NVIDIA Q3 Earnings, NeoCloud Financing Risks, and Jensen Huang's Departure
Summary
Zoe and Nash analyze NVIDIA's Q3 FY27 results, noting $115.8B in revenue and the rapid ramp of the Vera Rubin architecture, while highlighting the company's $550B in future commitments and contingent guarantees. They discuss the leadership transition with Jensen Huang's departure and Colette Kress's appointment as interim CEO. The hosts examine the financial stress facing NeoCloud providers like CoreWeave, citing rising interest rates and potential refinancing risks. They also cover Gartner's forecast of a $5 trillion gap between AI infrastructure spending and revenue, suggesting significant consolidation in the sector. Finally, they touch on AMD's competitive positioning and the potential impact of AI efficiency tools on hardware demand.
Key takeaways
- 1 Core Weave bonds issued in June at 9.625% yield saw their market yield move toward 15% by November 17th, reflecting increased credit risk perception in the AI infrastructure sector.
Nash · 5:59
Episode sections
- 0:00 – 3:23
Intro and NVIDIA Financials Overview
Zoe and Nash introduce NVIDIA's Q3 FY27 results, highlighting $115.8B in revenue and the rapid ramp of the Vera Rubin architecture. They contrast this operational success with the company's $550B in future commitments and the financial stress faced by some NeoCloud customers.
- 3:23 – 4:58
Jensen Huang Leadership Change
The hosts discuss Jensen Huang's departure from NVIDIA, noting the board's decision not to reinstate him and the appointment of CFO Colette Kress as interim CEO. They emphasize focusing on business fundamentals rather than rumors surrounding the personal conduct investigation.
- 4:58 – 13:40
Macro Radar: Rates and Financial Stability
Nash outlines the macro environment, citing the Fed's hold at 3.75–4%, a 10-year Treasury yield of 5.1%, and CoreWeave bond yields approaching 15%. He highlights the Fed's Financial Stability Report, where 87% of contacts identified AI-related risks as potential shocks.
- 13:40 – 20:44
NVIDIA Q3 FY27 Earnings Deep Dive
Detailed analysis of NVIDIA's Q3 results: $115.8B revenue, $107.6B Data Center revenue, and 74.8% gross margin. The hosts note the strong free cash flow ($43.8B) but warn about rising receivables ($81B) and inventory ($39.7B) linked to the Vera Rubin ramp.
- 20:44 – 24:55
The $550B Commitments and Guarantees
Breakdown of NVIDIA's $550B in future commitments, including $390B in supply contracts, $74B in leases, and $32B in conditional NeoCloud agreements. The hosts also discuss $118B in contingent guarantees, primarily related to an OpenAI data center project with SB Energy.
- 24:55 – 28:44
NeoCloud Business Model and Risks
Explanation of the NeoCloud model (CoreWeave, Nebius, IREN), which focuses on specialized GPU infrastructure. The hosts highlight the capital intensity, concentration risk, and sensitivity to interest rates and construction costs, contrasting them with hyperscalers like AWS.
- 28:44 – 30:53
Hypothetical NeoCloud Financial Model
Nash presents a simplified three-scenario model for a NeoCloud: optimistic (9% debt interest), higher-for-longer (elevated costs), and a third scenario to follow. The model illustrates how financing costs and depreciation impact profitability even with strong demand.
- 30:53 – 36:04
Neocloud Stress Scenarios and Refinancing Risks
Nash and Zoe model three financial scenarios for a hypothetical neocloud provider, showing how rising interest rates, falling revenue, and higher operating costs can turn a profitable project into a loss-making one. They emphasize that while the infrastructure may remain operational, the equity returns can deteriorate significantly due to refinancing squeezes and covenant breaches.
- 36:04 – 40:18
Gartner Impairment Warning and Consolidation Forecast
The speakers cite Gartner's October 14 research predicting a $5 trillion gap between required AI infrastructure revenue and projected enterprise/consumer AI revenue by 2030. Gartner estimates over $3.5 trillion in infrastructure spending could face impairment risk, and forecasts that over 80% of neocloud providers will cease to operate independently by 2030, likely through acquisition or merger rather than bankruptcy.
- 40:18 – 43:49
Hardware Treadmill and Custom Silicon Competition
Nash explains that newer GPU architectures like Rubin make older hardware less competitive, forcing rental price declines even if the older hardware remains technically functional. The discussion highlights the threat from hyperscaler custom silicon (Amazon Trainium, Google TPU, Microsoft Maia) and AMD, noting that while software compatibility is a barrier, cost efficiency drives migration for predictable workloads.
- 43:49 – 53:37
Nvidia's Financial Exposure and Strategic Dilemma
The hosts analyze Nvidia's $43.8 billion Q3 free cash flow against its $45 billion in conditional agreements and guarantees. They debate whether Nvidia should continue financing ecosystem partners to accelerate adoption or reduce exposure to credit risk. The segment concludes that the primary risk is to valuation, not solvency, as a credit squeeze could slow hardware orders and depress GPU rental prices.
- 53:37 – 59:36
AMD Stock Idea and Competitive Landscape
Nash presents AMD as a research idea based on market share gains from diversification, not as a bet against Nvidia. He notes AMD's Q3 revenue of $13.4 billion and data center growth of 97%, but cautions that switching costs and CUDA's dominance make share gains difficult. Three scenarios are outlined: Nvidia maintains lead, customers diversify to AMD, or AI spending slows across the board.
- 59:36 – 1:04:21
Startup Corner: Daisytuner
The episode highlights Daisytuner, a Darmstadt-based company that raised €6 million in seed funding. Their product, DaisyFlow, is a compiler-based AI inference platform that automates hardware optimization. The speakers note that while efficiency gains benefit customers, they create competitive pressure on hardware suppliers by potentially reducing the compute needed per task.
- 1:04:21 – 1:12:14
Podcast Conclusion and Disclaimer
Zoe and Nash conclude the episode by emphasizing that infrastructure must generate returns to justify its cost. They reiterate that the podcast provides educational commentary, not financial advice, and highlight the uncertainty surrounding AI financing and neocloud profitability.
Assets discussed
| Asset | Kind | Stance | Context |
|---|---|---|---|
| NVIDIANVDA | equity | mixed | $116 billion. That's roughly how much revenue NVIDIA generated in just three months. |
| MicrosoftMSFT | equity | neutral | buyers include. NVIDIA, AMD, Microsoft, Google, Broadcom, Micron, and a TSMC entity. |
| AlphabetGOOGL | equity | neutral | buyers include. NVIDIA, AMD, Microsoft, Google, Broadcom, Micron, and a TSMC entity. |
“Stance” is how the speakers talked about the asset, not a recommendation.
Claims worth verifying
NVIDIA generated $115.8 billion in quarterly revenue, with Data Center revenue up 110% year-over-year.
Zoe ·0:00
How to verify: NVIDIA Q3 FY27 Earnings Release, SEC Filing 10-Q or Press Release. For recent events, verify 2026-08-18 through 2026-11-18. Use another period only when the speaker explicitly names it.
Primary sourceVera Rubin is ramping faster than any architecture in NVIDIA's history.
Nash ·0:09
How to verify: NVIDIA Q3 FY27 Earnings Call Transcript or Press Release. For recent events, verify 2026-08-18 through 2026-11-18. Use another period only when the speaker explicitly names it.
Primary sourceNVIDIA has $550 billion in future contractual commitments and up to $118 billion in contingent guarantees.
Zoe ·0:27
How to verify: NVIDIA Q3 FY27 10-Q Filing, Notes to Financial Statements regarding commitments and guarantees. For recent events, verify 2026-08-18 through 2026-11-18. Use another period only when the speaker explicitly names it.
Primary sourceGartner believes more than 80% of today's NeoCloud providers may no longer operate independently by 2030.
Unidentified speaker (UNKNOWN) ·1:03
How to verify: Gartner Research Report or Press Release (Note: Requires subscription to Gartner database). For recent events, verify 2026-08-18 through 2026-11-18. Use another period only when the speaker explicitly names it.
Jensen Huang departed NVIDIA; Colette Kress is interim CEO; board does not intend to reinstate him.
Unidentified speaker (UNKNOWN) ·3:50
How to verify: NVIDIA Press Release or SEC Form 8-K regarding executive changes. For recent events, verify 2026-08-18 through 2026-11-18. Use another period only when the speaker explicitly names it.
Primary sourceFederal Reserve left the federal funds target range unchanged at 3.75% to 4% on October 28th.
Unidentified speaker (UNKNOWN) ·5:14
How to verify: Federal Reserve FOMC Statement and Press Conference Transcript. For recent events, verify 2026-08-18 through 2026-11-18. Use another period only when the speaker explicitly names it.
10-year Treasury yield stood at 5.1% on November 17th.
Unidentified speaker (UNKNOWN) ·5:36
How to verify: U.S. Department of Treasury Daily Yield Curve Report. For recent events, verify 2026-08-18 through 2026-11-18. Use another period only when the speaker explicitly names it.
CoreWeave bond yields moved toward 15% as reported by Reuters on November 17th.
Nash ·5:59
How to verify: Reuters Article dated Nov 17, 2026; Bloomberg Terminal for CoreWeave bond pricing. For recent events, verify 2026-08-18 through 2026-11-18. Use another period only when the speaker explicitly names it.
87% of 23 market contacts in the Fed's Financial Stability Report identified AI-related developments as possible shocks.
Unidentified speaker (UNKNOWN) ·6:55
How to verify: Federal Reserve Financial Stability Report, November 13, 2026. For recent events, verify 2026-08-18 through 2026-11-18. Use another period only when the speaker explicitly names it.
Anthropic's amended registration statement from November 10th describes a $26 billion private placement.
Unidentified speaker (UNKNOWN) ·11:45
How to verify: SEC EDGAR Filing, S-1/A Registration Statement for Anthropic AI. For recent events, verify 2026-08-18 through 2026-11-18. Use another period only when the speaker explicitly names it.
Primary sourceRevenue falls from $90 to $80 million, operating costs rise from $30 to $34 million, and debt interest rate rises from 9 to 12%, resulting in an annual interest bill of $8.4 million and pre-tax profit of $12.6 million.
Nash ·30:53
How to verify: Verify the arithmetic: 80 - 34 - 25 (depreciation) - 8.4 (interest) = 12.6. This is a hypothetical model calculation presented by the speaker. For recent events, verify 2026-08-18 through 2026-11-18. Use another period only when the speaker explicitly names it.
Gartner published a warning on October 14, 2026, expecting nearly $1 trillion in AI infrastructure spending during 2026 and approximately $6 trillion between 2027 and 2030.
Unidentified speaker (UNKNOWN) ·36:14
How to verify: Search Gartner's research database for reports published around October 14, 2026, regarding AI infrastructure spending forecasts. For recent events, verify 2026-08-18 through 2026-11-18. Use another period only when the speaker explicitly names it.
Gartner estimates AI data center owners would need $8.5 trillion in cumulative revenue through 2030 to cover planned spending, financing costs, and a 12% return on invested capital.
Zoe ·36:36
How to verify: Locate the specific Gartner report cited by the speaker to confirm the $8.5 trillion revenue requirement and 12% ROI assumption. For recent events, verify 2026-08-18 through 2026-11-18. Use another period only when the speaker explicitly names it.
