Newsletter · · Ashutosh Agarwal

Qualcomm Wins an AWS Custom Chip Deal With Warrants Attached - Custom Silicon vs Nvidia - Week of September 21, 2026

Custom Silicon vs Nvidia for the week of September 14 to September 21, 2026. Podcast synthesis on Qualcomm breaking into AWS custom silicon with a warrant-backed deal, the warrant playbook spreading to Generac, Google's TPU cost edge and Anthropic-heavy backlog, Marvell's Google upside, Nvidia anchoring Anthropic's IPO, and the first hard data on softening top-end AI demand.

Custom Silicon vs Nvidia

Week of September 21, 2026: Qualcomm Wins an AWS Custom Chip Deal With Warrants Attached

Last week the podcasts argued about who pays for all the chips. This week they told us who's getting paid to build them, and the surprise name is Qualcomm. The single most book-relevant chip development of the week is that Qualcomm, the smartphone company everyone had written off in the data center, has an actual custom-silicon deal with Amazon's AWS (the hardest hyperscaler on earth to break into), with real terms now on the table. And it uses the same "warrant" sweetener we flagged last week, which is quietly becoming the standard way hyperscalers pay their chip designers.

Underneath that, the money story from last week didn't go away. It got bigger and more specific: Oracle funding a build it can't cover with cash, Amazon handing a generator maker warrants, Nvidia writing a $10 billion check to anchor its own customer's IPO, and, new this week, the first hard demand-side data suggesting the AI spending machine may be sputtering at the very top.

A quick reminder for new readers: an ASIC, or "custom chip" (the industry calls it an XPU), is a chip a company designs for its own specific job (Google's TPU, Amazon's Trainium) instead of buying a general-purpose Nvidia GPU off the shelf. Firms like Broadcom, Marvell, and now Qualcomm don't own those chips; they're the engineering shops that turn a customer's design into a real, manufacturable chip at TSMC. A warrant is the right to buy a company's stock at a set price later, so when a hyperscaler takes warrants from its chip supplier, it's grabbing a slice of that supplier's stock upside as part of the deal. This letter tracks whether custom chips are eating into Nvidia's dominance and, increasingly, whether the whole boom can keep funding itself.

TL;DR

  • Qualcomm is now a real hyperscaler custom-silicon vendor, at AWS of all places. Industry analysts Patrick Moorhead and Daniel Newman detailed the multi-generation Amazon deal to build custom chips for "AI inference and connectivity," with Amazon able to buy 25 million Qualcomm shares tied to spending about $60 billion over a decade, roughly $9 billion of orders already committed, and (the tell) Qualcomm's Alphawave IP acquisition "was very much enabling AWS Silicon. They're already in there" (The Six Five). AWS was "the hardest company to get in."
  • The warrant playbook is now everywhere. Same week, Amazon signed a $2.4 billion generator supply deal with Generac and took warrants to buy up to $340 million of Generac stock; the stock ran 19% (Telltales). Last week we learned the mechanism; this week it showed up in two more deals.
  • The best single teardown of the week: Google's TPU economics. Google's own chips run at roughly a 40% discount to the Nvidia equivalent, are on their 7th generation, and were historically an internal cost-saving tool now sold "outright" to select outside customers, the vertical-integration edge that lets Google dodge Nvidia's 65% margins (The Intrinsic Value Podcast). But the same episode flags the catch: one customer (Anthropic) is ~40% of Google Cloud's half-trillion-dollar backlog.
  • Marvell's Google deal, reframed as the "pick of the decade." A portfolio manager on Full Signal says the Google warrant deal "implies $120 billion of revenue" against a ~$12 billion base, and that Marvell fell ~30 points for "no reason" after declining to guide to fiscal 2029 (Full Signal).
  • Nvidia is now anchoring its biggest customer's IPO. Nvidia is in talks to put up to $10 billion into Anthropic's roughly $2 trillion IPO, a supplier buying equity in its own buyer, which "round trips right back to their own revenue column" (Elon Musk Podcast). Anthropic's annualized revenue reportedly leaped to $65 billion, a 7x jump from $9 billion seven months earlier.
  • New and important: the first hard cracks in demand. Ramp's "Cracks in the AI Thesis" data shows frontier-model usage falling from 53% to 45% of usage, blended token prices down 41% to $0.68 per million (from a $1.15 March peak), and the critical top-1% of enterprise spenders cutting per-employee AI spend 9.7% in a single month (Big Technology Podcast).
  • The Fed hiked, roughly as expected. Brad Gerstner put ">90% chance" of a rate hike on the day of the decision (All-In); by mid-week, shows were discussing "post rate hike" volatility (The KE Report). Last week's imminent catalyst has now landed.

