TechOrNothing

TechOrNothing: Anthropic's $90B IPO Bet, CoreWeave Leverage, and AI Infrastructure Risks

60 min Listen to the original Episode page
AI InfrastructureIPO AnalysisCoreWeaveAnthropicApplied MaterialsFederal ReserveCybersecurityFinancial Metrics

Summary

Hosts Nash and Zoe analyze the AI infrastructure build-out using a gold rush analogy, contrasting model labs with hardware and financing providers. They critique CoreWeave's high leverage and reliance on customer prepayments, noting that adjusted EBITDA masks significant depreciation and interest costs. The episode details Anthropic's IPO pricing, its $900.9 billion in long-term purchase commitments, and a $26 billion strategic investment from major tech firms. Applied Materials is highlighted as a profitable 'shovel seller' with strong cash flow, while a security breach at Fireworks underscores the risks of transparency and trust in AI-as-a-service.

Episode sections

  1. 0:00 – 4:58

    Introduction and AI Gold Rush Analogy

    Hosts introduce the episode's theme using a California Gold Rush analogy to describe the current AI infrastructure build-out. They outline the week's key topics, including earnings from CoreWeave, Nebius, Cerebras, Cisco, and Applied Materials, as well as Anthropic's IPO pricing and Federal Reserve policy.

  2. 4:58 – 8:43

    Federal Reserve Policy and Interest Rate Environment

    Speakers discuss the Federal Reserve's recent rate decisions, noting a hold at 3.25%-3.75% with dissenting votes for a hike. They highlight Governor Christopher Waller's suggestion of a potential 50 basis point hike in December if core inflation remains sticky, and the market's pricing of such a move.

  3. 8:43 – 18:56

    CoreWeave Financial Analysis

    The hosts analyze CoreWeave's Q3 earnings, noting $3.66 billion in revenue and a $132.5 billion backlog. They critique the company's high leverage, with debt doubling to $46.9 billion, and highlight the significant interest expense ($846 million) that turns a 62% adjusted EBITDA margin into a net loss.

  4. 18:56 – 22:08

    Nebius Financial Analysis

    Nebius is discussed as a faster-growing but less profitable alternative to CoreWeave. The hosts note that Q3 revenue included a $325 million one-time payment, and that capital expenditures are expected to rise to $27 billion while revenue guidance is lowered. They highlight Nebius's reliance on customer prepayments to fund its $40-$45 billion 2027 CapEx.

  5. 22:08 – 24:12

    Cerebras Financial Analysis

    Cerebras is presented as a hardware and cloud service provider with $231 million in core revenue. The hosts point out that while the company reported a net loss under standard accounting due to stock-based compensation, it was profitable on its own 'core' measure. They note a $26.1 billion remaining performance obligation and a new hyperscaler cloud deal.

  6. 24:12 – 25:47

    Cisco Financial Analysis

    Cisco is analyzed as a 'shovel seller' benefiting from AI networking demand, with Q1 revenue up 24% and AI infrastructure revenue guidance raised to $9 billion. The hosts note a margin compression from 68% to 66% due to a shift toward lower-margin hardware sales.

  7. 25:47 – 26:34

    Anthropic IPO Details

    The segment concludes with details from Anthropic's updated IPO prospectus. The company plans to offer 825 million Class A shares priced at $70-$78, implying a gross offering value of up to $64.35 billion. The shares are expected to trade on NASDAQ under the ticker ANT.

  8. 26:34 – 27:02

    Anthropic's IPO Structure and Supply Chain Investments

    Nash details a $26 billion strategic private placement where Nvidia, Google, AMD, Broadcom, Micron, Microsoft, and a TSMC affiliate invest in Anthropic. Zoe notes this represents the AI supply chain buying into a major customer.

  9. 27:02 – 28:46

    Anthropic's Financials and Massive Commitments

    Nash reports Q3 revenue of ~$17 billion with a ~$0.5 billion operating loss. The prospectus highlights $900.9 billion in purchase commitments and leases for compute/infrastructure stretching to 2045, with ~$70.7 billion due within 12 months of Sept 2026.

  10. 28:46 – 30:07

    Anthropic vs. OpenAI Capital Access

    Nash compares Anthropic's IPO to OpenAI's funding, citing OpenAI's March round ($122B committed, $852B valuation) and a November 6 report of a new $11B tranche at a $1.25T valuation. They conclude public access provides benchmarking but doesn't force OpenAI's hand.

  11. 30:07 – 35:26

    API Pricing Models and Token Economics

    Nash and Zoe break down Anthropic's API pricing for Claude Opus 5.5, Sonnet, and Haiku, noting input vs. output costs and batch discounts. They discuss whether token price declines outpace production costs, emphasizing throughput, utilization, and reasoning tokens.

  12. 35:26 – 37:15

    Applied Materials Q4 Fiscal 2026 Results

    Nash reports Applied Materials Q4 fiscal 2026 revenue of $10.41 billion (+53%) and adjusted EPS of $4.16 (+92%). Gross margin rose to 50.6% and operating margin to 35.3%, driven by product mix, higher prices, and operating leverage.

  13. 37:15 – 41:43

    Applied Materials Business Segments and Growth Drivers

    Nash explains Applied's semiconductor systems revenue ($8B, +64%), driven by Foundry/Logic (58%), DRAM (33%), and Flash (9%). DRAM equipment sales nearly doubled due to AI-driven HBM demand. Advanced packaging revenue is expected to grow ~90% in calendar 2026.

  14. 41:43 – 59:57

    Applied Materials Cash Flow and Shareholder Returns

    Applied generated $3.3 billion in operating cash flow and $2.62 billion in free cash flow (a record). The company returned $2.42 billion to shareholders via $2 billion in buybacks and $421 million in dividends, contrasting sharply with CoreWeave's cash burn.

Assets discussed

AssetKindStanceContext
NVIDIANVDAequityneutralA group of chip and cloud companies, including NVIDIA, Google, and Microsoft, plans to buy $26 billion of anthropic stock.
AlphabetGOOGLequityneutralA group of chip and cloud companies, including NVIDIA, Google, and Microsoft, plans to buy $26 billion of anthropic stock.
MicrosoftMSFTequityneutralA group of chip and cloud companies, including NVIDIA, Google, and Microsoft, plans to buy $26 billion of anthropic stock.
Anthropicprivate companyneutralAnthropic and OpenAI.
OpenAIprivate companyneutralAnthropic and OpenAI.

“Stance” is how the speakers talked about the asset, not a recommendation.

Claims worth verifying

CoreWeave Q3 revenue was $3.66 billion, up 168% year-over-year, with a backlog of $132.5 billion.

Nash ·9:05

How to verify: CoreWeave Q3 2026 Earnings Report or Press Release, available on CoreWeave's investor relations page or SEC EDGAR if filed as a public company (note: CoreWeave is private, so check official company announcements or press releases). For recent events, verify 2026-08-14 through 2026-11-14. Use another period only when the speaker explicitly names it.

Primary source

CoreWeave's adjusted EBITDA margin was 62%, but operating income was ~$200 million and interest expense was $846 million, resulting in a net loss of $709 million.

Nash ·9:57

How to verify: CoreWeave Q3 2026 Earnings Report or Press Release. For recent events, verify 2026-08-14 through 2026-11-14. Use another period only when the speaker explicitly names it.

CoreWeave's debt increased from $21.4 billion at the start of the year to $46.9 billion at the end of September.

Nash ·11:23

How to verify: CoreWeave Q3 2026 Earnings Report or Press Release. For recent events, verify 2026-08-14 through 2026-11-14. Use another period only when the speaker explicitly names it.

CoreWeave's average cost of debt was 7.9% at the end of the quarter, down 2.5 percentage points from a year earlier.

Unidentified speaker (UNKNOWN) ·13:56

How to verify: CoreWeave Q3 2026 Earnings Report or Press Release. For recent events, verify 2026-08-14 through 2026-11-14. Use another period only when the speaker explicitly names it.

CoreWeave expects Q4 interest expense of $1.05 to $1.15 billion.

Nash ·16:41

How to verify: CoreWeave Q3 2026 Earnings Report or Press Release. For recent events, verify 2026-08-14 through 2026-11-14. Use another period only when the speaker explicitly names it.

Nebius Q3 revenue was $981 million, up 571% year-over-year, including a $325 million one-time payment.

Nash ·19:04

How to verify: Nebius Q3 2026 Earnings Report or Press Release, available on Nebius's investor relations page. For recent events, verify 2026-08-14 through 2026-11-14. Use another period only when the speaker explicitly names it.

Nebius expects 2026 CapEx of ~$27 billion and 2027 CapEx of $40-45 billion.

Nash ·20:23

How to verify: Nebius Q3 2026 Earnings Report or Press Release. For recent events, verify 2026-08-14 through 2026-11-14. Use another period only when the speaker explicitly names it.

Cerebras Q3 core revenue was $231 million, up 70% year-over-year, with a $26.1 billion remaining performance obligation.

Nash ·22:31

How to verify: Cerebras Q3 2026 Earnings Report or Press Release, available on Cerebras's investor relations page. For recent events, verify 2026-08-14 through 2026-11-14. Use another period only when the speaker explicitly names it.

Cisco Q1 revenue was $18.5 billion, up 24%, with AI infrastructure revenue guidance raised to $9 billion.

Nash ·24:30

How to verify: Cisco Q1 FY2027 Earnings Report, available on Cisco's investor relations page or SEC EDGAR (ticker: CSCO). For recent events, verify 2026-08-14 through 2026-11-14. Use another period only when the speaker explicitly names it.

Primary source

Anthropic plans to offer 825 million Class A shares at $70-78, trading under ticker ANT on NASDAQ.

Nash ·26:03

How to verify: Anthropic IPO Prospectus (S-1/F-1) filed with the SEC, available on SEC EDGAR (ticker: ANT). For recent events, verify 2026-08-14 through 2026-11-14. Use another period only when the speaker explicitly names it.

Primary source

Fireworks detected a security incident on October 6th, with attackers entering through a third-party cloud provider in late October.

Nash ·51:51

How to verify: Check Fireworks' official security blog or press release updates from October 2026 for incident timelines and third-party provider involvement. For recent events, verify 2026-08-14 through 2026-11-14. Use another period only when the speaker explicitly names it.

On November 9th, Fireworks published an update stating account/billing info, secrets, and incomplete model weights were affected for a limited set of customers.

Zoe ·52:02

How to verify: Review Fireworks' November 2026 public security advisory or customer notification letters for details on affected data types and scope. For recent events, verify 2026-08-14 through 2026-11-14. Use another period only when the speaker explicitly names it.

Fireworks' core inference services showed no impact from the breach, and payment card numbers were not affected.

Unidentified speaker (UNKNOWN) ·52:26

How to verify: Verify via Fireworks' post-incident investigation report or third-party security audit findings published after November 2026. For recent events, verify 2026-08-14 through 2026-11-14. Use another period only when the speaker explicitly names it.