Gartner forecasts no more than $3.6 trillion in enterprise and consumer AI revenue over the period 2027-2030, creating a nearly $5 trillion gap.
Zoe ·36:50
How to verify: Cross-reference Gartner's revenue forecast figures with the spending forecast in the same report to verify the $5 trillion gap calculation. For recent events, verify 2026-08-18 through 2026-11-18. Use another period only when the speaker explicitly names it.
More than $3.5 trillion of infrastructure spending could be exposed to impairment risk by the end of 2030 if investment continues at the current pace.
Unidentified speaker (UNKNOWN) ·37:14
How to verify: Check the Gartner report for the specific $3.5 trillion impairment exposure figure and the methodology used to derive it. For recent events, verify 2026-08-18 through 2026-11-18. Use another period only when the speaker explicitly names it.
Nvidia generated $43.8 billion in free cash flow in Q3 FY27.
Zoe ·47:55
How to verify: Verify against Nvidia's Q3 FY27 earnings release or SEC 10-Q/10-K filing for the free cash flow figure. For recent events, verify 2026-08-18 through 2026-11-18. Use another period only when the speaker explicitly names it.
Primary sourceNvidia has $32 billion in conditional NeoCloud agreements, $9.5 billion in power and systems guarantees, and $3.5 billion in NeoCloud infrastructure guarantees.
Unidentified speaker (UNKNOWN) ·48:02
How to verify: Verify these figures in Nvidia's Q3 FY27 earnings release or SEC filings under 'Commitments' or 'Contingencies'. For recent events, verify 2026-08-18 through 2026-11-18. Use another period only when the speaker explicitly names it.
Primary sourceAMD reported Q3 revenue of $13.4 billion, up 45% year-over-year, with data center segment revenue of $8.5 billion, up 97%.
Unidentified speaker (UNKNOWN) ·54:25
How to verify: Verify against AMD's Q3 FY26 earnings release or SEC 10-Q filing for revenue and segment breakdowns. For recent events, verify 2026-08-18 through 2026-11-18. Use another period only when the speaker explicitly names it.
Primary sourceDaisytuner announced a €6 million seed round on November 16, 2026, co-led by Eliya and LEA Ventures.
Zoe ·59:42
How to verify: Search for press releases or news articles from Daisytuner or venture capital databases (e.g., Crunchbase, PitchBook) confirming the funding amount, date, and lead investors. For recent events, verify 2026-08-18 through 2026-11-18. Use another period only when the speaker explicitly names it.
Primary sourceGartner forecasts that by 2030, more than 80% of NeoCloud providers will cease operating independently.
Unidentified speaker (UNKNOWN) ·38:20
How to verify: Locate the Gartner report section discussing neocloud market consolidation and the specific 80% forecast figure. For recent events, verify 2026-08-18 through 2026-11-18. Use another period only when the speaker explicitly names it.
Nvidia shares rose almost 3% in after-hours trading after closing at about $155, which is below the roughly $174 price before Jensen Huang's resignation.
Zoe ·1:04:31
How to verify: Check Nvidia's stock price history and after-hours trading data for the date of the episode (November 18, 2026) via Yahoo Finance or a brokerage platform; verify the resignation date and prior price via SEC 8-K filing. For recent events, verify 2026-08-18 through 2026-11-18. Use another period only when the speaker explicitly names it.
Primary sourceNvidia reported Q3 FY27 revenue of $115.8B with Data Center up 110% and provided Q4 guidance of $130B.
Zoe ·1:04:31
How to verify: Review Nvidia's Q3 FY27 Earnings Press Release and Supplemental Financial Information filed with the SEC via EDGAR. For recent events, verify 2026-08-18 through 2026-11-18. Use another period only when the speaker explicitly names it.
Primary sourceJensen Huang resigned as CEO, with Colette Kress named interim CEO, and the board stated it would not reinstate him.
Unidentified speaker (UNKNOWN) ·1:09:05
How to verify: Check Nvidia's Form 8-K filed with the SEC regarding the CEO transition and the Board Statement on Huang. For recent events, verify 2026-08-18 through 2026-11-18. Use another period only when the speaker explicitly names it.
Primary sourceThe Federal Reserve held rates at 3.75–4% and the 10-year Treasury yield was at 5.10% as of the episode date.
Unidentified speaker (UNKNOWN) ·1:09:05
How to verify: Check the Federal Reserve FOMC Statement dated October 28, 2026, and the US Treasury Daily Par Yield Curve for November 2026. For recent events, verify 2026-08-18 through 2026-11-18. Use another period only when the speaker explicitly names it.
Gartner warned that over $3.5 trillion of AI infrastructure spending could be exposed to impairment by the end of 2030.
Unidentified speaker (UNKNOWN) ·1:09:05
How to verify: Locate Gartner's press release dated October 14, 2026, regarding AI infrastructure impairment risks via Gartner's newsroom. For recent events, verify 2026-08-18 through 2026-11-18. Use another period only when the speaker explicitly names it.
Do your own research
- Verify NVIDIA's Q3 FY27 10-K or 8-K filings for exact revenue, gross margin, and contingent liability disclosures.
- Review the Federal Reserve's November 13 Financial Stability Report for details on AI-related financial stability risks.
- Check CoreWeave's bond trading yields and credit spreads on financial data platforms like Bloomberg or TRACE.
- Examine Gartner's October 14 research report on AI infrastructure spending versus revenue projections.
- Analyze AMD's Q3 earnings report to disaggregate data center revenue between EPYC processors and AI accelerators.
- Review SEC filings for Anthropic's post-IPO private placement details and investor list.
- Assess NVIDIA's supply chain commitments and lease obligations disclosed in recent financial statements.
- Monitor interest rate trends and 10-year Treasury yields to evaluate refinancing risks for capital-intensive AI providers.
Full transcriptexpand
0:00Zoe: $116 billion. That's roughly how much revenue NVIDIA generated in just three months. Nash, put that in perspective for us.
0:09Nash: Data center revenue more than doubled from a year ago, free cash flow more than doubled compared with the previous quarter. And Vera Rubin, NVIDIA's newest AI computing architecture, is ramping faster than any architecture in the company's history. Operationally, these results are extraordinary.
0:26Unidentified speaker (UNKNOWN): But
0:27Zoe: then you open NVIDIA's financial disclosures: $550 billion in future contractual commitments, up to another $118 billion in contingent guarantees, and a surprisingly direct acknowledgement that many NeoClouds and AI labs are growing faster than their balance sheets can support.
0:45Nash: Think about what that means. The company selling the infrastructure is also helping select. selected customers secure the financing and capacity they need to buy and operate it. That doesn't make the demand artificial, but it makes the financial relationships a lot more complicated.
1:03Unidentified speaker (UNKNOWN): And
1:03Zoe: there's one more twist. Research firm Gartner believes that more than 80% of today's NeoCloud providers may no longer operate independently by 2030.
1:12Nash: So we have a contradiction. NVIDIA is reporting extraordinary growth. Its latest architecture is scaling exceptionally well. But parts of its customer ecosystem face hard questions about financing, depreciation, electricity prices, and long-term profitability.
1:30Zoe: Can both things be true at once?
1:32Nash: They can. A technology boom can be real even when some of its investors lose money.
1:37Unidentified speaker (UNKNOWN): Today
1:37Zoe: we're following that conversation. contradiction all the way through. From NVIDIA's earnings and the departure of Jensen Huang to the Federal Reserve's latest financial stability warning. From the economics of GPU rentals to the future of companies like Coreweave and Nebius. And later we'll ask whether AMD could benefit from NVIDIA's changing strategy. Welcome to Tech or Nothing. I'm Zoe.
2:00Unidentified speaker (UNKNOWN): And
2:00Nash: I'm Nash.
2:01Unidentified speaker (UNKNOWN): It's
2:01Zoe: Wednesday, November. 18th. And today we're talking about what happens when an extraordinary technology revolution meets the much less glamorous reality of financing it. Before we get into this week's markets, a quick recap. In our previous episode, we explored one of the biggest misunderstandings in the AI infrastructure boom. Nash, Cash, what was it?
2:22Nash: Backlog is not cash. We looked at companies building enormous amounts of computing capacity. Core Weave had reported roughly $132 billion in revenue backlog. Nebius was growing rapidly and expanding its AI infrastructure plans. And other emerging cloud providers were announcing increasingly ambitious projects.
2:43Unidentified speaker (UNKNOWN): So
2:44Zoe: where was the catch?
2:45Nash: A contract for future computing capacity isn't the same as a profitable data center. First, you need
2:50Unidentified speaker (UNKNOWN): to be able to get the computer.
2:51Nash: land, then electricity, then transformers, cooling, servers, networking, and GPUs. You need financing. And finally, the facility has to run reliably enough to turn those contracts into cash, and every delay has a cost.
3:07Zoe: Which led us to an uncomfortable question.
3:09Unidentified speaker (UNKNOWN): What
3:10Zoe: happens if the technology changes faster than the infrastructure can repay its original investment? This week, we're looking at that question from the other side of the transaction. The supplier, NVIDIA.
3:23Unidentified speaker (UNKNOWN): And
3:23Zoe: we should address something else. Last week we didn't discuss Jensen Huang's departure from NVIDIA. That wasn't because the event was unimportant.
3:33Unidentified speaker (UNKNOWN): It
3:33Zoe: was because the resignation was immediately surrounded by rumors about the circumstances, a possible success. successor, and even a possible return.
3:41Nash: And we didn't want to turn tech or nothing into a collection of rumors.
3:45Zoe: Since then, the board has clarified two things. It does not intend to reinstate him.
3:50Unidentified speaker (UNKNOWN): And
3:50Zoe: CFO Colette Cress is serving as interim CEO while the board searches for a permanent successor.
3:57Unidentified speaker (UNKNOWN): The
3:57Zoe: board cited personal conduct inconsistent with the company's code of conduct and says an investigation is ongoing.
4:03Unidentified speaker (UNKNOWN): And
4:04Nash: that's where we'll leave the personal side. We won't speculate about the circumstances or the outcome of that investigation.
4:11Unidentified speaker (UNKNOWN): So,
4:11Zoe: what does the leadership change mean for the business?
4:14Nash: Less than the headlines suggest, at least for now. Jensen Huang helped build one of the most consequential technology companies in the world, but a great company consists of considerably more than one individual. NVIDIA still has the leading AI computing platform, an exceptional engineering organization, the Siuda software ecosystem, and deep customer relationships.
4:37Zoe: And this week's numbers back that up.
4:39Unidentified speaker (UNKNOWN): In
4:39Zoe: the same week as the leadership change, NVIDIA reported record revenue, more than doubled its data center business, and called Ruben its fastest ramp ever.
4:48Nash: So we'll analyze NVIDIA the way we would. any market leader, buy its products, its numbers, and its strategic choices. We'll return to that shortly.
4:58Zoe: First, the macro radar. Because the AI industry is increasingly sensitive to something that gets far fewer headlines than GPU performance. The price of money.
5:08Nash: And this time, let's put real market interest rates on the table, not just hypothetical ones.