What's new

Ranked by what actually moves a book: the genuinely new, design-win-tier item first, then the new framings and financing developments, then the demand-side data that changes the risk picture.

1. Qualcomm is a real custom-silicon vendor now, and it broke into AWS.

This is the most actionable new chip development of the week, and it graduates a story we flagged as unconfirmed last week into one with hard numbers. On The Six Five, longtime semiconductor analysts Patrick Moorhead (Moor Insights & Strategy) and Daniel Newman (Futurum Group), two of the best-connected industry watchers around though not Qualcomm or Amazon insiders, walked through the deal (The Six Five).

The facts, as they laid them out:

  • Qualcomm "entered a multi-generation collaboration with Amazon to build custom silicon for large scale data center, focused around their AI inference and connectivity."
  • The structure is the now-familiar warrant model: "Amazon has the opportunity to buy a large block of shares... 25 million shares," tied to Amazon "spending about $60 billion over a decade."
  • Early traction: "based on the initial issuance, about $9 billion of orders committed," with $5 billion expected in Qualcomm's fiscal year starting October 1 and $15 billion by 2029.
  • The strategic tell: the Alphawave IP company Qualcomm bought "was very much enabling AWS Silicon. They're already in there." In other words, Qualcomm didn't cold-call AWS; it acquired its way inside.

Why it matters. Moorhead's line is the whole thesis: AWS "would have been" the "hardest company to get in" of all the hyperscalers, because "they have competency on the CPU side. They have competency on the accelerator side." AWS builds its own Trainium and Graviton; it does not need help. That it is now paying an outside vendor for custom silicon says the demand is so far beyond what even the best in-house teams can serve that hyperscalers will buy quality capacity anywhere they can find TSMC wafers. Newman: "the demand is so outlandishly large that if you have TSMC capacity right now and you can build a quality product... you can sell it right now."

For the book, this puts Qualcomm (QCOM) on the custom-silicon map alongside Broadcom and Marvell, a fourth merchant name with a hyperscaler anchor. Note the skepticism, too: the stock "jumped pretty sharply immediately, but then tapered off pretty quickly," because sell-side "looks 18 months, maybe two years out" and the $60 billion is a ceiling, not a locked purchase order. Newman's read: "the market is underplaying this." (Independent corroboration of the ceiling: Telltales flagged the same deal as "up to $60 billion... Up to. That's a ceiling, not a commitment" (Telltales W2637.)

2. The warrant playbook spreads to the picks-and-shovels layer.

Last week we explained why hyperscaler custom-chip revenue may not translate into hyperscaler-designer profit: warrant structures let the buyer claw back the upside. This week the model showed up twice more, and once outside chips entirely, which tells you it's becoming a standard tactic, not a one-off.

The clean example is generators. On Telltales (a podcast produced entirely with AI tools, with AI-generated voices, so treat it as a well-organized reading of a "cash flow memo" rather than human reporting), the hosts detailed Amazon's $2.4 billion backup-generator supply deal with Generac, under which Amazon took warrants to buy up to $340 million of Generac stock. The stock "ran more than 19%" (Telltales W2638). Their framing is the sharpest one-liner of the week: "Amazon just bought a supplier's upside with the supplier's own order book." Generac makes ~$600 million of free cash flow a year, so a $2.4 billion order is "roughly four years of everything the company makes landing in one contract," and the warrant is "Amazon telling you it knows exactly how big that order is."

The point for our sector: the Amazon-Qualcomm warrant (25 million shares) and the Amazon-Generac warrant ($340 million) are the same move a level apart. When a hyperscaler is the whole market for your product, it can dictate that it shares in your stock's re-rating. That is the value-capture squeeze in action, and it now spans custom chips and power equipment.

3. The Google TPU teardown: a 40% discount, and a backlog leaning on one customer.

The richest single episode of the week was a two-hour Alphabet deep-dive on The Intrinsic Value Podcast, hosted by Shawn O'Malley and Kyle Grieve (The Intrinsic Value Podcast). It's the best plain-English explanation I've heard of why Google's custom silicon is a genuine structural edge, and where the risk hides.