Notion confirmed it was affected by the Fireworks breach, with fewer than 1% of its workspaces impacted.

Nash ·52:55

How to verify: Check Notion's official security incident report or blog post from late 2026 confirming the breach scope. For recent events, verify 2026-08-14 through 2026-11-14. Use another period only when the speaker explicitly names it.

Fireworks processes 40 trillion tokens a day.

Zoe ·55:52

How to verify: Verify against Fireworks' public usage statistics, press releases, or third-party API monitoring data for daily token volume. For recent events, verify 2026-08-14 through 2026-11-14. Use another period only when the speaker explicitly names it.

Mine Core Weave has a backlog of more than $100 billion and an interest bill four times its operating profit.

Nash ·55:01

How to verify: Check Mine Core Weave's financial filings (if public) or investor presentations for backlog figures and interest expense vs. operating profit ratios. For recent events, verify 2026-08-14 through 2026-11-14. Use another period only when the speaker explicitly names it.

Primary source

Anthropic has almost $900 billion of computing commitments, much of it with companies now buying its shares.

Nash ·56:02

How to verify: Review Anthropic's IPO prospectus or S-1 filing for disclosed infrastructure commitments and related-party transactions with investors. For recent events, verify 2026-08-14 through 2026-11-14. Use another period only when the speaker explicitly names it.

Primary source

OpenAI raised money at a valuation above a trillion dollars.

Zoe ·56:11

How to verify: Check OpenAI's latest private placement terms or press releases regarding valuation from 2026. For recent events, verify 2026-08-14 through 2026-11-14. Use another period only when the speaker explicitly names it.

Applied Materials is returning money to shareholders and has rising margins.

Unidentified speaker (UNKNOWN) ·55:41

How to verify: Review Applied Materials' quarterly earnings reports (10-Q/10-K) for share repurchase programs and gross margin trends. For recent events, verify 2026-08-14 through 2026-11-14. Use another period only when the speaker explicitly names it.

Primary source

Cisco raised its AI forecast again.

Zoe ·55:34

How to verify: Check Cisco's earnings call transcripts or press releases from 2026 for updated AI-related revenue guidance. For recent events, verify 2026-08-14 through 2026-11-14. Use another period only when the speaker explicitly names it.