5:14Unidentified speaker (UNKNOWN): Let's
5:14Zoe: start with the Federal Reserve. On October 28th, the Fed left the federal funds target range unchanged at 3.75 to 4%.
5:23Unidentified speaker (UNKNOWN): But
5:23Zoe: the vote was 10 to 2.
5:25Unidentified speaker (UNKNOWN): And
5:25Zoe: both dissenters wanted to raise rates by another quarter point.
5:29Nash: That's the overnight policy rate. It's not what a NeoCloud pays on a multi-year loan.
5:34Unidentified speaker (UNKNOWN): So
5:34Zoe: what about long-term borrowing?
5:36Unidentified speaker (UNKNOWN): The
5:36Nash: better reference is the 10-year treasury yield. On November 17th, it stood at 5.1%. A few days earlier, it touched 5.15%. And companies don't borrow at the Treasury rate, they pay a credit spread on top. Sometimes they also pay arrangement fees, or they borrow against specific assets like GPUs.
5:56Zoe: Do we have a real example of what that spread looks like right now?
5:59Nash: Core Weave is the clearest one. Reuters reported on November 17th that a core weave bond issued in June at a yield of 9.625% had fallen in price. Its yield moved toward 15%.
6:13Unidentified speaker (UNKNOWN): So
6:13Zoe: Core Weave now pays 15%?
6:16Nash: No. The coupon on that existing bond doesn't change. But the market yield tells you what investors would demand today. That's roughly the price of the next dollar of credit.
6:26Unidentified speaker (UNKNOWN): So
6:26Zoe: there's a difference between financing that's already arranged and refinancing that still has to be negotiated.
6:32Nash: Exactly. And higher yields can come from two places higher treasury rates or wider credit spreads. Right now, the AI infrastructure sector is getting a bit of both. Rising electricity and construction costs are a separate pressure on top of that.
6:48Unidentified speaker (UNKNOWN): On
6:48Zoe: November 13th, the Fed released its latest financial stability report. And one figure stood out.
6:55Unidentified speaker (UNKNOWN): In
6:55Zoe: a survey of 23 market contacts, 87% identified AI-related developments among possible shocks to U.S. financial stability over the next 12 to 18 months.
7:06Unidentified speaker (UNKNOWN): In
7:07Zoe: the previous survey, that figure was 50%.
7:10Nash: That's a substantial jump.
7:12Zoe: But let's be precise about what it means. These 23 contacts work at banks, investment funds, broker dealers, and advisory firms.
7:21Unidentified speaker (UNKNOWN): They
7:21Zoe: weren't predicting an AI crash. They were naming risks. Elevated equity valuations, AI-enabled cybersecurity threats, and increasingly complex financing around AI infrastructure.
7:33Nash: And that last point matters. AI investment is no longer confined to technology companies. It's spreading into private credit, infrastructure funds, construction, utilities, real estate, and the broader financial system.
7:48Zoe: What does that look like for a single project?
7:50Nash: Take a large AI data center. The technology company may not own the building. An infrastructure fund may finance the property. Another company may own the power connection. A specialized cloud provider may lease the site and install the computing hardware,
8:05Unidentified speaker (UNKNOWN): and
8:05Nash: a model developer may commit to buying that capacity for several years.
8:09Zoe: So everyone sees a different asset.
8:12Nash: And a different source of revenue, but they may all depend on the same facility being delivered on time. If it's delayed, several contracts are affected at once.
8:22Zoe: Let's bring inflation into the picture.
8:24Unidentified speaker (UNKNOWN): When
8:24Zoe: we hear about inflation, we usually think of consumer prices.
8:28Unidentified speaker (UNKNOWN): What
8:28Zoe: does a data center developer see?
8:31Nash: Inflation in construction labor, electrical systems, cooling equipment, and power infrastructure. Imagine signing a multi-year computing contract based on a projected construction cost. Two years later, transformers are more expensive, the grid connection is delayed, and lenders want a higher return. The customer contract has hasn't changed, but the economics have.
8:54Unidentified speaker (UNKNOWN): So
8:54Zoe: the investment gets worse without any drop in customer demand.
8:57Nash: Right. A company can grow exceptionally quickly and still deliver disappointing returns to its shareholders.
9:04Unidentified speaker (UNKNOWN): How
9:04Zoe: big is the interest rate effect on its own?
9:07Nash: Here's a simple illustration: a $100 hundred million dollar project financed with $70 million of debt
9:13Unidentified speaker (UNKNOWN): at
9:14Nash: 9%. the annual interest bill is $6.3 million. At 15%, it's $10.5 million. That's an extra $4.2 million every year. And that assumes the debt doesn't grow. If a project is delayed, companies often have to borrow more. We'll come back to this model in our deep dive.
9:34Unidentified speaker (UNKNOWN): And
9:34Zoe: the hyperscalers?
9:35Nash: Microsoft, Amazon, and Alphabet already generate enormous amounts of
9:38Unidentified speaker (UNKNOWN): the market.
9:39Nash: cash. They can fund at least part of their AI infrastructure from existing operations. A neo cloud usually doesn't have that cushion. Its financial success depends much more directly on the computing infrastructure it operates.
9:52Zoe: Are all NeoClouds financed the same way?
9:55Nash: No. Some use secured loans, others rely on customer prepayments, some fund expansion with equity. and many use a combination. So there's no single interest rate for the sector. What matters is the cost of financing the next project, and whether it's expected return justifies that cost.
10:14Unidentified speaker (UNKNOWN): And
10:14Zoe: there's a stock market side to this too.
10:17Nash: There is. When rates stay high, investors discount future earnings more heavily, and companies whose profits lie far in the future feel that the most.
10:26Zoe: Which describes a lot of AI infrastructure providers.
10:29Nash: Their biggest investments happen today.
10:39Unidentified speaker (UNKNOWN): Now,
10:40Zoe: two developments that tell us where AI demand might be heading. Microsoft is holding its Ignite conference in San Francisco this week.
10:47Unidentified speaker (UNKNOWN): One
10:48Zoe: of the big themes is bringing AI agents under enterprise. control. Microsoft is expanding Agent 365, including tools for observing and managing agent deployments. Nash, why does that matter for infrastructure investors?
11:03Nash: Because the question is shifting, it's no longer just whether an agent can do a task, it's how agents operate inside a business. Who is responsible for them?
11:13Unidentified speaker (UNKNOWN): Which
11:13Nash: systems can they access? How are they governed? And above all, what measurable value do they create? Enterprise AI has to become a return on investment story.
11:24Unidentified speaker (UNKNOWN): And
11:24Zoe: that brings us back to NVIDIA.
11:26Nash: Right, the whole infrastructure industry is investing on the assumption that enterprises will keep demanding more computing power. That may be correct, but somewhere in the system, enterprise revenue ultimately has to pay for. those costs.
11:40Zoe: There's another event on the horizon: Anthropic's expected public market debut.
11:45Unidentified speaker (UNKNOWN): We
11:45Zoe: won't give an unconfirmed valuation, but one detail in the IPO filing is directly relevant to today's story. Anthropic's amended registration statement from November 10th describes a $26 billion private placement. It would close right after the IPO.
12:01Unidentified speaker (UNKNOWN): And the
12:01Zoe: buyers include. NVIDIA, AMD, Microsoft, Google, Broadcom, Micron, and a TSMC entity.
12:11Unidentified speaker (UNKNOWN): So
12:11Nash: some of the companies that sell chips and cloud capacity to AI labs are also investing in those same labs.
12:18Zoe: Is that a problem?
12:19Nash: Not by itself. It's a proposed transaction, not money already raised or spent on GPUs. And strategic investment. investments like this can be perfectly rational. But it shows why we have to separate final demand from intermediate purchases. An AI lab raises capital, commits to buying compute, and a cloud provider uses that commitment to buy hardware. Every step can be legitimate, but the financial strength of each participant still matters.
12:46Zoe: Our final macro observation comes from the data center property market.
12:51Unidentified speaker (UNKNOWN): On
12:51Zoe: November 16th, Blackstone's data center vehicle agreed to buy a 75% stake in 13 data center buildings leased to Amazon Web Services.
13:01Unidentified speaker (UNKNOWN): The
13:01Zoe: deal is worth about $5 billion. And AWS leases the buildings back on long term contracts.
13:08Unidentified speaker (UNKNOWN): And
13:08Nash: what investors are buying there are the buildings and the power connections, not the GPUs.
13:14Unidentified speaker (UNKNOWN): Why does
13:14Zoe: that distinction matter?
13:16Nash: Because buildings and power connections can stay useful for decades, GPUs have much shorter technology cycles. That difference becomes central when we talk about depreciation.
13:26Unidentified speaker (UNKNOWN): So
13:27Zoe: here's the macro picture. AI infrastructure demand is still growing, but the financial system is paying much closer attention to valuations, debt, interest rates, and whether these investments can earn adequate returns.
13:40Unidentified speaker (UNKNOWN): Now
13:40Zoe: let's see what NVIDIA itself reported. Nvidia's third fiscal quarter ended on October 25th. The company reported revenue of $115.8 billion. That's up 103% from a year earlier and 20% from the previous quarter. Analysts had expected around $109 billion.
14:01Unidentified speaker (UNKNOWN): For
14:01Nash: a company this size, that growth is exceptional.
14:04Unidentified speaker (UNKNOWN): And
14:05Zoe: Data Center.
14:06Nash: Data Center revenue reached $107.6 billion, up 110% year over year, and 21% quarter over quarter.
14:15Zoe: What do those numbers tell us?
14:17Nash: Two things. First, the AI infrastructure expansion is clearly showing up in NVIDIA's revenue. Second, NVIDIA still has enormous operating leverage.
14:26Zoe: Let's go through the profit lines. GAAP operating income reached about $77.5 billion. GAAP net income was about $65.3 billion. Diluted GAAP earnings per share came in at $2.70. Non-GAAP GAAP earnings per share were $2.68. And GAAP GAP gross margin was $70. 74.8%.
14:52Nash: So almost 75 cents of gross profit for every dollar of revenue. That isn't the bottom line margin, but it shows how extraordinary the economics of NVIDIA's hardware and platform business are.
15:04Zoe: Now the technical story. Vera Rubin. Architecture ramp in the company's history. Nash, why is that such a big deal?
15:14Nash: Because every major GPU transition carries execution risk. Customers have to validate the new systems. Manufacturing capacity has to expand. Memory, networking, and cooling have to be available. The software has to support the new architecture. And the whole system has to work at scale. If any of that goes wrong, the financial consequences can be enormous.
15:36Unidentified speaker (UNKNOWN): And
15:37Zoe: according to NVIDIA, it's going right.
15:39Nash: So far, yes. And at the same time, demand for Blackwell Ultra and standalone Vera CPUs keeps growing. The previous generation is still generating revenue while the next one scales. That's very valuable if customers are adding capacity rather than just replacing old systems.
15:57Unidentified speaker (UNKNOWN): Can
15:57Zoe: we say how much of the revenue came from Ruben?
15:59Nash: No. Nvidia doesn't report a separate Ruben revenue. figure. We know the ramp is fast because management says so. We can't say how many billions it contributed.
16:09Unidentified speaker (UNKNOWN): But
16:09Zoe: it does challenge one bearish narrative.