The bull mechanics on TPUs:

  • The discount. "Google's Tensor processing units come in at about a 40% discount to the NVIDIA equivalent." That's the vertical-integration payoff: Google skips what the hosts call "this kind of NVIDIA tax that pretty much all of these other companies are forced to pay" (Nvidia runs "65% operating margins as well as pricing power").
  • The maturity. Google is "already on their seventh generation" of TPUs. For most of that history they were "an internal cost saving tool specifically for Google," letting it "bypass buying chips from NVIDIA or to at least reduce their dependency." Now "select companies... have actually been offered the opportunity to purchase Google's TPUs outright."
  • The utilization edge. Unlike a neocloud that "rents by the hour," Google "can allocate compute capacity toward Google Search and YouTube and Gemini as needed," so "very little capacity being wasted," like a hotel that's always full.

The result shows up in Google Cloud: revenue grew 63% then 82% year-over-year in the first two quarters of 2026, a ~$100 billion run rate, at ~36% operating margins, "on par with Amazon's 2025 cloud profitability." Cloud backlog (remaining performance obligations) is "more than $500 billion," up 5x from $106 billion a year earlier.

The catch, and it's a big one. That backlog is dangerously concentrated. Google signed "a five-year $200 billion deal with Anthropic to use Google Cloud," which "implies that Anthropic makes up somewhere around 40% of that backlog." And it's circular: Google "committed up to about $40 billion into Anthropic," so "Google invests in the customer, the customer buys Google Compute, then Google then books the backlog." If Anthropic stumbles, "a lot of that backlog growth would just appear to be kind of fugazi." Layer on Michael Burry's warning (relayed) that hyperscalers are "understating depreciation by as much as $175 billion over the next few years," and Google's own ~$122 billion of "assets not yet in service" that will eventually hit the depreciation schedule, and you have a business whose reported profits ($112 billion in Q2, of which $99 billion was a non-cash markup of its SpaceX and Anthropic stakes) need reading with care.

Why it matters for us: the 40% TPU discount is the single cleanest number quantifying the merchant-vs-custom cost gap, and it's the engine under the Broadcom/MediaTek/Marvell TPU-build story. But the concentration and circularity are exactly the fragility the financing bears keep pointing at.

4. Marvell reframed: the Google warrant deal as "pick of the decade."

On Full Signal, a portfolio manager (Jay, whose firm recommended Marvell in the $80s "on national television") made an unusually aggressive Marvell case (Full Signal). The math: Marvell is "only a $200 billion market cap company," and the Google deal ("there's warrants") "implies $120 billion of revenue. Well, guess what their revenue base is? 12." So a potential 10x on the revenue base if the program scales. He carries "$12 [EPS] for fiscal 29" and a $300 target at 25x, against a stock "trading right now at 230." The catalyst he's waiting for: Marvell "said, well, we're not going to guide to... 2029 fiscal" and "stock comes off 30 points for no reason"; he expects them to eventually guide to that number.

Worth holding this next to last week's more skeptical Marvell read (the same warrant that creates the upside also hands value back to Google, and Marvell has been "very precise" about framing the Google opportunity; read: it may not have the high-margin compute tile). Same deal, two very different spins. The honest synthesis: Marvell's revenue opportunity is enormous and real; how much profit it keeps is the open question, and the warrant cuts against it.

5. Nvidia anchors Anthropic's ~$2 trillion IPO as the circle tightens.

The Elon Musk Podcast (an AI-generated discussion show; treat as a walk-through of reported news, not original reporting) laid out the structure cleanly (Elon Musk Podcast). Anthropic is seeking to raise "up to $100 billion" at roughly a $2 trillion valuation, and Nvidia is in talks to anchor with "up to $10 billion." The hosts' framing is the one that matters for our thesis: Nvidia is "evolving from just supplying the hardware... to really holding the purse strings of their biggest buyers." The capital "essentially just round trips right back to their own revenue column," "a highly efficient way of ensuring your assembly lines never stop running." And once Nvidia holds $10 billion of your equity, "you're highly unlikely to start experimenting with alternative chips from competitors." That's the anti-ASIC angle hiding inside a financing headline: equity ties can lock a lab onto Nvidia silicon.

