The Moat Shows Up in the Price

Jensen Huang, Nvidia's chief executive, told Brad Gerstner on BG2Pod on October 15, 2026 that in a thought experiment, a $100 billion order for next-generation compute placed sight unseen could go to one company only, his own. Ren, who writes about investing in the AI buildout, had argued the reverse three days earlier: the openness that brought Meta onto Nvidia's Spectrum-6 networking makes it easier for Meta to leave, while the largest buyers design more of their own silicon.

A rival chip can win without taking an order from Nvidia, by lowering what Nvidia can charge for the rest.

Walter Goodwin, founder of the chip startup Fractile, repeated a joke on Sarah Guo's No Priors on October 2 that he said may hold some truth: the main purpose of the big buyers' in-house chips is to lower the price they pay Nvidia. Scott Guthrie, Microsoft's cloud chief, whose company builds its own chip, told Tae Kim of Key Context on October 12 that Microsoft will remain a big buyer of Nvidia and AMD and calls its approach multi-silicon. If Goodwin is right, the contest is already under way. A rival chip can win without taking an order from Nvidia, by lowering what Nvidia can charge for the rest.

Nvidia's CFO commentary of August 26 reported a non-GAAP gross margin of 75.0% for the quarter to July 26, 2026, as in the quarter before, on Data Center revenue of $89.0 billion, up 117%; hyperscale revenue more than doubled. Broadcom's release of September 2 reported AI semiconductor revenue of $16.7 billion for the quarter to August 2, up 221% from a year earlier and 54% from the quarter before, and guided $21.7 billion for the next. Its non-GAAP gross margin is also 75.0% by DeepStack's arithmetic, with software, 30% of revenue, inside; Broadcom does not report a margin for its AI chips alone. Alphabet's release of July 22 says Google Cloud, whose second-quarter revenue rose 82% to $24.8 billion, earns its product revenue primarily from selling TPU systems, without giving their size. AMD reported on August 4 Data Center revenue of $6.7 billion for the quarter to June 27, up 107%, and named Anthropic, Meta, Microsoft, OpenAI and Oracle among those deploying its Helios rack. Nvidia guided its margin to 74.0%, plus or minus 50 basis points, for the quarter that ended October 25, and reported commitments of $279 billion, up from $119 billion, which it said were primarily for memory.

Part twenty-one asks whether Nvidia can keep charging a premium once Google's TPUs, Broadcom's custom chips and AMD's racks arrive in volume. Part sixteen read Nvidia as a seller that finances part of its demand; this Part leaves the financing there and reads the price. DeepStack read 404 posts, letters, filings and episodes by 113 voices and verified 241 agreements and 59 disagreements, 13 of them direct contradictions. A disclosure: DeepStack is built with Anthropic's models; AMD's release names Anthropic among those deploying its Helios racks, the lenders' release of October 26 names it as the customer for racks built on Broadcom's chips, and Alphabet named it among its TPU customers on its October 28 call.

The ruler

DeepStack's ruler is the share of AI chip dollars on three filings. In the quarters to late June, July and early August 2026, Nvidia's $89.0 billion, Broadcom's $16.7 billion and AMD's $6.7 billion add up to $112.4 billion, and Nvidia's part is 79.2%. Broadcom's AI line and Nvidia's Data Center both include networking. The error runs both ways. AMD's segment includes server processors, which pulls Nvidia's share down. AMD's quarter closed a month earlier, and the total omits Google's TPU system sales, Marvell's custom chips and Amazon's Trainium, all of which push it up. Alphabet again gave TPU system sales no size on October 28: its release called them a meaningful contributor to the growth of Google Cloud revenue, which rose 103% to $30.7 billion in the third quarter, and its finance chief said on the call that Alphabet does not give financials at the product level. The 79.2% is neither floor nor ceiling.

Broadcom's AI revenue was 18.8% of Nvidia's Data Center revenue in the latest quarters, against about 14.4% a quarter earlier, which DeepStack derives from the 54% sequential growth Broadcom reported on September 2. Broadcom gained that ground while Nvidia's margin held at 75.0% in both quarters. Had Nvidia cut its price to defend the share, its margin would have shown it unless its costs fell as fast, which is why this Part reads the margin, and memory with it, before the share.