16:12Nash: Right. The idea that rapid GPU innovation must slow NVIDIA's own growth.
16:18Unidentified speaker (UNKNOWN): At
16:18Nash: least this quarter, NVIDIA is executing a major architecture transition without any visible break in its revenue trajectory.
16:26Zoe: What about gross margin? At 74.8%, it was slightly below the previous quarter's 75%.
16:34Nash: A decline of just 0.2 percentage points. NVIDIA says the early Ruben ramp caused that small dip. Compared with a year ago, a better Blackwell Ultra Mix helped. So we're watching one product generation mature while another scales up.
16:49Unidentified speaker (UNKNOWN): And
16:49Zoe: that matters even more. for the outlook. For the fourth fiscal quarter, NVIDIA expects revenue of $130 billion, plus or minus 3%. Analysts had expected about $125 billion.
17:04Unidentified speaker (UNKNOWN): And
17:04Zoe: on the earnings call, Colette Cress said NVIDIA now expects sales to grow about 75% next fiscal year, up from 70% previously.
17:13Unidentified speaker (UNKNOWN): But
17:14Nash: there's a calendar. detail in that Q4 guidance. The fourth quarter has an extra week, so part of the sequential jump simply reflects a longer reporting period.
17:23Unidentified speaker (UNKNOWN): And
17:23Zoe: the margin outlook?
17:25Nash: Lower. Nvidia expects GIAP and non-GIAP gross margin of 71 to 72%. At $130 billion of revenue, one percentage point of margin is worth about $1%. $1.3 billion of gross profit. Three points are worth almost $4 billion.
17:44Unidentified speaker (UNKNOWN): So
17:44Zoe: why is the margin falling?
17:46Nash: Nvidia hasn't given a full explanation for the Q4 guidance. It named the early Ruben Ramp as a factor in Q3. Beyond that, we shouldn't jump to conclusions. Not AMD, not memory costs, not discounts to NeoClouds. neo clouds. Those are questions to investigate, not answers.
18:06Unidentified speaker (UNKNOWN): And
18:06Zoe: does a lower margin mean lower profit?
18:08Nash: Not necessarily. If revenue keeps growing this fast, total gross profit can still rise even as the percentage falls.
18:16Unidentified speaker (UNKNOWN): So
18:16Nash: the Rubin ramp isn't failing. Quite the opposite. The real question is whether this new scale can sustain the profitability investors have come to expect.
18:26Unidentified speaker (UNKNOWN): Now
18:26Zoe: let's look at cash, because this is where NVIDIA differs most from many of the companies buying its infrastructure. Nvidia generated about $48.3 billion in operating cash flow during the quarter. Free cash flow was about $43.8 billion.
18:42Nash: That's after capital expenditures and the principal payments NVIDIA includes in its free cash flow definition. Almost $44 billion in a single quarter.
18:52Unidentified speaker (UNKNOWN): What
18:52Zoe: does that buy NVIDIA?
18:54Nash: Flexibility. It can invest in research, secure supply, support strategic partners, return capital to shareholders, and absorb some volatility in its ecosystem.
19:05Unidentified speaker (UNKNOWN): And
19:05Zoe: it did return capital. About $28.1 billion Dollars this quarter through buybacks and dividends, cash, cash equivalents, and marketable debt security. securities stood at about $53.5 billion.
19:18Unidentified speaker (UNKNOWN): And
19:19Nash: there's a second pool that's easy to miss. Nvidia also holds about $44 billion in publicly traded equity securities. That's largely stakes in other companies, some of them part of its own ecosystem. We'll come back to why that matters.
19:33Zoe: So this is a highly profitable cash-generating business.
19:37Unidentified speaker (UNKNOWN): Not
19:37Nash: a startup trying to build. its first data center.
19:40Unidentified speaker (UNKNOWN): But
19:40Zoe: there are two balance sheet figures worth watching. Accounts receivable reached about $81 billion. And days, sales outstanding rose from 60 to 64 days.
19:52Nash: In plain terms, NVIDIA is waiting a little longer to get paid. The company points to the timing of Rubin's Rubin shipments and to extended payment terms on certain large agreements with investment-grade investment grade customers.
20:03Unidentified speaker (UNKNOWN): Is
20:03Zoe: that a warning sign?
20:05Nash: Not automatically. Large enterprise contracts often have special terms. But one more detail stands out. Two direct customers account for 18 and 17% of those receivables. When a company grows as fast, you want to watch how quickly sales turn into cash and how concentrated the customers are.
20:24Zoe: Inventory also rose to about 30%. $39.7 billion.
20:29Nash: Nvidia says that supports the RubenRamp. During a fast product transition, that makes sense. If demand keeps growing, the extra inventory is justified. If demand slowed unexpectedly, the same inventory would become a different kind of problem.
20:43Unidentified speaker (UNKNOWN): So
20:44Zoe: the earnings report tells us two things at once.
20:46Unidentified speaker (UNKNOWN): The
20:47Nash: current business is exceptionally strong, and the next phase is becoming increasingly capital intensive.
20:53Zoe: Which brings us to the most surprising part of the prepared remarks the commitments. $550 billion. That's the combined amount of future commitments across two tables in Nvidia's latest CFO commentary.
21:07Unidentified speaker (UNKNOWN): The
21:07Zoe: first table totals $518 billion. The second adds up to $32 billion in conditional neoclassics. NeoCloud agreements. Nash,
21:17Unidentified speaker (UNKNOWN): is
21:17Zoe: that $550 billion of debt?
21:20Nash: No, and that's the most important distinction in this episode. It's an analytical sum of very different things. They aren't all debt, they aren't all immediate cash payments, and they're certainly not all losses.
21:33Zoe: Then let's break them down.
21:35Unidentified speaker (UNKNOWN): The
21:35Nash: largest category is supply and capacity commitments. $390 billion, up from $279 billion in the previous quarter. Nvidia says the increase mainly reflects multi-year memory procurement.
21:49Unidentified speaker (UNKNOWN): Why
21:49Zoe: would NVIDIA lock in that much supply?
21:52Nash: Because to sell enormous quantities of AI systems, it needs reliable access to scarce components, especially memory. If Ruben and the next platforms are going to scale, NVIDIA has to secure supply long before the final system ship.
22:06Unidentified speaker (UNKNOWN): So
22:07Zoe: these commitments protect future production.
22:09Nash: They do. They also create exposure if demand changes unexpectedly. But their purpose is very different from financing a neo cloud.
22:17Unidentified speaker (UNKNOWN): What
22:18Zoe: else is in the first table?
22:19Nash: $74 billion in data center leases that haven't started yet, $28 billion in cloud service agreements. $20 billion in planned equity investments, and $6 billion in capital expenditure commitments. Together, that's the $518 billion.
22:37Unidentified speaker (UNKNOWN): And
22:37Zoe: then the $32 billion in NeoCloud Agreements.
22:40Unidentified speaker (UNKNOWN): This
22:40Zoe: is where NVIDIA's language gets unusual.
22:43Unidentified speaker (UNKNOWN): The
22:44Zoe: company says that many NeoClouds and Frontier AI labs are growing faster than their balance sheets, and long-term credit profiles. can support.
22:52Unidentified speaker (UNKNOWN): And
22:52Zoe: that NVIDIA is helping selected partners secure the land, power, and compute they need.
22:58Nash: That's a remarkable thing for a hardware supplier to say. It's acknowledging that some customers can't finance their planned expansion on their own.
23:06Unidentified speaker (UNKNOWN): But
23:06Zoe: the agreements are conditional.
23:08Nash: Very much so. The partners have to secure financing, the infrastructure has to be built and become operational. And NVIDIA has to accept the capacity. $32 billion is the maximum under those conditions. And if certain criteria are met, NVIDIA may also share in the revenue these neo clouds earn from third-party customers.
23:27Unidentified speaker (UNKNOWN): So
23:27Zoe: walk us through the full loop.
23:29Nash: NVIDIA sells the infrastructure. Selected partners build cloud capacity with it. NVIDIA may commit to buying capacity from those facilities. and in some cases, it shares in the revenue they generate. That's a much more integrated relationship than simply selling chips.
23:46Unidentified speaker (UNKNOWN): It
23:46Zoe: may strengthen the ecosystem.
23:48Unidentified speaker (UNKNOWN): And
23:48Nash: it increases financial interdependence at the same time.
23:51Unidentified speaker (UNKNOWN): And
23:52Zoe: there's one more category: guarantees. NVIDIA discloses a maximum gross guarantee exposure of about $118 billion. Most of that, $105 billion, relates to an arrangement with SB Energy for a large OpenAI data center campus in Ohio. There are also $9.5 billion in power and systems credit support guarantees,
24:15Unidentified speaker (UNKNOWN): and
24:15Zoe: $3.5 billion related to land, power, and building infrastructure for neo clouds.
24:21Nash: And these guarantees don't all apply at once. The Ohio guarantees become effective in phases. building by building as each one is completed. The first phase is expected in fiscal 2029, and NVIDIA says it hasn't made any payments under these guarantees so far.
24:38Unidentified speaker (UNKNOWN): So
24:38Zoe: calling the whole amount a current liability would be misleading.
24:42Nash: Very misleading. But ignoring the guarantees would also be a mistake. If counterparties fail to meet their obligations, the exposure can become real.
24:51Zoe: Which brings us to the question at the center of this episode.
24:55Unidentified speaker (UNKNOWN): Is
24:55Zoe: NVIDIA becoming more than the supplier of the AI revolution?
24:59Unidentified speaker (UNKNOWN): Is
24:59Zoe: it also becoming one of the institutions financing the infrastructure that revolution depends on? Let's explore that through the NeoClouds. Nash, to understand the NeoCloud neo cloud business model, we need to go back a few years, don't we?
25:12Nash: We do. When generative AI does demand accelerated, advanced GPUs became extremely difficult to obtain. The large cloud providers were expanding, but demand for AI training capacity was growing faster than some of their infrastructure plans could accommodate.
25:28Unidentified speaker (UNKNOWN): And
25:28Zoe: that's where the specialists saw their opening.
25:30Nash: Right. Instead of building a complete public cloud with hundreds of services, they could focus on AI infrastructure, high high-performance performance. GPUs, fast networking, optimized storage, and large clusters for AI model developers.
25:46Zoe: Core Weave became one of the most prominent examples. Others entered or expanded into the market, including Nebius and IREN, although their activities and financing structures differ.
25:57Unidentified speaker (UNKNOWN): But
25:57Zoe: why would a customer choose one of them over AWS or Azure?
26:02Unidentified speaker (UNKNOWN): Because
26:02Nash: a customer doesn't necessarily need the entire AWS or Azure platform. It may just need a large amount of reliable GPU computing power.
26:11Unidentified speaker (UNKNOWN): If
26:12Nash: a Neo Cloud can deliver that faster, more efficiently, or at a competitive price, customers have a reason to use it. Specialization helped them address a very specific shortage.
26:23Unidentified speaker (UNKNOWN): So
26:23Zoe: the business case is not imaginary.
26:25Unidentified speaker (UNKNOWN): And the
26:25Zoe: rapid revenue growth. of several neo-cloud providers shows that substantial demand exists. Where's the catch?