The eye-watering number underneath: Anthropic's annualized revenue "leaped to $65 billion... a seven-fold jump from $9 billion just seven months prior." Nvidia has "$99 billion in equity investments right now" plus "$25 billion in investment commitments"; it can, as the hosts put it, "essentially just make their own markets."

6. The financing plumbing: Apollo's "calling all precincts," Oracle's equity-not-debt build, and the $5 trillion capex wall.

This is the theme that dominated last week and stayed loud. The most authoritative voice was an actual financier of these chips. Jim Zelter, President of Apollo Global Management, an operator whose firm is directly financing this build, told Bloomberg that Apollo has done "the Intel financing, the Broadcom financing, the NVIDIA financing" (Bloomberg Talks). His key points:

  • Scale forces diversification of funding. Nvidia's top four or five investors already hold "anywhere from 2%, 3%... to almost 9% of the equity. That's a $500 billion exposure." No single asset class can fund this: "they're going to need any and all. And I call it calling all precincts." He expects the "broad ecosystem of AI" to become "10% of the IG [investment-grade bond] market."
  • Access to capital is now a moat: "having access to capital in scale is certainly a competitive advantage... the winners and the haves and the have-nots."
  • A neat aside supporting the picks-and-shovels bull case: "the gross margin is highest away from the models," i.e., in the chips and infrastructure, not the LLMs themselves. And, notably, "companies like Broadcom and many others are doing very, very well, notwithstanding who's the winner of the LLM race."

The clearest single case study is Oracle, dissected across two Telltales episodes. Oracle printed revenue of $19.3 billion (+30%), cloud infrastructure +121%, and remaining performance obligations up $209 billion year-over-year to $664 billion, but free cash flow of negative $5 billion on a record $23 billion operating cash quarter, because capex was $28.5 billion. It funded the gap not with bonds but with "$19.9 billion of net proceeds from an at-the-market equity program," against "$126 billion of net debt" already on the books (Telltales W2637). This week Larry Ellison called off a planned "$7.5 billion" personal stock sale and Oracle signed a "433-megawatt power purchase agreement with RWE" (Telltales W2638). The hosts' read: the equity channel "is still the open channel, and the people closest to it are not leaning on it," while locking up power years ahead "is not what a company about to slow down does."

And the macro frame, from ad-industry veteran Sir Martin Sorrell (S4 Capital), relaying a breakfast with an Alphabet EMEA engineer (The Geopolitics of Business): the four biggest spenders (Microsoft, Amazon, Alphabet, Meta) "are going to spend $5 trillion between 2025 and 2030." The 2026 forecast for those four rose from ~$630 billion to ~$830 billion after Q2; 2027 is forecast at $1.1 trillion, of which roughly "$530 billion" (about half) would be "raised in the debt markets," a burden big enough to "crowd out government debt." His verdict, waving a book about the 1873 railway boom: "yes, there will be failures."

7. New: the first hard demand-side cracks.

This is the development that most changes the risk picture, because the bull case has always rested on token demand growing faster than prices fall. Two shows independently flagged the Ramp "Cracks in the AI Thesis" data:

  • On Big Technology, hosts Alex Kantrowitz and Ranjan Roy walked through it (Big Technology Podcast): frontier-model share of usage fell from 53% to 45% in a month as routing services push users to cheaper standard models; the top-1% of enterprise spenders (who "make up 80% of the spend for OpenAI and Anthropic") cut per-employee AI spend 9.7%, from $7,976 to $7,205; and blended token prices fell 41% to $0.68 per million, from a $1.15 March peak. Kantrowitz: "if there's a red flag, this is a red flag."
  • Investor Dan Niles (Niles Investment Management) corroborated on RiskReversal (RiskReversal Pod): cost per token down "50% since the end of May," though "the number of tokens being produced quadrupled." He layered on three worries: Texas putting "a moratorium on data center expansion," Altman/Amodei/Musk all publicly calling to "slow down the pace of innovation," and architectural shifts that "reduce the need for some levels of compute." His Cisco analogy is the one to sit with: in May 2001, Cisco went from "orders being up 50% year over year to down 30 in several months," and "Cisco was NVIDIA back then." His net call, though: "another year or two before we hit the top," because "if you fundamentally believe... AI is the most transformational technology... then by definition, you're overbuilding."

