asurion): 9. Is anyone underwriting data-center hardware warranty-liability risk transfer today, anywhere? Battery analog (Munich Re + TWAICE) proves the structure exists; we need to test whether a specialist has already entered the compute layer [[reverse-logistics-warranty-tam-2026-05-29]] §6.4].

What we are not yet sure of

Does Phase 1 SAM/TAM exceed the cost of running an independent operator long-term? The triangulated SAM range is $30M–$320M with growth — that’s a narrow-fast niche, not a billion-dollar standalone software market today [[reverse-logistics-warranty-tam-2026-05-29]] §5]. Phase 2 financialization is what makes this a billion-dollar story — but only if Phase 1 generates the proprietary data layer the underwriter needs.
Buyer-base concentration risk. If the financial warranty burden stays NVIDIA-concentrated (AMD ~10x smaller, Intel/Broadcom/Marvell disclose nothing), the platform may effectively be a 2-customer business [[reverse-logistics-warranty-tam-2026-05-29]] §7]. The Phase 1 NVIDIA engagement does not resolve this — only sequentially acquiring AMD does.
Hyperscaler willingness to share data with an independent operator. Joe’s argument assumes hyperscalers will participate; whether Meta/Microsoft/Google specifically will sign with us is the most consequential Phase 1 risk.

Sources for §6: 2026-06-17-logistics-catch-up-wgreg-lonny-nvidia; 2026-06-17-lonny-greg-debrief; 2026-06-17-joe-malchow-bliss-n-dustin-advice-re-nvidia-design-partnersh; nvidia-latent-problem-2026-06-17; nvidia-engagement-bullets-2026-06-17; nvidia-engagement-structure-2026-06-17; reverse-logistics-warranty-tam-2026-05-29; financialization-primer-2026-05-29.

Surprises and contradictions (per RDI methodology)

NVIDIA already owns the forward fleet-management layer (Mission Control + Run:ai). We sit in the reverse layer, downstream of where customer-side telemetry stops. The boundary is structural (customers strip data), not a gap NVIDIA can close by buying us. [Synthesis from §1 + §5]
Shanker Trivedi retired April 2026 — the warm-intro asset is an outside voice, not an active insider. The pitch should respect that. [Public: SDxCentral 2026]
The widely-circulated $8.22B warranty reserve figure is wrong. The filed 10-K number is $2.81B FY26. Tech-press repetition does not validate the number. [Public: NVIDIA FY26 10-K; WarrantyWeek correction in [[reverse-logistics-warranty-tam-2026-05-29]]]
NVIDIA paying SAP $2M+ for planning automation while Greg explicitly described that he was transforming away from siloed ERP instances of the same vintage at Cisco 2007–09. A live tension: the existing procurement language anchors to SAP, but the structural critique in Greg’s voice anchors to the opposite of an ERP-centric solution. [Interview: Greg + Lonny, 2026-06-17; Alex Zhu, 2026-05-27]
System OEMs (Dell, HPE, Supermicro) carry flat-to-declining warranty reserves despite booming AI-server revenue. The cost is not distributing down the chain — it concentrates at NVIDIA. [Public: Dell/HPE/SMCI 10-Ks; [[reverse-logistics-warranty-tam-2026-05-29]] §3]
NVIDIA’s repair workflow is structurally similar to Cisco’s pre-2007 ERP state, per Greg’s own framing. The historical analog suggests a 2–3 year transformation timeline, not a 6-month one — Phase 1 dashboard delivery in 45 days is aggressive but it is the start, not the finish. [Interview: Greg + Lonny, 2026-06-17]
The OCP RAS v1.7 spec (Oct 2025) is hyperscaler-defined, not NVIDIA-defined. Microsoft, Google, Meta wrote the standard NVIDIA’s hardware has to meet — a structural shift in who sets serviceability requirements. Worth tracking for whether our reverse-flow tool eventually needs to meet OCP RAS interface specs to be acceptable to hyperscaler customers. [Public: OCP GPU & Accelerator RAS Requirements v1.7, 2025-10-23]

Confidence summary

Claim	Confidence	Basis
NVIDIA FY26 data-center revenue $193.7B; total $215.9B	High	Primary 10-K
Two customers ≈ 36% of FY26 revenue; top single ≈ 22%	High	Primary 10-K
Product warranty reserve $2.81B FY26; 0.46→1.15% accrual rate	High	Primary 10-K + cross-check in reverse-logistics-warranty-tam-2026-05-29
$8.22B WarrantyWeek figure is wrong	High	Directly contradicted by filing
NVIDIA secured ~60% of TSMC CoWoS 2026 capacity	Medium-High	Multiple trade-press sources triangulate; primary TSMC data not public
Mission Control / Run:ai is forward-layer only; reverse layer is open	Medium-High	NVIDIA marketing materials describe forward role; Lonny’s “telemetry won’t come back” matches
Shanker Trivedi retired April 2026	High	SDxCentral primary
Greg/Lonny pain quotes (“we don’t even know the score,” “not even armed”)	High	Direct interview
60-of-100 repair rate; 90% remanufacturing	Medium-High	Direct interview (Alex Zhu); single source
SLA spec 30 days, not currently measured	High	Direct interview (Greg)
Dallas/Wistron July 2026 — repair vs. production	Medium	Interview says repair, public reporting says production; possibly both
Tier 1 hyperscalers = 4 unnamed >10% customers	High	10-K + public triangulation
RFP finalists: Palantir, Accenture, build-internal	High	Direct debrief
Munich Re aiSure does NOT cover data-center hardware	High	Munich Re primary materials
No specialist underwrites data-center hardware warranty risk transfer today	Medium	Absence of evidence in market survey

Sources

Internal (vault):

Greg + Lonny @ NVIDIA, 2026-06-17 — anchor
Bliss + Dustin debrief, 2026-06-17
Lonny Orona, 2026-05-12
Alex Zhu, 2026-05-27
Joe Malchow, 2026-06-17
nvidia-latent-problem-2026-06-17
nvidia-engagement-bullets-2026-06-17
nvidia-engagement-structure-2026-06-17
reverse-logistics-warranty-tam-2026-05-29 — critical for warranty figure correction
reverse-supply-chain-research-2026-05-13
independent-distributors-research-2026-05-13
financialization-primer-2026-05-29
dram-market-deep-dive
substrate-research-2026-04-17

External — NVIDIA financials:

External — warranty financials (cross-firm):

NVIDIA FY2026 10-K product-warranty footnote
[Dell FY2026 10-K — SEC accession 0001571996-26-000008]
[HPE FY2025 10-K — accession 0001645590-25-000130]
[Supermicro FY2025 10-K — accession 0001375365-25-000027]
WarrantyWeek — Discrete GPU Warranty Expenses (Apr 9 2026) — use with §5 correction
WarrantyWeek 23rd Annual Product Warranty Report (Apr 16 2026)

External — supply chain (forward):

External — reverse-flow ecosystem:

External — hyperscaler / customer:

External — NVIDIA software & M&A:

External — competitive landscape:

External — financialization adjacencies:

External — NVIDIA exec coverage:

Project TBD Memory Vault

Explorer

nvidia-primer-2026-06-18