26:32Unidentified speaker (UNKNOWN): The
26:33Nash: catch is the order of the cash flows. The industry requires enormous capital investment before much of the resulting revenue arrives. The company may need to acquire GPUs, arrange facilities, install networking equipment, secure electricity, and hire technical staff.
26:49Unidentified speaker (UNKNOWN): Then
26:50Nash: it has to operate
26:50Unidentified speaker (UNKNOWN): the company.
26:51Nash: capacity and earn enough over time to recover the cost.
26:54Unidentified speaker (UNKNOWN): And
26:54Zoe: unlike a diversified hyperscaler, a neo cloud may depend on a relatively narrow set of products and customers.
27:01Nash: Which creates concentration risk. If a major customer changes its infrastructure plans, delays deployments, or shifts workloads elsewhere, the financial impact can be substantial, especially when the infrastructure was built around that customer.
27:16Unidentified speaker (UNKNOWN): But
27:17Zoe: many of these companies have long-term contracts. Doesn't a five-year contract solve that problem?
27:22Nash: It helps. It doesn't solve it. We need to distinguish between a customer contract and an economically successful project. A five-year contract can provide visibility, it can help secure financing, it can reduce some demand uncertainty,
27:37Unidentified speaker (UNKNOWN): but
27:37Nash: it doesn't guarantee that. construction costs, electricity prices, hardware depreciation, and interest expense will remain within the original budget.
27:46Zoe: That's the central tension. And scale makes it sharper. AWS alone generates tens of billions of dollars in quarterly revenue. Microsoft and Alphabet also have exceptionally large, diversified revenue streams. Core Weave generated roughly 3.6% $3.7 billion in its latest reported quarter. Nebius reported just under $1 billion.
28:09Nash: Both are growing quickly, but they operate at a very different financial scale. And that affects how much financial stress they can absorb.
28:17Unidentified speaker (UNKNOWN): What
28:17Zoe: does that mean in practice?
28:19Unidentified speaker (UNKNOWN): A
28:19Nash: hyperscaler may be able to fund a major AI expansion from existing operating cash flows.
28:25Unidentified speaker (UNKNOWN): A
28:25Nash: NeoCloud neo cloud may rely much more heavily. on customer prepayments, debt, new equity, or project financing.
28:32Unidentified speaker (UNKNOWN): Is
28:32Zoe: that a red flag?
28:33Nash: Not necessarily. Infrastructure businesses have used external financing for decades, but it means the terms of the financing become a central part of the investment thesis.
28:44Unidentified speaker (UNKNOWN): And
28:44Zoe: this is where inflation and interest rates become potentially decisive. Nash, you've built a simplified financial model. for us. Walk us through it.
28:53Nash: We'll use the same hypothetical company in three different economic environments. That way we can see how the business model changes even when the underlying technology remains useful. Our fictional NeoCloud purchases $100 million of computing infrastructure. It finances $70 million through debt and $30 million through equity.
29:12Unidentified speaker (UNKNOWN): And
29:13Zoe: how quickly does that equipment lose value? in your model?
29:16Nash: For simplicity, we'll depreciate it over four years.
29:19Unidentified speaker (UNKNOWN): That
29:19Nash: means $25 million of annual depreciation
29:22Unidentified speaker (UNKNOWN): will
29:23Nash: also assume interest-only debt and ignore taxes and changes in working capital.
29:28Zoe: So, to be clear for our listeners, real companies have more complex capital structures.
29:34Unidentified speaker (UNKNOWN): This
29:34Zoe: is an educational model, not a forecast.
29:37Unidentified speaker (UNKNOWN): What
29:37Zoe: does the optimistic scenario? look like?
29:40Unidentified speaker (UNKNOWN): Let's
29:40Nash: call it the blue sky scenario. Demand remains strong, the company gets the infrastructure operational on time. Utilization is high. GPU rental pricing remains attractive. Annual revenue reaches $90 million. Operating costs, excluding depreciation and interest, are $30 million, and the debt carries 9% annual interest.
30:01Unidentified speaker (UNKNOWN): That's
30:02Zoe: $6.3%. million a year.
30:04Nash: So $90 million in revenue, minus $30 million in operating costs, minus $25 million in depreciation, minus $6.3 million in interest. That gives us a simplified pre-tax profit of $28.7 million.
30:20Unidentified speaker (UNKNOWN): A
30:20Zoe: highly attractive result.
30:22Unidentified speaker (UNKNOWN): And
30:22Zoe: it shows why investors can be enthusiastic about this market.
30:26Nash: If capacity is deployed efficiently and customers pay attractive prices, infrastructure assets can produce strong operating returns.
30:35Unidentified speaker (UNKNOWN): Now
30:35Zoe: let's make life harder.
30:36Nash: The second scenario, we call it higher for longer. Inflation remains elevated, construction and electricity costs are higher, new GPU capacity becomes available across the industry, putting some pressure on rental pricing, and financing becomes more expensive.
30:52Zoe: Give us the numbers.
30:53Nash: Revenue falls from $90 to $80 million. Operating costs rise from $30 to $34 million. And the debt interest rate rises from 9 to 12% when financing is refinanced. The annual interest bill becomes $8.4 million.
31:09Unidentified speaker (UNKNOWN): So
31:09Zoe: $80 in revenue, minus $34 in operating costs, minus $25 in depreciation. minus 8.4 in interest.
31:18Nash: And pre-tax profit falls to $12.6 million.
31:22Zoe: Still profitable, but much less attractive.
31:25Unidentified speaker (UNKNOWN): And
31:25Nash: here's why that matters: the company hasn't failed. The facility is still operating. Customers still use the infrastructure. But returns have deteriorated. If investors originally valued the project, assuming almost $30 million. in annual pretax profits, the revised economics could justify a substantially lower valuation.
31:45Unidentified speaker (UNKNOWN): In
31:45Zoe: the third scenario?
31:47Nash: The refinancing squeeze. Some capacity remains underutilized. Rental prices are weaker, inflation increases operating expenses, and new debt is expensive.
31:58Zoe: How expensive?
32:00Nash: Revenue falls to $65 million. Operating expenses rise to $38 million. Depreciation remains $25 million. And the interest rate on the debt reaches 15%. That's $10.5 million in annual interest. Subtract those costs, and the simplified annual pre-tax result becomes negative $8.5 million.
32:21Zoe: Same original infrastructure investment. Completely different financial outcome.
32:26Nash: That's our point. The technology keeps working while the accounting profit turns negative.
32:32Zoe: Does a loss of $8.5 million mean the company is about to default?
32:36Nash: No, and that's an important distinction. Look at the operating side. $65 million in revenue minus $38 million in operating costs still leaves $27 million before interest and depreciation. That easily covers $10.5%. million in interest. A negative pre-tax result isn't the same as negative free cash flow or a default. What decides that are principal repayments, cash reserves, the cost of replacing equipment, and the terms of the loans?
33:05Unidentified speaker (UNKNOWN): Let
33:05Zoe: me push back on one thing though. We've talked about interest rates all episode, but in your model, going from 9 to 15% costs $4.2 million. a year. Losing $25 million in revenue costs? $25 million.
33:20Nash: Fair challenge. In this model, utilization and rental prices are the bigger lever. The interest rate is what turns a weak year into a refinancing problem. But whether you have a weak year in the first place depends on how full the clusters are and what customers pay.
33:37Unidentified speaker (UNKNOWN): And I
33:38Zoe: assume these scenarios. aren't forecasts for Core Weave or Nebbius.
33:42Nash: They aren't.
33:43Unidentified speaker (UNKNOWN): We
33:43Nash: haven't tried to reproduce their actual cost structures. Real companies also use different depreciation schedules, leases, hedges, take-or-pay contracts, and customer prepayments.
33:55Unidentified speaker (UNKNOWN): And
33:55Nash: a fixed-rate borrower doesn't suddenly pay 15% just because market rates rise.
34:00Unidentified speaker (UNKNOWN): So
34:01Zoe: the refinancing date matters.
34:03Nash: It does, but it's not just the interest rate, the loan terms matter too.
34:07Unidentified speaker (UNKNOWN): What
34:08Zoe: do you mean by that?
34:09Nash: Lenders attach conditions to their loans. They're called covenants: a minimum amount of cash, a minimum ratio between cash flow and debt payments, limits on additional borrowing, and requirements for the collateral that secures the loan. The details vary from deal to deal, and to be clear, we're describing a mechanism, not claiming that Core Weave or Nebius has breached anything.
34:32Unidentified speaker (UNKNOWN): So
34:33Zoe: a loan can become more restrictive long before it matures.
34:36Nash: Exactly, and loans secured by GPUs add a second risk.
34:40Zoe: Walk us through it.
34:41Nash: Stay with our model. Say a lender values the GPU CPU cluster at $100 million against $70 million of debt. That's a loan. to value ratio of 70%. Now newer Rubin systems arrive, and the older cluster could only be sold for $60 million. The debt is still $70 million, so the loan to value ratio jumps to about 117%.
35:05Zoe: Before the company has missed a single payment.
35:08Nash: Right. That 60 million is purely hypothetical, and not every loan triggers an automatic demand for more. collateral. That depends on the contract. But lenders could ask for more security, faster repayment, a higher interest rate, or a lower credit line, or a new lender might simply refuse to finance the next cluster.
35:28Zoe: So that's the double hit. The old GPUs are worth less as collateral at exactly the moment new capital gets more expensive.
35:36Unidentified speaker (UNKNOWN): And
35:36Nash: that's the irony. Ruben Rubin can be an excellent. product for NVIDIA and at the same time make life harder for owners of older GPUs.
35:45Zoe: Let's sum up the scenarios.
35:46Nash: Several moderate changes can combine: a lower rental price, less utilization, higher costs, more expensive financing.
35:55Zoe: Each of those may look manageable on its own.
35:58Nash: Together, they can eliminate the expected return.
36:01Unidentified speaker (UNKNOWN): And
36:01Zoe: that's why a strong backlog doesn't settle the investment case.
36:04Unidentified speaker (UNKNOWN): Now
36:04Zoe: let's add Gartner's latest research. On October 14th, Gartner published a warning about the amount of capital being invested in AI infrastructure.
36:14Unidentified speaker (UNKNOWN): And
36:14Zoe: the figures are enormous. Gartner expects nearly $1 trillion in AI infrastructure spending during 2026, and approximately another $6 trillion between 2027 and 2030.
36:27Nash: Around $7 trillion in total. That's a forecast, not a completed investment program. And the required return makes the calculation even more demanding.
36:36Zoe: Gartner estimates that AI data center owners would need around $8.5 trillion in cumulative revenue through 2030 to cover planned spending, financing costs, and a 12% return on invested capital.
36:49Unidentified speaker (UNKNOWN): But it
36:50Zoe: forecasts. no more than $3.6 trillion in enterprise and consumer AI revenue over that period.
36:56Nash: That's a very large projected gap, nearly $5 trillion.
37:01Zoe: Which raises an uncomfortable question.
37:03Unidentified speaker (UNKNOWN): If
37:04Zoe: the businesses that use AI don't generate sufficient revenue, how does the infrastructure support the return investors expect?
37:11Nash: Some of the expected return may have to be written down.