Brad Gerstner (Altimeter) framed the same tension as a math problem on All-In (All-In): to fund "$1.5 trillion a year in CapEx," the offtake revenue has to climb from a ~$200 billion exit run rate this year toward "$450... to $800 or a trillion dollars." He thinks the demand (TAM) exists ("you only have to get to about 4%" of the knowledge-work market), but the physical build is the bottleneck: he thinks the "43 gigawatts" of compute the market expects to add in 2027 is "too aggressive" and pegs the real number "closer to 25 gigawatts." His portfolio stance: "we're medium position... mentally flexible."

The debate

The motion: do hyperscaler custom chips structurally take share and cap Nvidia's market and margins, or do CUDA, systems integration, and Nvidia's pace keep merchant GPUs on top?

For the custom-silicon side (ASICs take share). This week strengthened it. Qualcomm just proved a fourth merchant vendor can win custom-silicon business at the hardest hyperscaler to crack, AWS, evidence the pie is big enough for many designers. Google's TPUs deliver a hard 40% cost discount to Nvidia and now sell to outside customers; Marvell's Google program alone "implies $120 billion of revenue"; and even Apollo's Zelter says the infrastructure layer wins "notwithstanding who's the winner of the LLM race," with "gross margin highest away from the models." The volume story is intact and broadening.

For the merchant side (Nvidia keeps the pie). Two threads run its way. First, value capture: the warrant model, now visible in Amazon-Qualcomm (25M shares) and Amazon-Generac ($340M), means custom-chip revenue can boom while the profit accrues to the hyperscaler buyer, not the designer. Second, lock-in through capital: Nvidia anchoring Anthropic's IPO with up to $10 billion is a supplier using its balance sheet to guarantee its customer stays on Nvidia silicon, the money "round-tripping right back to their own revenue column." The TechSurge episode adds nuance (TechSurge): even OpenAI, designing its own blank-sheet "jalapeno" chip, is "buying compute from... the big vendors, mostly GPUs," while its own team co-designs for its specific workload; the two approaches coexist rather than one killing the other. And Nvidia keeps absorbing adjacent sockets (it acquired Grok for the decode step), reinforcing the "buy a system, not a chip" pull.

The most useful new frame this week is that the fight is no longer really a chip fight; it's a financing and demand fight, and both moved. On the financing side, the warrant/anchor structures show hyperscalers hold the bargaining power. On the demand side, the Ramp data (frontier usage 53%→45%, token prices down 41%, top-1% spend down 9.7%) is the first quantified sign that the token-growth flywheel might not outrun price declines forever. If demand cracks, it doesn't matter who makes the better chip; the whole complex de-rates together, exactly as Niles' Cisco-2001 analogy warns.

Where I come out this week: the silicon call is unchanged: own the volume growth on both sides, and add Qualcomm to the merchant-ASIC watchlist as a genuine new entrant. But the two swing factors both got sharper, and they point in opposite directions. The bull swing factor is that custom silicon keeps winning sockets at every hyperscaler, even AWS. The bear swing factor is new and more concerning than the financing worry: the first real evidence that top-end token demand is softening while prices fall 40%+. I'd rather own the picks-and-shovels names with the warrant terms understood, size medium like Gerstner, and watch the Ramp demand data and the frontier-usage share as the leading indicators that matter more than any single design win.

Stocks in play

Every ticker with fresh podcast commentary this week, with bull, bear, and the next thing to watch. Most of this is analyst and pundit opinion, not operator disclosure.