Do your own research

  • Review CoreWeave's Q3 10-Q filing for debt maturity schedules and recourse provisions.
  • Verify Anthropic's S-1 registration statement for the $900.9 billion in purchase commitments and related party transactions.
  • Check Applied Materials' Q4 earnings transcript for China revenue exposure and export control risks.
  • Analyze Nebius's financial disclosures for the $325 million one-time payment and CapEx funding structure.
  • Assess Fireworks' public disclosures or press releases regarding the scope of the security breach and customer impact.
  • Compare CoreWeave's adjusted EBITDA to net income to understand the impact of depreciation and interest expenses.
  • Review Federal Reserve meeting minutes for Governor Waller's dissenting vote and inflation forecasts.
  • Examine Anthropic's API pricing pages for input vs. output token costs and batch discounts.
Full transcriptexpand
0:00
Zoe: Welcome to Tech or Nothing, where technology meets markets, business models, and occasionally a balance sheet that could use a therapist. I'm Zoe.
0:08
Unidentified speaker (UNKNOWN): And
0:08
Nash: I'm Nash, and this week the therapist is fully booked.
0:11
Zoe: Last Wednesday, we explored the blockbuster economy.
0:15
Unidentified speaker (UNKNOWN): We
0:15
Zoe: asked whether spectacular technology automatically creates a spectacular business.
0:20
Unidentified speaker (UNKNOWN): And
0:20
Nash: we compared technology ETFs, because sometimes the most interesting investment question isn't which company. has the best technology.
0:28
Zoe: It's which part of the industry actually captures the value.
0:32
Unidentified speaker (UNKNOWN): And
0:32
Nash: this week the AI industry sent us about 500 pages of answers.
0:37
Zoe: Core Weave, Nebius, Cerebras, Cisco, and applied materials, all reported earnings.
0:43
Nash: Anthropics set the price range for what could become one of the largest IPOs in history.
0:48
Unidentified speaker (UNKNOWN): And
0:48
Zoe: the Federal Reserve is walking into its December. meeting with inflation still refusing to behave. One Fed governor is openly talking about a half-point rate hike.
0:58
Nash: Which matters a lot more than you'd think. If your business model involves borrowing tens of billions of dollars to build data centers.
1:05
Zoe: Not exactly a quiet week.
1:07
Nash: Our stock idea is applied materials, after a quarter in which adjusted earnings per share grew 92%.
1:14
Unidentified speaker (UNKNOWN): And in
1:14
Zoe: Startup Corner, we're looking at Fireworks AI,
1:17
Unidentified speaker (UNKNOWN): a
1:17
Zoe: company serving more than 40 trillion tokens a day.
1:21
Nash: Which sounds enormous.
1:23
Zoe: Until I ask my favorite question:
1:25
Nash: How much money does it actually keep?
1:27
Zoe: Precisely. Because the entire AI industry seems to be celebrating bigger numbers.
1:33
Nash: More GPUs, more tokens, more contracts, more capital expenditure.
1:37
Unidentified speaker (UNKNOWN): But
1:37
Zoe: I'm interested in the number that comes. after all of those.
1:40
Nash: Profit.
1:42
Zoe: Preferably followed by cash.
1:44
Nash: Ambitious standards for a technology podcast.
1:47
Zoe: Someone has to have them.
1:48
Nash: So to make sense of all this, we're going back in time.
1:52
Zoe: California, 1849.
1:55
Unidentified speaker (UNKNOWN): The
1:55
Nash: gold rush.
1:56
Zoe: Tens of thousands of people are heading west. Some leave everything behind, others borrow money, buy equipment. and join the rush. Mining camps turn into towns almost overnight.
2:07
Nash: And here's the part people forget: the miners weren't the only ones making money. Often they weren't even the ones making the most.
2:15
Zoe: Merchants sold shovels and pans. Transport companies moved supplies. Refiners Miners turned rough ore into something you could actually sell. And banks lent everyone money.
2:26
Unidentified speaker (UNKNOWN): One
2:26
Nash: boom. completely different businesses.
2:29
Zoe: With completely different risks. The miner has to find gold and get it out of the ground profitably. The shovel seller needs customers who can pay.
2:37
Unidentified speaker (UNKNOWN): And
2:37
Zoe: the banker needs borrowers who pay back their loans.
2:40
Unidentified speaker (UNKNOWN): So
2:40
Nash: the same goal rush produces winners, losers, and a lot of people who were very busy and ended up with nothing.
2:47
Zoe: And that almost one one-to-one, to one is the AI economy in twenty twenty. 2026.
2:52
Unidentified speaker (UNKNOWN): The
2:52
Nash: gold is intelligence.
2:54
Zoe: The prospectors are the model labs, Anthropic, OpenAI, and others, staking claims on hundreds of billions of dollars of future computing power.
3:03
Unidentified speaker (UNKNOWN): The
3:03
Nash: mines are companies like Coreweeb and Nebious. They dig the actual compute out of the ground, one data center at a time.
3:11
Unidentified speaker (UNKNOWN): The
3:11
Zoe: shovel sellers make the machines that make the chips. Today that's applied materials.
3:16
Unidentified speaker (UNKNOWN): The
3:16
Nash: refineries turn raw compute into answers customers will pay for. That's Fireworks AI.
3:22
Zoe: And the banks? Bond investors, lenders, equity investors, and more and more often, the customers themselves paying years in advance under long-term Term contracts.
3:34
Nash: Except our modern miners aren't buying a few shovels.
3:37
Unidentified speaker (UNKNOWN): We're
3:37
Nash: talking about infrastructure. costing billions and commitments reaching into the hundreds of billions.
3:43
Unidentified speaker (UNKNOWN): And
3:43
Zoe: here's the twist that makes this gold rush different.
3:46
Unidentified speaker (UNKNOWN): Some
3:46
Zoe: of the companies selling the equipment are also buying shares in the companies buying it.
3:57
Nash: A group of chip and cloud companies, including NVIDIA, Google, and Microsoft, plans to buy $26 billion of anthropic stock.
4:05
Unidentified speaker (UNKNOWN): So
4:05
Zoe: the shovel sellers are investing in the prospectors.
4:08
Unidentified speaker (UNKNOWN): And
4:09
Nash: the same filing shows the prospector has signed enormous, mostly long-term purchase commitments with several of those same companies.
4:17
Unidentified speaker (UNKNOWN): The
4:17
Nash: share purchase itself doesn't create those obligations, but the money clearly flows in both directions.
4:23
Unidentified speaker (UNKNOWN): It's
4:23
Zoe: a very small town.
4:25
Unidentified speaker (UNKNOWN): And
4:25
Nash: there's one more piece to this. the price of the gold itself.
4:29
Zoe: Tokens.
4:30
Nash: Exactly. AI providers increasingly charge by usage, measured in tokens, and different models and different ways of serving them come at very different prices.
4:40
Zoe: Which means more demand doesn't automatically mean more revenue at the same rate.
4:45
Unidentified speaker (UNKNOWN): So
4:45
Nash: today we're asking two questions. How expensive is it to build the AI economy?
4:50
Unidentified speaker (UNKNOWN): And how
4:50
Zoe: profitable
4:50
Unidentified speaker (UNKNOWN): is the same.
4:51
Zoe: it to sell what that infrastructure produces?
4:53
Nash: Let's start with the part of the economy that makes almost everything more expensive.
4:58
Unidentified speaker (UNKNOWN): The
4:58
Zoe: bond market.
5:00
Nash: You make it sound personal.
5:01
Unidentified speaker (UNKNOWN): After
5:02
Zoe: reading Corweave's interest bill, it is personal.
5:05
Nash: Quick context first, because it matters for everything that follows. In September, the Federal Reserve raised its policy rate by a quarter point to a range of 3.2%. 3.75 to 4%.
5:16
Unidentified speaker (UNKNOWN): And
5:17
Zoe: on October 28th, it held there, but not unanimously.
5:21
Nash: 10 to 2. Governor Lisa Cook and Cleveland Fed President Beth Hammock both wanted another quarter point hike right away.
5:29
Zoe: So the debate isn't whether the Fed is tightening, it's how fast.
5:33
Unidentified speaker (UNKNOWN): And
5:33
Nash: then came the October inflation report. Headline consumer prices rose 0.4%.
5:39
Nash: from the previous month.
5:41
Unidentified speaker (UNKNOWN): And
5:41
Zoe: the core number?
5:42
Nash: 0.3%.
5:44
Nash: For the third month in a row, economists had expected 0.2.
5:48
Zoe: Core meaning without food and energy, because those two jump around so much that they can drown out the underlying trend.
5:55
Nash: Right. And here's a strange detail: the annual inflation rate couldn't even be published this time.
6:01
Zoe: Why not?
6:02
Nash: Because during the government shutdown last year, the Bureau of Labor statistics didn't collect prices in October 2025. So there's nothing to compare October 2026 against.
6:13
Zoe: Economics is a very glamorous profession.
6:16
Unidentified speaker (UNKNOWN): Now
6:16
Nash: the interesting part is what Governor Christopher Waller said just hours before that report. On November 10th, in a speech in London, he said he was already leaning toward a quarter-point quarter point hike in December.
6:27
Zoe: Even if inflation had improved?
6:29
Nash: Even then, and he went further, if core inflation came in around 0.3% or higher, he said a 50 basis point hike should be considered. His argument was that hiking only a quarter point could put the Fed's credibility at risk.
6:43
Unidentified speaker (UNKNOWN): And
6:44
Zoe: it came in at 0.3?
6:46
Nash: Right at the threshold, yes.
6:48
Zoe: So, 50 basis points in December. Done deal.
6:52
Nash: Not so fast. Waller is one voice on a committee of 12 voters, and Fed Chair Kevin Walsh has been famously reluctant to give any guidance at all about the rate path.
7:02
Unidentified speaker (UNKNOWN): The
7:02
Zoe: Fed chair who doesn't like to talk about rates, very relaxing.
7:06
Unidentified speaker (UNKNOWN): And
7:06
Nash: New York Fed President John Williams struck a slightly different tone. Speaking in Frankfurt on November 6th, he said inflation had been too high for
7:14
Unidentified speaker (UNKNOWN): the years.
7:15
Nash: too long, but he also backed the idea of taking time to gather more information.
7:19
Unidentified speaker (UNKNOWN): So
7:20
Zoe: the honest summary is inflation is still a problem, most officials agree.
7:25
Unidentified speaker (UNKNOWN): And
7:25
Zoe: they disagree about speed.
7:27
Nash: And markets have made up their minds faster than the Fed. After the inflation report, futures were pricing an almost certain hike in December, to four 4 to four 4. and a quarter percent.
7:38
Unidentified speaker (UNKNOWN): Okay.
7:38
Zoe: Okay, but why does any of this matter for an AI data center?
7:42
Nash: Because a data center is basically a giant pile of expensive hardware, often bought with borrowed money, and the price of that money moves with interest rates.
7:51
Unidentified speaker (UNKNOWN): But
7:51
Zoe: not one to one with the Fed rate.
7:53
Nash: Not even close. A company's borrowing cost depends on a reference rate, plus a spread, plus the terms of the loan, maturity, collapse. collateral, hedging.
8:04
Zoe: Explain the spread for everyone who doesn't speak bond.
8:07
Nash: The spread is the extra interest lenders demand on top of a reference rate to compensate them for the risk that you might not pay them back. The riskier you look, the wider the spread.
8:18
Unidentified speaker (UNKNOWN): So
8:18
Zoe: the government borrows cheaply, a solid company borrows a bit more expensively, and a fast-growing fast growing company with a lot of debt pays a premium.
8:26
Nash: Exactly. And even inside one company, an old fixed-rate loan and a brand new loan can have very different costs.
8:33
Zoe: Which is especially uncomfortable when you have to keep building.
8:37
Nash: Because your financing needs don't end with the first data center.
8:40
Zoe: Which brings us to our first mine.
8:43
Nash: Core weave.
8:44
Unidentified speaker (UNKNOWN): If
8:44
Zoe: the AI Gold Rush has a mine mind that never sleeps, it's this one.
8:49
Unidentified speaker (UNKNOWN): One
8:49
Nash: quick housekeeping note before we dig in, and then I promise we'll talk like normal people. Everything in this segment comes from Core Weave's own filings, and anything forward-looking is management's forecast, not a fact.