37:14Unidentified speaker (UNKNOWN): And
37:14Zoe: Gartner puts a number on that. More than $3.5 trillion of infrastructure spending could be exposed to impairment risk by the end of 2030 if investment continues at the current pace. Over a longer horizon through 2032, the exposed amount could exceed $4 trillion.
37:33Unidentified speaker (UNKNOWN): Now
37:33Nash: let's be careful. Gartner isn't saying those losses have already happened, nor does potential impairment exposure mean every dollar will be written off. This is a model of what could happen under certain assumptions, and it's a model that investors should test rather than simply repeat.
37:50Unidentified speaker (UNKNOWN): So,
37:50Zoe: what's the case against Gartner?
37:52Nash: There are reasonable counterarguments. AI adoption could accelerate faster than Gartner currently anticipates. More efficient inference could create entirely new applications. Enterprises might deploy far more AI agents.
38:05Unidentified speaker (UNKNOWN): And
38:06Zoe: beyond the data center, robotics, autonomous vehicles, and scientific computing could expand the addressable market.
38:12Unidentified speaker (UNKNOWN): New
38:12Nash: infrastructure might also remain economically useful for longer than some forecasts assume. The underlying demand curve is still uncertain.
38:20Unidentified speaker (UNKNOWN): But
38:21Zoe: Gartner makes another prediction that's directly relevant to our topic. By 2030, it expects more than 80% of NeoCloud providers to cease operating independently. Nash, does that mean 80% go bankrupt?
38:35Unidentified speaker (UNKNOWN): No,
38:36Nash: and that's an important nuance. Some may be acquired, others may merge, some may sell their infrastructure and exit the market, and some may fail. A company can disappear as an independent business while its assets continue serving customers.
38:50Zoe: That's a fascinating possibility. The AI infrastructure can remain valuable even if the original financial structure was unsustainable.
38:58Unidentified speaker (UNKNOWN): What
38:59Zoe: would that look like in practice?
39:01Nash: Imagine a neo cloud that invested heavily in new GPU clusters. The business grows, but the debt becomes too expensive, and the company cannot finance the next expansion. A larger cloud provider acquires its assets at a discount.
39:16Unidentified speaker (UNKNOWN): And
39:16Zoe: from the customer's point of view?
39:18Nash: Very little changes. The infrastructure continues operating. Customers retain access to computing services, and the acquiring company can finance operations more cheaply.
39:28Unidentified speaker (UNKNOWN): So
39:29Zoe: the technology hasn't failed.
39:31Nash: But the original equity investment may have.
39:34Zoe: Which makes the distinction between enterprise value and shareholder value essential.
39:39Nash: A company can own valuable infrastructure and still deliver poor returns to its shareholders, especially if new financing, dilution, or restructuring shifts much of that value to creditors or new investors.
39:52Zoe: Gartner also forecasts AI optimized infrastructure as a service spending rising from $66 billion in 2027 to $174 billion. in 2030. That's substantial growth.
40:06Nash: But the research argues that it may still be insufficient for many AI compute providers to recover their investments before the current generation of accelerators is replaced.
40:15Zoe: Which brings us to another uncomfortable question.
40:18Unidentified speaker (UNKNOWN): What
40:18Zoe: happens to the economics of an older GPU when newer GPUs can perform the same work more efficiently? Nvidia's Ruben success is excellent news. for NVIDIA. But is it good news for the companies that already own large fleets of older GPUs?
40:33Nash: Not automatically. For some infrastructure owners, every major new architecture creates a new economic calculation. A NeoCloud may have financed a large cluster of older GPUs. Those GPUs can remain technically useful, but their competitive position changes when newer systems arrive.
40:50Zoe: Make that concrete for us.
40:52Nash: Imagine two computing clusters. One uses an older GPU generation, the other uses Ruben. If the new system can perform a given workload at significantly lower total cost, customers have an incentive to migrate.
41:07Unidentified speaker (UNKNOWN): So
41:07Zoe: the older GPU becomes worthless?
41:10Nash: No. Its rental price may fall to a level where it remains attractive for less demanding. workloads. It may continue running smaller models. It may serve customers who value lower prices more than maximum performance. And if the original equipment has already been substantially depreciated, lower rental rates can still generate positive operating cash flow.
41:31Unidentified speaker (UNKNOWN): So
41:31Zoe: technological obsolescence isn't a binary event?
41:34Nash: Right. But the financing can make it much more difficult. Imagine an older cluster that still owes lenders money based on the original acquisition cost. Its rental revenue may decline as newer systems become available, but the debt doesn't automatically decline at the same pace.
41:52Unidentified speaker (UNKNOWN): And
41:52Zoe: that's the hardware treadmill. Now add custom silicon to the picture. Amazon is developing trainium. Google operates its TPU family. Microsoft is investing in Maya. And AMD is offering an expanding alternative GPU platform.
42:09Unidentified speaker (UNKNOWN): Can
42:09Zoe: workloads simply move between all these systems?
42:12Nash: Not every workload can move easily. Software compatibility matters. Memory capacity matters. Networking matters. And the economics depend on the actual task. But the hyperscalers have a strong incentive to reduce their long-term cost per useful AI result.
42:29Unidentified speaker (UNKNOWN): So
42:29Zoe: if custom accelerators become more competitive for certain workloads, a portion of demand may migrate away from general purpose GPU rentals.
42:38Nash: Especially for predictable high-volume inference workloads.
42:41Zoe: That's a reasonable thesis.
42:43Unidentified speaker (UNKNOWN): But
42:43Zoe: isn't there a powerful counter argument?
42:46Nash: There is. Cheaper inference can expand the total market.
42:49Unidentified speaker (UNKNOWN): When
42:50Nash: serving an AI request becomes less expensive, new applications become commercially viable. Customers use AI more frequently. Developers integrate models into more products.
43:02Unidentified speaker (UNKNOWN): And
43:02Zoe: agents perform increasingly complex workflows.
43:05Nash: So the total amount of compute demanded can rise even while the cost per individual task falls.
43:49Unidentified speaker (UNKNOWN): Did
43:49Zoe: that already show up this quarter?
43:52Nash: It did. The publicly traded holdings produced net gains of about $1.8 billion. The private holdings took about $1.8 billion in impairments and unrealized losses. partly offset by gains. A falling market can reverse those gains without changing how many Rubin systems and video ships.
44:10Zoe: Fifth, the capacity agreements we just discussed.
44:14Nash: If the capacity arrives just as rental rates fall, those purchase commitments become less attractive. The revenue share could provide upside in good times, but we don't know enough about the contracts to put a number on either side.
44:27Zoe: And six?
44:28Nash: Supply and prices. The $390 billion in supply commitments secure NVIDIA's own production, mostly memory. They're not loans to NeoClouds. But if demand disappointed, those commitments and the $39.7 billion in inventory could weigh on margins.
44:45Zoe: And on prices?
44:47Unidentified speaker (UNKNOWN): A
44:47Nash: NeoCloud neo cloud in distress might sell used GPUs or undercut. current rental prices. That lowers the value of older hardware, and that can influence how much customers are willing to pay for new systems. It's a plausible feedback loop. We haven't seen it happen at scale.
45:03Unidentified speaker (UNKNOWN): And
45:03Zoe: there's one company that runs through almost all of these channels: OpenAI.
45:08Nash: Right. OpenAI buys compute from the big clouds. It's a customer behind neo-cloud neo cloud capacity contracts. and it's the future tenant of the Ohio campus behind NVIDIA's largest guarantee.
45:20Unidentified speaker (UNKNOWN): So
45:20Zoe: the exposures aren't really diversified?
45:23Nash: Not entirely. They don't share the same legal counterparty, and they won't necessarily fail together. But several relationships that look separate depend on the health of the same AI ecosystem. Adding up every headline exposure as if it were independent exaggerates. the risk. Ignoring their common drivers understates it.
45:43Zoe: Which is why the latest CFO commentary is so important.
45:46Unidentified speaker (UNKNOWN): It
45:47Zoe: puts those relationships in a much clearer financial perspective. Here's a strategic question for the next NVIDIA CEO. Should NVIDIA keep supporting the financing and development of selected AI infrastructure projects, or should it become more selective? Nash, make the case for continuing.
46:05Nash: NVIDIA has a powerful interest in making sure customers can deploy its technology. If land, electricity connections, or financing are bottlenecks, a technically superior GPU may not generate revenue as quickly as it could. Helping selected partners overcome those constraints could accelerate adoption. It may also strengthen customer relationships.
46:27Unidentified speaker (UNKNOWN): and
46:27Nash: it could preserve NVIDIA's position as the preferred infrastructure platform.
46:32Unidentified speaker (UNKNOWN): An
46:32Zoe: ecosystem strategy, not just a hardware strategy.
46:35Nash: And NVIDIA has the financial capacity to pursue that strategy on a scale most semiconductor companies couldn't contemplate. It generated almost $44 billion in free And cash flow in one quarter. So strategic capital allocation can be a meaningful competitive. advantage.
46:52Unidentified speaker (UNKNOWN): Now
46:52Zoe: the opposing argument.
46:54Nash: If NVIDIA begins supporting too many customers financially, it could increase exposure to precisely the industry risks we've been discussing.
47:02Zoe: Project delays, customer concentration, credit deterioration,
47:07Nash: lower GPU rental prices in technology transitions. A downturn in part of the AI infrastructure market could affect NVIDIA. not only through hardware demand, but also through selected investments and contractual arrangements.
47:21Zoe: So the downside becomes more interconnected,
47:23Unidentified speaker (UNKNOWN): which
47:24Zoe: raises another question. How do we know whether the additional sales justify the financial support?
47:29Nash: That's difficult to answer from the headline revenue numbers alone. Investors would ideally want to know the incremental revenue created, the amount of capital. committed, the probability of guarantees being called, and the expected return on strategic investments.
47:45Unidentified speaker (UNKNOWN): Some
47:45Zoe: of those details may be commercially sensitive.
47:48Nash: True, but the larger the commitments become, the more important the disclosure quality becomes.
47:55Zoe: Let's put the scale in perspective. Nvidia generated $43.8 billion of free cash flow in Q3.
48:02Unidentified speaker (UNKNOWN): the
48:02Zoe: $32 billion in conditional NeoCloud agreements, the $9.5 billion in power and systems guarantees,
48:09Unidentified speaker (UNKNOWN): and
48:09Zoe: the $3.5 billion in NeoCloud infrastructure guarantees add up to $45 billion.
48:15Nash: That's roughly one quarter of free cash flow, but they aren't comparable cash payments due in one quarter. Order, They differ in timing and conditions, and they aren't expected losses. The 105% billion Ohio guarantee is a separate, much larger maximum exposure.
48:32Unidentified speaker (UNKNOWN): So
48:32Zoe: NVIDIA isn't facing an insolvency risk.
48:35Nash: No. The real risk is to valuation. A credit squeeze could slow hardware orders, stretch collections, depress the value of NVIDIA's investments, and In pressure GPU rental prices all at the same time. Nvidia's cash flow is a for formidable buffer, but a stock is priced on expected growth and margins, not just on survival.
48:57Unidentified speaker (UNKNOWN): And
48:57Zoe: this is where NVIDIA's leadership transition matters. Jensen Huang was central to the company's long-term strategy.