  • QCOM (Qualcomm). Bull: broke into AWS custom silicon (inference + connectivity) on a ~$60B/decade framework, ~$9B orders already committed, $5B expected this fiscal year, $15B by 2029; Alphawave IP already "in there" at AWS; "this will not be their only massive partnership with a hyperscaler" (The Six Five). Bear: $60B is a ceiling, not a purchase order; stock jumped then faded as sell-side waits for granularity; still "single low-digit market share." Watch: confirmation of the terms in filings; whether the ~$9B committed converts and whether a second hyperscaler signs.
  • GOOGL (Alphabet). Bull: TPUs at a ~40% discount to Nvidia (7th gen, now sold to outside customers) give a real cost moat; cloud growing 63%→82% YoY at ~36% margins on a ~$100B run rate; >$500B backlog, up 5x YoY (The Intrinsic Value Podcast). Bear: backlog is ~40% Anthropic and circular ($40B Google→Anthropic→Google Cloud); first negative-FCF quarter in Q2 2026; buybacks stopped; $99B of Q2's $112B "profit" was a non-cash markup; Burry says depreciation understated by up to $175B industry-wide; >$800B off-balance-sheet liabilities; analyst avg target ~$430 vs ~$340 (no variant view left). Watch: Anthropic's ability to actually pay cash against that backlog; the depreciation schedule as "assets not yet in service" ($122B) switch on.
  • MRVL (Marvell). Bull: the Google warrant deal "implies $120 billion of revenue" vs a ~$12B base; a bull carries $12 EPS for FY29 and a $300 target (25x) vs ~$230 (Full Signal). Bear: declined to guide to FY29 and fell ~30 points; the same warrant hands value back to Google; likely doesn't hold the high-margin compute tile. Watch: whether management eventually guides to the FY29 number, the catalyst the bull is waiting for.
  • NVDA (Nvidia). Bull: still the architecture everyone builds around; acquiring adjacent sockets (Grok for decode); 65% operating margins and pricing power; anchoring Anthropic locks its biggest buyer onto Nvidia silicon (Elon Musk Podcast, TechSurge). Bear: the $10B Anthropic anchor is "customer acquisition cost" dressed as an investment (circular demand); DOJ is investigating whether the $20B Grok licensing deal was structured to dodge antitrust review; semis down ~25% from July highs; Cisco-2001 risk if orders turn. Watch: the Ramp frontier-usage data; DOJ/Grok inquiry; whether hyperscaler capex growth decelerates in 2027.
  • AMZN (Amazon). Bull: using warrants to lock in supply cheaply (Qualcomm chips, Generac generators), buying suppliers' upside "with the supplier's own order book"; trailing capex ~$173B shows unmatched build scale (The Six Five, Telltales W2638). Bear: free cash flow roughly negative on that capex. Watch: Trainium 3 ramp detail; whether the Qualcomm relationship is incremental to in-house silicon or a substitute.
  • ORCL (Oracle). Bull: $664B backlog (+$209B YoY), cloud infra +121%, 850 MW and 300k GPUs delivered in a quarter, locking up 433 MW of RWE power years early, behaving like it believes its order book (Telltales W2637). Bear: negative $5B FCF on a record cash quarter; $126B net debt (~20% of EV); funding the gap with equity ($19.9B ATM); Ellison subpoenaed over a $27B VA contract and calling off a $7.5B stock sale. Watch: the funding line on the next print (equity vs debt); whether the backlog converts to cash.
  • CRDO (Credo). Bull: fastest-growing of the connectivity names: revenue +115% YoY last quarter, 25% operating margin, ~17-18% FCF conversion, $764M net cash, guiding 70-80% revenue growth for the new fiscal year; a "new Baby Broadcom" (Chip Stock Investor). Bear: sold off hard despite the growth; cyclical business whose margins "are probably not going to stand" a sharp slowdown in the data-center buildout; stock ~$171. Watch: whether guidance holds; any sign of the buildout pace slowing.
  • GNRC (Generac). Bull: $2.4B Amazon supply deal (~4 years of FCF in one contract) turns a lumpy storm-demand business into a contracted infrastructure supplier (Telltales W2638). Bear: already 29x FCF before Amazon walked in; must build capacity for an order "nobody has shown us what that costs yet"; Amazon's $340M warrant caps the upside. Watch: the 2027 delivery date and the capex to serve it.
  • AVGO (Broadcom), ALAB (Astera Labs). Broadcom surfaced via Apollo's financing mention ("Broadcom and many others are doing very, very well"). Astera was named in passing as a "new Baby Broadcom" and a connectivity/rack-scale name (Chip Stock Investor).