9:02
Zoe: Noted.
9:03
Unidentified speaker (UNKNOWN): Now
9:03
Zoe: let's start with the good news.
9:05
Nash: Third quarter revenue, $3.66 billion, up 168% from a year earlier.
9:13
Unidentified speaker (UNKNOWN): And
9:13
Zoe: revenue grew by more than a billion dollars. dollars in just one quarter.
9:17
Unidentified speaker (UNKNOWN): And
9:17
Nash: the backlog, meaning contracted future revenue, about $132.5 billion.
9:24
Zoe: That's the slide every CEO dreams about.
9:27
Nash: It is. But let's read what it actually is. It's revenue core weave expects to recognize under committed customer contracts if it delivers the capacity on time as promised.
9:38
Unidentified speaker (UNKNOWN): So
9:39
Zoe: not $132 billion sitting in a bank account.
9:42
Nash: No, it's $132 billion of homework.
9:46
Zoe: Okay, but here's my read.
9:49
Unidentified speaker (UNKNOWN): This
9:49
Zoe: is the miner who found the vein.
9:51
Unidentified speaker (UNKNOWN): Look
9:51
Zoe: at the adjusted EBITDA margin. 62%. That's not a struggling business.
9:57
Nash: It's a miner who found the vein, yes, with a very expensive excavator bought on credit.
10:03
Zoe: Go on, ruin it.
10:05
Nash: Adjusted EBITDA leaves out depreciation and interest. And for this particular business, those are exactly the two things it's made of.
10:13
Zoe: Expensive hardware that wears out.
10:15
Unidentified speaker (UNKNOWN): And
10:15
Nash: the borrowed money that paid for it. So let's walk down the income statement. Depreciation and In amortization, almost $1.9 billion in the quarter. Operating income. about $200 million.
10:29
Zoe: From a 62% margin to $200 million. That's quite a drop.
10:33
Nash: And then interest expense, $846 million.
10:38
Zoe: Wait, in one quarter?
10:40
Nash: In one quarter.
10:41
Unidentified speaker (UNKNOWN): So
10:41
Zoe: the interest bill is four times what the business actually earns from operations.
10:46
Nash: About 4.1 times, which is how you get a net loss of $7.5%. $709 million out of a 62% margin.
10:54
Zoe: That's not a margin. That's a magic trick.
10:57
Nash: It's what happens when adjusted leaves out the parts that cost the most. Different profit measures answer different questions. Adjusted EBITDA tells you how the operations run before the big structural costs. The full income statement tells you what's left after them.
11:12
Unidentified speaker (UNKNOWN): And
11:12
Zoe: if you build data centers full of hardware, those costs. aren't decorative.
11:16
Nash: They're the whole building.
11:18
Zoe: All right. Show me the balance sheet. How much borrowed money are we talking about?
11:23
Nash: About $46.9 billion of debt at the end of September.
11:27
Zoe: And at the start of the year?
11:29
Nash: About $21.4 billion.
11:32
Unidentified speaker (UNKNOWN): So
11:32
Zoe: the debt more than doubled in nine months.
11:34
Nash: Up about $25.5 billion, roughly $199. 119%.
11:40
Zoe: Somewhere a bond investor just felt a chill.
11:43
Nash: And the market noticed. In the weeks before these results, the stock had lost about 40% of its value, and some core weave bonds were trading at yields in the mid-teens.
11:53
Zoe: Mid-teens. That's not a bank loan, that's a credit card.
11:57
Nash: And about $10.6 billion of that debt is classified as current, meaning due within
12:02
Unidentified speaker (UNKNOWN): the $10%.
12:03
Nash: 12 months.
12:04
Unidentified speaker (UNKNOWN): That
12:04
Zoe: sounds alarming.
12:05
Nash: It's worth watching, but not a panic button. Repayments can come from different sources, not only new borrowing. The company's repayment schedule shows about $2.4 billion due in the rest of this year and about $11 billion in 2027.
12:21
Zoe: How much cash do they have?
12:23
Nash: About $4.1 billion in unrestricted cash. And shareholders' equity on the books, about $4 billion.
12:30
Unidentified speaker (UNKNOWN): So
12:30
Zoe: debt is more than 11 times book equity.
12:33
Nash: On the accounting figures, yes. Book equity isn't market value, and that ratio alone doesn't tell you whether a default is coming, but it tells you how thin the cushion is.
12:43
Unidentified speaker (UNKNOWN): Now
12:43
Zoe: I remember there's an important distinction inside that debt.
12:47
Nash: Recourse versus non-recourse.
12:49
Unidentified speaker (UNKNOWN): In
12:49
Zoe: normal English, please.
12:51
Nash: With recourse debt, lenders can go after the parent company, Core Weave itself. With non-recourse debt, they can generally only go after specific assets or subsidiaries, apart from some contractual exceptions.
13:04
Unidentified speaker (UNKNOWN): So
13:04
Zoe: if one data center project goes wrong, a non-recourse lender can't grab the whole company.
13:09
Nash: That's the idea. And here's the split: about 41.6 billion is recourse.
13:15
Nash: and only about 5.3 billion is non-recourse.
13:19
Unidentified speaker (UNKNOWN): So
13:19
Zoe: almost 90% of the debt ultimately points back to the parent company.
13:23
Nash: Right. Which is why the parent company's cash flow matters so much.
13:28
Unidentified speaker (UNKNOWN): And
13:28
Zoe: the loan covenants?
13:29
Unidentified speaker (UNKNOWN): The
13:29
Zoe: rules borrowers have to stick to?
13:32
Unidentified speaker (UNKNOWN): Core
13:32
Nash: Weav says it was in compliance with all of them at the end of September.
13:36
Zoe: Important.
13:37
Unidentified speaker (UNKNOWN): A lot of
13:38
Zoe: debt is not. the same thing as a broken loan agreement.
13:41
Unidentified speaker (UNKNOWN): And
13:41
Nash: large future obligations are not a default today.
13:44
Zoe: And to be fair, management is working on the cost. They say their average cost of debt was about 7.9% at the end of the quarter, down two and a half percentage points from a year earlier.
13:56
Unidentified speaker (UNKNOWN): And
13:56
Nash: they've said they won't use the unsecured bond market at current yields. They'd rather borrow against specific customer contracts.
14:03
Zoe: Which brings us to the part I really want to understand: cash flow.
14:07
Nash: Here's the good news first. Operating cash flow was positive, about $1.98 billion in the quarter.
14:15
Zoe: That's progress.
14:16
Nash: It is. Now compare it with the shovel spending. Cash spent on property and equipment, including capitalized software, $9.44 billion.
14:27
Zoe: In three months?
14:29
Nash: In three months.
14:30
Unidentified speaker (UNKNOWN): So
14:30
Zoe: operating cash flow covered only about a fifth of the investment.
14:34
Nash: About 21%. Subtract one from the other, and you get roughly minus $7.5 billion.
14:40
Unidentified speaker (UNKNOWN): A
14:40
Zoe: simplified free cash flow?
14:42
Nash: Simplified, yes. And if anything, it understates the build. CoreWeb's Core Weave's own CapEx measure for the quarter was closer to $14.5%.
14:51
Nash: billion because part of the equipment is financed through leases and payables rather than paid in cash right away.
14:58
Unidentified speaker (UNKNOWN): So
14:58
Zoe: how do you keep a build out this size going?
15:01
Nash: Debt, other capital, and increasingly, customers paying up front.
15:05
Unidentified speaker (UNKNOWN): And
15:06
Zoe: here's where I'll push back on you because this is genuinely clever. Core Weave isn't digging blind. Management says customers prepaid on about 70% of this year's. deals, some of them up to the full contract value.
15:18
Unidentified speaker (UNKNOWN): And
15:18
Zoe: they expect around $9 billion in prepayments and deposits over the next three months alone.
15:23
Nash: That's real money. And more than $3 billion of it comes from re-renting older capacity at higher prices than originally.
15:30
Unidentified speaker (UNKNOWN): So
15:30
Zoe: the customers are funding the mine?
15:32
Nash: Partly, yes. You can see it in the cash flow statement. Deferred revenue, meaning money received before the service. is delivered, added about $2.1 billion to operating cash flow this quarter.
15:44
Zoe: More than the operating cash flow itself.
15:47
Nash: Right. But here's a line in the quarterly report I'd underlined twice. During the quarter, about $1.9 billion of those advanced payments were reclassified as customer liabilities because the contracts behind them became cancelable.
16:02
Zoe: Wait.
16:03
Unidentified speaker (UNKNOWN): So
16:03
Zoe: some of the gold they've already been paid for could still walk out the door?
16:07
Nash: Some of it, potentially. That's why the total deferred revenue balance only grew from about 9.7 to $10 billion, even though the cash flow line looks much bigger.
16:17
Zoe: Okay.
16:18
Unidentified speaker (UNKNOWN): And
16:18
Zoe: I don't want to oversimplify this. We can't just subtract deferred revenue from operating cash flow and call that the real number.
16:26
Nash: No? lots of other working capital items move too, but the direction is clear. A meaningful part of the financing is coming from customers, and not all of it is locked in.
16:37
Zoe: Fair. Now, December. The interest rate question.
16:41
Nash: Core Weave expects fourth-quarter interest expense of $1.05 to $1.15 billion.
16:48
Unidentified speaker (UNKNOWN): More
16:48
Zoe: than a billion dollars in one quarter.
16:51
Unidentified speaker (UNKNOWN): And
16:51
Nash: that guidance came before the Fed's December decision.
16:54
Unidentified speaker (UNKNOWN): So
16:54
Zoe: we can't blame the increase on a rate hike that hasn't happened yet.
16:58
Nash: No, it's mostly the growing amount of debt and its terms. But the quarterly report does give us something better than guessing: an actual interest rate sensitivity.
17:08
Zoe: Go ahead.
17:09
Nash: Based on its floating rate debt at the end of September, Core Weave says a one percentage point change in rate. rates would change interest expense by about $41 million per quarter.
17:19
Zoe: So a quarter point hike is roughly $10 million a quarter.
17:23
Nash: About $10.25 million, and a half-point hike about $20.5 million, assuming the debt stays the same, which it won't because they keep borrowing.
17:32
Zoe: And not every dollar of that $47 billion floats with the Fed.
17:36
Nash: Right. Some debt has fixed rates. some floats, and the company also uses interest rate swaps to hedge part of it. So the direct effect on existing debt is modest. The bigger issue is the price of all the new money they still need.
17:50
Zoe: Okay.
17:51
Unidentified speaker (UNKNOWN): But
17:51
Zoe: let me make the bull case properly.
17:54
Unidentified speaker (UNKNOWN): This
17:54
Zoe: is how infrastructure has always been built: railroads, power grids, fiber networks. You borrow, you build, and you collect. for decades.
18:04
Nash: Railroads had thousands of customers. Core Weav's two largest customers made up about 54% of revenue this quarter.
18:11
Zoe: More than half from two customers.
18:14
Nash: And we don't know their exact contract terms, so we shouldn't guess. But concentration is concentration.
18:21
Zoe: Okay. That's less railroad and more private rail line for two very rich passengers.
18:26
Unidentified speaker (UNKNOWN): That's a
18:26
Nash: fair. summary.
18:28
Zoe: So, verdict. Me, brilliant miner, real gold, extraordinary demand, and a financing machine that is smarter than people give it credit for.
18:38
Nash: Me, brilliant miner, real gold, and the bank owns the pickaxe.
18:43
Unidentified speaker (UNKNOWN): Which
18:43
Zoe: makes Core Weave look less like a software company and more like a huge, highly leveraged infrastructure business.
18:49
Unidentified speaker (UNKNOWN): With
18:49
Nash: a backlog that is only as strong as the customers inside it.
18:53
Unidentified speaker (UNKNOWN): And
18:53
Zoe: that brings us to the next mine down the road.
18:56
Nash: Nebius.
18:57
Zoe: Nebius is another AI cloud provider based in Amsterdam, and it's growing even faster than CoreWave.
19:04
Nash: Third quarter revenue, about $981 million, up 571% from a year earlier.
19:11
Zoe: These growth rates are getting hard to say without sounding impressed.
19:15
Nash: Then let me help you sound a little less impressed. About $325 million of that revenue was a one-time payment. It came from a new so-called asset light model, where partners pay Nebius to roll out as technology and cloud platform on capacity, the partners finance themselves.