The platform holds the price

The first reading holds that the moat has moved from the chip to the system. Huang argued on October 15 that GPUs which run any workload and stay useful for years give the lowest total cost of ownership, as long as Nvidia keeps proving its performance, and that specialization is the enemy of both. Ian Buck, an Nvidia vice president, told Tae Kim of Key Context the same day that Nvidia answers Microsoft's Maia 300, Amazon's next Trainium and OpenAI's Jalapeño with more tokens and a lower cost per token each generation. Both men speak for the seller.

Voices outside Nvidia put the moat in the rack. Ren, who discloses an unpaid partnership with Nvidia that includes access to its press briefings, wrote on October 12 that Nvidia now ties GPU, CPU, networking, power, cooling, software and facility design into one specification and wins while it sits at its center. Jay Goldberg said on Ben Bajarin's The Circuit on October 5 that Nvidia sets the terms of competition and forces AMD and Broadcom to follow. Baillie Gifford, a holder of the stock, told clients on October 15 that Nvidia is a durable node rather than the beneficiary of a temporary shortage.

The side's hardest evidence is a price paid. CoreWeave's investor presentation, filed with the SEC on September 17, described a research lab that renewed Nvidia H200 capacity for three years at a premium to its original contract; CoreWeave rents those GPUs and has reason to praise them. Colette Kress, Nvidia's chief financial officer, said on the August 26 call that memory prices have risen faster than expected and are headed higher, that the margin bottoms at 71% to 72% in the fourth quarter and that it settles at 72% to 73% in fiscal 2028 as price increases already executed take effect in the first quarter. Gaetano drew the same path from the call on August 28.

DeepStack's reading: the 75.0% is this side's real argument, and Nvidia held it in a quarter when Broadcom grew faster. The renewal is its best witness, a premium paid for chips a generation old, but the witness has an interest in the answer and gives no margin.

Buyers build their way out

The second reading holds that the biggest buyers design chips to bargain down Nvidia's price, then move any workload large enough to justify a specialist. Amin Vahdat, who runs Google's AI infrastructure, told Sonya Huang on Sequoia's Training Data on October 6 that Google built its TPU 8i for inference because serving looked like 30% to 60% of the market over the chip's life. Lisa Su, AMD's chief executive, said that Helios entered production in September with OpenAI, Meta, Microsoft, Oracle and Anthropic as customers, as Kim reported on October 13. Like Vahdat and Guthrie, she speaks for an alternative.

On October 2 Goodwin said everyone deploying AI at scale now tries to run as many separate chip platforms as it can. OpenAI said on August 25 that Jalapeño did 1.5 to 1.9 times more AI work per watt at peak throughput than Nvidia's GB200 and GB300 systems, on InferenceX runs it made itself, normalized to each chip's rated power and with no cost per token. Kakashii reported on August 25 that Nebius customers take GB300 capacity on three- to six-month terms for training, at a premium, and judged that they will probably move inference to cheaper GPUs afterward; whether that means TPUs or custom chips, Kakashii left open. In an October 27 filing Broadcom disclosed that it backstops the residual value of AI racks built on its custom accelerators and financed by partners, with a maximum potential liability of about $43 billion once the racks are deployed, and the four lenders said the $63 billion transaction is for Anthropic, with deliveries expected from calendar 2027.

Goodwin also supplied this side's limit. An Nvidia system combines six to nine chips Nvidia designs itself, he said on the same episode. Ben Bajarin said on The Circuit on October 12 that even successful neoclouds may be only largely Nvidia shops, and that it remains unclear how vendors will run one workload across Nvidia and AMD GPUs without splitting the model.