49:04Unidentified speaker (UNKNOWN): But
49:04Zoe: the next CEO will inherit a much larger and more financially interconnected organization than the NVIDIA of even five years ago. What does that job require?
49:14Unidentified speaker (UNKNOWN): A
49:14Nash: great CEO must understand technology, but also capital allocation, risk management, governance, supply chain concentration, and the relationship between near-term growth and long-term returns.
49:27Unidentified speaker (UNKNOWN): So
49:27Zoe: where's the opportunity and where's the risk?
49:30Nash: The opportunity is to preserve NVIDIA's technological leadership while making its financial strategy more transparent and disciplined. The risk is that a new leadership team either becomes too conservative or extends commitments without sufficient economic discipline.
49:46Unidentified speaker (UNKNOWN): And
49:46Zoe: these decisions now sit with interim leadership. Colette Cress is interim CEO, and she has been NVIDIA's CFO for many years.
49:54Nash: Her experience is an asset, but investors will still want clear governance over any new guarantees and strategic commitments.
50:01Unidentified speaker (UNKNOWN): A
50:01Zoe: transition can improve governance. but it doesn't guarantee better decisions.
50:05Unidentified speaker (UNKNOWN): And
50:05Nash: there is something else worth remembering. Nvidia's long-term ambition extends beyond the data center.
50:11Zoe: Robotics, autonomous vehicles, physical AI, and emerging applications in space computing.
50:17Nash: These businesses are at very different stages of commercial development, but they underline that NVIDIA is trying to build a broader, accelerated computing platform. The company has developed robotic software and simulation platforms. It is working with automotive partners on autonomous driving technology, and it has introduced computing technology intended for satellites and autonomous systems operating in space.
50:41Zoe: But none of that replaces today's data center revenue.
50:45Nash: No, but it could broaden the company's future opportunities.
50:48Unidentified speaker (UNKNOWN): So
50:48Zoe: the question isn't whether NVIDIA should abandon. the data center market.
50:52Nash: That would make little sense given its current position.
50:55Unidentified speaker (UNKNOWN): The
50:56Nash: question is whether NVIDIA needs to finance as much of the surrounding infrastructure ecosystem as it currently plans, or whether its superior technology and software platform can support the business with less exposure to customer financing risk.
51:10Zoe: That's a legitimate strategic debate, and one that will become more important as the industry matures. Let's bring the pieces together.
51:17Unidentified speaker (UNKNOWN): We
51:18Zoe: have NVIDIA reporting extraordinary growth, a new computing architecture scaling exceptionally quickly, hyperscalers continuing to invest enormous sums, and AI native cloud companies expanding capacity.
51:31Nash: We also have high capital requirements, potential refinancing stress, and growing financial commitments across the supply chain.
51:39Zoe: So, Nash, is the AI infrastructure boom too large?
51:43Nash: That depends on how much economically valuable demand develops.
51:47Zoe: And when
51:48Nash: that's an important distinction. A large infrastructure investment may eventually be justified, but still generate poor returns if revenue arrives much later than expected.
51:58Zoe: Give us an example.
51:59Nash: A data center scheduled to begin operating in 2027. might make excellent economic sense if it opens on time. But if the power connection slips into 2029, the financing and depreciation assumptions can change substantially. And by then, the hardware roadmap may have advanced.
52:17Zoe: Which is one reason to distinguish between land, power, buildings, and GPUs.
52:22Nash: Land and power infrastructure can have very long economic lives. Buildings can remain useful across multiple hardware generations. GPUs generally have much shorter economic and technological cycles. So treating all AI infrastructure assets as though they depreciate at the same rate would be a mistake.
52:40Unidentified speaker (UNKNOWN): So
52:40Zoe: the same project can contain both attractive and unattractive components.
52:44Nash: A strategic power connection might become more valuable over time, while an older GPU cluster at the same site. becomes less competitive. The infrastructure stack has different economics at every layer.
52:56Unidentified speaker (UNKNOWN): And
52:56Zoe: that makes consolidation more complicated. What would a buyer actually want from a distressed NeoCloud?
53:03Nash: Power connections, land, customer relationships, experienced technical teams, and existing computing capacity. Even if shareholders lose money, those assets may continue. contributing to the AI economy.
53:17Unidentified speaker (UNKNOWN): Which is
53:17Zoe: why Gartner's consolidation forecast deserves attention without being treated as a prediction of universal failure.
53:24Unidentified speaker (UNKNOWN): A
53:24Nash: smaller number of better capitalized infrastructure providers could emerge, and the surviving companies might be stronger.
53:31Unidentified speaker (UNKNOWN): But
53:31Zoe: the transition could be painful for equity holders.
53:34Nash: That's the paradox of capital-intensive innovation.
53:37Unidentified speaker (UNKNOWN): A
53:37Zoe: technology can transform the world, and the investors who financed its first wave can still earn disappointing returns. After all that discussion about NVIDIA, it's time for our stock idea. Today we're looking at AMD. Nash, does AMD automatically benefit when NVIDIA runs into trouble?
53:56Nash: No. If NVIDIA's growth slows because the entire AI market is weakening, AMD faces the same
54:02Unidentified speaker (UNKNOWN): thing.
54:03Nash: demand pressure. The more interesting scenario is different. AI demand keeps growing. Customers want a second supplier, and AMD becomes more competitive on performance, software, and total cost of ownership.
54:17Zoe: So it's a market share opportunity, not a bet against AI.
54:21Nash: That's the idea.
54:22Zoe: Let's start with AMD's latest results.
54:25Unidentified speaker (UNKNOWN): Third
54:26Zoe: quarter revenue. reached $13.4 billion, up 45% from a year earlier.
54:32Unidentified speaker (UNKNOWN): The
54:32Zoe: data center segment generated $8.5 billion, up 97%,
54:38Unidentified speaker (UNKNOWN): and
54:38Zoe: AMD guided to about $17 billion of revenue for the fourth quarter, plus or minus 3%.
54:45Nash: That's serious growth, but we have to be careful with the data center number. It includes EPYC service. server processors as well as AI accelerators. So 8.5 billion isn't pure AI GPU revenue.
54:57Zoe: What's driving the AI side?
54:59Nash: AMD's instinct roadmap, and especially the MI-450 series, together with ROCM, the software ecosystem around the hardware.
55:08Unidentified speaker (UNKNOWN): And
55:08Zoe: that's where the competition gets interesting.
55:11Nash: Because Nvidia's advantage isn't just as chips, it's the combination. of hardware, software, networking, system design, and developer familiarity. CUDA is deeply embedded in many AI workflows. Moving a large production deployment to another platform can take significant engineering work.
55:29Zoe: So customers don't switch just because a competing GPU is cheaper.
55:33Nash: Right. They need the whole system to deliver performance, reliability, and software service. support.
55:39Unidentified speaker (UNKNOWN): Could
55:40Zoe: NVIDIA's leadership change open a door for AMD?
55:43Nash: Imagine new leadership takes a more conservative approach to strategic investments. Some customers would have to become more financially independent. That could push some of them to diversify their hardware, but it could also reduce how much they can invest overall. So the effect on AMD is ambiguous.
56:02Unidentified speaker (UNKNOWN): It
56:02Zoe: depends on whether NVIDIA's problem is company-specific or industry-wide.
56:07Nash: That's the key question, and it gives us three scenarios.
56:10Zoe: Scenario one:
56:12Nash: NVIDIA keeps executing exceptionally well, and Ruben keeps a substantial technological lead. AMD can still grow with a larger AI market, but winning share will be hard, especially where NVIDIA's software and system integration are strong. strongest.
56:27Zoe: Scenario 2.
56:29Nash: Hyperscalers and other customers actively diversify. AMD wins more deployments, its software improves, and customers get comfortable running mixed infrastructure. That's the bullish case for AMD.
56:41Unidentified speaker (UNKNOWN): And
56:42Zoe: scenario three:
56:43Nash: AI infrastructure spending slows substantially. Customers postpone projects. Both NVIDIA and AMD face weaker demand. In that case, a small gain in market share may not mean strong earnings growth.
56:57Unidentified speaker (UNKNOWN): So
56:57Zoe: NVIDIA's problems aren't automatically AMD's opportunity. Let's compare profitability. AMD reported a GAAP gross margin of 54% in the third quarter. Nvidia reported 74.8%.
57:12Nash: Those are company-wide figures with very different business mixes, so it's not a direct comparison of AIGPU profitability, but it shows how different the two economic profiles are. Nvidia operates at extraordinary scale and profitability. AMD is still building its position in several markets.
57:31Zoe: Which creates both opportunity and risk for investors.
57:35Nash: If AMD gains share and improves margins, earnings growth could be substantial. But if the market has already priced in years of aggressive expansion, even strong results may not deliver attractive stock returns. The expectations matter as much as the execution.
57:52Zoe: An AMD has its own capital needs.
57:55Nash: It does. Expanding a data center business takes inventory, engineering investment, and manufacturing partnerships. And a big customer win isn't automatically profitable. AMD has to deliver on time and add an adequate margin.
58:10Zoe: What about valuation?
58:12Nash: That's the piece we haven't covered today. We've looked at revenue and margins. But a fair verdict also needs the current share price, the forward earnings, multiple and analyst expectations. So we see AMD as a research idea. not a price target call.
58:29Unidentified speaker (UNKNOWN): Then
58:29Zoe: what makes AMD interesting over the longer term?
58:32Unidentified speaker (UNKNOWN): As
58:33Nash: the market moves beyond a single dominant platform, software portability could improve. More customers may use different accelerators for different workloads, and AI infrastructure may become more diverse. That creates room for AMD without NVIDIA losing its leadership position. Both companies can grow. just at different rates and with different economics.
58:55Zoe: So AMD is our stock idea because it offers another way to participate in AI infrastructure. Not because we expect NVIDIA to disappear, and not because a change of CEO automatically changes technical competitiveness.
59:10Nash: The investment case depends on customer adoption, accelerator revenue. ROCM maturity, margins, and valuation. Over a six to twelve month horizon, we'd also watch how much growth is already in the share price and how AMD trades relative to the broader semiconductor market.
59:29Unidentified speaker (UNKNOWN): A
59:29Zoe: research thesis, not a buy signal.
59:32Nash: But an important company to watch as the AI computing market evolves.
59:36Unidentified speaker (UNKNOWN): Our
59:36Zoe: startup corner takes us to Darmstadt in Germany. to a company called DaisyTuner.
59:41Unidentified speaker (UNKNOWN): On
59:42Zoe: November 16th, it announced a 6 million euro seed round. Eliya and LEA Ventures co-led the round, with participation from the Amadeus Apex Technology Fund and Seed Camp. Its product, DaisyFlow, aims to improve how AI models run across different hardware.
59:59Nash: That may sound less exciting than building a giant GPU cluster. But after everything we've discussed today, it's a very relevant business. The next phase of AI competition may not be decided only by who owns the most chips, it may also be decided by who uses them most efficiently.
1:00:17Zoe: Explain the problem it's solving.
1:00:19Nash: AI models are essentially long chains of mathematical operations. The hardware has to execute those operations efficiently. But different accelerators have different architectures, memory systems, and performance characteristics. A model tuned for one platform may run much less efficiently on another. Developers can spend a lot of time and money tuning the software by hand.