19:33
Unidentified speaker (UNKNOWN): So
19:33
Zoe: roughly a third of the quarter was a one-off.
19:35
Nash: Roughly. Nebia says it expects that kind of revenue to become more regular, but for now the underlying quarter was closer to 650 million.
19:45
Zoe: Still extraordinary growth.
19:47
Nash: Still extraordinary. Just not quite 571% extraordinary.
19:52
Unidentified speaker (UNKNOWN): And
19:52
Zoe: there's a second wrinkle. Delays.
19:55
Nash: Right. Some capacity that third-party data center partners were supposed to deliver this year slipped into the first half of 2027. including part of a big site in New Jersey that serves Microsoft.
20:06
Unidentified speaker (UNKNOWN): So
20:07
Zoe: demand isn't the problem. Getting the buildings and the power ready is
20:11
Nash: Exactly. And because of that, Nebias now expects its 2026 revenue at the low end of its $3 to $3.4 billion range.
20:21
Zoe: And spending
20:23
Nash: going the other way, this year's capital expenditure is now expected. at around $27 billion, above the earlier $20 to $25 billion range.
20:32
Zoe: Revenue at the low end, spending above the top end. That's the most AI infrastructure sentence of the year.
20:39
Nash: And then the big one: the first outlook for 2027, revenue of $12 to $13 billion.
20:45
Zoe: Against how much investment?
20:47
Nash: $40 to $45 billion of capital expenditure.
20:51
Zoe: Let's put that into perspective.
20:53
Nash: Using the midpoints, about $42.5 billion of planned investment against about $12.5 billion of expected revenue.
21:01
Zoe: About $3.40 of CapEx for every dollar of expected revenue that year.
21:06
Nash: With the usual caveat, CapEx Apex buys assets that earn money over several years. It isn't an annual operating expense.
21:14
Unidentified speaker (UNKNOWN): And
21:14
Zoe: here's the clever part. Nabius says customers are now paying a lot more upfront. Prepayments on the deals it signed in the third quarter covered more than 100% of the related CapEx.
21:27
Nash: That's a big jump from 50 to 60% in the quarter before. And management expects prepayments Payments plus borrowing against contracts with investment grade customers to cover 75 to 85% of network. next year's CapEx.
21:40
Unidentified speaker (UNKNOWN): And
21:40
Zoe: prices are rising too.
21:42
Unidentified speaker (UNKNOWN): They
21:42
Zoe: say the price per GPU hour on new deals went up more than 20% compared with the previous quarter.
21:48
Nash: Which tells you demand is real. It also tells you the whole model depends on customers being willing and able to pay years in advance.
21:57
Zoe: They're even targeting pre-tax breakeven by the end of 2027.
22:01
Unidentified speaker (UNKNOWN): A
22:01
Nash: target, not a response. result, but at least it's a profit target, not just a growth target. I'll take that.
22:08
Unidentified speaker (UNKNOWN): Now
22:08
Zoe: Cerebrus, a completely different kind of mine.
22:12
Nash: Different hardware anyway. Cerebrus builds giant chips, each one made from an entire silicon wafer, and it increasingly sells access to them as a cloud service for very fast AI responses.
22:24
Zoe: Third quarter revenue under standard accounting. about $211 million, up 55%.
22:31
Nash: And the company's own measure, which it calls core revenue, about $231 million, up 70%.
22:39
Zoe: Why the gap?
22:40
Nash: Mainly two things. Sarah Brays gave some customers warrants, a kind of option on its shares, and accounting treats that as a reduction in revenue, and it strips out. pass-through data center costs it bills onto customers. So core revenue is the company's preferred lens, not the official one.
22:59
Unidentified speaker (UNKNOWN): And
22:59
Zoe: the cloud business?
23:01
Nash: Core cloud revenue up 272%.
23:04
Zoe: That's the inference boom, right there.
23:06
Unidentified speaker (UNKNOWN): But
23:06
Nash: the company still reported a net loss of about $161 million under standard accounting.
23:13
Zoe: So again, growth doesn't answer the profit question.
23:16
Nash: Partly, yes, but here the loss is mostly stock-based compensation after its IPO, around $155 million in the quarter. On its own core measure, Cerebrus actually made a small profit.
23:29
Unidentified speaker (UNKNOWN): A
23:29
Zoe: loss on paper, a small profit on its own definitions. That's a very 2026 sentence.
23:36
Unidentified speaker (UNKNOWN): And
23:36
Nash: the pipeline is huge. $26.5%. $26.1 billion of remaining performance obligations, meaning contracted work not yet delivered.
23:45
Zoe: Plus 700 gross megawatts of data center capacity under contract for delivery by the end of 2027.
23:52
Unidentified speaker (UNKNOWN): And
23:52
Nash: a new cloud deal with one of the three largest hyperscalers. Management expects to triple revenue next year.
23:59
Unidentified speaker (UNKNOWN): Which
23:59
Zoe: is a forecast, not a fact.
24:01
Nash: And as with everyone today, the capacity still has to be built and delivered before any of that becomes revenue.
24:08
Zoe: All right. We've talked about the mines. What about the companies selling them equipment?
24:12
Nash: Cisco is an interesting contrast.
24:15
Zoe: The networking company. Not the first name people think of when they hear AI.
24:20
Nash: Until the processors need to talk to each other, a GPU that can't move data fast enough is just an expensive. space heater.
24:28
Unidentified speaker (UNKNOWN): So
24:28
Zoe: Cisco sells the plumbing.
24:30
Nash: Very profitable plumbing. First quarter revenue, $18.5 billion, up 24%. Adjusted earnings per share, $1.40, up 40%.
24:41
Zoe: Cisco says that's its fastest growth in more than 15 years.
24:45
Nash: And networking product sales alone grew 43%.
24:49
Unidentified speaker (UNKNOWN): And
24:49
Zoe: the AI number?
24:51
Nash: Cisco now expects $9 billion of AI infrastructure revenue from hyperscalers this fiscal year,
24:57
Unidentified speaker (UNKNOWN): up
24:57
Nash: from $7.5 billion just three months ago.
25:00
Zoe: And last year?
25:01
Nash: About $4 billion, so more than double.
25:04
Unidentified speaker (UNKNOWN): That's
25:04
Zoe: what I'd call a shovel seller having a very good year.
25:07
Nash: With one catch, the skeptic in me noticed gross margin fell from about 68% to about 66%.
25:15
Nash: on an adjusted basis.
25:17
Zoe: Why?
25:18
Nash: Because AI networking is mostly hardware, and hardware earns lower margins than Cisco's software. So the more AI it sells, the more the mix shifts. Cisco makes up for it with tight cost control, but it's worth knowing.
25:32
Zoe: Still, expensive GPUs only work as part of a working system, and somebody gets paid for every part of that system.
25:39
Unidentified speaker (UNKNOWN): Which
25:39
Nash: is why the supply chain matters.
25:41
Unidentified speaker (UNKNOWN): And
25:42
Zoe: we'll come back to that with our stock idea. But first, the prospectors.
25:47
Nash: Anthropic and OpenAI.
25:49
Zoe: And the actual business of selling intelligence.
25:52
Nash: Same note as before. The anthropic figures come from its updated IPO filing, and a filing is not yet a finished IPO.
26:00
Zoe: So, Anthropic's updated prospectus.
26:03
Nash: 825 million Class A shares at a range of $70 to $78 each, trading on NASDAQ under the ticker ANT, reportedly starting as early as November 19th.
26:16
Unidentified speaker (UNKNOWN): At
26:16
Zoe: the top of the range, that's $64.35 billion.
26:20
Nash: Before the bank's option to sell extra shares, and not all of it goes to Anthropic, about $801 million. of those shares are new, and about 24 million come from existing shareholders.
26:31
Unidentified speaker (UNKNOWN): And
26:32
Zoe: then there's the private placement we mentioned.
26:34
Nash: $26 billion sold at the IPO price. Nvidia and Google $10 billion each, AMD $2.5 billion, Broadcom, Micron, and Microsoft $1 billion each, and half a billion from an affiliate of TSMC.
26:49
Zoe: Basically, the entire AI supply. chain buying a piece of one customer.
26:53
Nash: And the filing finally shows us the business behind it. Third quarter revenue was about $17 billion, with an operating loss of only around half a billion.
27:02
Zoe: That's a lot closer to break-even than people might assume.
27:05
Nash: It is. But the largest number in the prospectus isn't revenue, and it isn't the IPO.
27:11
Zoe: It's the commitments.
27:13
Nash: Approximately 900.5%. $900.9 billion of purchase commitments and leases. Almost all of it for compute and infrastructure.
27:20
Zoe: 900 billion. That sounds like the most terrifying invoice in history.
27:25
Nash: Except it isn't one invoice due next week. It stretches all the way to 2045. About 70.7 billion falls due in the 12 months after September, 2026.
27:37
Unidentified speaker (UNKNOWN): And
27:37
Zoe: by the end of the decade?
27:39
Nash: About $451.6 billion.
27:42
Zoe: And not all of that $900 billion is the same kind of promise, right?
27:47
Nash: No. About half of it is signed data center leases that haven't even started yet. About $186 billion can still be canceled or depends on conditions, like a partner securing its own financing. And on top of the $900 billion, anthropic estimates another $150 billion to $175 billion
28:07
Nash: from deals without a fixed dollar amount.
28:09
Unidentified speaker (UNKNOWN): And
28:09
Zoe: who's on the other side of those contracts?
28:12
Nash: Some very familiar names. The filing lists minimum commitments of around $121 billion to Google, at least $45 billion to Microsoft, largely non-cancelable, and around $160 billion to Broadcom.
28:27
Zoe: So Google, Microsoft, and Broadcom are buying shares in Anthropic, while Anthropic has committed to spend far more than that with them.
28:35
Nash: That's the small town again, money flowing in a circle. The share purchase doesn't create those contracts, but it does mean the suppliers now have two reasons to want Anthropic to succeed.
28:46
Unidentified speaker (UNKNOWN): Now,
28:46
Zoe: does Anthropic going public put OpenAI at a disadvantage?
28:50
Unidentified speaker (UNKNOWN): Not on
28:50
Nash: the evidence we have. In March, OpenAI closed a funding round with $122 billion of committed capital at an $852 billion valuation.
29:01
Zoe: And this month?
29:02
Nash: Reuters and Bloomberg reported on November 6th that OpenAI closed a first $11 billion of a new round at a valuation of $1.25 trillion.
29:12
Zoe: Even higher?
29:13
Nash: Higher than March. but lower than the 1.4 to 1.5 trillion it reportedly wanted. And it's based on anonymous sources, not a company announcement. It's also unclear whether that's before or after the new money.
29:27
Unidentified speaker (UNKNOWN): So
29:28
Zoe: both have extraordinary access to capital.
29:30
Nash: Which means the real competition isn't who announces the bigger funding round.
29:35
Zoe: It's who turns that capital into products, customers, and returns.
29:39
Nash: And going public change is one thing for Anthropic. Every quarter, investors will see standardized numbers: revenue, costs, cash flow, commitments.
29:49
Zoe: Which creates a public benchmark for the whole industry.
29:52
Nash: But it doesn't force OpenAI into an IPO of its own.
29:55
Unidentified speaker (UNKNOWN): Or
29:55
Zoe: mean that Microsoft's small stake in Anthropic changes anything about its contracts with OpenAI?
30:01
Nash: Those are separate questions.
30:03
Unidentified speaker (UNKNOWN): So
30:03
Zoe: let's move to something we can actually see with our own eyes.
30:07
Nash: API prices.
30:08
Zoe: Because when a company uses an AI model through an API, it usually pays per token.
30:14
Nash: Input tokens are what you send the model, output tokens are what it writes back, and output is usually much more expensive.
30:21
Zoe: Let's use Anthropic's current published prices.
30:25
Nash: Claude Opus 5.5 $4 per million input tokens, $20 per million output tokens.
30:32
Zoe: Claude Sonnet $5.5,
30:34
Unidentified speaker (UNKNOWN): $2
30:35
Zoe: in, $10 out.
30:37
Unidentified speaker (UNKNOWN): And
30:37
Nash: Claude Haiku $5.5 for normal length prompts, $10 in, $0.50 out. For very long prompts over 100,000 tokens, it goes up to $0.50 50 cents and $2.50.
30:50