Read-throughs

  • Custom-silicon IP (Arm). The relevant read-through is Qualcomm's Alphawave acquisition "enabling AWS Silicon," a reminder that the IP and interconnect layer inside every hyperscaler ASIC is where quiet share gets won (The Six Five).
  • Connectivity (Credo, Astera Labs, Marvell, Broadcom). The clearest connectivity signal is Credo's +115% revenue growth and the framing of Credo and Astera Labs as the "new Baby Broadcoms" in AI data-center networking, with very different gross-margin profiles despite the same sub-industry (Chip Stock Investor). Qualcomm's AWS deal is explicitly "inference and connectivity," so watch connectivity attach as a shared tailwind.
  • New chip architectures (a watch item for the merchant/ASIC balance). TechSurge is worth reading in full: OpenAI's blank-sheet "jalapeno" chip (unveiled at Hot Chips, giving each accelerator "its own little HBM slice"), Etched's low-voltage inference bet, and Tensordyne's "log math" approach all point to a proliferation of custom designs, and AI is cutting chip-design timelines "from three years to one year" and design cost "by a third," lowering the barrier to entry (TechSurge). More entrants is bullish for design-services and IP shops, bearish for the "one GPU fits all" story.
  • HBM / memory (SK Hynix, Micron, Samsung). Memory is flagged as a bottleneck and a cost driver: SK Hynix is "in talks with Intel to make memory chips in the United States," and Apple's iPhone 18 took a $100 price increase against "a 400% increase in the cost of memory," with gross-margin guidance cut from 51% to 48% (Practical News, The Six Five). Memory pricing remains the swing variable in every accelerator's cost stack.
  • TSMC foundry capacity. TSMC posted "record August revenue of about $16 billion, up 53% year over year, and still can't keep pace with AI chip demand" (Telltales W2637). Wafers remain the governor on both GPUs and ASICs; Newman's "wafers are fungible at the same node level" point is why any vendor with TSMC allocation and a quality design can sell today.
  • Power and the grid. Power is repeatedly named the binding constraint. Oracle's 433 MW RWE deal, the new 18-member AI Energy Management Alliance (Nvidia, Google, Emerald AI, Anthropic, National Grid, AES, NRG), and xAI's Memphis site "making a billion dollars a month... roughly renting" compute to Anthropic all underline it (Telltales W2638, Practical News). Generac ($2.4B Amazon deal) is the newest power read-through.
  • The financing plumbing (the master dial). Apollo "calling all precincts," AI heading to 10% of the IG bond market, Nvidia anchoring its own customer, Oracle funding with equity, and $530B of 2027 capex expected to hit the debt markets: this is still the fault line under every name in this letter (Bloomberg Talks, The Geopolitics of Business).

What changed vs last week

The tempo flipped back toward chips, but with a twist, and a new worry.

  1. A real new chip development. This week, Qualcomm's AWS custom-silicon deal got hard terms (25M-share warrant, ~$60B/decade, ~$9B committed, Alphawave inside AWS). That's the first design-win-tier story in two weeks.
  2. The Amazon-Qualcomm deal went from unconfirmed to confirmed. Last week we flagged (via Vaughan Nelson) an Amazon-Qualcomm deal that "happened this past week" as unconfirmed, chase against filings. This week two shows detailed the actual structure. That's the exact confirmation we asked for.
  3. The warrant thesis kept compounding. Last week we introduced the warrant mechanism (Google-Marvell, Amazon-Qualcomm). This week it appeared in two more forms (the confirmed Qualcomm terms and a brand-new Amazon-Generac warrant, $340M), cementing it as the standard hyperscaler tactic, not an anomaly.
  4. Marvell flipped from a value-capture worry to a "pick of the decade" pitch. Last week's read stressed that the Google warrant hands value back to Google. This week a PM reframed the same deal as a potential 10x on the revenue base ($120B implied vs $12B). Both can be true; the tension is now explicit.
  5. The financing circle tightened at the top. Last week: Blackstone underwriting a TPU cloud (Crux AI, $10-20B). This week: Nvidia anchoring Anthropic's ~$2T IPO with up to $10B, a supplier funding its own customer, an even more direct circular loop.
  6. New this week: the first quantified demand cracks. Last week the bears' worry was still mostly about financing. This week the Ramp "Cracks in the AI Thesis" data hit the demand side: frontier usage 53%→45%, token prices down 41%, top-1% spend down 9.7%. That is a different and arguably more important warning than the debt story.
  7. The Fed catalyst resolved. Last week the market was at ~60% odds of a hike; this week Gerstner put it at >90% on decision day and mid-week shows referred to "post rate hike" volatility. The nearest-term macro catalyst has passed, and rates are now a live headwind, not a coming one.