DeepStack's reading: Goodwin's mechanism reaches the margin even when the custom chip loses, because a buyer with a credible second source pays less for the first. The filings for the quarters to early August 2026 show the second sources are real: Google has begun selling TPU systems, with most of that revenue due in 2027, Broadcom's AI line grew 54% in a quarter and AMD names its rack buyers. They do not yet show a lower Nvidia price. Against chip designs and racks still ramping, the margin printed 75.0% twice; the results due November 17 are the first to cover a quarter in which Helios was in production.

The shortage set the price

The third reading holds that today's margin is a scarcity rent. Michael Burry, who disclosed a short and put position in the stock in the same post, wrote on August 26 that his estimate of Nvidia's value sits far below the market's because he expects its monopoly power, and with it its margins, to fall in the short run. On October 1 he wrote that once the memory shortage turns into a glut, GPU server prices should fall sharply, possibly as early as next year. Aristotle Capital Management called Nvidia's earnings transient scarcity profits in its second-quarter 2026 letter. Glen Kacher, founder and chief investment officer of Light Street Capital, told Barry Ritholtz on Masters in Business on September 18 that the compute bottleneck pushes prices above what they would otherwise be.

Invest In Assets read the same margin path on September 5 as the work of memory prices. Damnang, in Edelbridge Alpha's memory masterclass of September 8, described the counterweight: buyers like Nvidia have real bargaining power, so memory prices cannot climb without limit.

DeepStack's reading: this side names the right risk and, for now, the wrong cause. On Kress's own account, the decline Nvidia guided is a memory cost, and the price increases she described are already executed. A scarcity rent shows when a cost eases and the price falls with it. A moat shows when a cost rises and the price follows, and those increases, in effect from the first quarter of fiscal 2028, are the first chance in this record to measure it.

What would settle it

The first deciding document is Nvidia's outlook for the fourth quarter of fiscal 2027, due with its third-quarter results on November 17, the date Nvidia gave on its August 26 call and confirmed on October 28. The largest buyers reported first and settled nothing about the price. Microsoft said on October 28 that it scales its fleet with Nvidia and AMD alongside its own Maia systems and has taken delivery of its first Nvidia Vera Rubin GPUs. Amazon said on its October 29 call that it will begin selling Trainium outside its cloud late in 2027, and Alphabet's quarterly report says the significant majority of revenue from its agreements to supply TPU systems for use on customers' premises will be recognized in 2027. For capital spending in calendar 2026, Microsoft expects about $170 billion, finance leases included, Amazon about $220 billion in cash and Alphabet $195 billion to $205 billion. AMD reports on November 3. DeepStack will read four lines, each from a filing: whether Nvidia guides the fourth quarter's non-GAAP gross margin at or above 71.0% on November 17, a range counting at its midpoint; whether Broadcom guides AI semiconductor revenue for its first quarter of fiscal 2027 at or above $25.0 billion on December 9, 15% above its September 2 guide; whether that quarter's AI revenue reaches 20.0% of Nvidia's Data Center revenue for its fourth quarter, read in March 2027; and whether Nvidia's non-GAAP gross margin for the first quarter of fiscal 2028, reported in May 2027, is back at or above 72.0%.

DeepStack's Call is that Nvidia keeps its premium and the dip is memory's, as its CFO says. The margin held at 75.0% in the quarter to July 26, 2026, while Broadcom's AI revenue grew 54% and Google began selling TPU systems. The November guide will probably land inside the 71% to 72% bottom Kress gave in August, and DeepStack would not read that alone as lost pricing power. A margin at or above 72.0% in the first quarter of fiscal 2028, when the executed price increases take effect, sits at the bottom of the 72% to 73% Kress gave for fiscal 2028; DeepStack reads it in that quarter as the least a moat requires; a margin held below that line would mean the premium has gone. Memory prices tell the two apart: if they are still climbing when Nvidia prints 72% or more, as Micron's reports will show, the price followed the cost; if they have eased by then, the recovery says less about pricing power. Two readings would change the Call: a fourth-quarter guide below 71.0% on November 17, or a first-quarter margin below 72.0% in May 2027 with the executed price increases in effect.

Until then, read the margin outlook before the revenue line. The rivals have arrived. The price will say whether they won.