Unidentified speaker (UNKNOWN): That's a
30:50
Zoe: 40-fold price range between the top and the bottom.
30:54
Nash: But they're not the same product. Bigger models can do harder things, and they cost more to run.
30:59
Unidentified speaker (UNKNOWN): Like
31:00
Zoe: comparing a race car and a delivery van and asking which one is overpriced.
31:05
Nash: Exactly. And there's batch pricing too. If you don't need the answer immediately and let the request run in bulk, Sonnet 5.5 costs $1 in and $5.50. $5 out.
31:16
Zoe: Half price?
31:17
Nash: Half price, same model, just less urgency.
31:20
Zoe: So even one model has several prices depending on how you use it.
31:24
Nash: And that leads to the central economic question of this whole episode.
31:28
Zoe: Is the price per token falling faster than the cost of producing it?
31:32
Nash: Because that decides what happens to the margin.
31:35
Zoe: Let's make this easy to follow.
31:37
Unidentified speaker (UNKNOWN): A
31:37
Zoe: simple illustration. not a forecast for any company.
31:41
Nash: A provider charges $1 per unit and sells 100 units.
31:45
Zoe: $100 of revenue.
31:47
Nash: Then the price falls by half.
31:50
Zoe: $50.
31:51
Nash: And usage rises by 30%.
31:54
Zoe: 130 units.
31:55
Unidentified speaker (UNKNOWN): So
31:56
Nash: 50 cents times 130.
31:59
Zoe: 65 dollars.
32:00
Nash: Usage went up, revenue went down. 35%.
32:04
Unidentified speaker (UNKNOWN): So
32:04
Zoe: to make up for a 50% price cut, usage would have to double.
32:08
Nash: Assuming the same mix of customers and products,
32:11
Unidentified speaker (UNKNOWN): and
32:11
Nash: that's only revenue, we haven't even talked about profit yet.
32:15
Zoe: Okay, profit.
32:16
Nash: If the cost of producing a token falls faster than its price, the margin per token gets better. If the price falls faster than the cost, the margin per token shrinks.
32:27
Zoe: Which makes the cost of computing very important.
32:30
Nash: And very hard to put into one number.
32:33
Zoe: Because a GPU hour price isn't the same as a cost per token.
32:37
Nash: Right. You also need throughput, utilization, the model's architecture, how fast the answers need to be, power, and operating costs.
32:47
Zoe: Let's unpack two of those.
32:49
Nash: Throughput is how much work a system. can process in a given amount of time.
32:53
Zoe: And utilization is how much of the available capacity is actually being used.
32:58
Unidentified speaker (UNKNOWN): And
32:58
Nash: an idle GPU still costs money. It's like a taxi waiting at the airport with the meter off.
33:04
Unidentified speaker (UNKNOWN): So
33:04
Zoe: keeping the hardware busy can improve the economics a lot.
33:07
Nash: Depending on the workload, some applications need instant answers, others can wait and run in batches.
33:14
Unidentified speaker (UNKNOWN): Which
33:14
Zoe: is exactly. why batch processing is cheaper.
33:17
Nash: Exactly.
33:18
Unidentified speaker (UNKNOWN): And
33:18
Zoe: now there's another twist: reasoning models.
33:22
Nash: Models that think through a problem step by step before they answer. And that thinking uses tokens and computing power too.
33:29
Unidentified speaker (UNKNOWN): So
33:29
Zoe: the answer you see isn't the whole workload.
33:32
Unidentified speaker (UNKNOWN): And
33:32
Nash: depending on the provider's billing rules, those thinking tokens can show up on your bill as output.
33:37
Unidentified speaker (UNKNOWN): So the
33:38
Zoe: price per million. tokens doesn't tell you what it costs to actually finish a task.
33:42
Nash: Right. A pricier model might use more tokens, but get a hard task right the first time.
33:48
Unidentified speaker (UNKNOWN): And
33:48
Zoe: a cheaper model might need three tries and a human to check the result.
33:52
Nash: Which brings us to a much better business metric.
33:55
Zoe: Cost per successfully completed task.
33:58
Unidentified speaker (UNKNOWN): At
33:58
Nash: a quality level, the customer actually accepts.
34:01
Zoe: That's far more useful than crowning. whoever has the lowest token price.
34:05
Nash: But it's also much harder to measure because it's different for every application.
34:09
Zoe: And let's be honest about what we don't know.
34:12
Unidentified speaker (UNKNOWN): We
34:12
Zoe: don't have the internal serving costs of anthropic or OpenAI for each model.
34:17
Nash: Or the discounts big enterprise customers negotiate.
34:21
Unidentified speaker (UNKNOWN): So
34:21
Zoe: we're not going to invent their margins.
34:23
Nash: Or claim a precise industry-wide industry wide price decline. by comparing today's very different models with each other. A price list shows you prices and service options. It's not a clean historical record of performance adjusted prices.
34:37
Unidentified speaker (UNKNOWN): And
34:38
Zoe: that distinction matters.
34:39
Nash: Because the opportunity depends on both sides how much people want the gold and what it costs Cost to get it out of the ground.
34:46
Unidentified speaker (UNKNOWN): So
34:47
Zoe: Anthropics IPO will give investors a much better look at one. major prospector.
34:52
Nash: But it won't answer every question about how profitable each model is.
34:56
Unidentified speaker (UNKNOWN): And
34:57
Zoe: it doesn't tell us who wins the price for.
34:59
Nash: The public evidence simply doesn't settle that yet.
35:02
Zoe: Which brings us to a very different way of making money from this gold rush. Not mining, not refining,
35:09
Nash: selling the shovels,
35:10
Unidentified speaker (UNKNOWN): our
35:11
Zoe: stock idea, applied materials.
35:13
Nash: Except these shovels cost. tens of millions of dollars each, and they work at the scale of atoms.
35:19
Zoe: Same deal as before. Company figures and anything about the future is management's view. Not a promise. Done.
35:26
Nash: Done. And honestly, after Corweeb's interest bill, this quarter feels like a spa day.
35:31
Unidentified speaker (UNKNOWN): A
35:32
Zoe: company that makes substantial profits and real free cash flow without running a chatbot.
35:37
Nash: How refreshing.
35:38
Zoe: So the headline numbers. Fourth quarter revenue for fiscal 2026, 10.41 billion dollars, up 53%.
35:48
Nash: And adjusted earnings per share, $4.16, up 92%.
35:53
Zoe: Earnings growing almost twice as fast as revenue.
35:57
Unidentified speaker (UNKNOWN): And
35:57
Nash: that's the number I care about, because it means they're not just selling more, they're keeping more of every dollar.
36:03
Unidentified speaker (UNKNOWN): So
36:03
Zoe: what happened to the margins?
36:05
Nash: Adjusted gross margin went from 48.1% a year ago to 50.6%.
36:11
Zoe: And operating margin?
36:13
Nash: From 28.6 to 35.3%. That's 6.7 percentage points in one year.
36:19
Unidentified speaker (UNKNOWN): So
36:19
Zoe: more of every revenue dollar actually reaches operating profit.
36:23
Nash: And the company points to three reasons: a better product. mix, higher prices, and operating leverage. Revenue simply grew much faster than costs.
36:32
Zoe: That's a very different story from a company that just reports more orders.
36:36
Nash: It's realized profit in the quarter, not a promise.
36:40
Unidentified speaker (UNKNOWN): And
36:40
Zoe: under standard accounting?
36:42
Nash: Earnings per share of $3.40. Up 43%.
36:47
Zoe: That's quite a gap to the adjusted number.
36:49
Unidentified speaker (UNKNOWN): It
36:49
Nash: is, and it's worth explaining.
36:51
Nash: Most of it, about 68 cents per share, comes from a paper loss of around $650 million on the value of stakes applied holds in other companies. Operating profit is almost identical on both measures.
37:05
Unidentified speaker (UNKNOWN): So
37:05
Zoe: the gap isn't the business getting worse, it's the value of an investment portfolio moving.
37:11
Nash: Exactly, but we still keep the two numbers separate.
37:14
Unidentified speaker (UNKNOWN): Okay.
37:15
Zoe: quick picture for everyone. What does applied materials actually do?
37:19
Nash: Building a modern ship is like building a skyscraper out of hundreds of layers, each thinner than a virus.
37:26
Unidentified speaker (UNKNOWN): And
37:26
Zoe: applied builds the cranes.
37:28
Nash: Nicely put, its machines handle several of the most important steps: deposition, which means adding ultra-thin ultra thin layers of material.
37:37
Zoe: Etching.
37:38
Nash: Carving away exactly the material you don't want to form tiny structures.
37:43
Zoe: Chemical mechanical planarization, which sounds like a medical procedure.
37:48
Nash: It's basically polishing, grinding each layer perfectly flat before the next one goes on top.
37:53
Unidentified speaker (UNKNOWN): And
37:54
Zoe: process control?
37:55
Nash: Measuring and inspecting everything along the way to catch errors before they ruin a whole batch of chips.
38:02
Unidentified speaker (UNKNOWN): Which comes
38:02
Zoe: out down to yield.
38:04
Unidentified speaker (UNKNOWN): The
38:04
Nash: share of chips that actually come out working. And when a single wafer of advanced chips is worth a fortune, every percentage point of yield is worth a lot of money.
38:13
Unidentified speaker (UNKNOWN): So
38:13
Zoe: where did the growth come from?
38:15
Nash: Mainly from the chipmaking equipment business, which applied calls semiconductor systems, about $8 billion of revenue in the quarter.
38:23
Zoe: Up how much?
38:25
Nash: About 64%. And it's roughly 77% of the whole company.
38:30
Unidentified speaker (UNKNOWN): And
38:30
Zoe: here's what I find most interesting.
38:32
Unidentified speaker (UNKNOWN): Look
38:32
Zoe: at who's buying.
38:34
Nash: 58% from Foundry and Logic, so the chipmakers producing processors.
38:39
Zoe: 33% from DRAM, the memory business.
38:42
Nash: And 9% from Flash Memory.
38:45
Unidentified speaker (UNKNOWN): A
38:45
Zoe: year ago, DRAM was only 28% of that business.
38:49
Unidentified speaker (UNKNOWN): So
38:49
Nash: memory became a bigger slice. of a much bigger pie.
38:52
Unidentified speaker (UNKNOWN): Do
38:52
Nash: the math with those rounded shares, and memory equipment sales went from about $1.37 billion to about $2.64 billion.
39:01
Zoe: Almost doubled.
39:03
Unidentified speaker (UNKNOWN): And
39:03
Nash: that's AI showing up in the order book. Large AI models need enormous amounts of very fast memory sitting right next to the processor.
39:11
Zoe: High bandwidth memory, HBM.
39:14
Unidentified speaker (UNKNOWN): Right.
39:14
Nash: but to be precise, HBM is one type of DRAM, and applied's DRAM category includes more than just HBM. So we can't label all of it as AI revenue.
39:25
Zoe: Fair.
39:25
Unidentified speaker (UNKNOWN): But
39:26
Zoe: management did report record DRAM revenue.
39:29
Nash: And record revenue in advanced packaging.
39:31
Zoe: Which is the next big story. Explain why it matters.
39:35
Nash: Modern AI chips aren't one piece of silicon anymore. They're several components, like processors and memory, stacked and wired together in one package.
39:45
Unidentified speaker (UNKNOWN): Like
39:45
Zoe: a high rise where the floors have to connect perfectly.
39:48
Nash: And those connections are microscopic. So stacking and wiring chips together has suddenly become as important as making the chips themselves.
39:56
Zoe: How fast is it growing?
39:58
Nash: Management expects advanced packaging revenue to grow around 90% in what applies. Applied calls calendar 2026.
40:05
Zoe: What applied calls calendar 2026?
40:09
Nash: Applied's calendar year is shifted by a quarter. It runs from early this year to the end of January, 2027. So it's partly still a forecast.
40:18
Zoe: And no exact dollar amount for it.
40:20
Nash: Not in what they published, so we won't invent one.
40:24
Unidentified speaker (UNKNOWN): And
40:24
Zoe: there's a partner in this area?
40:26
Unidentified speaker (UNKNOWN): B
40:26
Nash: BSI, Sub. a Dutch company specializing in hybrid bonding. That's a way of connecting chips with extremely fine, direct connections. Applied has agreed to increase its stake in BEC to about 20%.
40:39
Unidentified speaker (UNKNOWN): So
40:39
Zoe: the shovel seller is buying a stake in another shovel seller.
40:42
Nash: Everyone's buying stakes in everyone this week.
40:45
Unidentified speaker (UNKNOWN): Now,
40:45