1:00:40Unidentified speaker (UNKNOWN): And
1:00:40Zoe: Daisy Tuner wants to automate that.
1:00:43Nash: Exactly. It's building compiler technology that automates part of that optimization. Its inference platform adapts how a model. runs to the hardware underneath.
1:00:53Zoe: Does it work?
1:00:54Unidentified speaker (UNKNOWN): The
1:00:54Nash: company says its platform runs some models up to 18% faster than a widely used open source inference engine on AMD instinct GPUs. That's a company benchmark for a specific setup. It still needs independent testing across different models, hardware, and production workloads.
1:01:11Zoe: But even a few percent can matter.
1:01:13Nash: Absolutely. When a company spends millions of dollars on compute, even modest efficiency gains have real financial value.
1:01:21Unidentified speaker (UNKNOWN): How
1:01:21Zoe: does that connect to today's theme?
1:01:23Nash: We've spent this episode on how expensive it is to finance new infrastructure. Daisy Tuner comes at the problem from the opposite direction. What if you can get more useful work from infrastructure that's already installed? What if better software lets a customer postpone some hardware purchases? or run more inference on alternative accelerators like AMDs?
1:01:43Unidentified speaker (UNKNOWN): That
1:01:44Zoe: could shift the balance between NVIDIA, AMD, and Custom Silicon.
1:01:48Nash: And it could improve the economics of some existing GPU clusters.
1:01:52Zoe: Is there a downside?
1:01:54Nash: For the suppliers, yes. If each task needs less compute, infrastructure providers may see lower prices for some workloads. Optimization
1:02:02Unidentified speaker (UNKNOWN): is a very important thing to do.
1:02:03Nash: helps customers, but it creates new competitive pressure for suppliers. Better technology doesn't distribute the financial benefits equally.
1:02:11Zoe: Which may be the unofficial theme of today's episode. Daisy Tuner is early stage, funding isn't proof of commercial success,
1:02:19Unidentified speaker (UNKNOWN): and
1:02:20Zoe: its business model still has to prove itself.
1:02:23Nash: But it represents an interesting part of the AI industry. the software layer that makes the hardware layer more efficient.
1:02:30Zoe: And after a week of hundreds of billions in capital commitments, it's refreshing to find a company whose ambition is to make existing systems do more. So where do we finish? Nvidia has just reported another extraordinary quarter. Revenue more than doubled from a year ago. Data center revenue reached almost $108 billion. Free cash flow approached $44 billion. And Vera Rubin is ramping faster than any previous NVIDIA architecture.
1:02:58Nash: That's the operational picture, a remarkably strong business.
1:03:02Zoe: But the financial picture around the AI infrastructure ecosystem is getting more complicated.
1:03:08Nash: Nvidia has hundreds of billions of dollars in future commitments, and they're very different things. Supply purchases. leases, conditional capacity agreements, equity commitments, and credit guarantees. It supports selected partners through capacity arrangements, strategic investments, and guarantees. And the company itself says many of those customers are expanding faster than their balance sheets can support.
1:03:32Unidentified speaker (UNKNOWN): That
1:03:32Zoe: doesn't make the demand imaginary.
1:03:35Nash: But it raises the question of who carries the financing risk.
1:03:38Zoe: Gartner's warning reminds us that a technology boom doesn't guarantee attractive returns for every infrastructure owner.
1:03:45Unidentified speaker (UNKNOWN): And
1:03:45Zoe: our three NeoCloud scenarios showed how interest rates, operating costs, utilization, and rental prices can transform the economics of the same investment.
1:03:55Nash: That's the difference between strong demand and a strong balance sheet.
1:03:59Unidentified speaker (UNKNOWN): And
1:03:59Zoe: the most interesting conclusion is that seemingly contradictory outcome. outcomes can happen at the same time.
1:04:05Nash: Nvidia can keep growing, Ruben can become a major technological success. Enterprises can keep adopting AI, and some NeoCloud shareholders can still lose money.
1:04:15Unidentified speaker (UNKNOWN): Which
1:04:15Zoe: is why investors have to look beyond the headline backlog, beyond the announced gigawatts,
1:04:21Unidentified speaker (UNKNOWN): and
1:04:21Zoe: beyond the GPU benchmarks.
1:04:23Nash: The real questions are who funds the assets, who uses them. how much revenue they produce, and who ultimately earns the return.
1:04:31Zoe: Between today and Saturday, there are three things we'll be watching. First, the market's verdict on NVIDIA's results. In after hours trading, the shares rose almost 3%, after closing at about $155. That's still well below the roughly $174 before Jensen Huang's resignation.
1:04:51Nash: The question is whether that gain holds in regular trading and how much weight investors put on the lower Q4 margin guidance.
1:04:58Zoe: Second, any new details about Nvidia's NeoCloud agreements, its strategic financing, and the search for a permanent CEO.
1:05:05Nash: And third, whether hyperscalers and AI native infrastructure companies keep showing healthy demand while the market becomes more selective about balanced. balance sheet strength.
1:05:16Zoe: Because the next phase of the AI boom may be less about announcing capacity and more about proving returns.
1:05:22Unidentified speaker (UNKNOWN): The
1:05:22Nash: infrastructure still has to pay for itself.
1:05:25Unidentified speaker (UNKNOWN): And
1:05:25Zoe: that's a good place to leave
1:05:26Unidentified speaker (UNKNOWN): it. A
1:05:27Zoe: successful technology isn't automatically a successful investment.
1:05:31Unidentified speaker (UNKNOWN): And
1:05:31Nash: sometimes the winners aren't the companies that build the most. They're the ones that build what customers need at a cost the business can. sustain.
1:05:40Zoe: Thanks for joining us on Tech or Nothing. Before we go, our usual disclaimer: Tech or Nothing is an independent podcast focused on technology, financial markets, business models, and the economic consequences of innovation. The information presented in this episode is intended exclusively for general information, education, and editorial discussion.
1:06:03Unidentified speaker (UNKNOWN): It
1:06:03Zoe: does not constitute personalized financial advice, investment advice, legal advice, tax advice, or a recommendation to buy, sell, or hold any security or financial product. We discuss publicly traded companies, private businesses, and emerging technologies because they illustrate broader developments in the market. Mentioning a company does not constitute an endorsement of its products, management, shares, or investment strategy. All financial information is based on sources available at the time of research and recording. These sources may include company earnings releases, regulatory filings, investor presentations, central bank publications, industry research, and reputable financial news reporting.
1:06:48Unidentified speaker (UNKNOWN): While
1:06:49Zoe: we aim to use reliable sources, and distinguish reported facts from our interpretations, data may be incomplete, subsequently revised, or subject to different definitions and accounting treatments. Any opinions, scenarios, estimates, or forward-looking statements reflect analytical judgments rather than established future outcomes.
1:07:09Unidentified speaker (UNKNOWN): They
1:07:09Zoe: involve uncertainty and should not be treated as guarantees. In particular, our discussion of a new NVIDIA, AMD, AI infrastructure financing, neo-cloud consolidation, and future industry profitability involves substantial uncertainty about market demand, technological progress, interest rates, financing conditions, customer behavior, and competition. Illustrative calculations used in this episode are educational examples. All modeled NeoCloud revenues, costs, interest rates, and collateral values are hypothetical and are not forecasts for any specific company. Market bond yields should not be confused with the contractual interest rates on existing debt. Company benchmarks, including those from Daisy Tuner, are company claims that have not been independently verified. Investments in technology. companies can involve substantial volatility. Businesses with significant financing needs, customer concentration, high capital expenditure requirements, or rapidly depreciating assets may be exposed to additional risks. These include refinancing risk, liquidity risk, dilution, changes in asset valuations, execution delays, and potential loss of invested capital.
1:08:27Historical performance does not guarantee future performance. Listeners should evaluate their own financial circumstances, objectives, and risk tolerance, and consult qualified independent professionals where appropriate.
1:08:41Unidentified speaker (UNKNOWN): The
1:08:41Zoe: views expressed on Tech or Nothing are those of the speakers and do not necessarily reflect the views of any employer, financial institution, or company mentioned. Tech or nothing does not guarantee the accuracy, completeness, or continuing relevance of the information discussed.
1:08:57Nash: And after today's episode, perhaps one final reminder: extraordinary revenue growth doesn't eliminate the need to understand where the cash comes from.
1:09:05Unidentified speaker (UNKNOWN): Or
1:09:06Zoe: where it goes. Thanks for listening. See you on Saturday. Nvidia 8K, A K CEO C O Transition, at 12 November 2026, HTTPS, WW W at Seed.com. C Eidgov Archives Edgar. Data 010545A0602A NVIDIA announces C O transit EI. HTM NVIDIA Q3FY227 press release. HTTPS uh egov uh archives Edgar Data 0104 F801045 Busomaro 2608 Q3Y27 PR HTMID HTM NVIDIA Q3FY227 CFO commentary uploaded source. NVIDIA Supplemental Financial Information T5 October 2026 to TPS WWSIC Govhar Archives Edgar Data 10458 00 58 Dao 608. Supplemental Financial Info HTM Board Statement on Huang Consensus Figures FYC28 Outlook After Hours Move GPS Finance.yahoo.com Technology. Live NVIDIA. Earnings. Um, live NVIDIA. Posts. Q3 Beatboard says it will not. Reinstate Jensen. Huang as C O two four to eight four theater three. HTML Federal Reserve FOMC statement twenty-eight October twenty-six. HTTPS on Jadu Federal Reserve.gov. Sa news events on press releases. Leases Monetary twenty 2610 six one zero msede. MCDA HTM US Treasury Daily. Par
1:10:25yield curve, November 2026. HTPS Home.treasury.gov resource center data chart center interest rates. Mun type daily treasury. Yield curve Reuters 17 November 26. NVIDIA turmoil marks the end of easy money for the AI trade. HTTPS www.reuters.com business. NVIDIA turmoil marks, and easy money AI trade. 2026 to 11 SETA 17, Federal Reserve Financial Stability Report, 13th November 2026. HTTPS SOW dot federalreserve.gov. Publications are files Financial Stability Report TNOT6 Muawamon one 3. three PDF Anthropic S1A, s one A 10 ten November 2026. twenty twenty six A summary, HTTPS. SeerletsData Science dot com news Anthropic sets IPO range and ads twenty six billion private placement F six Savs six one nine Blackstone BXDC AWS sixteen november twenty twenty six HTTPS the nextweb.com R News Blackstone BXDC A bias data centers five BN Gartner 14 october twenty 26. six at GTPS, TPS wgartner wgartner.com dot com
1:11:38and and newsroom press releases 2026 to 10. 14 Gartner says over us dollars three and a half trillion of artificial intelligence. Infrastructure spending could be exposed to impairment by the end of 2030 A. MDQ3 2026. 8K Sh TPS
1:11:55Unidentified speaker (UNKNOWN): W
1:11:56Zoe: dot sec.govon Sec.gov on archives eggger. Data 000482 zero zero zero four eight two 6LOY six cello wad three two zero six nine one HTM Daisy Tuner seed round sixteen november twenty twenty six HTPS TTEC.eu twenty six I eleven sixteen zero AI inference startup Daisy Turner raises your six M seed.