Zoe: the equipment business didn't just grow, it became much more profitable.
40:49
Nash: Adjusted operating. margin in that segment, 39%. A year ago, 29.6.
40:56
Zoe: Almost 10 points.
40:58
Nash: 9.4 to be exact. And about $3.1 billion of operating income from that one segment.
41:05
Unidentified speaker (UNKNOWN): And
41:05
Zoe: then there's the part you love.
41:07
Nash: Services, applied global services.
41:09
Zoe: The boring part.
41:11
Unidentified speaker (UNKNOWN): The
41:11
Nash: beautiful boring part. Every machine applied cells needs maintenance. spare parts, and upgrades for years.
41:17
Unidentified speaker (UNKNOWN): So
41:17
Zoe: the relationship doesn't end when the machine leaves the factory.
41:21
Nash: Exactly. The installed base keeps paying. About $1.9 billion in the quarter, up 26%.
41:27
Zoe: At what margin?
41:29
Nash: 30.5%, up from about 28.8% a year earlier.
41:34
Unidentified speaker (UNKNOWN): So
41:34
Zoe: it's a shovel seller who also gets paid every time someone sharpens the shovel.
41:39
Nash: That's the business model in one sentence.
41:41
Zoe: And cash.
41:43
Nash: About $3.3 billion of operating cash flow in the quarter. Capital expenditure of about $684 million.
41:50
Unidentified speaker (UNKNOWN): So
41:51
Zoe: free cash flow of about $2.6 billion.
41:54
Nash: $2.62 billion, a record.
41:56
Unidentified speaker (UNKNOWN): Let's
41:57
Zoe: compare that with Core Weave for a second.
41:59
Nash: Carefully, they're very different businesses.
42:02
Unidentified speaker (UNKNOWN): Of
42:02
Zoe: course.
42:03
Unidentified speaker (UNKNOWN): But
42:03
Zoe: Core Weave spent more than $9 billion of cash on equipment in one quarter against about $2 billion of operating cash flow.
42:11
Nash: And Applied produced more than $2.5 billion of free cash.
42:15
Zoe: One builds and runs the mine, the other sells the equipment.
42:18
Nash: Different investment needs, different asset lives, different revenue models, and very different cash flow statements.
42:25
Unidentified speaker (UNKNOWN): And
42:25
Zoe: what did Applied do with all that cash?
42:27
Unidentified speaker (UNKNOWN): It
42:28
Nash: returned about $2.42 billion to shareholders in the quarter. $2 billion in share buybacks, the largest in more than eight years, and about $421 million in dividends.
42:41
Unidentified speaker (UNKNOWN): So
42:41
Zoe: most of the cash went straight back to the owners.
42:44
Nash: In this quarter, yes.
42:46
Unidentified speaker (UNKNOWN): Now
42:46
Zoe: let's look ahead using management's actual guidance.
42:49
Unidentified speaker (UNKNOWN): For the
42:49
Nash: first quarter of February, fiscal 2027, revenue of about $11.1 billion, plus or minus $500 million.
42:58
Zoe: And adjusted earnings per share of about $4.35, plus or minus 20 cents.
43:04
Unidentified speaker (UNKNOWN): And
43:04
Nash: there's a little trap hidden in there.
43:06
Unidentified speaker (UNKNOWN): The
43:06
Zoe: bonus week.
43:08
Nash: The bonus week. That quarter has 14 weeks instead of 13. And the extra week mainly boosts the services business.
43:15
Unidentified speaker (UNKNOWN): So
43:15
Zoe: if you compare it with the previous quarter, you have to adjust for that.
43:19
Nash: Otherwise, the growth rate looks better than it really is. Every analyst's favorite trap.
43:24
Unidentified speaker (UNKNOWN): And
43:25
Zoe: what does management say about the bigger picture?
43:27
Nash: Applied expects the market for chipmaking equipment, the industry calls it wafer fab equipment, to grow more than 30% in 2027.
43:36
Unidentified speaker (UNKNOWN): And
43:36
Zoe: applied wants to grow faster than that market.
43:39
Nash: That's the claim. Management also says it's tracking more than 20 new fabs and production line expansions scheduled to ramp up in 2027.
43:48
Unidentified speaker (UNKNOWN): But a
43:49
Zoe: fab ramp on a planning slide isn't revenue yet.
43:52
Nash: No. Those buildings have to be finished, the machines installed and qualified. And management itself says that clean room space, meaning finished buildings where the equipment can actually go, is one
44:02
Unidentified speaker (UNKNOWN): of the things that we're going to
44:03
Nash: limiting factors right now.
44:05
Unidentified speaker (UNKNOWN): So
44:05
Zoe: the customers want the cranes, but the construction site isn't ready.
44:09
Nash: In some cases, yes. And Applied is preparing its own capacity too. It's investing to be able to double its system output by 2028, compared with the level of the last quarter.
44:19
Zoe: Capacity, not guaranteed sales.
44:22
Nash: Exactly.
44:23
Unidentified speaker (UNKNOWN): There's
44:23
Zoe: also Applied's Epic Center.
44:25
Unidentified speaker (UNKNOWN): A
44:26
Nash: huge new research. building in Silicon Valley where chipmakers, equipment partners, and universities develop new manufacturing processes together with Applied's engineers under one roof.
44:37
Zoe: The idea being develop in parallel instead of one after the other and get new technology to market faster.
44:44
Nash: And this quarter, four new partners join: Intel, Keoxia, Kioxia, BSI, and UCLA. That makes 15 announced partners. partnerships.
44:51
Unidentified speaker (UNKNOWN): And
44:52
Zoe: now my favorite connection to the rest of this episode?
44:55
Nash: Applied materials isn't just selling to the AI industry.
44:58
Zoe: It's also an AI customer.
45:00
Nash: Management says it's rapidly expanding its own use of AI in research, manufacturing, service, and software development, and it talks about significant productivity gains.
45:12
Zoe: Those are the company's own claims, and we couldn't independently verify specific figures.
45:17
Nash: Right.
45:17
Unidentified speaker (UNKNOWN): But
45:18
Nash: the connection is fascinating. Applied sells the equipment that makes the chips and then buys AI services that run on those chips.
45:26
Zoe: Supplier and customer at the same time.
45:29
Nash: But more AI usage at Applied doesn't tell us how much it spends or how much margin the model providers earn on it.
45:35
Unidentified speaker (UNKNOWN): And
45:35
Zoe: there's our central theme again:
45:37
Nash: usage revenue. and profit are three different things.
45:41
Zoe: So, strong quarter. Does that make the stock a buy?
45:45
Nash: No. How well a company performs and what its shares are worth are two separate questions.
45:51
Zoe: Okay. Then let me give you my bull case. This is the cleanest way to own the AI build out.
45:57
Unidentified speaker (UNKNOWN): You
45:57
Zoe: don't have to pick which model lab wins. Whoever wins needs chips, and whoever makes chips needs. applied.
46:04
Unidentified speaker (UNKNOWN): And
46:04
Nash: here's my bear case. This is one of the most cyclical industries on earth. When chipmakers cut their budgets, equipment orders don't dip. They fall off a cliff.
46:13
Zoe: Not this year.
46:15
Nash: Not this year, but there are more risks. About a quarter of Applied's revenue in the quarter still came from China. Export rules can change quickly, and earlier this fiscal year, Applied paid two
46:26
Unidentified speaker (UNKNOWN): years.
46:27
Nash: $253 million to settle an export control matter with the U.S. Commerce Department.
46:33
Zoe: That's a very concrete reminder that geopolitics isn't theoretical here.
46:37
Nash: Then there's competition from other equipment makers. The clean room bottleneck we just mentioned, and the bonus week flattering next quarter.
46:45
Zoe: And the big one.
46:46
Unidentified speaker (UNKNOWN): The
46:47
Nash: big one isn't in the earnings report at all.
46:49
Unidentified speaker (UNKNOWN): How
46:49
Nash: much of this future growth? is already in the share price?
46:53
Unidentified speaker (UNKNOWN): So
46:53
Zoe: a great earnings report isn't the same thing as a great price to buy at.
46:57
Nash: Which is why you look at valuation alongside revenue, margins, cash, and demand.
47:03
Zoe: So, verdict. Me,
47:05
Unidentified speaker (UNKNOWN): the
47:05
Zoe: best shovel shop in town, and the rush is just getting started.
47:09
Nash: Me, the best shovel shop in town. But shovel shops have bad years too, and you don't want to buy at the top of the rush.
47:16
Unidentified speaker (UNKNOWN): And
47:16
Zoe: that's exactly why applied materials is our stock idea.
47:20
Nash: Not our stock tip. Its latest results give us measurable evidence of how one critical part of chip manufacturing is doing, not a guarantee of what the shares will do.
47:30
Unidentified speaker (UNKNOWN): And
47:30
Zoe: now from the company making equipment for advanced chips to a company trying to serve AI models efficiently.
47:36
Nash: Startup corner.
47:38
Unidentified speaker (UNKNOWN): And
47:38
Zoe: this week's startup is the gold rush's most fascinating character:
47:42
Unidentified speaker (UNKNOWN): the
47:42
Zoe: one who doesn't own a mine and doesn't sell shovels.
47:45
Unidentified speaker (UNKNOWN): He
47:45
Nash: runs the refinery.
47:47
Zoe: Fireworks AI.
47:49
Nash: Usual note: these are company figures and press reports. Fireworks is private, so there are no audited financials, which, spoiler, will matter in a minute.
47:58
Zoe: Noted.
47:59
Unidentified speaker (UNKNOWN): So
47:59
Zoe: here's the question:
48:01
Unidentified speaker (UNKNOWN): Can
48:01
Zoe: an independent company build a lasting business serving AI models right next to OpenAI, Anthropic, and the big cloud providers?
48:10
Nash: Let's start with what Fireworks actually sells:
48:13
Zoe: Infrastructure and software to deploy, run, and customize AI models, very often open weight models.
48:20
Nash: Meaning models whose train parameters, the weights, are published under a license, so companies can download them, adapt
48:26
Unidentified speaker (UNKNOWN): to the company.
48:27
Nash: them, and run them under the license terms.
48:29
Unidentified speaker (UNKNOWN): So
48:30
Zoe: a company brings a raw model and Fireworks tunes it, runs it, and serves the answers at enormous scale.
48:36
Nash: Either as a serverless service, where you just send requests and pay per token, or as a dedicated deployment, where you rent the hardware and pay for GPU time.
48:46
Zoe: Two different ways of paying for the same refinery.
48:49
Unidentified speaker (UNKNOWN): Right,
48:49
Nash: and the scale is wild.
48:51
Unidentified speaker (UNKNOWN): In
48:51
Zoe: July, Fireworks raised $1.5 billion in a Series D round at a $17.5 billion valuation.
49:00
Nash: Nine months earlier, in October 2025,
49:03
Unidentified speaker (UNKNOWN): it
49:03
Nash: raised $250 million at a $4 billion valuation.
49:07
Unidentified speaker (UNKNOWN): So
49:07
Zoe: the valuation more than quadrupled in under a year.
49:10
Unidentified speaker (UNKNOWN): And
49:11
Nash: the company says it has passed $1 billion in annualized revenue run rate. five times more than a year earlier.
49:17
Zoe: That's substantial.
49:19
Nash: It is. But an annualized run rate isn't the same as a completed audited year. It usually takes recent revenue and multiplies it up to a full year.
49:28
Unidentified speaker (UNKNOWN): And
49:28
Zoe: the valuation is about 17 and a half times that run rate.
49:32
Nash: Which is rich, even by AI standards. Its competitors are richly valued too. Together, AI rate raised at 8.3 billion this year, and Bassettin at 13 billion.
49:43
Unidentified speaker (UNKNOWN): And
49:43
Zoe: then the headline number: more than 40 trillion tokens a day.
49:48
Nash: Up from 15 trillion.
49:50
Zoe: For scale. According to a comparison by CNBC, that's more than Google and OpenAI each reported for their developer platforms earlier this year.
49:59
Unidentified speaker (UNKNOWN): And
49:59
Nash: there it is. My favorite question of the episode.
50:03
Unidentified speaker (UNKNOWN): I
50:03
Zoe: knew it was coming.
50:05
Nash: 40 trillion tokens. How much money does it keep?
50:08
Zoe: We don't know.
50:09
Nash: We don't know. No published margins, no profit, no cash flow. A token counter is a speedometer, not a bank statement.
50:17
Zoe: Okay, but let me make the bull case, because it's a good one. Fireworks says 95% of the tokens it serves come from customized models.
50:26
Nash: Models tuned on a customer's own data.
50:28
Zoe: Companies like Uber and Shopify don't want a generic chatbot.
50:32
Unidentified speaker (UNKNOWN): They
50:33
Zoe: want a model trained on their own information that does one job brilliantly and cheaply.
50:37
Nash: That's a real niche, I'll give you that.
50:40
Zoe: It's more than a niche.
50:42
Unidentified speaker (UNKNOWN): It's
50:42
Zoe: a bet that the future of AI isn't one giant brain, but thousands of specialists.
50:48
Unidentified speaker (UNKNOWN): And
50:48
Zoe: if that's right, Fireworks owns the specialist factory.
50:52
Nash: Counterpoint. Amazon, Google, and Microsoft all sell this kind of service too. And they own the data centers. Fireworks mostly rents its compute.
51:02
Unidentified speaker (UNKNOWN): So
51:02
Zoe: it's a refinery that leases its furnaces from its competitors.
51:06
Nash: Which is why differentiation matters so much here. Speed, price, how easy it is to customize a model.
51:14
Unidentified speaker (UNKNOWN): So how would
51:14
Zoe: we actually judge whether the business model works?
51:17
Nash: You'd separate what each workload brings in from what it costs to serve, meaning the computing and operating costs. That gives you the contribution margin.
51:26
Unidentified speaker (UNKNOWN): And
51:27
Zoe: then look at how those margins differ between serverless and dedicated, between big and small customers.
51:33
Nash: And customer retention, because fast usage growth doesn't tell you how loyal the customers are.
51:38
Unidentified speaker (UNKNOWN): And we
51:38
Zoe: can't invent customer losses, price cuts, or how the big clouds might respond.
51:43
Nash: None of that has been disclosed, which brings us to the part of the story that aged it by about 10 years in one month.
51:49
Zoe: The security incident.
51:51
Nash: Incidents plural. The first was detected on October 6th. Then in late October, attackers got in through a third-party cloud provider and reached fireworks development machines.
52:01
Unidentified speaker (UNKNOWN): And on
52:02
Zoe: November 9th, the company published an update on what was actually affected.
52:06
Nash: Account and billing information, and for a limited set of customers, secrets like access keys, data sets, and incomplete model weights.
52:16
Zoe: Model weights. The trained brain of a customer's AI.
52:19
Nash: Partial ones. And to be fair, payment card numbers weren't affected because Fireworks says it doesn't store them.
52:26
Unidentified speaker (UNKNOWN): And
52:26
Zoe: the company says that, based on its investigations so far, its main services that keep no customer data, including core inference, show no impact.
52:35
Nash: That qualification matters. It's what the investigation found up to that point. Not a guarantee about anything discovered later.
52:43
Zoe: Fireworks has rotated credentials, brought in outside forensic experts, and is notifying affected customers directly.
52:50
Unidentified speaker (UNKNOWN): It
52:50
Zoe: has linked the attacks to criminals who hijack AI computing power and resell it.
52:55
Nash: And at least one well-known customer has confirmed it was affected. Notion says fewer than 1% of its workspaces were impacted.
53:03
Unidentified speaker (UNKNOWN): And
53:03
Zoe: here's the irony:
53:05
Nash: Fireworks' whole pitch is bring us your proprietary data and we'll build you a model only you own.
53:11
Zoe: So the product is trust, and trust just got a dent.
53:15
Nash: A dent, not necessarily a crack. No one has disclosed how much revenue, if any, this will cost.
53:21
Unidentified speaker (UNKNOWN): But
53:21
Zoe: for any company handling customers' business data, security isn't an extra. It's part of the product.
53:27
Nash: Right next to reliability, performance, and the economics of serving the workload.
53:33
Unidentified speaker (UNKNOWN): So
53:33
Zoe: can fireworks survive?
53:35
Nash: What we know, a lot of financing, fast-growing reported revenue. and enormous usage.
53:40
Unidentified speaker (UNKNOWN): What
53:40
Zoe: we don't know: workload level margins, overall profitability, and cash flow.
53:46
Nash: So the survival question stays open. That's the honest evidence-based answer.
53:51
Zoe: My verdict: the right idea at the right moment with a growth curve most companies only see in their dreams.
53:57
Nash: Mine, impressive engine, no visible fuel gauge, and a security scare right. where it hurts most.
54:04
Unidentified speaker (UNKNOWN): So
54:04
Zoe: can it make it?
54:06
Nash: Ask me again when we see a margin.
54:08
Zoe: Which brings us back to inference pricing one last time.
54:11
Nash: Because the profitability of serving models depends on far more than the number of tokens.
54:16
Unidentified speaker (UNKNOWN): It
54:17
Zoe: depends on the price charged.
54:19
Nash: The cost of processing the work.
54:21
Unidentified speaker (UNKNOWN): How
54:21
Zoe: busy the hardware is.
54:23
Nash: Which models and which tasks.
54:25
Unidentified speaker (UNKNOWN): And
54:25
Zoe: whether customers find it useful enough. to keep paying.
54:28
Nash: Exactly.
54:29
Unidentified speaker (UNKNOWN): So
54:30
Zoe: is inference a high volume business with falling prices or a premium product with lasting pricing power?
54:36
Nash: Right now it's both. There are premium models and dirt cheap models, instant answers and batch processing.
54:43
Unidentified speaker (UNKNOWN): But
54:43
Zoe: that doesn't add up to one single price path or margin structure for the whole industry.
54:48
Unidentified speaker (UNKNOWN): It
54:48
Nash: differs by model, by task, by provider and by contract.
54:52
Unidentified speaker (UNKNOWN): So
54:52
Zoe: we'll need better disclosures to see where this goes.
54:55
Nash: On prices, costs, usage, and cash.
54:58
Unidentified speaker (UNKNOWN): So
54:58
Zoe: let's take one last walk through town.
55:00
Unidentified speaker (UNKNOWN): The
55:01
Nash: mine core weave, real gold, record demand, a backlog of more than $100 billion, and an interest bill four times its operating profit.
55:11
Zoe: Brilliant miner.
55:12
Unidentified speaker (UNKNOWN): And
55:12
Zoe: the bank owns the pickaxe.
55:14
Nash: You stole my line.
55:16
Unidentified speaker (UNKNOWN): It
55:16
Zoe: was a good line.
55:18
Unidentified speaker (UNKNOWN): Then
55:18
Zoe: Nebious, also digging fast.
55:20
Unidentified speaker (UNKNOWN): But
55:21
Zoe: a big part of this quarter's growth was a one-time payment and some new capacity slipped into next year.
55:26
Nash: Cerebrus. Cloud revenue growing quickly, still a loss on paper, and a big pile of contracted work to deliver.
55:34
Zoe: Cisco. Quietly selling the plumbing every AI data center needs. and raising its AI forecast again.
55:41
Unidentified speaker (UNKNOWN): The
55:41
Nash: shovel shop, applied materials, rising margins, real cash, money going back to shareholders, and a business that has very bad years when the rush slows down.
55:52
Zoe: The refinery. Fireworks. 40 trillion tokens a day. No visible profit and a security scare right in the middle of its trust business.
56:01
Unidentified speaker (UNKNOWN): And
56:02
Nash: the prospector. prospectors. Anthropic with almost $900 billion of computing commitments, much of it with the very companies now buying its shares.
56:11
Zoe: And OpenAI, raising money at a valuation above a trillion dollars.
56:16
Nash: So the AI economy isn't one business model.
56:19
Zoe: It's a whole town of businesses with different capital needs and different ways of making money.
56:24
Unidentified speaker (UNKNOWN): And
56:24
Nash: the gold rush comparison holds up surprisingly. well. Miners, shovel sellers, refiners, and bankers are all part of the same boom.
56:33
Unidentified speaker (UNKNOWN): But
56:34
Zoe: they don't automatically earn the same return.
56:36
Unidentified speaker (UNKNOWN): And
56:36
Nash: the amount of gold being found isn't the only number that matters.
56:40
Zoe: There's the cost of finding it.
56:42
Nash: The cost of getting it out of the ground.
56:45
Unidentified speaker (UNKNOWN): And
56:45
Zoe: the price someone is willing to pay for it.
56:47
Nash: Which in AI brings us back to token prices. computing costs, and the value delivered to customers.
56:54
Zoe: Different parts of the same machine.
56:56
Nash: With very different financial statements.
56:59
Unidentified speaker (UNKNOWN): So
56:59
Zoe: what did we learn?
57:00
Nash: A funding round isn't revenue.
57:02
Zoe: Revenue isn't profit.
57:04
Nash: Profit isn't necessarily cash.
57:07
Unidentified speaker (UNKNOWN): And
57:07
Zoe: 40 trillion tokens don't tell you how much money is left at the end of the day.
57:11
Nash: That's going on a t-shirt.
57:13
Unidentified speaker (UNKNOWN): You
57:13
Zoe: hate t-shirts.
57:14
Nash: I'd make an exception if it came with a cash flow statement.
57:18
Unidentified speaker (UNKNOWN): Of
57:19
Zoe: course it would.
57:20
Unidentified speaker (UNKNOWN): That's
57:20
Nash: all for this week's Tech or Nothing.
57:22
Zoe: Thanks for listening.
57:23
Unidentified speaker (UNKNOWN): If
57:24
Zoe: you enjoyed it, follow the show and send it to the one friend who keeps saying, It's not a bubble, it's infrastructure.
57:30
Nash: And before we go, the part our lawyers like best, our usual important reminder.
57:35
Zoe: Tech or nothing is an independent podcast discussing technology.
57:39
Zoe: companies, financial markets, business models, and investment-related subjects.
57:45
Nash: Everything discussed is provided for general information, education, and commentary only.
57:50
Zoe: Nothing in this episode constitutes personalized financial, investment, legal, or tax advice.
57:56
Nash: Nor does Or it represent an offer, solicitation, or recommendation to buy, sell, or hold any security. investment product, or financial instrument?
58:05
Zoe: We don't know your personal financial circumstances, investment objectives, time horizon, or tolerance for risk.
58:12
Nash: Investing involves risk, including market volatility, operational and financial uncertainty, and the potential loss of some or all invested capital.
58:21
Unidentified speaker (UNKNOWN): The
58:22
Zoe: companies discussed may face technological, competitive, regulatory, geopolitical,
58:27
Zoe: financing, and execution risks.
58:30
Nash: Management guidance, company forecasts, analyst estimates, and other forward-looking statements describe uncertain future outcomes.
58:37
Zoe: Actual results may differ materially from those expectations.
58:45
Nash: Financing rounds, company valuations, and past share price movements don't guarantee future financial results or investment returns.
58:52
Zoe: Some financial figures discussed are non-GAAP measures, company-defined metrics or arithmetic calculations based on published disclosures.
59:02
Nash: Definitions may differ between companies, and the resulting figures may not be directly comparable.
59:07
Zoe: Mathematical examples are used to explain economic relationships.
59:11
Unidentified speaker (UNKNOWN): They
59:11
Zoe: are not forecasts of specific companies, markets, or investment returns.
59:16
Nash: Information reflects sources available at the time of research and recording, as of November 14th, 2026.
59:23
Zoe: While we aim for accuracy, information may be incomplete, change after publication, or contain errors.
59:30
Nash: Always verify important financial information through original company disclosures, regulatory filings, and reliable independent sources.
59:38
Unidentified speaker (UNKNOWN): Conduct
59:38
Zoe: your own. research and consider consulting a qualified independent professional before making financial decisions.
59:44
Nash: I'm Nash.
59:45
Unidentified speaker (UNKNOWN): And
59:46
Zoe: I'm Zoe.
59:47
Nash: Stay curious.
59:48
Zoe: Follow the numbers.
59:50
Unidentified speaker (UNKNOWN): And
59:50
Nash: keep asking where the economic value actually goes.
59:54
Zoe: See you on Wednesday, Tech or Nothing.
This page was produced automatically: the audio was transcribed with nvidia/parakeet-tdt-0.6b-v3+sortformer-v2.1 and interpreted by Qwen3.6-35B-A3B-UD-IQ3_S.gguf. Transcription and interpretation can contain errors; always confirm against the original episode. Nothing here is investment advice.
All insights